Journal of Communication and Computer（Issue 6，2013）

Volume 10, Number 6, June 2013 (Serial Number 103)

Journal of

Communication and Computer

David Publishing Company

www.davidpublishing.com

PublishingDavid

Publication Information: Journal of Communication and Computer is published monthly in hard copy (ISSN 1548-7709) and online (ISSN 1930-1553) by David Publishing Company located at 3592 Rosemead Blvd #220, Rosemead, CA 91770, USA. Aims and Scope: Journal of Communication and Computer, a monthly professional academic journal, covers all sorts of researches on Theoretical Computer Science, Network and Information Technology, Communication and Information Processing, Electronic Engineering as well as other issues. Contributing Editors: YANG Chun-lai, male, Ph.D. of Boston College (1998), Senior System Analyst of Technology Division, Chicago Mercantile Exchange. DUAN Xiao-xia, female, Master of Information and Communications of Tokyo Metropolitan University, Chairman of Phonamic Technology Ltd. (Chengdu, China). Editors: Cecily Z., Lily L., Ken S., Gavin D., Jim Q., Jimmy W., Hiller H., Martina M., Susan H., Jane C., Betty Z., Gloria G., Stella H., Clio Y., Grace P., Caroline L., Alina Y.. Manuscripts and correspondence are invited for publication. You can submit your papers via Web Submission, or E-mail to [email protected]. Submission guidelines and Web Submission system are available at http://www.davidpublishing.org, www.davidpublishing.com. Editorial Office: 3592 Rosemead Blvd #220, Rosemead, CA 91770, USA Tel：1-323-984-7526, Fax: 1-323-984-7374 E-mail: [email protected]; [email protected] Copyright©2013 by David Publishing Company and individual contributors. All rights reserved. David Publishing Company holds the exclusive copyright of all the contents of this journal. In accordance with the international convention, no part of this journal may be reproduced or transmitted by any media or publishing organs (including various websites) without the written permission of the copyright holder. Otherwise, any conduct would be considered as the violation of the copyright. The contents of this journal are available for any citation. However, all the citations should be clearly indicated with the title of this journal, serial number and the name of the author. Abstracted / Indexed in: Database of EBSCO, Massachusetts, USA Chinese Database of CEPS, Airiti Inc. & OCLC Chinese Scientific Journals Database, VIP Corporation, Chongqing, P.R.China CSA Technology Research Database Ulrich’s Periodicals Directory Summon Serials Solutions Subscription Information: Price (per year): Print $520; Online $360; Print and Online $680 David Publishing Company 3592 Rosemead Blvd #220, Rosemead, CA 91770, USA Tel：1-323-984-7526, Fax: 1-323-984-7374 E-mail: [email protected]

David Publishing Company

www.davidpublishing.com

DAVID PUBLISHING

D

Journal of Communication and Computer

Volume 10, Number 6, June 2013 (Serial Number 103)

Contents Computer Theory and Computational Science

731 Proposal for an Optimal Job Allocation Method Based on Multiple Costs Balancing in Hybrid Cloud

Yumiko Kasae and Masato Oguchi

743 Probabilistic Health-Informatics and Bioterrorism

Ramalingam Shanmugam

748 A Hybrid Method to Improve Forecasting Accuracy—An Application to the Canned Cooked Rice and the Aseptic Packaged Rice

Hiromasa Takeyasu, Daisuke Takeyasu and Kazuhiro Takeyasu

759 System Reconstruction in the Field-Specific Methodology of Economic Subjects

Katarina Krpalkova Krelova and Pavel Krpalek

769 The Research of Wind Turbine Fault Diagnoses Based on Data Mining

Yu Song and Jianmei Zhang

Network and Information Technology

772 A Recommendation System Keeping Both Precision and Recall by Extraction of Uninteresting Information

Tsukasa Kondo, Fumiko Harada and Hiromitsu Shimakawa

783 Comparison of Contemporary Solutions for High Speed Data Transport on WAN 10 Gbit/s Connections

Dmitry Kachan, Eduard Siemens and Vyacheslav Shuvalov

796 Awaken the Cyber Dragon: China’s Cyber Strategy and its Impact on ASEAN

Miguel Alberto N. Gomez

806 Inter-MAC Green Path Selection for Heterogeneous Networks

Olivier Bouchet, Abdesselem Kortebi and Mathieu Boucher

815 Web Block Extraction System Based on Client-Side Imaging for Clickable Image Map

Hiroyuki Sano, Shun Shiramatsu, Tadachika Ozono and Toramatsu Shintani

Communications and Electronic Engineering

823 Multiple Chaos Generator by Neural-Network-Differential-Equation for Intelligent Fish-Catching

Mamoru Minami, Akira Yanou, Yuya Ito and Takashi Tomono

832 An Efficient Algorithm for the Evaluate of the Electromagnetic Field near Several Radio Base Stations

Algenti Lala, Sanije Cela and Bexhet Kamo

844 Composing Specific Domains for Large Scale Systems

Asmaa Baya and Bouchra EL Asri

857 Mobile Station Speed Estimation with Multi-bit Quantizer in Adaptive Power Control

Hyeon-Cheol Lee

863 On the Comparison Analysis of Two 4G-WiMAX Base Stations in an Urban Sub-Saharan African Environment

Eric Tutu Tchao, Kwasi Diawuo1 and Willie Ofosu

873 Error Analysis for the Three-Dimensional Detection and Reconstruction of the Road Surface

Youquan He and Jian Wang

Journal of Communication and Computer 10 (2013) 731-742

Proposal for an Optimal Job Allocation Method Based

on Multiple Costs Balancing in Hybrid Cloud

Yumiko Kasae and Masato Oguchi

Department of Information Sciences, Ochanomizu University, Tokyo 112-8610, Japan

Received: February 07, 2013 / Accepted: March 13, 2013 / Published: June 30, 2013.

Abstract: Due to the explosive increase in the amount of information in computer systems, we need a system that can process large amounts of data efficiently. Cloud computing system is an effective means to achieve this capacity and has spread throughout the world. In our research, we focus on hybrid cloud environments, and we propose a method for efficiently processing large amounts of data while responding flexibly to needs related to performance and costs. We have developed this method as middleware. For data-intensive jobs using this system, we have created a benchmark that can determine the saturation of the system resources deterministically. Using this benchmark, we can determine the parameters in this middleware. This middleware can provide Pareto optimal cost load balancing based on the needs of the user. The results of the evaluation indicate the success of the system. Key words: Hybrid cloud, load balancing, data processing, performance, cost balance.

1. Introduction

In recent years, large amounts of data, referred to as

big data, have become more common with the

development of information and communications,

creating the need for efficient data processing. As a

platform for processing these data, hybrid cloud

environments have become a focus of attention. In

hybrid cloud environments, users can access public

clouds and private clouds; private clouds are secure

clouds built using the secure resources of the user

company, and public clouds can provide scalable

resources if the user pays metered rates. Combining

these clouds can address shortcomings related to

safety and scalability. For data-intensive jobs, hybrid

clouds are appropriate. For increasing amounts of data,

hybrid clouds can provide secure and scalable

processing.

However, performance and costs must be balanced.

When we want to process large amounts of data more

rapidly, using many resources that are provided by

Corresponding author: Yumiko Kasae, master, research

field: information science. E-mail: [email protected].

public clouds, in addition to those provided by private

clouds, will increase speed, but the metered cost will

also be greater. In contrast, if these jobs are processed

using private cloud resources almost exclusively,

users will not have to pay metered rates, but the job

execution time will be longer. Thus, we need a system

that can determine optimal job placement based on

cost limitations and necessary performance to ensure

efficient processing in hybrid cloud environments.

Therefore, in this research, we proposed a method

for providing optimal job placement in hybrid cloud

environments in terms of monetary costs and

performance. We have developed this system as

middleware. In addition, the middleware provides

optimal job placement for both CPU-intensive

applications and data-intensive applications. In

general, unlike in CPU-intensive applications, which

can accurately determine the load using the CPU

usage, efficient resource use in data-intensive

applications is difficult to determine. In the proposed

method, we created a benchmark that can be used to

change the extent of the load of CPU processing and

Proposal for an Optimal Job Allocation Method Based on Multiple Costs Balancing in Hybrid Cloud

732

I/O processing, and measured the performance of

hybrid clouds as an execution environment using this

benchmark. Based on the results obtained using this

benchmark, we propose a method of determining job

execution status based on the status of the I/O

resources.

In this paper, we will describe the details of the

middleware that can be used to implement the method

proposed in this study. We have evaluated the balance

of performance and costs by using this middleware

with data-intensive applications. We examine the

evaluation axis for performance and monetary costs

and show that this middleware can provide optimal

job placement for efficient job processing. The

monetary cost is the sum of the power consumption

cost for private clouds and the metered costs

associated with public clouds.

The remainder of this paper is organized as follows:

Section 2 introduces cloud computing; Section 3

describes the proposed method of determining the

load; Section 4 describes the middleware that can be

used for optimal job allocation; Section 5 introduces

the evaluation results for their middleware; Section 6

comments on related research studies; and Section 7

presents concluding remarks and suggestions for

future work.

2. Cloud Computing

2.1 Overview of Cloud Computing and Classification

Cloud computing is a service through which users

can use necessary software and hardware resources

from servers through networks. If a user uses cloud

computing services, without having the physical

computer resources, the user can receive various

services.

The types of services include SaaS (software as a

service), PaaS (platform as a service) and IaaS

(infrastructure as a service).

Recently, because of the diversity of services that

can be provided, these services have been collectively

called XaaS (X as a service).

This study will consider IaaS. The types of

platforms for IaaS are private clouds and public

clouds. Public clouds can be used through the Internet,

and users can use cloud services scalable if they pay

metered rates to the cloud provider. However, in

public clouds, it is necessary to leave the data with the

cloud provider (albeit temporarily) during processing

jobs, which generates some security concerns.

Using private clouds can solve these problems.

A private cloud is a cloud that is built using

resources that users already have. The user can

construct the cloud taking security into account.

However, private clouds lack scalability relative to

public clouds.

Hybrid clouds can address the shortcoming of each

of the cloud types. These clouds can be both secure

and scalable. In this research, which is focused on

hybrid cloud environments, we proposed a method for

ensuring efficient processing.

2.2 The Trade-off between Cost of the Evaluation Axis

for Hybrid Clouds

When people use hybrid cloud environments, there

will be a trade-off relationship between performance

and necessary costs. When we want to process a large

amount of data more rapidly, using many resources

that are provided by the public cloud, in addition to

the resources of the private cloud, and the associated

metered cost will make the job more expensive. In

contrast, if these jobs are processed using private

cloud resources with little or no use of public cloud

resources, users will not have to pay metered costs,

but the job execution time will be longer. Thus, we

need a system that can determine optimal job

placement based on the equilibrium between

necessary cost and performance to ensure efficient

processing in a hybrid cloud environment.

This research proposes a method of providing

optimal job placement in hybrid cloud environments

in terms of monetary costs and performance. This

method has been operationalized as middleware. This


733

middleware will consider job processing time and

monetary costs. Monetary cost is the sum of the

charge for power consumption in private clouds and

the metered costs associated with public clouds. The

recent environmentalism in global affairs makes it

especially important to reduce power consumption

when processing large amounts of data. It is important

that we not waste power. In addition, the monetary

cost of private clouds may include fixed costs

associated with the system installation. However, such

amounts are difficult to define categorically. Thus, we

assume that the equipment has already been

depreciated, and the fixed costs were evaluated as

zero.

The monetary cost does not include the fee for the

power consumption associated with the public cloud.

This is because it is difficult for the user to know

the price of the power consumption by each resource

in the public cloud. For the public cloud, the power

consumption charges are assumed to be included in

the metered costs.

2.3 Eucalyptus

Eucalyptus [1] is open source software that can

create cloud infrastructure. Eucalyptus is compatible

with the Amazon EC2 API [2]; the Amazon EC2

(Amazon elastic compute cloud) is a cloud service

that is provided by the U.S. company Amazon.com.

Using a cloud built in Eucalyptus, you can port a

service on this cloud as if the service was on Amazon

EC2. Fig. 1 shows the architecture of Eucalyptus.

Eucalyptus is composed of three components. It is

treated as a public network to the upper layers of the

CLC (cloud controller) from the CLC (cluster

controller), and it is treated as a private network to the

lower layers of the NC (node controller).

CLC (cloud controller)

manages the information in the entire cloud. Equipped

with compatible interfaces for Amazon EC2; a web

management screen provides an API for the user.

CC (cluster controller)

manages the node controller, the state of instances

(virtual machines) and the virtual network for the

instances.

NC (node controller)

controls the instance. When the program needs to run

multiple instances, the virtualization software runs on

the node controller.

2.4 Building a Hybrid Cloud Environment

In this paper, we have used the Eucalyptus to build

two cloud systems. By connecting with Dummynet to

generate an artificial delay between them, we have

built an emulated hybrid cloud environment in their

laboratory. In the configuration of server in

Eucalyptus, it is possible to run CC and CLC in a

single server. In this study, in each cloud, CC and

CLC operated in a single server as Frontend Server.

Number of node server running in NC, as shown in

Fig. 2, was four in each cloud.

Servers that constitute this hybrid cloud are shown

in Tables 1-4.

Fig. 1 Architecture of Eucalyptus.

Fig. 2 Hybrid cloud environment.


734

Table 1 Private cloud frontend.

OS Linux 2.6.38 / / Debian GNU / Linux 6.0

CPU Intel® Xeon® CPU @ 3.60GHz 1 core

Memory 4 GByte

Table 2 Private cloud node.

OS Linux 2.6.32-xen-amd 64 and xen-4.0-amd64 / / Debian GNU / Linux 6.0


Memory 8 GByte

Table 3 Public cloud frontend.

OS Linux 2.6.38 / / Debian GNU / Linux 6.0


Memory 1 GByte

Table 4 Public cloud node.

OS Linux 2.6.32-xen-amd 64 and xen-4.0-amd64 / / Debian GNU / Linux 6.0


Memory 4 GByte

3. Proposed Method for Determining the Load

The proposed middleware in this paper processes

not only CPU-intensive applications but also

data-intensive applications. In both these jobs in a

hybrid cloud environment, to obtain high-speed and

low-cost processing after all of the resources have

been used in the private cloud, the next tasks should

be processed using in the public cloud. In addition, in

the public cloud, after the borrowed resources have all

been used, new resources will be needed. Therefore,

even when it is important to utilize resources without

waste, it is also important to properly determine the

load, and all of each resource should be used.

Therefore, when used for CPU processing and disk

processing, this middleware determines when the

resource has been saturated. Based on this information,

this middleware will determine the resource load. The

methods of determining the load for each type of

processing are as follows.

3.1 Method for Determining the Load of CPU

Processing

Load balancing for CPU-intensive jobs has been

investigated in many past studies. In this research,

CPU usage is the focus. This proposed middleware

also determines the load based on CPU usage. This

method is same as that used in other studies; if the

usage reaches 100%, the resource has been saturated,

and the middleware does the load balancing. The

method of optimally balancing CPU-intensive jobs is

not a feature of this proposal because it does not

fundamentally change the techniques used in other

studies.

3.2 Method for Determining the Disk Processing Load

3.2.1 Disk Performance Measurement

Unlike in CPU intensive-jobs, it is difficult to make

a definitive decision about whether the disk load has

reached the saturation point during data-intensive jobs.

For data-intensive jobs, because the system is often

waiting for I/O processing, it is difficult to determine

the CPU load. Thus, in the proposed method, we use

each cloud resources /proc/disk stats file to obtain the

length of the queue for the current disk. Then, we

estimate the number of jobs that are running in these

disks. Therefore, in this method, it is also necessary to

know the length of the queue, which indicates the

saturation of the disk resources.

Therefore, we have created benchmarks that can

change the balance of I/O and CPU processing.

By using the benchmark Disk Bench, which

performs read-only processing, we measured disk

performance using the execution environment of the

middleware.

In Fig. 3, as an example of a job by Disk Bench,

people can see a state of transition for the CPU load

and the number of disk accesses. Disk Bench is a

simple benchmark that performs read processing for

the disks in the instance. This figure shows that Disk

Bench processing is not performed when there is little

CPU processing and will become I/O bound if many

jobs are processed at the same time.

Using Disk Bench, we have measured the

performance of the disk. We have made this performance


735

Fig. 3 One example of the load transition of Disk Bench.

measurement for the instances of performance in

Table 5 for the hybrid cloud. In this measurement

process, we measured the execution time for the jobs

and the queue that is accumulated for the disk during

the processing of multiple simultaneous jobs using

Disk Bench. We then compare the processing time

when these jobs are processed sequentially and the

processing time using this measurement.

In general, if there are sufficient disk resources,

simultaneously processing the jobs can be more rapid

than sequentially processing them. However, when we

increase the number of jobs to be processed

simultaneously, there is a point at which processing

time will be slow as a result. This method determines

when there are no more disk resources, and the length

of the queue that has accumulated for the disk at that

time is defined as the “conditions in which the disk

resources have run out”.

Furthermore, in Disk Bench, there are two

parameters. One of parameters specifies the amount of

reading at a particular time, and the other specifies the

number of times this reading has been repeated. In this

performance measurement process, we create a job

that we intended to access a variety of patterns using

the disk; these parameters were varied. We have

measured performance changing these parameters.

Figs. 4 and 5 show examples of the comparison

results for processing times and the length of the

queue at the time of this performance measurement

experiments.

First, in this performance measurement process, we

compare the processing time for simultaneous

processing and sequential processing. In Fig. 4, the

vertical axis represents a ratio that indicates the

Table 5 Instance.

OS Linux 2.6.27-21-0.1-xen / x86_64 GNU / CentOS 5.3

CPU Intel® Xeon® CPU @ 3.60 GHz 1 core

Memory 1 Gbyte

Disk 20 Gbyte

Fig. 4 One example of comparison of processing times (block size: 512 Kbytes, repetition rate: 1,024 times).

Fig. 5 One example of the length of the queue (block size: 512 Kbytes, repetition rate: 1,024 times).

comparison results for the processing time. This ratio

was obtained by dividing the sequential processing

time into the simultaneous processing time. The

vertical axis at the value of 1 is indicated by a red

dashed line. If the value is below the dashed line, then

simultaneous processing is faster than sequential

processing. If the value is above the dashed line, then

sequential processing is faster than simultaneous

processing. In other words, the disk resource has been

exhausted. Thus, in Fig. 4, people can see that in job

number 4, the disk resource was exhausted.

Fig. 5 shows the transitions in the length of the

queue for each number of concurrent jobs at this

measurement. As shown in Fig. 5, if the number of

concurrent jobs is increased, the length of the queue is

increased. For this state, people can use the following

queuing model: The disk access requests from

multiple jobs arrive at random, the processing time for

the job is nearly constant, block size is constant and

the window for each disk is one. Therefore, the degree


736

of congestion of I/O requests from the job, that is, the

length of the queue, accurately reflects the degree of

saturation of the input and output.

By analyzing the relative processing time and the

length of the queue with some parameter setting, in

this experimental environment, we found that the

lengths of the queues are between 2,000 and 2,700

when the disk resources run out. However, clearly,

this is a range of values. If we analyze the physical

disk in detail, these values may be uniquely

determined. However, in general, the accuracy of the

actual job will not be exact. Therefore, in this method,

we determine this range as the saturated disk load.

We discuss their preliminary experiments in the

next section.

3.2.2 Preliminary Experiments: Experiment in

Controlling Load Balancing

In our preliminary experiments, we process

data-intensive jobs using Disk Bench in this hybrid

cloud environment. In this experiment, as their

threshold for load balancing, we use the length of the

queue for the disk resource. Using this threshold, by

choosing a value in the range of values determined in

the performance measurements, we have examined the

evaluation of performance and cost.

In this experiment, we received Disk Bench jobs

every 2 seconds 100 times. These experiments were

load balancing experiments intended to determine

where to place jobs: whether in private clouds or

public clouds. For this experiment in hybrid cloud

environments, we ran eight instances with

performance as indicated in Table 5: four instances for

each cloud. In addition, this range of values for the

length of the queue (which was determined by

measuring performance) depends on the physical

machine. However, because all of the physical servers

built as hybrid cloud environments had the same

performance, this range is unified at the

above-mentioned value.

First, jobs are placed in one instance in a private

cloud. If the length of the queue for that instance is

equal to or greater than the threshold value, the next

jobs will be distributed under the conditions that the

length of the queue is less than this threshold or that

has not been used within the private cloud. If all of the

queue lengths in a private cloud are equal to or greater

than the threshold value, the public cloud begins to be

used. Then, the next jobs are similarly distributed until

the queue length is greater than or equal to the

threshold.

In this experiment, as the threshold for load

balancing, we varied this value from a small value to

large value from 500 to 12,500. By using a value

within the range of values obtained in the performance

measurement process, we verified whether this load

distribution could provide an optimal balance between

monetary costs and performance as described in

Section 2.2. During this experiment, we measured the

processing time for the jobs, the power consumption

rate for the private cloud and the metered rates for the

public cloud. To measure power consumption in this

environment, we have used a watt-hour meter

SHW3A [3], which is a high-precision power meter

produced by the System Artware Company in Japan.

After one plugs an electric product into the SHW3A,

the power consumption is instantly measured and

displayed. In this study, we measure only the private

cloud’s node power consumption.

Fig. 6 shows the evaluation results for the

experiment.

The horizontal axis is the processing time cost, and

the vertical axis is the monetary cost. Monetary costs

are calculated using the following equation:

Monetary cost: TR · NR · CR + PL · CL

Fig. 6 Evaluation results for experiment in controlling load balancing.


737

TR: execution time for public cloud (h);

NR: number of instances of use of public cloud;

CR: charges for public cloud use ($/hour);

PL: power consumption in a private cloud (kWh);

CL: charges for power consumption in a private

cloud ($/kWh).

In this evaluation, the metered unit price is $0.5

based on the price of Amazon EC2 and the unit price

of power consumption is set at $0.24 based on the

price charged by the Tokyo electric power company.

As shown in Fig. 6, there is no configuration that

optimally balances both time costs and monetary costs.

However, in selecting a threshold for load balancing,

if we choose a value from 2,000 to 2,700 based on the

performance measurement that indicates the saturation

of the disk resource, we find that load balancing can

be provided based on a Pareto optimal cost balance. In

other words, if we set the threshold near 2,000,

although efficiency will be ensured and the load

balancing will occur quickly, the monetary costs will

increase slightly. In contrast, if we set the threshold

near 2,700, while efficiency will be ensured, it will

take a little time to perform load balancing and ensure

a low monetary cost. The balance of time costs and

monetary costs should be based on the needs of the

user.

Thus, in these preliminary experiments, we could

not find a point that best balances time cost and

monetary cost because there is a range in which the

disk resource is exhausted. However, by setting a

threshold value in response to a user request within

this range, we found that a processing cost balance

can be obtained without wasting resources.

3.2.3 Method of Controlling the Load Distribution

in Disk Processing

Based on the discussion in Sections 3.2.1 and 3.2.2,

in the proposed method of load determination for disk

processing, first, by measuring the performance of the

disk, we determine the range for queue length that

indicates disk saturation. This phase is regarded as a

learning phase. The threshold for load balancing in

middleware is the length of the queue for the disk

resource, and the user can select a threshold within

that range, which is determined by the performance

measurement process. This middleware can be used to

control the Pareto optimal cost balance load

distribution without wasting resources.

4. The Pareto Optimal Job Allocation Middleware

4.1 The Structure of the Middleware

Fig. 7 shows the behavior of the middleware. This

middleware consists of a dispatch unit and monitor

unit. The monitor unit in this middleware (for instance,

in hybrid clouds) uses the priority of job placements

to check the status of the resource on a regular basis.

As mentioned in the previous section, to check the status

Fig. 7 Behavior of this middleware.


738

of the resources requires measuring CPU utilization

for CPU processing and the length of the queue for

disk processing. In addition, the middleware evaluates

the load status of these resources, determining CPU

utilization and disk processing at the same time, and if

the processing becomes saturated, the middleware

determines that. The dispatch unit receives and

distributes jobs based on the information from the

monitor unit.

4.2 An Algorithm for Middleware

This middleware algorithm is as follows.

Additionally, when running this middleware in a

hybrid environment cloud for the first time, as

mentioned in Section 3.2.3, you must determine the

range of the queue length to identify disk resource

saturation.

(1) Based on the range of threshold values

determined in the learning phase, the user sets the

threshold for load balancing, which can be used to

obtain the desired cost balance, and runs the

middleware;

(2) Middleware receives the submitted job;

(3) In order of placement priority in private cloud

in-stances, check the load state of the resource to

determine whether it is greater than or equal to the

threshold. If the resource is at a value that is less than

the threshold value, execute the job using that instance,

and then return to (2). If the load states of all

resources in the private cloud are equal to or greater

than the threshold value, go to (4);

(4) In order of placement priority for public cloud

in-stances, check the load states of the resources to

determine whether they are greater than or equal to the

threshold. If the load state is less than the threshold

value, execute the job using that in-stance; then return

to (2). If all of the resource load states are equal to or

greater than the thresh-old value at that time, proceed

to (5);

(5) In the public cloud, select a new instance and

execute the submitted job using that instance; then

return to (2).

5. Experiments Using This Middleware

In this chapter, we describe examples of the results

of evaluations conducted using this middleware.

Conducting load balancing experiments using this

middleware for CPU-intensive applications in hybrid

cloud environments is not fundamentally different

from the process used in earlier studies of load

balancing. Therefore, in these examples, we consider

the evaluation results obtained for data-intensive

applications using this middleware.

5.1 Overview of Experiments

As shown in Figs. 2 and 7, as the experimental

environment for this middleware, we have built a

hybrid cloud environment. The node servers that

constitute each cloud are single-core CPUs, and all

servers have the same performance. For this reason,

we will generate one instance of performance from the

Table 5 for each node server. There are four instances

in the private cloud. If all instances are saturated, we

conduct load balancing using the public cloud

resources.

The experiment procedures are as follows:

First, in a learning phase, we measured the

performance of the disk. However, this experimental

environment is the environment in which the

performance measurement was carried out in Section

3.2. Therefore, for all instances in this hybrid cloud,

the queue length range that indicates disk saturation is

between 2,000 and 2,700. Next, based on this range,

we execute this middleware by varying the value of

the threshold for load balancing. During these

experiments, we measured the processing time for the

jobs, the cost of power consumption when the private

cloud was used and the metered rates for the public

cloud.

In this experiment, we evaluated these three types

of costs by varying the threshold for load balancing.

In particular, setting the threshold in the range


739

determined by the performance measurement process,

we evaluated whether Pareto optimal load balancing is

possible without wasting resources.

5.2 Data-Intensive Applications That Were Used in

the Experiment

In these two experiments, we have evaluated

middleware used with two different data-intensive

applications.

As the first, we used pgbench, which is the

PostgreSQL benchmark. Pgbench is a simple tool

benchmark that is bundled with PostgreSQL. Tatsuo

Ishii created the first version, published in 1999 by the

PostgreSQL mailing list in Japan. Pgbench was

created based on the TPC-B, which mimics the online

transaction process and can measure the number of

transactions that can be processed per unit of time. We

received pgbench’s jobs every two seconds 200 times;

the middleware processed these jobs.

For the second, we used queries from DBT-3.

DBT-3 is a simplified version of the TPC-H and

performs complex select statement queries in large

databases. The TPC-H and DBT-3 are decision

support benchmarks and consist of ad-hoc queries and

concurrent data modifications. In the DBT-3, there are

22 search queries. However, because the processing

time may be long because of the number of queries, in

this experiment, we select 11 queries for shorter

processing times. Then, we submitted these queries

repeatedly for a total of 110 jobs, and the middleware

processed the jobs. The DBT-3 database was built

using MySQL.

The major difference between these two types of

data intensive applications is the difference in the

CPU processing load. In executing pgbench jobs, we

confirmed that CPU processing is generally not

performed. In contrast, the search queries for DBT-3

were processed to some extent with the CPU.

However, all the search queries were mainly

executed using disk processing; thus, these are

data-intensive applications. For each of these two

data-intensive applications, using this proposed

method, we show that the method does not depend on

the nature and type of application.

5.3 Data Placement in Experiments

In a cloud environment, especially for

data-intensive jobs, considering data placement is

very important. For data placement in a hybrid

environment cloud, we can consider using the block

storage associated with each cloud or using remote

access to local storage from the public cloud. In these

experiments, it is assumed that due to remote backup,

the necessary data are already located in some

instances. In Ref. [4], using middleware, remote

access to the local storage is attained using iSCSI

from a public cloud. We wish to consider this method

of data placement in the future.

5.4 Evaluation of the Results Obtained Using

Middleware

5.4.1 An Example of the Use of Pgbench

Fig. 8 shows the results of the cost evaluation

obtained using the middleware, which processed

pgbench’s jobs.

As people can see from this figure, if we set the

correct threshold at the relevant queue length based on

the performance measurement process (i.e., between

2,000 and 2,700), this middleware provides a

Pareto-optimal cost balance and uses resources

efficiently. Conversely, some points are on a

Pareto-optimal curve even though they are out of the

range of values determined by the performance

measurement process.

These points are examples that indicate when the

load balance is too focused on processing performance

and too many resources are used or when a

tremendous burden has been placed on the available

resources so as not to raise the monetary cost. For

points that are not listed on the Pareto optimal curve

and that for values other than the threshold value, it is

possible that a better cost balance exists.


740

Fig. 8 Cost evaluation of processing pgbench jobs.

Fig. 9 Cost evaluation of processing DBT-3 queries.

Based on the above, we found that it is a necessary

condition for the point on the Pareto optimal curve to

set the threshold for load balancing based on the

saturation of the disk resource.

5.4.2 Example Using Search Queries of DBT-3

Fig. 9 shows the results of the cost evaluations

obtained using the middleware with DBT-3

processing queries.

In Fig. 9, as in Section 3.2.2, the vertical axis shows

the monetary cost, and the horizontal axis shows the

time cost. From this figure, as well as the DBT-3

processing queries, people can see that if we set the

correct threshold at the queue length determined by

the performance measurement process (i.e., from

2,000 from 2,700), this middleware provides a

Pareto-optimal cost balance in which resources are

used efficiently. Additionally, in other respects,

results similar to the ones found using pgbench are

obtained.

5.4.3 Observations from These Experiments

Based on the evaluation results for the processing

search queries for DBT-3 and pgbench as examples of

data-intensive applications, people can conclude that

this middleware provides a Pareto-optimal cost

balance while using resources efficiently when we set

the threshold to the queue length determined by the

performance measurement process.

In these experiments, we deliberately added a delay

of 20 msec by using Dummy net between the clouds.

However, because only certain jobs are transferred to

the remote cloud in this middleware, some of the

evaluations were barely influenced by the differences

in the delay time. In addition, because the unit price of

the metered cost for public cloud use was large, the

influence of the differences in power consumption in

the private cloud was limited. However, for technical

and social reasons, these monetary costs may vary

significantly. Even when the proposed method is used,

when the charge for power consumption is more

significant, this factor must be kept in mind during

load balancing. However, we can make this

modification by simply changing the cost calculation

expression.

6. Related Works

Previous researchers have discussed load balancing

in cloud computing [5, 6]. In these papers, however,

CPU-intensive applications were used as the targets of

load balancing jobs rather than data-intensive

applications. In computing-centric applications,

similarly to some scientific calculations, it is possible

to perform appropriate load balancing based on the

CPU load of each node. In this research, however, we

have used a data-intensive application for the jobs. In

such cases, because the CPU is often in the I/O

waiting state, load balancing is almost impossible

based on CPU load. In this research, we have used the

disk I/O as a load indicator. In data-intensive

applications, load balancing middleware has also been

developed that uses the amount of disk access for load

decisions [4]. This middleware based on disk access

provided dynamic load balancing between public

clouds and a local cluster and ensured optimal job


741

placement. Because we have further developed

middleware by introducing user-specified parameters,

it will be possible to reduce the monetary costs of load

balancing, including the cost of power consumption.

Power saving in cloud computing has also been

actively investigated. Unlike in this study, researchers

have discussed an approach to power saving that

involves the use of CPU-intensive applications in a

cloud [7]. Other studies examined power saving

efforts for a cloud datacenter [8, 9]. Our study aims to

save power in all clouds, including private clouds.

Researchers [10] proposed a scheduling algorithm that

could be used to evaluate power consumption and job

execution time. However, these studies differ from

their study, especially because we have focused on

total costs in hybrid clouds, including job execution

time, public cloud charges at a metered rate, and

power consumption charges for private clouds. In

addition, we have used data-intensive applications as

the target jobs.

7. Conclusions

We proposed a method of determining the load

based on the required cost and performance to ensure

efficient processing load balancing in a hybrid

environment cloud. We have implemented this

procedure using middleware. This middleware uses

information about CPU processing and disk

processing to provide efficient load balancing if the

resources needed to perform a job are scarce. In

particular, in the proposed method, we determine the

load from CPU usage and the length of the queue for

disk processing. First, in a learning phase, by

measuring the performance of the disk, we determine

the range of queue lengths that indicate disk

saturation.

The user can select a threshold within that range,

which is determined by the performance measurement

process. Using this middleware, tone can control the

Pareto optimal cost balance load distribution without

wasting resources.

Future research should be focused on improving

data placement. In the experiments in this paper, data

placement was the task of interest. This data

placement is not realistic as a model for real situations.

A realistic model might aggregate local and remote

storage and synchronize these forms of storage.

Therefore, we are considering introducing network

storage. In their system, iSCSI has already been

introduced, and we plan to carry out an experiment

using iSCSI in the future. In addition, in the current

implementation, based on the values obtained in the

learning phase, the threshold should be set for load

balancing before running the middleware. In the

future, we would like to develop an automating

learning phase as a part of this middleware.

Acknowledgments

This work is partly supported by the Ministry of

Education, Culture, Sports, Science and Technology,

under Grant 22,240,005 of Grant-in-Aid for Scientific

Research. We would like to thank to Drs. Atsuko

Takefusa, Hidemoto Nakada, Ryousei Takano,

Tomohiro Kudoh at the National Institute of AIST

(Advanced Industrial Science and Technology),

Project Associate Professor Miyuki Nakano, Assistant

Professor Daisaku Yokoyama, Senior Researcher

Norifumi Nishikawa at the Institute of Industrial

Science, the University of Tokyo, and Associate

Professor Saneyasu Yamaguchi at Kogakuin

University for the conscientious advice and help with

this work.

References

[1] Eucalyptus, http://www.eucalyptus.com/. [2] Amazon EC2, http://aws.amazon.com/jp/ec2/. [3] SHW3A, http://www.system-artware.co.jp/shw3a.html. [4] S. Toyoshima, S. Yamaguchi, M. Oguchi, Middleware

for load distribution among cloud computing resource and local cluster used in the execution of data-intensive application, Journal of the Database Society of Japan 10 (2011) 31-36.

[5] G. Jung, K.R. Joshi, M.A. Hiltunen, R.D. Schlichting, C. Pu, Generating adaptation policies for multi-tier applications in consolidated server environments, in: Proc.


742

5th IEEE International Conference on Autonomic Computing (ICAC2008), US, 2008.

[6] E. Kalyvianaki, T. Charalambous, S. Hand, Self-adaptive and self-configured CPU resource provisioning for virtualized servers using Kalman filters, in: Proc. 6th International Conference on Autonomic Computing and Communications (ICAC2009), Barcelona, Spain, 2009.

[7] C.Y. Tu, W.C. Kuo, W.H. Teng, Y.T. Wang, S. Shiau, A power-aware cloud architecture with smart metering, in: Proc. Parallel Processing Workshops (ICPPW), 2010 39th International Conference, San Diego, 2010.

[8] C. Peoples, G. Parr, S. McClean, Energy-aware data

centre management, in: Proc. Communications (NCC), 2011 National Conference, Bangalore, India, 2011.

[9] J. Baliga, R.W.A. Ayre, K. Hinton, R.S. Tucker, Green cloud computing: Balancing energy in processing, storage and transport, in: Proceedings of the IEEE, China, 2011.

[10] L.M. Zhang, K. Li, Y.Q. Zhang, Green task scheduling algorithms with speeds optimization on heterogeneous cloud servers, in: Proc. Green Computing and Communications (GreenCom), 2010 IEEE/ACM Int’l Conference on and Int’l Conference on Cyber, Physical and Social Computing (CPSCom), San Diego, California, 2010.


Probabilistic Health-Informatics and Bioterrorism

Ramalingam Shanmugam

School of Health Administration, Texas State University, San Marcos, TX 78666, USA

Received: April 22, 2013 / Accepted: May 07, 2013 / Published: June 30, 2013.

Abstract: In this article, new probabilistic health-informatics indices connecting probabilities: P r( ), P r( ), P r( )A B A B

and

P r( )A B are discovered, where A and B denote respectively the “ability of a hospital to treat anthrax patients” and “whether

a hospital drilled to be prepared to deal with an adverse bioterrorism”. These probabilistic informatics are not seen in any textbooks or journal articles and yet, they are too valuable to be unnoticed to comprehend the hospitals’ preparedness to treat anthrax patients in an outbreak of bioterrorism. A demonstration of this new probabilistic informatics is made in this article with the data in the U.S. Government’s General Accounting Office’s report GAO-03-924. Via this example, this article advocates the importance of the above mentioned probabilistic-informatics for health professionals to understand and act swiftly to deal with public health emergencies. Key words: Conditional, marginal, total probability, hospital’s drilling, inequalities.

1. Motivation

Historically, probability tools are used to

understand the regularity in the uncertain environment.

Probability is a foundation on which informatics about

the uncertainty could be built to address the cyber

security, the importance of the electronic health

records, the efficacy of the medical drugs, the

designing of biomedical engineering devices, and the

decision making in public health crises among others.

Many notable heroes in the humanity like the famous

astronomer Christian Huygens utilized probability

informatics in his quest for securing the physical laws

of the universe. See de Finetti [1] and Jaynes [2] for

the history and basics of probability information. Yet,

some important probability informatics does not

appear in any textbooks or journal articles. These

probability informatics inequalities are too important

to be unnoticed. They are first derived in this article

and then are applied to comprehend the hospital’s

Corresponding author: Ramalingam Shanmugam, Ph.D.,

research fields: informatics, modeling infectious diseases, diagnostic methodology and modeling. E-mail: [email protected].

preparedness to treat anthrax in an outbreak of

terrorism in USA.

In Section 2, three new probability based

informatics inequalities are derived; the new

inequalities are illustrated in Section 3 with the data

about the U.S. Hospitals’ preparedness to treat anthrax

patients in an event of bioterrorism; In Section 4, a

few final thoughts are given. In the demonstration, the

preparedness of US hospitals in 34 states to a survey

conducted by the CDC (center for the disease control)

in Report GAO-03-924 by the federal Government

Accounting Office are analyzed and interpreted. The

remaining sixteen states in the USA did not respond to

the CDC’s survey; in the end, some conclusive

thoughts for future research work probability

informatics are included.

2. Probability-Informatics Inequalities

In this section, we first derive three new probability

informatics inequalities and use them later in Section

3 to comprehend how prepared are the states in the

USA to treat anthrax patients who might flock into a

hospital in an outbreak of a bioterrorism. To begin


744

with, let us introduce the following notations:

_ _ _A hospitals Can treat anthrax

_ _B hospital drilled bioterrorism

_ _ _B hospital not drilled bioterrorism

Among other illnesses, an infliction of anthrax

spores could turn out fatal, if it is not treated quickly

and it might happen in an outbreak of bioterrorism.

The public health agencies need to have worked out

the “informatics” early on before the occurrence of the

bioterrorism. There are uncertainties about the

hospitals’ readiness in a state. Had the hospitals

“drilled” on their emergency plans, it would help to

quickly treat anthrax patients. Since the terrorism on

September 11, 2001 in the New York City, several

hospitals in the USA and elsewhere are worried about

the likelihood of treating massive influx of anthrax

patients who might arrive for treatment in an event of

bioterrorism. Many hospitals, if not all the hospitals,

do drill and practice with preparedness plan to deal

bioterrorism including the treating effectively the

anthrax patients. In this regard, let Pr( )A be the

probability that a hospital is prepared to treat the

anthrax patients, whether or not the hospital has

drilled to deal any bioterror event. Realize that the

event A can happen with or without an associated

event B . Suppose that Pr( )B be the probability

that the hospital has drilled to deal bioterrorism.

Consequently, Pr( ) Pr( )

Pr( ) Pr( ) [1 Pr( )] Pr( )

A ABUAB

B A B B A B

.

The above statement can be rearranged as: Pr( ) Pr( ) Pr( )

Pr( )1 Pr( )

A B A BA B

B

(1)

The Eq. (1) means that the conditional probability

P r( )A B for a hospital to treat anthrax patients even

without a drilling experience is comparable to the

conditional probability P r( )A B for a hospital to

treat anthrax patients with a drilling experience. In this

comparison, the impact of drilling on the hospital’s

ability to treat anthrax patients can be felt and it is

brought out in our demonstrations. However, there

could be three possibilities in the scenario to capture

the impact of drilling and they are be felt and it is

brought out in our demonstrations. Now, let us discuss

one after another possibility.

The first possibility is that the cure of anthrax

patients could be certain because of the drilling and it

is echoed by the conditional probability P r( ) 1A B .

In this Scenario 1, by substituting P r( ) 1A B in

Eq. (1), the Eq. (1) transforms to Eq. (2) below:

Pr( ) [Pr( ) Pr( )] /[1 Pr( )]A B A B B (2)

Notice that Pr( ) Pr( )0 1

1 Pr( )

A B

B

, because the

probability P r( )A B has to obey an axiom

0 P r( ) 1A B of the probability. Consequently,

the probability informatics in Eq. (3) exists.

0 Pr( ) Pr( ) 1B A (3)

The second possibility is that the cure of anthrax

patients is unlikely in spite of the drilling and it is

indicated by the conditional probability

Pr( ) 0A B . In this Scenario 2, with the substitution

of Pr( ) 0A B in Eq. (1), the Eq. (1) transforms to

Eq. (4) below:

P r( ) P r( ) /[1 P r( )]A B A B (4)

Consequently, another probability informatics in Eq.

(5) is noticed.

0 Pr( ) Pr( ) 1B A (5)

The third possibility is that the cure of an anthrax

patient is likely but not surely or unlikely and it is

indicated by the conditional probability statement

0 Pr( ) 1A B . It is so because the drilling helps

but does not necessarily guaranty the successful


745

treatment of anthrax patients. In this Scenario 3, the

Eq. (1) implies a probability informatics in Eq. (6). P r ( ) P r ( ) P r ( )

1 P r( ) [1 P r( ) ]

B A B A

B A B

(6)

The probability for any hospital in a US state to be

prepared to treat anthrax patients is indicated by one

of the mutually exclusive probability informatics Eqs.

(3), (5) and (6) depending on the prevailing scenario

in the state’s hospitals. However, there is way to

visualize the probability informatics Eqs. (3), (5) and

(6) in a cubic box.

For this purpose, we sketch probability informatics:

Eqs. (3), (5) and (6) with the coordinate Pr( )x A ,

P r( )y B and P r( )z A B . The unit volume is

partitioned into mutually exclusive scenarios as seen

in the Fig. 1. Each state’s in the USA would fall in

only one scenario depending on its x, y and z

coordinates. Lastly, the importance of drilling to treat

anthrax patients can be captured by plotting

P r( )A B in terms of P r( )A B depending on

whether the Scenario 3, Scenario 2 or Scenario 1

prevails. The correlation between Pr( )A B and

P r( )A B portray a relationship between the presence

and absence of the drilling to treat anthrax patients.

Next, all these probability informatics are illustrated.

3. Preparedness for Anthrax Patients

In this section, the results of the Section 2 are

illustrated using 34 states’ data in the report

GAO-03-924 by the US General Accounting Office.

Fig. 1 Compartments in the unit cube due to probability informatics Eqs. (3), (5) and (6).

Fig. 2 (a) Eastern zone.


746

Fig. 2 (b) Central zone.

Fig. 2 (c) Mountain zone.

Fig. 2 (d) Pacific zone.


747

Table 1 The correlation between Pr( )A B and Pr( )A B in four time zones.

_Tim e Z one

C orr

Eastern zone Central zone Mountain zone Pacific zone

[Pr( ), Pr( )]Corr A B A B 0.43 0.12 -0.86 -0.98

The remaining sixteen states in the USA did not

respond to the survey which was conducted by the

CDC. For the sake of comparisons on how the states

perform, the thirty four states are grouped according

to their time zones. The rationale for grouping is that

the states are homogeneous within a time zone but not

across the time zones. The preparedness and drilling

data for the fifteen states in the Eastern Time zone, the

thirteen states in the Central Time zone, the three

states in the Mountain Time zone and the three states

in the Pacific Time zone are calculated and compared

using Pr( ),Pr( ),Pr( )A B A B and Pr( )A B as in

Figs. 2a-2d below. The correlation between Pr( )A B

and Pr( )A B are calculated and displayed in the

Table 1. The correlation captures the intricate balance

between the presence and absence of the drilling on

the ability to treat anthrax patients.

4. Conclusions

The states in the Eastern and Central time zones are

smaller in size in comparison to the states in the

Mountain or Pacific Time zones. Consequently, the

number of states is lesser in the Mountain and Pacific

Time zones. However, the pattern in the Eastern time

zone is different from the pattern in the Central time

zone and it is noticed by comparing the Fig. 2a with

the Fig. 2b. Likewise, the pattern in the Mountain

Time zone is different from the pattern in the Pacific

Time zone and it is realized by comparing the Fig. 2c

with the Fig. 2d. The preparedness in the Pacific states

is more collinear (see Fig. 2d). The preparedness in

the Mountain states is not collinear (see Fig. 2c). The

preparedness in the Central Time zone (see Fig. 2b) is

more diversified in comparison to the preparedness in

the Eastern Time zone (see Fig. 2a). Furthermore,

their correlation between Pr( )A B and Pr( )A B

reveals interesting similarities and differences among

the 34 states in the four time zones of USA (see Table

1). The correlation between Pr( )A B and Pr( )A B

is highest in the Eastern states, gradually decreasing

as one moves westward, and becomes the lowest in

the Pacific states. Surprisingly, the correlation

between Pr( )A B and Pr( )A B is negative only in

the Mountain and Pacific states. More research

investigations for the non-trivial reasons of such

remarkable differences among these time zones ought

to be done in a future project. The findings will be

valuable to the public health administrators. This new

informatics knowledge is made possible with the help

of new probability informatics in Eqs. (3), (5) and (6)

which involve the marginal, complementary and

conditional probabilities.

References

[1] B.D Finetti, Logical foundations and measurement of

subjective probability, Acta Psychologica 34 (1970)

129-145.

[2] E.T. Jaynes, Probability Theory: The Logic of Science,

Cambridge University Press, UK, 2003.

[3] Hospital Preparedness: Most Urban Hospitals Have

Emergency Plans but Lack Certain Capacities for

Bioterrorism Response Report GAO-03-924, Government

Accounting Office, Federal Government Press,

Washington, 2003.


A Hybrid Method to Improve Forecasting Accuracy—An

Application to the Canned Cooked Rice and the Aseptic

Packaged Rice

Hiromasa Takeyasu1, Daisuke Takeyasu2 and Kazuhiro Takeyasu3

1. Faculty of Life and Culture, Kagawa Junior College, Kagawa 769-0201, Japan

2. Graduate School of Culture and Science, The Open University of Japan, Chiba City 261-8586, Japan

3. College of Business Administration, Tokoha University, Fuji City 417-0801, Japan

Received: March 24, 2013 / Accepted: April 17, 2013 / Published: June 30, 2013.

Abstract: In industries, how to improve forecasting accuracy, such as sales, shipping, is an important issue. In this paper, a hybrid method is introduced and plural methods are compared. Focusing that the equation of ESM (exponential smoothing method) is equivalent to (1, 1) order ARMA Model (autoregressive moving average model) equation, new method of estimation of smoothing constant in exponential smoothing method is proposed before by us which satisfies minimum variance of forecasting error. Trend removing by the combination of linear and 2nd order non-linear function and 3rd order non-linear function is executed to the original production data of two kinds of cooked rice (canned rice and aseptic packaged rice). Genetic algorithm is utilized to search the optimal weight for the weighting parameters of linear and non-linear function. For the comparison, monthly trend is removed after that. The new method shows that it is useful for the time series that has various trend characteristics and has rather strong seasonal trend.

Key words: Minimum variance, exponential smoothing method, forecasting, trend, bread.

1. Introduction

Many methods for time series analysis have been

presented such as AR Model (autoregressive model),

ARMA Model (autoregressive moving average model)

and ESM (exponential smoothing method) [1-4].

Among these, ESM is said to be a practical simple

method.

For this method, various improving methods, such

as adding compensating item for time lag, coping with

the time series with trend [5], utilizing Kalman Filter

[6], Bayes Forecasting [7], adaptive ESM [8],

exponentially weighted moving averages with

irregular updating periods [9], making averages of

forecasts using plural method [10], are presented. For

Corresponding Author: Hiromasa Takeyasu, professor, research field: time series analysis. E-mail: [email protected].

example, Maeda [6] calculated smoothing constant in

relationship with S/N ratio under the assumption that

the observation noise was added to the system. But he

had to calculate under supposed noise because he

could not grasp observation noise. It can be said that it

does not pursue optimum solution from the very data

themselves which should be derived by those

estimation. Ishii [11] pointed out that the optimal

smoothing constant was the solution of infinite order

equation, but he did not show analytical solution.

Based on these facts, we proposed a new method of

estimation of smoothing constant in ESM before [12].

Focusing that the equation of ESM is equivalent to

(1, 1) order ARMA model equation, a new method of

estimation of smoothing constant in ESM was

derived.

In this paper, utilizing above stated method, a

A Hybrid Method to Improve Forecasting Accuracy—An Application to the Canned Cooked Rice and the Aseptic Packaged Rice

749

revised forecasting method is proposed. In making

forecast such as production data, trend removing

method is devised. Trend removing by the

combination of linear and 2nd order non-linear

function and 3rd order non-linear function is executed

to the original production data of two kinds of cooked

rice (canned rice and aseptic packaged rice). Genetic

algorithm is utilized to search the optimal weight for

the weighting parameters of linear and non-linear

function. For the comparison, monthly trend is

removed after that. Theoretical solution of smoothing

constant of ESM is calculated for both of the monthly

trend removing data and the non-monthly trend

removing data. Then forecasting is executed on these

data. This is a revised forecasting method. Variance of

forecasting error of this newly proposed method is

assumed to be less than those of previously proposed

method.

The rest of the paper is organized as follows: In

Section 2, ESM is stated by ARMA model and

estimation method of smoothing constant is derived

using ARMA model identification; The combination

of linear and non-linear function is introduced for

trend removing in Section 3; The Monthly ratio is

referred in Section 4; Forecasting is executed in

Section 5, and estimation accuracy is examined.

2. Description of ESM Using ARMA Model

In ESM, forecasting at time t + 1 is stated in the

following equation:

tttt xxxx ˆˆˆ 1 (1)

tt xx ˆ1 　　　　 (2)

Here,

:ˆ 1tx forecasting at 1t ;

:tx realized value at t ;

: smoothing constant 10 ;

Eq. (2) is re-stated as:

ltl

lt xx

01 1ˆ (3)

By the way, we consider the following (1, 1) order

ARMA model:

11 tttt eexx (4)

Generally, qp, order ARMA model is stated

as:

jt

q

jjtit

p

iit ebexax

11

(5)

Here,

:tx Sample process of Stationary Ergodic

Gaussian Process tx ,,,2,1 Nt ;

te : Gaussian White Noise with 0 mean 2e

variance; MA process in Eq. (5) is supposed to satisfy

convertibility condition. Utilizing the relation that

0,, 21 ttt eeeE

We get the following equation from Eq. (4):

11ˆ ttt exx (6)

Operating this scheme on t + 1, we finally get:

ttt

ttt

xxx

exx

ˆ1ˆ

1ˆˆ 1

(7)

If we set 1 , the above equation is the same

with Eq. (1), i.e., equation of ESM is equivalent to (1,

1) order ARMA model, or is said to be (0, 1, 1) order

ARIMA model, because 1st order AR parameter is

1 . Compared with Eqs. (4) and (5), we obtain

1

1 1

b

a

From Eqs. (1)-(7), 1

Therefore, we get:

1

1

1

1

b

a (8)

From above, we can get estimation of smoothing

constant after we identify the parameter of MA part of

ARMA model. But, generally MA part of ARMA

model becomes non-linear equations which are

described below.

Let Eq. (5) be:

it

p

iitt xaxx

1

~ (9)


750

jt

q

jjtt ebex

1

~ (10)

We express the autocorrelation function of tx~ as

r k~

and from Eqs. (9) and (10), we get the following non-linear equations which are well known.

q

jje

jk

kq

jjek

br

qk

qkbbr

0

220

0

2

~

)1(0

)(~

(11)

For these equations, recursive algorithm has been

developed. In this paper, parameter to be estimated is

only b1 , so it can be solved in the following way.

From Eqs. (4), (5), (8) and (11), we get:

2

11

2210

1

1

~1~

1

1

1

e

e

br

br

b

a

q

(12)

If we set

0~

~

r

rkk (13)

the following equation is derived:

21

11 1 b

b

(14)

We can get b1 as follows:

1

21

1 2

411

b (15)

In order to have real roots, 1 must satisfy:

2

11 (16)

From invertibility condition, 1b must satisfy:

11 b

From Eq. (14), using the next relation:

01

012

1

21

b

b

Eq. (16) always holds.

As

11 b

1b is within the range of

01 1 b

Finally we get:

1

211

1

21

1

2

4121

2

411

b

(17)

which satisfied above conditions. Thus we can obtain

a theoretical solution by a simple way. Focusing on

the idea that the equation of ESM is equivalent to (1,

1) order ARMA model equation, we can estimate

smoothing constant after estimating ARMA model

parameter. It can be estimated only by calculating 0th

and 1st order autocorrelation function.

3. Trend Removal Method

As trend removal method, we describe the

combination of linear and non-linear function.

(1) Linear function

We set

11 bxay (18)

as a linear function.

(2) Non-linear function

We set

222

2 cxbxay (19)

332

33

3 dxcxbxay (20)

as a 2nd and a 3rd order non-linear function. ),,( 222 cba and ),,,( 3333 dcba are also parameters

for a 2nd and a 3rd order non-linear functions which

are estimated by using least square method.

(3) The combination of linear and non-linear

function.

We set

　33

23

333

222

22111

dxcxbxa

cxbxabxay

(21)

1,10,10,10 321321 (22)

as the combination linear and 2nd order non-linear

and 3rd order non-linear function. Trend is removed

by dividing the original data by Eq. (21). The optimal


751

weighting parameter 321 ,, are determined by

utilizing GA. GA method is precisely described in

Section 6.

4. Monthly Ratio

For example, if there is the monthly data of L years

as stated bellow:

12,,1,,1 jLixij where, Rxij ,

in which j means month and i

means year and ijx is a production data of i -th year,

j -th month. Then, monthly ratio jx~ 12,,1j is

calculated as follows:

L

i jij

L

iij

j

xL

xL

x

1

12

1

1

1211

1

~ (23)

Monthly trend is removed by dividing the data by

Eq. (23). Numerical examples both of monthly trend

removal case and non-removal case are discussed in

Section 7.

5. Forecasting Accuracy

Forecasting accuracy is measured by calculating the

variance of the forecasting error. Variance of

forecasting error is calculated by:

N

iiN 1

22

1

1 (24)

where, forecasting error is expressed as:

iii xx ˆ (25)

N

iiN 1

1 (26)

6. Searching Optimal Weights Utilizing GA

6.1 Definition of the Problem

We search 321 ,, of Eq. (21) which minimizes

Eq. (24) by utilizing GA. By Eq. (22), we only have to

determine 1 and 2 . 2 (Eq. (24)) is a function

of 1 and 2 , therefore we express them as

),( 212 . Now, we pursue the following:

Minimize: ),( 212

subject to: 1,10,10 2121 　　　　 (27)

We do not necessarily have to utilize GA for this

problem which has small member of variables.

Considering the possibility that variables increase

when we use logistics curve, etc., in the near future,

we want to ascertain the effectiveness of GA.

6.2 The Structure of the Gene

Gene is expressed by the binary system using 0, 1

bit. Domain of variable is [0, 1] from Eq. (22). We

suppose that variables take down to the second

decimal place. As the length of domain of variable is

1 – 0 = 1, seven bits are required to express variables.

The binary bit strings (bit 6, ～, bit 0) is decoded to

the [0, 1] domain real number by the following

procedure [13].

Procedure 1: Convert the binary number to the

binary-coded decimal:

X

bit

bitbitbitbitbitbitbit

i

ii

10

6

0

20123456

2

,,,,,,

(28)

Procedure 2: Convert the binary-coded decimal to

the real number:

The real number = (Left hand starting point of the

domain) + 'X ((Right hand ending point of the

domain)/( 127 )) (29) The decimal number, the binary number and the

corresponding real number in the case of 7 bits are

expressed in Table 1.

1 variable is expressed by 7 bits, therefore 2

variables needs 14 bits. The gene structure is

exhibited in Table 2.

6.3 The Flow of Algorithm

The flow of algorithm is exhibited in Fig. 1.

6.3.1 Initial Population

Generate M initial population. Here, 100M .

Generate each individual so as to satisfy Eq. (22).


752

Table 1 Corresponding table of the decimal number, the binary number and the real number.

The decimal number

The binary number

The Corresponding real number Position of the bit

6 5 4 3 2 1 0

0 0 0 0 0 0 0 0 0.00

1 0 0 0 0 0 0 1 0.01

2 0 0 0 0 0 1 0 0.02

3 0 0 0 0 0 1 1 0.02

4 0 0 0 0 1 0 0 0.03

5 0 0 0 0 1 0 1 0.04

6 0 0 0 0 1 1 0 0.05

7 0 0 0 0 1 1 1 0.06

8 0 0 0 1 0 0 0 0.06

… …

126 1 1 1 1 1 1 0 0.99

127 1 1 1 1 1 1 1 1.00

Table 2 The gene structure.

1 2

Position of the bit

13 12 11 10 9 8 7 6 5 4 3 2 1 0

0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1 0-1

Fig. 1 The flow of algorithm.

6.3.2 Calculation of Fitness

First of all, calculate forecasting value. There are 36

monthly data for each case. We use 24 data (1st to

24th) and remove trend by the method stated in

Section 3. Then we calculate monthly ratio by the

method stated in Section 4. After removing monthly

trend, the method stated in Section 2 is applied and

exponential smoothing constant with minimum

variance of forecasting error is estimated. Then the 1st

step forecast is executed. Thus, data is shifted to 2nd

to 25th and the forecast for 26th data is executed

consecutively, which finally reaches forecast of 36th

data. To examine the accuracy of forecasting, variance

of forecasting error is calculated for the data of 25th to

36th data. Final forecasting data is obtained by

multiplying monthly ratio and trend. Variance of

forecasting error is calculated by Eq. (24). Calculation

of fitness is exhibited in Fig. 2.

Scaling [14] is executed such that fitness becomes

large when the variance of forecasting error becomes

small. Fitness is defined as follows:

),(),( 212

21 Uf (30)

where U is the maximum of ),( 212 during the

past W generation. Here, W is set to be 5.

6.3.3 Selection

Selection is executed by the combination of the


753

Fig. 2 The flow of calculation of fitness.

general elitist selection and the tournament selection.

Elitism is executed until the number of new elites

reaches the predetermined number. After that,

tournament selection is executed and selected.

6.3.4 Crossover

Crossover is executed by the uniform crossover.

Crossover rate is set as follows:

7.0cP (31)

6.3.5 Mutation

Mutation rate is set as follows:

05.0mP (32)

Mutation is executed to each bit at the

probability mP , therefore all mutated bits in the

population M becomes 14 MPm .

7. Numerical Example

7.1 Application to the Original Production Data of

Processed Cooked Rice

The original production data of processed cooked

rice for two cases (data of canned rice and those of

aseptic packaged rice: Annual Report of Statistical

Research, Ministry of Agriculture, Forestry and

Fisheries, Japan) from January 2008 to December

2010 are analyzed. Furthermore, GA results are

compared with the calculation results of all

considerable cases in order to confirm the

effectiveness of GA approach. First of all, graphical

charts of these time series data are exhibited in Figs. 3

and 4.

7.2 Execution Results

GA execution condition is exhibited in Table 3. GA

convergence process is exhibited in Figs. 5-8.

We made 10 times repetition and the maximum,

average, minimum of the variance of forecasting error

and the average of convergence generation are

exhibited in Table 4 and 5.

The case that monthly ratio is not used is smaller

than the case that monthly ratio is used concerning the

variance of forecasting error for canned rice. It may be

because canned rice does not have definite seasonal

trend towards the data for 2008.

The minimum variance of forecasting error of GA

coincides with those of the calculation of all

considerable cases and it shows the theoretical

solution. Although it is a rather simple problem for

GA, we can confirm the effectiveness of GA approach.

Further study for complex problems should be

examined hereafter.

Next, optimal weights and their genes are exhibited

in Tables 6 and 7.

In the case that monthly ratio is not used, the linear

function model is best in both cases. In the case that

monthly ratio is used, the linear function model is best

in both cases. Parameter estimation results for the


754

Canned Cooked Rice

0

50

100

150

200

250

1 2 3 4 5 6 7 8 9 101112 1 2 3 4 5 6 7 8 9 101112 1 2 3 4 5 6 7 8 9 101112

2006 2007 2008

Fig. 3 Data of canned cooked rice.

Aseptic Packaged Cooked Rice

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

1 2 3 4 5 6 7 8 9 101112 1 2 3 4 5 6 7 8 9 101112 1 2 3 4 5 6 7 8 9 101112

2006 2007 2008

Fig. 4 Data of aseptic packaged cooked rice.

Table 3 GA execution condition.

GA Execution Condition

Population 100

Maximum generation 50

Crossover rate 0.7

Mutation ratio 0.05

Scaling window size 5

The number of elites to retain 2

Tournament size 2

Table 4 GA execution results (monthly ratio is not used).

Food The variance of forecasting error

Average of convergence generation Maximum Average Minimum

Canned cooked rice 4048.947313 3233.943529 3139.748713 9.4

Aseptic packaged cooked rice 815965.2277 684614.796 668794.6832 14.3

Table 5 GA execution results (monthly ratio is used).

Food The variance of forecasting error

Average of convergence generation Maximum Average Minimum

Canned cooked rice 5278.497079 3418.722411 3174.249743 13.2

Aseptic packaged cooked rice 323388.4005 214841.1023 206470.6901 9.7


755

Canned Cooked Rice

Minimum Variance

3110

3120

3130

3140

3150

3160

3170

3180

3190

3200

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Fig. 5 Convergence process in the case of canned cooked rice (monthly ratio is not used).

Canned Cooked Rice

Minimum Variance

3000

3100

3200

3300

3400

3500

3600

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Fig. 6 Convergence process in the case of canned cooked rice (monthly ratio is used).


Minimum Variance

667000

667500

668000

668500

669000

669500

670000

670500

671000

671500

672000

672500

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Fig. 7 Convergence process in the case of aseptic packaged cooked rice (monthly ratio is not used).


Minimum Variance

205500

206000

206500

207000

207500

208000

208500

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Fig. 8 Convergence process in the case of aseptic packaged cooked rice (monthly ratio is used).


756

Table 6 Optimal weights and their genes (monthly ratio is not used).

Data 1 2

Position of the Bit

13 12 11 10 9 8 7 6 5 4 3 2 1 0

Canned cooked rice 1 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0

Aseptic packaged cooked rice 1 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0

Table 7 Optimal weights and their genes (monthly ratio is used).

Data 1 2

Position of the Bit

13 12 11 10 9 8 7 6 5 4 3 2 1 0

Canned cooked rice 1 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0

Aseptic packaged cooked rice 1 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0

Table 8 Parameter estimation results for the trend of Eq. (21).

Data 1a 1b 2a

2b 2c 3a

3b 3c

3d

Canned cooked rice -0.38 130.17 0.24 -6.4 156.27 -0.004 0.39 -7.9 159.69

Aseptic packaged cooked rice 43.8 7410.01 0.69 26.45 7485.17 0.57 -20.85 246.28 6981.12

Fig. 9 Trend of canned cooked rice.

Fig. 10 Trend of aseptic packaged cooked rice.

Table 9 Parameter estimation result of monthly ratio.

Data 1 2 3 4 5 6 7 8 9 10 11 12

Canned cooked rice 1.38 1.39 1.43 0.98 0.69 0.68 0.43 0.87 0.6 0.98 1.23 1.3

Aseptic packaged cooked rice 0.89 0.94 1.1 1.06 0.92 0.98 1.01 1.0 0.98 1.04 1.0 1.07


757

Table 10 Smoothing constant of minimum variance of Eq. (17) (monthly ratio is not used). Data ρ1 α

Canned cooked rice -0.280995 0.692422

Aseptic packaged cooked rice -0.206225 0.784169

Table 11 Smoothing constant of minimum variance of Eq. (17) (monthly ratio is used).

Data ρ1 α

Canned cooked rice -0.475263 0.274765

Aseptic packaged cooked rice -0.302345 0.663399

Fig. 11 Forecasting result of canned cooked rice.

Fig. 12 Forecasting result of frozen cooked rice.

trend of Eq. (21) using least square method are

exhibited in Table 8 for the case of 1st to 24th data.

Trend curves are exhibited in Figs. 9 and 10.

Calculation results of monthly ratio for 1st to 24th

data are exhibited in Table 9.

Estimation results of the smoothing constant of

minimum variance for the 1st to 24th data are

exhibited in Tables 10 and 11.

Forecasting results are exhibited in Tables 10 and 11,

Figs. 11 and 12.

7.3 Remarks

In the case of canned rice, that monthly ratio was

not used had a better forecasting accuracy. On the

other hand, aseptic packaged rice had a better

forecasting accuracy in the case that monthly ratio

was used. Both cases had a good result in the

combination of linear and non-linear function model.

The minimum variance of forecasting error of GA

coincides with those of the calculation of all

considerable cases and it shows the theoretical


758

solution. Although it is a rather simple problem for

GA, we can confirm the effectiveness of GA approach.

Further study for complex problems should be

examined hereafter.

8. Conclusions

Focusing on the idea that the equation of ESM was

equivalent to (1, 1) order ARMA model equation, a

new method of estimation of smoothing constant in

exponential smoothing method was proposed before

by us which satisfied minimum variance of

forecasting error. Generally, smoothing constant was

selected arbitrary. But in this paper, we utilized above

stated theoretical solution. Firstly, we made estimation

of ARMA model parameter and then estimated

smoothing constants. Thus theoretical solution was

derived in a simple way and it might be utilized in

various fields.

Furthermore, combining the trend removal method

with this method, we aimed to improve forecasting

accuracy. An approach to this method was executed in

the following method. Trend removal by a linear

function was applied to the original production data of

two kinds of cooked rice (canned rice and aseptic

packaged rice). The combination of linear and

non-linear function was also introduced in trend

removal. Genetic algorithm was utilized to search the

optimal weight for the weighting parameters of linear

and non-linear function. For the comparison, monthly

trend was removed after that. Theoretical solution of

smoothing constant of ESM was calculated for both of

the monthly trend removing data and the non-monthly

trend removing data. Then forecasting was executed

on these data. The new method shows that it is useful

for the time series that has various trend

characteristics. The effectiveness of this method

should be examined in various cases.

References

[1] B. Jenkins, Time Series Analysis, 3rd ed., Prentice Hall, US, 1994.

[2] R.G. Brown, Smoothing, Forecasting and Prediction of Discrete–Time Series, Prentice Hall, US, 1963.

[3] H. Tokumaru, Analysis and Measurement–Theory and Application of Random data Handling, Baifukan Publishing, Tokyo, Japan, 1982.

[4] K. Kobayashi, Sales Forecasting for Budgeting, Chuokeizai–Sha Publishing, Japan, 1992.

[5] P.R. Winters, Forecasting sales by exponentially weighted moving averages, Management Science 6 (1984) 324-343.

[6] K. Maeda, Smoothing constant of exponential smoothing method, Seikei University Report Faculty of Engineering 38 (1984) 2477-2484.

[7] M. West, P.J. Harrison, Baysian Forecasting and Dynamic Models, Springer-Verlag, Germany, 1989.

[8] S. Ekern, Adaptive exponential smoothing revisited, Journal of the Operational Research Society 32 (1982) 775-782.

[9] F.R. Johnston, Exponentially weighted moving average (EWMA) with irregular updating periods, Journal of the Operational Research Society 44 (1993) 711-716.

[10] S. Makridakis, R.L. Winkler, Averages of forecasts;

some empirical results, Management Science 29 (1983) 987-996.

[11] N. Ishii, Bilateral exponential smoothing of time series, Int. J. System Sci. 12 (1991) 997-988.

[12] K.Takeyasu, K. Nagao, Estimation of smoothing constant of minimum variance and its application to industrial data, Industrial Engineering and Management Systems 7 (2008) 44-50.

[13] M. Sakawa, M. Tanaka, Genetic Algorithm, Asakura Publishing Co., Ltd., Japan, 1995.

[14] H. Iba, Genetic Algorithm, Igaku Publishing, Japan, 2002.


System Reconstruction in the Field-Specific

Methodology of Economic Subjects

Katarina Krpalkova Krelova1 and Pavel Krpalek2

1. Department of Economic Teaching Methodology, University of Economics in Prague, Prague 130 67, Czech

2. Department of Finance and Accounting, University of Business, Prague 130 67, Czech

Received: May 19, 2013 / Accepted: June 09, 2013 / Published: June 30, 2013.

Abstract: The proposal is focused on reconsideration and total reconstruction of recent economic education on the basis of different teaching styles, real competence-based learning, education for entrepreneurship and other trends in life-long education. These qualities are closely and mainly linked to target competencies which are the results of education and are reconditioned by these qualities. A modernized taxonomy of teaching methods and styles and instructions for their application in a diverse form will be proposed. The project applicants believe that the conceptual changes should be addressed by basic research as an innovation of methodological support in the form of teaching aids stratification. Teachers of economic subjects must be able to create their own teaching aids independently. They should be field methodology innovators, too. The new proposed taxonomy should be reflected in the training of teachers so that they avail of the new paradigm. This article is provided as one of the outputs of the research project of the Faculty of Finance and Accounting, which is realized in the framework of institutional support University of Economics IP100040.

Key words: Field-specific methodology, economic education, competences, curriculum, teacher training.

1. Introduction

In the modern, information- and knowledge-based

society, according to Vasutova [1], the entire human

life is the time to learn and each area of knowledge

penetrates and enriches all the other areas. Therefore,

the perception of education as a preparation for life is

replaced by the concept of education as a concomitant

life phenomenon. Thus, learning is acquiring

experience, shaping and molding an individual

throughout his lifetime. And, according to Prucha et al.

[2] and other authors, the learned is the opposite of the

inborn. Moreover, Prucha [3] also points out that the

archetype of human learning is acquiring

preconditions for successful adaptation, active

adjustment to the natural and social environment. In

Corresponding author: Katarina Krpalkova Krelova, Ing.,

Ph.D., Ing.-Paed, research fields: methodology of education, education of vocational subjects. E-mail: [email protected].

this perspective the education is seen as a process in

which an individual within the system of education

masters a certain sum of information and activities

which he subsequently through an internal

process-learning transforms into knowledge, through

application in practice into skills and through repeated

activities into habits.

Later on there were doubts about whether the

formal educational systems matched the current

requirements and whether there would be a possibility

to adjust them. They seem to suffer from a lack of

flexibility, especially with respect to the acquisition of

competences that would be useful in the future.

Cognitive skills, work skills, the art of living together

with others and life skills, the so-called “four pillars of

education” create four closely interrelated aspects of

the same reality. The idea of multidimensional

education through a person’s life, i.e., the idea that

follows from the legacy of outstanding pedagogical

System Reconstruction in the Field-Specific Methodology of Economic Subjects

760

thinkers of various cultures of the past, was now,

according to Vasutova [1], readopted and there are

efforts being exerted to implement it in the

educational systems since it appears to be ever more

necessary to bring the multidimensional education

in the pedagogical practice. Formal education has

traditionally focused primarily on teaching

of knowledge rather than on learning how to act. The

remaining two pillars of education were mostly left to

chance or considered to be a natural outcome of the

first two pillars. But, according to Vasutova [1], equal

attention shall be paid to all the four “pillars of

education” as long as the education is seen as a

comprehensive experience, obtained all through one’s

life, as an experience related to both the knowledge

and skills and the ability to apply them and at the

same time to the personality of an individual and

his/her inclusion in society.

In February 2001 the Government of the Czech

Republic adopted the National Programme for the

Development of Education in the Czech Republic,

often referred to as the “White Paper” [4]. In line with

the EU developmental trends in education it offers a

basic system vision of education in the Czech

Republic. It defines the main starting points, general

objectives and developmental programs crucial for the

mid-term development of the educational system.

Apart from other things the document declares “the

transformation of the traditional school”, with the

main aim being to build firm foundations for lifelong

learning. Learners should acquire abilities and

motivation to autonomously learn and work with

information.

This shift in the recent paradigm is all-embracing,

covering the area of objectives and content of

education, but particularly the modern methods, forms

and styles of education and other teaching aids, the

school climate and environment which should be

based on openness, partnership and mutual respect.

An important aspect is also the stress put on

professionalism and responsibility of teaching staff

who, apart from their expertise, are expected to be

able to motivate, diagnose and efficiently manage the

learning processes of students.

At present, the final stage of curricular reform,

conceived in the Czech Republic as an internal

transformation of schools, is under way. It resulted in

the school educational programs which have been

developed by schools based on the framework

educational programs. The school educational

programs in secondary vocational education put an

emphasis particularly on the following aspects:

(1) preparedness of young people for lifelong

learning (provision of tools for effective work with

information, motivation, active involvement,

creativity, positive value orientation);

(2) employability of graduates throughout their life

and promotion of entrepreneurial mindsets;

(3) broad general and basic vocational educational;

(4) target competencies (key and vocational

competencies) [5].

Such a concept of education does not build on rote

learning (memorizing) of the largest possible quantity

of facts. The role of the school should be the provision

of a systematic structure of elementary concepts and

relationships of each taught discipline as a basis for

creation and development of the cognitive systems of

students. Mutual links between the objectives, the

content of education and target competences should be

enhanced. The presence of practical activities in

instruction should be substantially increased,

interdisciplinary bonds, teaching in integrated units

and new forms of instruction facilitating internal

differentiation and individualization of education

should be developed. It comprises especially the

project teaching, based on active involvement and

autonomy of students by which they are given an

opportunity to get a deeper insight in the covered

subject matter. It is essential to prepare the teacher

methodologically for the new concept of curriculum

and for the introduction of corresponding teaching

aids.


761

The above referred to principles support the

research conducted under Comenius 2.1. project

focusing on taking the responsibility for one’s own

learning by the learner [6]. The learner will

understand the importance of learning for his

professional development, will participate in his

education, will be motivated and mobilized. The

system behind the concept relies on the strengthening

of importance of learning process in the traditional

teaching scheme: teacher—subject matter

(knowledge)—student (learner). The learning

sub-system in the system of instruction assumes

pivotal role. The teacher moves from the position of a

mentor to the position of a facilitator who supports the

learner in his work with information, when the

information is received, mastered, integrated in the

knowledge system, used and interpreted. The learner

acquires information skills, under the teacher’s

guidance and masters the principles of rational work

with information. The teacher monitors and facilitates

the transformation of information into knowledge. The

direct teaching undergoes fairly dramatic changes,

especially with regard to teaching methods and forms

that are more open to students, and methods and forms

which are activity oriented and have stronger elements

of problem teaching. The focus of the above

mentioned principles is analogous to the integrated

forms of teaching of economic subjects. Also, the

work in practice firms, student companies and practice

offices reckons with the autonomy of students, their

independent processing of information and decision

making without any direct involvement of the teacher.

Activity teaching methods and teaching integration

help the learners get much better prepared for the

period when they will have to receive and process the

information on their own. In the course of direct

teaching activities the team work, cooperative

techniques and the project method, which is based on

communication between learners and their

autonomous work, are employed [7]. Where the

referred to principles are properly employed in

educational practice and under favorable conditions,

the competences of learners to work autonomously

with information, their ability to make decisions and

to act independently, their ability to deal with

problems and be flexible in new situations should

grow because the key competences are developed by

doing activities and students gain knowledge, skills

and experience not only by memorizing. The

following structure of competences was established

for the field of studies of M category (secondary

education with the GCE-school leaving examination):

competences for learning; competences for problem

solving; communication competences; personal and

social competences; civic competences and cultural

awareness; competences for employment and the

conduct of business; competences in mathematics;

competences to use information and communication

technologies and to work with information [8].

An important complement to key competences is

information literacy and education for

entrepreneurship. Preparation for life in the

information society requires the integration of the

foundations of information science in the education

system. School graduates should be able to define and

satisfy their information needs and demands made on

them by information- and knowledge-based society.

Apart from vocational and key competences they

should also achieve information literacy. That

presupposes an acquisition of a range of information

skills. Available information resources offer multiple

classifications and interpretations of information skills.

One of the most accurate concepts for the support of

information literacy in education is “The Big Six”

concept which understands information skills to be

information literacy curriculum, the process of

information problem solving based on work with

information. The strategy of six fundamental

information skills is applicable to numerous situations,

when students get the assignment with incomplete

information and have to find the “hidden” information

by themselves. Education for entrepreneurship is seen


762

as a purposeful effort of educators to shape the

attitudes learners towards entrepreneurship and to

create abilities which would help them successfully

join the business community, i.e., especially creativity,

independent critical thinking, responsibility, and

willingness and ability to take reasonable risks. The

referred to qualities are closely linked to key

competences which are to be obtained in the education

process

2. Structure of the Project, Its Objectives, Method of Research, Timetable and Stages of Research

The main objective of the proposed project is a

major paradigmatical innovation in the field of

methodology of teaching, development and

finalization of new taxonomy of teaching methods and

other means, adequate to the current curricular theory

in the Czech Republic, based on the target vocational

and key competencies. This new taxonomy should be

reflected in undergraduate and postgraduate training

of teachers of vocational subjects so that they avail of

the teachers’ portfolio.

The teachers portfolio is seen as a set of teaching

aids available to teachers who can combine the

individual methods, forms, styles, conception and

material teaching aids in real life instruction to make it

activity oriented and to ensure the integral and

harmonious development of all target competences of

students. The above referred to problems and ideas

concerning the desired changes have been discussed

for quite some time, but any projections into practice

are lacking. The issues of activity methods and

integration of subject matter was investigated by

numerous international research projects. Now it is

necessary to build on them and to reflect this concept

in training of teachers of vocational subjects and to

identify an appropriate manner of their

implementation in teaching practice. The educational

approaches of teachers are strongly affected by

methodological stereotypes, the teachers often give

preference to frontal teaching dominated by

whole-class presentation and do not integrate the

subject matter. If the activities of students are

managed and guided in this way, problems emerge in

the application of teaching methods based on

autonomous work of students with information.

Therefore, for methodological training of vocational

subject teachers the following recommendations are

made:

(1) To strengthen the comprehensive training in the

field of definition and implementation of pedagogical

and educational objectives targeted at competences, to

innovate the styles of teachers’ work with

instructional content and the development of

up-to-date classroom management projects (specific

and integrated models);

(2) to put stress on creative drafting and use of

curricular documents;

(3) to enhance the skills necessary for the use of

computers, the Internet and high tech educational

technology;

(4) the teaching aids of teachers portfolio—its

permanent creation and application in different

teaching units;

(5) as concerns the organizational forms of teaching

to devote special attention to problem teaching and

integrated forms of teaching;

(6) to improve the facilitation skills—to support

autonomous work of the learner with information

sources and to promote work with information

sources;

(7) to work with cross-cutting topics: man and the

world of work, citizen in a democratic society, man

and environment, information and communication

technologies;

(8) methodology of conducting and analyzing

evaluation activities and educometry.

The submitted project is an inter-university project

and relies on many years of experience of the

applicant and co-applicants with the training of

vocational subject teachers. In terms of transfer of


763

experience, extremely important for the research of

the proposed project is the participation of higher

education teachers of the University of Economics,

Prague and Mr. Rotport, the Associate Professor at the

Metropolitan University Prague, the founder of the

concept of practice firms in the Czech Republic. This

concept has undoubtedly helped improve the training

of future economists. In this field the economic

education has had a long tradition. In terms of transfer

of experience, extremely important for the research of

the proposed project is the participation of higher

education teachers of the University College of

Business in Prague, the Metropolitan University in

Prague, the University of Hradec Kralove and the

Institute of Hospitality Management in Prague. The

cooperation of all universities will undoubtedly help

improve the training of future economists. At the

Department of Economic Teaching Methodology of

the Faculty of Finance and Accounting of the

University of Economics, Prague, where the economic

education has had a long tradition, the teacher training

successfully incorporates integrated forms of

instruction, particularly the practice firms, student

companies (junior achievement), practice offices,

practicums, all sorts of practical training in industry

and project teaching. That can become a valuable

source of inspiration for the transfer of experience into

the innovated methodology of teaching of vocational

subjects.

First, a detailed description of the baseline situation

will be given, i.e., hypotheses will be formulated and

verified by conducting quantitative research.

Researchers shall focus on which teaching methods

are used, and in what structure they are used in

teaching practice, and to what extent they correspond

to the curriculum and the objectives of education. This

step will be followed by a thorough science based

projection of the innovated structure of teaching

methods for the purpose of teacher training and

development of the field-specific methodology

(didactics) as a scientific discipline. The innovated

taxonomy of teaching aids shall be subsequently

elaborated in terms of strategy of their application by

teachers of vocational subjects (to ensure that that are

able to use them in a qualified manner in their

teaching practice). Teachers of vocational subjects

should avail of methodological competences, directed

not only at expertise, but also at the development of

key competencies, information skills, creativity,

autonomy and entrepreneurial mindset of students.

The teaching styles should be differentiated, the

methodological stereotypes prevailing in teaching

units dominated by teacher-led monologue should be

overcome, and the students should be active and

should learn to autonomously work with information

under an indirect guidance of the teacher (preparation

for lifelong learning process).

The project’s output will be the innovation of

curriculum of field-specific methodologies at the

workplaces of the applicants achieved through the

addition of a new teachers portfolio, a scientific

monograph, which synthesizes the achieved results

and gives recommendations for pedagogical practice,

an international scientific conference for the sake of

dissemination of research outcomes and further

development of scientific and research cooperation.

Table 1 shows the three basic stages of research and

an indicative of the project:

(1) Preparatory stage—theoretical and

methodological preparation (setting up of an expert

group, situation analysis in the field of taxonomy and

application of teaching methods, forms, styles and

material means)—output: publications in professional

periodicals and on the occasion of conferences,

preparatory of surveys tools: 2014;

(2) Implementation (conceptual) stage—surveys,

analysis of the surveys outputs, and development of

teachers’ portfolio: 2014-2015;

(3) Finalization and validation stage—integration of

methodological approaches in a field-specific

methodology, a monograph, an international scientific

conference: 2016.


764

Table 1 Timetable of research activity.

Period Activity

February 2014-September 2014

Theoretical and methodological preparation, development of research strategy.

October 2014-September 2015

Questionnaire and qualitative survey, evaluation, modification, final survey, data collection.

October 2015-

July 2016 Evaluation, data interpretation, synthesis and dissemination.

September 2016-December 2016

Optimization of vocational subject teacher training and final international dissemination.

Timetable of conceptual steps of the project (within

the three year period):

3. Justification of the Need and Necessity to Do Research Concerning the Issues in the Given Time Frame

The methodology used in the pedagogical practice

of secondary technical and vocational schools fails to

reflect the achieved level of knowledge in

field-specific methodologies (didactics), while the

theory of instruction not always addresses the issues

essential for the pedagogical practice. There is a

number of perpetuating problems cutting across the

entire system of education. The Czech education has

been repeatedly criticized for overloading the students

with too many facts and forced memorizing of the

subject matter. The students are introduced to an

excessive sum of information, in a pre-processed form,

and the schools seek to imbue the students with a solid

system of knowledge which in the rapidly changing

society (learning society) has rather a negative effect.

The school education thus ironically contributes to the

fact that the school graduates lack the elementary key

competencies. They leave the school equipped with a

solid base of factual knowledge, but lacking the ability

of critical analysis, having a poorer ability of

independent decision making and expressing their

own opinions, lacking in sufficient predisposition for

team work and missing the ingrained urge for

continuous self-education.

The pedagogical practice continues to be dominated

by the teaching style relying on one-way

communication from the teacher to students and on an

external management that supports passive reception of

presented knowledge. It is not an exception that

teachers resort to frontal teaching comprising only a

whole-class presentation and dictation of notes to

students (the so-called “note taking in exercise books

style of teaching”) which subsequently serve the

students in their preparation, in which they do not use

the technical textbooks, let alone any other sources of

information. It has also been proven by surveys and

evaluations carried out at training facilities providing

the vocational subject teachers with pedagogical

competencies [9]. This situation was brought about by

the past educational policy.

A long-term priority of the Czech system of

education in the field of secondary vocational

education has been a one-sided training for future

career, emphasizing the educational content, whereas

throughout the world ever more preference was given

to the highest possible degree of autonomy of students

in the acquisition of information, to group work,

higher attractiveness of instruction, support for

creativity, own opinions of students and

individualization of instruction bearing in mind the

interests and needs of students. Majority of the current

trends in education highlights learning or teaching

based on activities of learners, on mutual

communication and activities. Learning by doing is

deemed to be the most effective way of acquiring and

using knowledge. The findings of educational

psychology explicitly prove the need to increase the

autonomy in learning, the need of the shift from

external management to self-regulation and

self-construction.

Educational research underlines the need of a

teaching dialogue and a broader use of activity and

cooperation based methods, i.e., methods enabling the

student to create his own knowledge by means of his

active involvement and communication, not only to

receive the knowledge in a passive manner. It is also


765

important in the course of education to create

conditions for the development of students’ abilities to

communicate, cooperate and learn. There have been

clear signs of a shift from behavioral models of

authoritarian education to cognitive models of

education, building on discovery, cooperative or

collaborative learning. Elements of constructivism

shall be integrated into the subsystem of learning,

learning shall be more quickly and efficiently induced

by doing and project oriented methods in the form of

cross-curricular projects shall be exploited. This is the

direction we wish to follow in the development and

implementation of the teacher's portfolio.

4. Description of the Proposed Conceptual and Methodological Procedures Necessary for the Research Conducted within the Project and for Achieving the Expected Results, and Their Analysis

The above referred to requirements will be

examined by means of triangulation of research

methods, with the use of methodological analysis of

basic curricular documents, research of approaches

taken abroad, questionnaire surveys and controlled

interviews with vocational subject teachers at

secondary schools, statistical evaluation and synthesis

of ascertained findings, proposals of system solutions,

development of a scientific monograph, organizing of

an international scientific conference and

dissemination of knowledge.

An overview of individual conceptual steps:

(1) theoretical and methodological preparation

(setting up of an expert group, research in the field of

taxonomy and application of teaching

methods)—outputs: six publications in professional

periodicals and on the occasion of conferences;

(2) development of a research strategy—output:

research techniques (the educational content analysis,

a concrete concept of a questionnaire survey, the

content and structure of controlled interviews and

definition of target group);

(3) research survey, evaluation, modifications, data

collection—output: data;

(4) evaluation, data interpretation, synthesis and

dissemination—outputs: scientific monograph;

(5) optimization of the model of teacher training

and final international dissemination—outputs: six

publications in professional periodicals and in

conference proceedings, modification of the

curriculum of teaching studies programs, organization

of an international scientific conference.

5. Cooperation of the Applicant with International Scientific Institutions

The proposed project is designed as an

inter-university project. It is based on the cooperation

between the Department of Economic Teaching

Methodology, the Faculty of Finance and Accounting

of the University of Economics in Prague and the

University College of Business in Prague, the

Metropolitan University in Prague, the University of

Hradec Kralove and the Institute of Hospitality

Management in Prague.

In 2008-2010 the applicant solved in the field of

research of education for entrepreneurship with the

Faculty of Material Sciences and Technology, Slovak

University of Technology in Trnava on the MoE SR

KEGA 3/62116/08 grant project addressing the

introduction of the “Encouraging the spirit of

enterprise” subject into the “Teaching studies of

vocational subjects” study program. Another

prominent foreign partner with whom the applicant

has many years of successful cooperation is the

Department of Pedagogy of the Faculty of National

Economy of the University of Economics in

Bratislava. Apart from other scientific and research

activities, they have been consistently endeavoring for

the integration of vocational subject teaching,

especially the practice firms. It is a sophisticated

integrated form of instruction, implicitly developing

the area of key competencies, which can be a source

of inspiration and experience for the research of the


766

submitted proposal of a grant project.

6. Information on Readiness of the Applicant, Co-applicants and Their Workplaces, Equipment of the Relevant Workplaces to be used in the Course of Research

The primary mission of the applicant is the training

of teachers for secondary and post-secondary

economic vocational schools in order to acquire the

pedagogical competence of a vocational subject

teacher. The teaching practice is done by trainee

teachers at laboratory schools. At present, the

applicant’s workplace collaborates with many

secondary and post-secondary vocational schools in

the Czech Republic serving as clinical schools. This

network of schools can be utilized for the conduct of

field surveys and experimental validation of proposed

approaches.

The co-applicant Mr. Rotport as the founder and top

methodologist of practice firms in the Czech Republic

is in contact with the headquarters of practice firms

and schools where this form of instruction is

successfully implemented. The co-applicants

workplace is equipped with quality multimedia

educational technology and information technologies,

including statistical software and know-how for the

research of the given area. Therefore, there is no need

of any capital investment, only limited material costs

are envisaged for the purchase of small pieces of

equipment. Communication and cooperation between

participating workplaces, including the foreign

partners, has already been under way. It is of benefit

across the full scope of conducted activities, including

exchange study stays and participation in the

organized international conferences.

7. Introducing Co-applicants and Partners; Specification of Their Involvement in the Research

The project will involve four major universities in

the Czech Republic. Top professionals who deal with

the issues discussed in the research were selected to

represent each participating university.

University of Economics, Prague

Applicant: Katarina Krpalkova Krelova, Ph.D.,

Ing.-Paed.—she will coordinate and guarantee all

team work under the project, she is methodologist of

vocational subjects, in the field of research she

focuses on the pedagogy, pedagogical practice and

information and communication technologies in the

education, under this project she will guarantee the

methodological aspects of the investigated area and

the field of educational goals;

Team of partners:

Jiri Dvoracek, CSc.—educator with many years of

experience, 17 years with the Faculty of Education of

the Charles University in Prague, later on as Associate

Professor at the ETMD FFA UOE, he has been

involved in international projects, focusing on general

and comparative pedagogy;

Alena Kralova, Ph.D.—methodologist of vocational

subjects, as a researcher she concentrates on

state-of-the-art teaching methods and their application

in the referred to subjects, she will be involved in the

concept of the research, creation of research tools and

interpretation of results;

Katerina Berkova, Ph.D.—methodologist of

economic subjects, as a researcher she concentrates on

accounting methodology, IFRS and its application in

the pedagogical practice, she will be involved in the

statistical support of the research and interpretation of

results;

Libor Klvana—methodologist of vocational

subjects, Ph.D. student, under this project he will

participate on creation of research tools and data

collection.

University College of Business, Prague

Co-applicant: Pavel Krpalek, CSc.—field

methodologist focusing on economic subjects, he will

be the guarantor of work progress in individual stages

of the project and the coordinator of project

CSc. is the older equivalent of Ph.D. in the Czech Republic.


767

cooperation with co-applicants;

Team of partners:

Karel Bruna, Ph.D.—Vice-Rector for Research, is

an university teacher of economics and tourism,

researcher focusing on the effectiveness and efficiency

of travel business and marketing activities. He has

successfully organized international scientific

conferences on economics and tourism. He will

participate in the preparation and conducting of the

quantitative research;

Pavel Neset, Ph.D.—Vice-Rector for Academic

Affairs, majored in economics, researcher focusing on

the applications of economics in higher education and

related social competences. For this reason he will be

an extremely useful participant in the research

conducted under the project;

Antonin Kulhanek, CSc.—for many years he has

concentrated on the finance in business and banking

(didactics), he is an expert in the financial aspects of

project activities, he will get involved in the research

financial aspects and evaluation processes.

Metropolitan University, Prague

Co-applicant: Miloslav Rotport,

CSc.—methodologist of vocational subjects focusing

on accounting, well known founder of practice firms

in the Czech Republic, author of the concept of

integrated forms of instruction and their incorporation

in curriculum, he will guarantee the transfer of

experience with the integration of subject matter and

mobilization of students into the concept of

methodology portfolio field methodologist focusing

on economic subjects, he will be the guarantor of

work progress in individual stages of the project and

the coordinator of project cooperation;

Partner:

Radek Maxa—university teacher with experience in

secondary education and higher education, he will

participate in the field of methodology, he will involve

in the concept of research and data collection, too.

Institute of Hospitality Management in Prague

Co-applicant: Jan Chromy, Ph.D.—university

teacher and researcher focusing on media

communications, technical, didactic, psychological,

social and other aspects of the use of media, and on

auditoriology;

Partner:

Donna Dvorak—senior lecturer, teaching English

language and leading the Institute of Hospitality

Management team in the international WelDest

project, she will be responsible of translation and

international communication.

University of Hradec Kralove

Co-applicant: Rene Drtina, Ph.D.—Associate

Professor at the Department of Technical Subjects in

the Faculty of Education at the University of Hradec

Kralove, he is focusing on auditoriology, especially

classroom acoustics, audiovisual equipment

installations and development in power electronics.

8. Conclusions

We recognize the need to change and prepare total

reconstruction of recent economic education on the

basis of different teaching styles, real

competence-based learning, education for

entrepreneurship and others trends in life-long

education. The currently used model is outdated. The

main objective is conceptual innovation in the

field-methodology of economic education: goals

revision, development of teaching aids incl.

Educational contents, adequate to the current

curricular needs. This new taxonomy should be

reflected in the teacher training so that they avail of

the new paradigm.

References

[1] J. Vasutova, Transformation of the educational context and competence of teachers on study program in training, Study Texts for Pilots Project (2005) 22.

[2] J. Prucha, E. Walterova, J. Mares, Pedagogical dictionary, Portal. Prague (2013) 400.

[3] J. Prucha, Modern education, Portal, Prague (2002) 495. [4] National Programme for the Development of Education

in the Czech Republic, Institute for Information on Education, Prague, (2001)98.


768

[5] P. Krpalek, K.K. Krelova, Methodology of economic subjects, Oeconomica: University of Economics (2012) 184.

[6] M. Vidal, Self-responsible learning, Comenius 2.1. Project Enabling the Learner to be Responsible for his own Learning, Pedagogical guide, Enter, Centre Experimantation Pedagogique of Florac, France, 2005.

[7] A. Kralova, Increasing the competencies of teachers in the subject of economics teaching in secondary schools, Analysis of Competencies of Vocational Subject

Teachers, Oeconomica, VSE, Praha, 2007. [8] K.K. Krelova, P. Krpalek, J. Chromy, M. Sobek,

Research in area of students entrepreneurship at middle and university schools: Recent advances in educational methods, in: 10th WSEAS International Conference on Engineering Education, WSEAS Press, Cambridge, 2013.

[9] P. Krpalek, Selected findings from the field of development of information skills, [Online], Media4u Magazine, 2010, pp. 34-37, http://www.media4u.cz/.

Contact:

K. Krpalkova Krelova is from the University of Economics, Prague, W. Churchilla 4, 130 67, Prague 3, Czech (corresponding

author to provide phone: +420 737930200; e-mail: [email protected]).

She is a member of International Society for Engineering Education (IGIP) in Graz, International Association for Continuing

Engineering Education, Atlanta, USA, TTnet Slovakia and Society for Information Science and Research (Slovakia).

P. Krpalek is from the University of Economics, Prague, W. Churchilla 4, 130 67, Prague 3, Czech (e-mail: [email protected]). He

is a teacher and guarantor of the Methodology of economic subjects and Economy of Enterprise. He is also a member of the board for

the defense of dissertations and a trainer in doctoral study programs, namely not only at the University of Economics, Prague, but

also at the Constantine the Philosopher University in Nitra (Slovakia).


The Research of Wind Turbine Fault Diagnoses Based

on Data Mining

Yu Song and Jianmei Zhang

School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China

Received: June 10, 2013 / Accepted: June 20, 2013 / Published: June 30, 2013.

Abstract: Wind turbine is the key equipments of the wind farm, the distribution area is wide, and the working condition is poor, in the event of failure, caused by the loss and consequences are incalculable. Therefore, in order to ensure the safe and stable operation, it is necessary to research the fault diagnosis for the wind turbine. As the wind turbine produces vast amounts of data in the daily monitoring process, these data imply a large number of potential rules, and meet the necessary conditions for data mining. So the authors can use data mining technology to diagnose potential faults. This paper introduces methods and process of data mining and the rough set technology in wind turbine fault diagnosis applications, finally an example is analyzed to verify. Key words: Data mining, rough set, wind turbine.

1. Introduction

Wind power is recognized as the closest

commercialization of renewable energy technologies

in the word [1]. In the past 10 years, wind power has

become the fastest growing renewable energy [2].

With the rapid development, and because of many

units are installed in remote areas, resulting in load

instability and other factors, most of wind turbines

have experienced operational failure, which directly

affect the safety of wind power and economy.

Therefore, it is particularly important to study fault

diagnosis of wind turbines. It is the key to ensure

long-term operation of the unit and safety generation.

In recent years, the rise of the data mining

technology can solve the problem of automatic

acquisition of knowledge. Data mining dig out hidden

information that has important reference value for

people to make decision from a large, incomplete,

noisy, fuzzy, random data, so that it can overcome the

bottleneck of knowledge acquisition effectively.

Corresponding author: Jianmei Zhang, postgraduate,

research field: database and management information system. E-mail: [email protected].

Applying data mining techniques to wind power

equipment fault diagnosis system, the authors can

transform volumes of raw data into valuable

knowledge, discover useful information and predict

the evolution of trends in mechanical equipment

failures, providing decision support information [3].

2. Data Mining

Data mining is a process that use of artificial

intelligence methods to analyze data in the database in

order to gain knowledge. Data mining is usually

extracted knowledge can be expressed as the concept,

rules, regularity, patterns and other forms that can be

used for information management, query optimization,

decision support and process control as well as the

maintenance of the data itself [4]. In recent years,

research and application of data mining technology is

developing rapidly, and a variety of data mining

algorithm emerge.

Rough set approach is a new mathematical analysis

tools that can deal with uncertain, imprecise,

incomplete and inconsistent information. Its basic

methods is to first use the approximation method of

The Research of Wind Turbine Fault Diagnoses Based On Data Mining

770

Rough Set to discrete attribute value in the

information system (relationship); then divide each

attribute into equivalence class, and then reduce the

information system (relationship) by equivalence

relation of set; final the authors will get a minimal

decision relationship, thereby facilitate access rules.

The main advantage is that rough set does not require

any initial or additional information about the data.

Rough set theory is founded on the basis of

classification mechanism that links knowledge

description with classification of things together. A

knowledge representation system can be expressed as:

S = (U, C, D, V, F).

where U represents universe, C ∩ D = A is a set of

attributes, subset C is called condition attributes set

and subset D is called decision attribute set. V =

aAaV

is a set of attribute values, Va represents the

range of attribute a A; f: U × A → V is an

information function, which specifies attribute values

of object x, x U. In this way, knowledge

representation system can be used to express as the

two-dimensional form, this form is called a decision

table [5].

A data mining system based on rough set theory is

generally composed of data preprocessing, data

reduction which is based on rough set theory or

extended, decision algorithms and other components.

Rough set deal with the data does not need to

understanding of the data, which no longer require a

priori information on the data.

3. Data Mining Application in Fault Diagnosis

The diagnosis of wind turbines should be found

potential rule from a lot of fan operation monitoring

information, to extract useful knowledge, intelligently

judge current fan operation state, and found that the

implicit or existing fault. Fault diagnosis is mainly

divided into four steps: signal acquisition, feature

extraction, pattern recognition and diagnosis decisions.

For equipment fault diagnosis, the most important and

the most difficult problem is the failure feature

information extraction.

In order to fundamentally solve the key problem of

feature extraction, people mainly get more

information from the depth of the signal processing.

As the measured data increasing, in order to be more

effective and more convenient to obtain fault

information, the authors need a newer perspective to

observe the huge amounts of data, which is an

important way to research and develop the method of

fault feature extraction based data mining.

Data mining is a process of repeated treatment, it

can be used to appropriate diagnostic algorithm

gradually dig real failure mechanism and diagnostic

rules. Its implementation process is: first, the original

data is arranged into mining theme-related

information; then, according to the theme as well as

the characteristics of a variety of learning algorithms

to design data mining algorithm, and knowledge of

specific data sets are extracted. Checking the mining

results are consistent, reasonable, and compare it with

the expected target. If the results with the expected

target deviation greater, you should return algorithm

design stage, adjust or redesign mining algorithms; if

the deviation is small, adjust the mining algorithm; if

satisfactory, you should return data phase, expand

data sets and restart a mining process. Repeat the

above steps until you reach the ultimate goal [6].

4. Experiment

As an example, Table 1 shows some analysis data

of the wind turbine gearbox fault.

In Table 1, I1, I2, I3… I7 and Fi represents peak,

kurtosis, margin, skewness, spectral, frequency

variance, harmonic factor and fault number,

respectively.

First of all, adopting fuzzy clustering method

discrete continuous attributes, and obtain discrete

diagnosis decision table; then take advantage of the

improved rough set theory and the maximum

clustering principle to attributes reduction; and select

The Research of Wind Turbine Fault Diagnoses Based On Data Mining

771

Table 1 Gearbox fault sample table.

I1 I2 I3 I4 I5 I6 I7 Fi

0.1302 0.1111 0.1275 0.2127 0.6449 0.925 0.298 F1

1 0.6 0.7761 0.5778 0.5735 0.9215 0.139 F2

0.4472 0.2873 0.3812 0.3871 0.6415 0.9269 0.1677 F2

0.62 0.2922 0.4557 0.321 0.4617 0.9842 0 F3

0.1725 0.126 0.1452 0.2190 0.5379 0.9987 0.1593 F4

0.5221 0.26 0.4122 0.3433 0.5489 0.9952 0.1874 F5

0.9574 1 1 1 0.7344 0.792 0 F6

Table 2 Gearbox fault diagnosis rule set.

R Rule Description

R1 If x4 ≤ 0.3422 and x5 > 0.71, then normal

R2 If x4 ≤ 0.3422 and x5 ≤ 0.585, then bearing cage is broken

R3 If 0.3422 ≤ x4 ≤ 0.7351 and x1 > 0.765, then gear outer peeling

R4 If x4 > 0.7426, then gear tooth collapse

the best reduction set; finally, construct diagnostic

decision tree, get diagnosis rules. The results are

shown in Table 2.

5. Conclusions

For wind turbine equipment, rough set theory

provides a new way of fault diagnosis; promote

research and application of data mining technology in

this area. However, there are many problems and

challenges to be solved, such as not properly handle

redundant information and noise data mining results

invalidity and so on. Rough set theory in the

application of data mining is becoming a hot research

topic in information science, and its broad

development space.

References

[1] H.X. Xu, Energy utilization in China and countermeasures, Sino-Global Energy 15 (2010) 3-6.

[2] P.Y. Zhou, The fault diagnosis of gear box for wind urbine generator system, Thesis, Xinjiang University, Urumqi, 2010.

[3] J. Hou, Application investigation of data mining on fault diagnosis for rotary machinery, Master Thesis, Dalian University of Technology, 2007.

[4] X. Li, Q.W. Gong, Q.Y. Yang, Fault information management and analysis system based on data mining technology for relay protection devices, Electric Power Automation Equipment 31 (2011) 88-91.

[5] J.L. Chu, B.Y. Chen, The research of machinery fault diagnoses based on data mining, Microcomputer Information 23 (2007) 208-209.

[6] H. Li, Vibration condition monitoring system development of large-scale wind turbine, Thesis, North China Electric Power University, Beijing, 2009.


A Recommendation System Keeping Both Precision and

Recall by Extraction of Uninteresting Information

Tsukasa Kondo1, Fumiko Harada

2 and Hiromitsu Shimakawa

2

1. Data Engineering Laboratory, Graduate School of Science and Engineering, Ritsumeikan University, Shiga 525-8577, Japan

2. Department of Information Science and Engineering, Ritsumeikan University, Shiga 525-8577, Japan

Received: March 21, 2013 / Accepted: April 17, 2013 / Published: June 30, 2013.

Abstract: A recommendation system which recommends interesting information to the target user must guarantee high precision and

recall. However, there is trade-off between precision and recall. In this paper, we propose a web page recommendation method

balancing both of them by take advantage of uninteresting information. The proposed method extracts the interest and uninterest

indicators from not only historical interesting web pages but also uninteresting ones in a target genre. The historical interesting and

uninteresting information is derived based on the browsing time and bookmarking. The proposed method can keep precision and

recall by excluding the uninteresting information from the recommended ones based on the interest and uninterest indicators. The

experimental result proved that the proposed method can improve the precision and recall than an existing method.

Key words: Recommendation system, uninterest indicator, precision, recall, trade-off.

1. Introduction

Huge numbers of web pages exist in resent WWW

(world wide web). Since it is difficult for users to pick

up his interesting web pages, web information

recommendation systems have been developed to

support a user to find individual interesting web pages

based on estimated user interest. Many such existing

recommendation systems find the web pages that are

interesting for a target user based on his interest

estimated from his WWW browsing history. For

example, a system derives the historical web pages

which may have been interesting for him from its

browsing time [1] or mouse operations [2]. Detecting

interesting historical web pages, the recommendation

system extracts the indicator, which enables the

system to determine whether a novel web page will be

interesting for the user or not by calculating the match

between the page and indicator. The recommendation

Corresponding author: Fumiko Harada, lecturer, research

fields: data engineering and real-time embedded systems. E-mail: [email protected].

system decides each of the

recommendation—candidate (target) web pages will

be interesting for him or not with the extracted

indicator. The target web pages which are decided to

match the indicator are recommended to the target

user.

The performance of such recommendation system

is measured by precision and recall. Though both of

them are important for accurate recommendation,

there is trade-off between them in general. Thus, this

paper focuses on the additional indicator which

indicates whether a web page is uninteresting for the

user or not. Moreover, this paper proposes a

recommendation method to keep both of precision and

recall based on interest and uninterest indicators.

The rest of this paper is organized as follows:

Section 2 explains needs for uninterest indicator from

the viewpoint of balancing precision and recall;

Section 3 proposes the recommendation method based

on interest and uninterest indicators. The experimental

result of verifying the validity of the proposed method

A Recommendation System Keeping Both Precision and Recall by Extraction of Uninteresting Information

773

is given in Section 4; Section 5 concludes this paper.

2. Need for Uninterest Indicator

2.1 Trade-off between Precision and Recall

Though recommendation systems must guarantee

high precision and recall, there is trade-off between

them [3]. An example of considering them in web

page recommendation is shown in Fig. 1. Suppose that

the historical interesting web page has the content

―Keisuke Honda scored the goal in the international

game‖. From the viewpoint of the included words or

content meaning, the word ―Keisuke Honda‖, that is

the name of a famous Japanese football player, can be

derived as the interest indicator (Indicator 1 in Fig. 1).

This means that a web page whose content matches

―Keisuke Honda‖ may be interesting for the user.

However, only ―Keisuke Honda‖ as the interest

indicator is a weak constraint. Here, suppose also 5

target web pages A-E to be the candidate of the

recommendation (Fig. 1). We assume that the pages A,

B and C will be interesting for the target user while

the pages D and E will be uninteresting because they

are not related to any excitement of the target user.

Though all of the web pages A-E are recommended

since they all match the interest indicator ―Keisuke

Honda‖, Web pages D and E are out of the target

user’s preference. This example implies that the

information matching the indicator is not always

interesting for him. This phenomenon decreases the

precision of recommendation. On the other hand, we

can consider extracting stricter indicator in order to

guarantee high precision by not recommending

uninteresting web pages. Suppose the indicator

―Keisuke Honda scored the goal‖ and ―Keisuke

Honda, a member of the national team‖ as a strict

indicator (Indicator 2 in Fig. 1). With Indicator 2, the

recommendation system excludes the web pages D

and E in the recommendation. However, it also

excludes the web page C. This is because the

historical web page does not enable extracting the

indicator matching the Web page C. This phenomenon

decreases the recall of the recommendation. In this

way, using strict interest indicator, any

recommendation system can not recommend all

target web pages that will be interesting for the target

user.

Fig. 1 Recommendation example that can not keep the precision and recall.


774

Thus, both of weak and strict interest indicator

decreases either precision or recall of the

recommendation. Balancing both of them is difficult if

the system is based on only interest indicator extracted

from the historical interesting web pages.

2.2 Interest and Uninterest Indicators

There are two kinds of information in the

informations which match the indicator and will be

interesting for the target user. The first kind of such

information is those relevant to and implied by the

browsing history. In Fig. 1, the target user is aware of

his preference on Japanese team’s international games

which Keisuke Honda plays, which is implied by and

relevant to the content of the historical page ―Keisuke

Honda scored the goal in the international game‖. The

second kind of such information is those irrelevant to

and not implied by the browsing history. For such

information, the target user finds his preference for its

content after he browses it actually. The target user is

not aware of that it is interesting before the browsing.

This implies that such information does not appear in

the target user’s historical web pages. In Fig. 1,

suppose the target user does not know existence of the

information about the reason of success of Keisuke

Honda. In this case, the target user can not be aware of

that it is interesting for him. No matter how strictly

the recommendation indicator is set, the

recommendation system can not give priority to the

page C, whose relevant pages do not appear in the

historical web page. Like as to the page C, the

recommendation system can not generate the indicator

leading to specific web pages about ―Keisuke Honda‖

whose content is not similar to any pages which

appear in the browsing history.

However, if we can estimate the uninteresting

historical web pages relevant to weak indicator and

can make additional indicator from them, it enables

the recommendation system to exclude the

uninteresting ones from the recommendation result.

The recommendation system could cover the all web

pages interesting for the target user, including ones

which content is not similar to any pages appearing in

the browsing history (the web page C in Fig. 1). The

web pages D and E are not interesting for the target

user in Fig. 1. Suppose an additional indicator which

prevents the recommendation of the information

relevant to D and E in the recommendation on

―Keisuke Honda‖. With such indicator, a

recommendation system can recommend the all and

only web pages that are interesting for the target user

such as A, B and C, while D and E can be excluded in

recommendation.

If the recommendation system can extract the

indicator not only from the historical interesting pages

but also uninteresting information, such as the

historical uninteresting pages, it can exclude only

uninteresting ones from recommendation candidates.

It can guarantee high precision and high recall.

2.3 Related Work

There are some researches to use information

uninteresting for the target user to improve

performance of the information search engine and

information recommendation system. The methods

proposed in Refs. [4, 5] improve the result of search

engines by using the word which the target user wants

to exclude from search results. The method in Ref. [6]

could be more accurate for recommendation by using

an indicator extracted from news articles which match

the condition, that is, ―When the target user does not

view a news article even though it is presented in the

first time from an online news service, we can assume

that the article is uninteresting for him‖.

However, a target user has to choose words that he

wants to exclude from the search results by himself in

Refs. [4, 5]. The method in Ref. [6] can extract the

indictor from only online news subscribers.

Furthermore, the method in Ref. [6] uses the all words

appearing in titles of detected uninteresting articles. It

implies that the recommendation result excludes the

information relevant to the words which the target


775

user are interested in. Actually, a part of words in an

uninteresting article can be inessential for his

uninterest and are rather than essential for his interest.

If we aim to use the uninteresting information as an

indicator for recommendation, we not only have to

extract the words essential to uninterestingness for the

target user as long as we should automate the

extraction of indicator corresponding to the

uninterestingness word.

3. Recommendation Method to Keep

Precision and Recall

3.1 Method Overview

We propose a recommendation system based on

interest and uninterest indicators to keep high

precision and recall. The proposed method

automatically extracts interest and uninterest

indicators. It gives positive weights to the target web

pages matching the interest indicator while it does

negative weights to the target web pages matching the

uninterest indicator. The target web pages are ranked

towards recommendation based on their total weights.

Fig. 2 shows the flow of the proposed method

consisting of the following 4 steps. Their details are

discussed in Subsections 3.2-3.5, respectively.

Step 1: The target user chooses a word for

recommendation. Then it extracts the historical web

pages whose contents are relevant to the chosen word

from the target user’s browsing history. The extracted

historical web page is called as historical related web

page;

Step 2: The method derives the interest indicator,

which is the indicator to decide whether a web page is

interesting for the target user or not. It is based on the

browsing time and bookmarking of each of the

historical related web pages;

Step 3: The method also extracts the uninterest

indicator by the similar way. The uninterest indicator

is for decision whether a web page is uninteresting or

not;

Step 4: The method determines the target web

pages to be recommended to the target user based on

interest and uninterest indicators.

3.2 Finding Historical Related Webpages

The target user chooses a word, which is called as a

related word, for recommendation. The method

recommends the web pages which have contents

relevant to the related word. It extracts the historical

related web pages from target user browsing history.

Fig. 2 Method overview.


776

The historical web pages whose titles or body texts

include the related word are supposed to be historical

related pages.

3.3 Extraction of Interest Indicator

The historical interesting web pages are picked up

from the historical related web pages. If a historical

related web page satisfies either of the following

conditions, the proposed method picks it up as a

historical interesting one.

Condition 1: The web page is browsed for more

than 1 seconds, where 1

is a threshold.

Condition 2: The web page is bookmarked.

Condition 1 comes from that an article interested by

the target user is browsed for long period of time [1].

For Condition 2, it is reasonable to assume that

bookmarking a web page is the target user’s

declaration to revisit the web page in the future

because its content is interesting for the target user.

The interest indicator X is generated from the

historical interesting web pages as a set of words

according to the following idea. The target user judges

whether a web page is interesting or not based on its

content. Interest indicator has to be set as

characteristic words that can represent the contents of

the historical interesting web pages. Therefore, the

proposed method extracts the characteristic words of

the historical interesting web pages based on TFIDF.

We can suppose that characteristic words of an

historical interesting web page have high TFIDF

values with high probability. In the proposed method,

the words with the 5 highest TFIDF values are

extracted for each historical interesting web page. The

extracted words for all historical interesting web pages

are included in the interest indicator X. The

related word chosen in Step 1 is also included

into the interest indicator X regardless of its TFIDF

value.

3.4 Extraction of Uninterest Indicator

The historical uninteresting web pages are picked

up from the historical related web pages. If a historical

related web page satisfies both of the following

conditions, the proposed method derives it as an

historical uninteresting web page.

Condition 3: The web page is browsed for less than

2 seconds, where 2

is a threshold;

Condition 4: The web page is not bookmarked.

For Condition 3, a person stops browsing a web

page at the moment when he finds its

uninterestingness [1]. Thus, a historical uninteresting

web page may have a relatively short browsing time.

The limitation of the browsing time is represented by

Condition 3. For Condition 4, any bookmarked web

page should be out of the candidate of uninteresting

ones even if its browsing time is short.

The uninterest indicator Y is extracted as follows.

We consider that an uninteresting web page have

characteristic words which do not appear in the

interesting web pages, which are the factors of

uninterest. Thus, the characteristic words included in

the historical uninteresting web pages but the

historical interesting web pages can be extracted as

the uninterest indicator. Note that, however, a person

stops browsing a web page immediately when he finds

its uninterestingness. Even if he looked a partial

content of a web page and stopped browsing because

of its uninterestingness, the following content could

be interesting in fact. Thus, we consider extracting

characteristic words from only the browsed area in

each uninteresting web page. The browsed area of a

historical uninteresting web page Wi is determined

from the estimated number of read letters m (Wi). It is

given by the following equation:

)( )( ii WtRRWm (1)

where RR and t(Wi) are the reading rate and the

browsing time of Wi. The proposed method firstly

finds the words in the browsed areas with the 5

highest TFIDF values among the historical

uninteresting web pages. Secondly, the proposed

method extracts the words which do not appear in all

historical interesting web pages but in the browsed


777

area of a historical uninteresting web page. The set of

such extracted words are the uninterest indicator Y.

3.5 Recommendation of Target Web Pages

The proposed method decides the target web pages

to be recommended based on their scores. Denoted by

W, M and N are a target web page which the target

user has not browsed, the number of words included

in both of the interest indicator X and the page W, and

the number of words included in both of the uninterest

indicator Y and the page W, respectively. The score

S(W) for each W is defined by Eq. (2).

M

m

N

n

muniWS11

)()1()()( (2)

Note that is a given parameter and 10 .

In Eq. (2), the word score of the k-th word included in

both of X and W is represented with )1)(( Nkki .

The word score of the l-th word included in Y is

represented with )1)(( Mllu . The Parameter

is constant to control the weights for the interest

indicator X and the uninterest indicator Y. The word

score of a word is given by its IDF value for the page

W. Eq. (2) estimates the interest on W for the target

user. If W includes a word in the interest indicator X,

the score S(W) increases by the IDF value of the word.

If W includes a word in the uninterest indicator Y,

S(W) decreases by the IDF value of the word. In this

way, the more words in the interest indicator W has,

the larger S(W) is. Similarly, the more words in the

uninterest indicator W has, the less S(W) is. The

proposed method do not use TFIDF value but the IDF

value as the word score because the possible range of

TFIDF value varies according to the length of the

body text of W. TFIDF value cannot fairly evaluate

the target web pages with different lengths of the body

texts.

Finally, the proposed method recommends the

target web pages according to their ranking in the

descending order of S(W).

4. Experimental Evaluation

4.1 Verification Contents

We verified the validity of the proposed method

through an experiment. We conducted two kinds of

verifications.

Verification 1: Thresholds decision and accuracy

The proposed method detects the historical

interesting and uninteresting web pages with

Conditions 1 and 3 in Subsections 3.3 and 3.4, which

evaluate the browsing time of web pages. These

conditions work under setting appropriate values of

the thresholds 1 and 2

.

Therefore, we investigated the browsing times of

interesting and uninteresting web pages to decide

appropriate values of the thresholds 1 and 2

. We

also investigated the detection accuracy for the

decided 1 and 2

.

Verification 2: Precision and recall improvement

The proposed method aims to keep both precision

and recall in recommendation by introducing

uninterest indicator.

We investigated whether the proposed method can

achieve balancing precision and recall than an existing

method or not. We also investigated the variation of

precision and recall with respect to the parameter ,

which determines the weight of uninterest indicator

relative to interest indicator.

4.2 Data Collection for Verification

We collected the data for verification through the

following steps. The examinees, called Examinees

A-L, were 12 university students in an IT department.

Step 1: Each examinee chose a related word. We

searched for the web pages whose titles or body texts

include the related word on the Hatena Bookmark [7].

We selected from 81 to 199 pages from the search

result and presented them to him;

Step 2: Each examinee browsed every presented

web page. He was not required to browse a presented

web page completely, where he could finish the


778

browsing of uninteresting web page at any time

according to his interest. The browsing time of every

presented page was also measured;

Step 3: Each examinee evaluated every presented

web page through the following evaluation items:

Item A: ―Are there any interesting or useful

information in the presented web page?‖

It refers whether the web page is interesting or not.

The answer is by the following 4-grade evaluation: ―4:

It is very interesting web page.‖, ―3: It is interesting

web page.‖, ―2: It is uninteresting web page.‖, or ―1: I

hate this web page.‖

Item B: ―Do you want to bookmark the presented

web page?‖

It refers whether the web page is worth

bookmarking for him or not. The answer is binary,

that is, ―1: Yes, I want to bookmark it.‖ or ―0: No, I

do not bookmark it.‖

Steps 1–3 generate the pair of each presented web

page and its evaluation (interestingness and worth of

bookmarking) for each examinee.

We call the generated data as the verification data.

In the verification data, we call a presented web page

evaluated as 4 or 3 for Item A as an interesting web

page. We also do that evaluated 2 or 1 for Item A as

an uninteresting web page. Similarly, we do that

evaluated as 1 for Item B as a bookmarking

page.

4.3 Verification 1: Thresholds Decision and Accuracy

4.3.1 Thresholds Decision

Table 1 shows the average browsing times of the

interesting, uninteresting and all presented web pages

for each examinee. The interesting web pages are

browsed for longer times than all pages in every

examinee. The uninteresting web pages are done for

shorter time than all pages in every examinee. Since

the browsing times differ significantly among the

examinees, same values of thresholds 1 and 2

may not be applied to all examinees. We decided 1

and 2 for each examinee individually.

We firstly normalized each browsing time as the

standard score among all browsing times of the

corresponding examinee. It normalizes the browsing

times such that the mean and standard deviation are

changed to 50 and 10, respectively. We call the

standard score of each browsing time as the

normalized browsing time in below. Table 2 shows

the mean and the variance of the normalized browsing

times of the interesting web pages and the

uninteresting web pages among all examinees. The

normalized browsing times have a little variance in

the interesting web pages while the variance is almost

zero in the uninteresting ones. The reason of such

variance in the interesting web pages may be the

length of the body text. An examinee consumed

longer time to browse an interesting web page as its

body text was longer because he tended to browse

interesting web pages completely. Different lengths of

the body texts of the interesting web pages may bring

a little variance of the normalized browsing times. On

the other hand, because an examinee stopped

browsing a web page immediately after he found that

it was uninteresting, the variance may be almost zero

in the browsing times of the uninteresting web pages.

Table 1 Average browsing times (unit: s).

Examinee Interesting Uninteresting All

A 113 31 113

B 80 28 84

C 118 41 188

D 32 21 23

E 51 29 -

F 41 15 56

G 128 48 126

H 88 43 78

I 112 77 112

J 31 17 38

K 49 9 84

L 43 23 24

Table 2 The mean and variance of normalized browsing

times.

Interesting Uninteresting

Mean 54.91 45.84

Variance 5.72 0.8


779

We decided the values of the thresholds 1 and

2 for each examinee such that the normalized

browsing times corresponding to 1 and 2

are the

mean of the normalized browsing times in the

interesting and uninteresting web pages, respectively.

ABAD

10

5091.541 (2)

ABAD

10

5084.452 (3)

Denoted by AD and AB are the standard deviation

and the mean of the browsing times of the presented

web pages for the examinee.

4.3.2 Verification of Thresholds Accuracy

We verified the validity of the values of 1 and

2 decided as mentioned in 4.3.1. We detected the

interesting and uninteresting web pages through

Conditions 1-4 in Subsections 3.3 and 3.4 by using the

decided values of 1 and 2

. Detection of

interesting web pages was conducted by 3 ways: only

Condition 1 (the condition on browsing times), only

Condition 2 (the condition on bookmarking), and both

Conditions 1 and 2. we call a detected web page as an

interesting and uninteresting one as detected

interesting and detected uninteresting web pages,

respectively. If the set of the interesting/uninteresting

web pages is the same of that of the detected ones, the

values of 1 and 2

are appropriate.

Table 3 shows the precision and recall of detection

of interesting web pages and uninteresting ones. The

result for interesting ones by only Condition 2 in

Examinee E is empty because no bookmarking page

for him existed. Table 3 shows that the deviation

between precisions of the interesting web pages is

smaller than that of the uninteresting web pages. It

may be caused by difference of browsing behavior for

interesting and uninteresting web pages.

There were four types of browsing behavior.

Generally, a person firstly looks the exordium of a

web page and determines to continue browsing or not

to do according to his interest. Examinees A, C, H, J

and L determined to continue browsing if they found

any little interesting content in the exordiums of a

presented web page. They saw that the web page was

uninteresting after the end of its browsing. It leads that

the browsing times of the uninteresting web pages are

longer and that the precisions for uninteresting web

pages are poor. Examinees B, D, F, G and K

determined to continue browsing or not definitely

when they browsed the exordiums. The browsing

times of the uninteresting web pages are short because

they stopped browsing a web page immediately after

they detect uninterestingness from the exordium. For

them, the precisions of detecting uninterest web pages

Table 3 Detection accuracies of interesting and uninteresting web pages (unit:%, P: Precision, R: Recall).

Exam.

Interesting web pages Uninteresting web pages

Only Cond. 1 Only Cond. 2 Cond. 1 & 2

P R P R P R P R

A 87 44 94 36 88 67 95 65

B 78 46 100 11 79 50 95 68

C 100 38 100 23 100 48 62 72

D 76 29 98 56 88 76 84 53

E 68 27 - - 68 27 76 63

F 83 52 100 15 84 58 93 71

G 76 41 67 4 73 41 91 68

H 80 29 100 44 90 63 88 57

I 69 32 88 83 79 86 80 48

J 91 33 75 3 89 35 69 68

K 96 56 100 22 96 56 97 61

L 94 28 75 6 90 33 52 55

Ave. 83 38 91 28 85 53 82 62


780

tend to be higher than other examinees. Examinees E

and I browsed the web pages by remarkable ways,

which lowered the detection accuracy. Examinee E

did not browse an interesting web page carefully. He

browsed the web page roughly, searched for a

significantly interesting part, and browsed carefully

only found significantly-interesting parts with mouse

dragging. The browsing time depends on the length of

the body text because he searches for

significantly—interesting parts from the body text.

Examinee I browsed a web page completely even after

he found it is uninteresting from the exordium. The

browsing time depends on the length of the body text

even for the uninteresting web pages. Note that,

however, the average browsing time of the

uninteresting web pages is relatively shorter than that

of the interesting ones because he browsed the

uninteresting web page roughly. The proposed method

assumes that the browsing times of uninteresting web

pages are short while those of interesting ones are

long. It could not detect interesting and uninteresting

web pages accurately for the users such as Examinees

E and I, whose browsing times are determined with

dependent of the lengths of the body texts.

Table 3 shows that the precision and recall of

detecting interesting web pages by using Conditions 1

and 2 are 85% and 53%, respectively. Those for

uninteresting web pages are 82% and 62%,

respectively. In the proposed method, high precisions

for detection of interesting and interesting web pages

are required because invalid inclusion of words in

interest and uninterest indicators gives negative effect

to recommendation. The precisions exceed 80% in the

experimental result. Moreover, the recalls exceed 50%,

which means more than half of the interesting and

uninteresting web pages are detected validly.

Eventually, Verification 1 proves that decided values

of the thresholds 1 and 2

achieve valid and

exhaustive detection of the historical interesting and

uninteresting web pages.

4.5 Verification 2: Precision and Recall Improvement

4.5.1 Verification Purpose

The proposed method gives ranks to target web

pages from the descending order of their scores

derived by Eq. (2). General recommendation methods

do not recommend all target web pages but certain

numbers of them with the highest ranks. If a

recommendation method generates a ranking such that

many interesting web pages are at higher ranks while

uninteresting ones are at lower ranks, it achieves high

precision and recall toward good recommendation.

Verification 2 compares goodness of the rankings

between the proposed method and an existing one. If

the ranking of the proposed method is better than that

of the existing one, it can be proved that the proposed

method keeps both of precision and recall better.

4.5.2 Evaluation Method

We set the case where = 1.0 in Eq. (2) as an

existing method, which does not consider uninterest

indicator at all. The proposed method is the cases

where is each of 0.1, 0.2… 0.8, 0.9.

We firstly generated the training and test data for

each examinee. We applied the proposed method to

the verification data and detected the historical

interesting and uninteresting web pages. The detection

used the thresholds 1 and 2

set according to

Verification 1. We generated the training data by

randomly selecting 15 detected historical interesting

and 15 detected historical uninteresting web pages

from the verification data. The rest of the verification

data was set as the test data.

We secondly extracted the interest and uninterest

indicators from the training data for each examinee.

The reading rate RR in Eq. (1) was set to 8.

We thirdly applied the interest and uninterest

indicators to the test data by the proposed method

with each . We derived the ranking L of the test

data in the descending order of the score S(W). The

ranking K by the existing method was also generated

similarly.


781

Fig. 3 RS improvement with respect to alpha.

We compared goodness of the rankings L and K for

each examinee by Rankscore [8]. The Rankscore RS

for L is calculated by Eqs. (5, 6) in below. is a

given parameter. I(L) is the set of interest web pages

in L. )( iWr is the rank of each interesting web page

Wi. idx(Wi) is the rank of each interesting web page

iW in an ideal ranking where all interesting web

pages are placed at the highest ranks.

)(1)(

)(1)(

2

1,

2

1

LIi

WWidxm

LIi

WWrp

iiRSRS

(5)

m

p

RS

RSRS (6)

RSp means the entire highness of the ranks of the

interesting web pages in L. It increases as more

interesting pages place in higher rank. RSm does that

in the ideal ranking. RSp/RSm measures how L is close

to the ideal ranking. We used 5 as the value of .

The Rankscore of the ranking K is calculated

similarly.

4.5.3 Verification Result

Fig. 3 shows the Rankscore improvement of the

proposed method compared with the existing one. It is

the value of RS for = 1.0 (the existing method)

subtracted from RS for each in 0.1, 0.2… 0.8, 0.9

with respect to . The positive value means the

proposed method improves RS than the existing one.

The average RS among all examinees improves as

decreases and it takes the highest value at =

0.3. However, it decreases for smaller . This is

because the uninterest indicator includes noisy words.

If the proposed method extracts the uninterest

indicator perfectly, the proposed method should

obtain the highest RS in the case = 1.0. However,

synonym words of the words in the interest indicator

and general words were included in the uninterest

indicator when the proposed method extracted the

uninterest indicator. Because of this fact, the weight

for uninterest indicator is too strong when the

proposed method works with = 1.0.

We can see the optimal value of is 0.3. A

one-tailed t-test finds statistical significance with the

significance level less than 5% (p = 0.0446) in the

difference of RSs between the proposed method with

= 0.3 and the existing method. This result shows

the utility of the proposed method. It also shows that

the proposed method can achieves balancing both of

the precision and recall by = 0.3.

6. Conclusions

This paper proposed a method to recommend web

pages under balancing both precision and recall based

on interest and uninterest indicators. The proposed

method extracts the interesting web pages and

uninteresting web pages automatically to determine

the interest indicator and uninterest indicator.

We have compared the goodness of rankings by

existing method and the proposed method. The

proposed method under = 0.3 improves the

Rankscore significantly compared with an existing


782

method, where 1 is the relative weight of

uninterest indicator with respect to interest indicator.

This proves that the proposed method can recommend

interesting web pages with higher precision and recall.

Future work aims to improve extraction accuracy of

uninterest indicator. By considering the actual

browsed area of web pages and user’s individual

browsing behavior, the accuracy will improves.

References

[1] M. Morita, Y. Shinoda, Information filtering based on

user behavior analysis and best match text retrieval, in:

Proceedings 17th Annual international ACM-SiGiR

Conference on Research and Development in Information

Retrieval, Dublin, 1994.

[2] H. Sakagami, T. Kamba, Learning personal

preferenceson online newspaper articles from user

behaviors, in: Proceedings of the Sixth International

World Wide Web Conference, In Computer Networks

and ISDN Systems, Santa Clara, CA, 1997.

[3] M. Buckland, F. Gey, The relationship between recall and

precision, Journal of the American Society for

Information Science 45 (1994) 12-19.

[4] Google Home Page, https://www.google.co.jp/.

[5] T. Yamamoto, S. Nakamura, K. Tanaka, Rerank

everything: A reranking interface for exploring search

results, in: Proceedings of the 20th ACM Conference on

Information and Knowledge Management, Glasgow,

2011.

[6] K. Otsuki, G. Hattori, H. Haruno, K. Matsumoto, F.

Sugaya, A preference cluster management method based

on user access/non-viewed logs for online news delivery

toward portable terminals, DBSJ Letters 6 (2007) 37-40.

[7] Hatena Bookmark, http://b.hatena.ne.jp

[8] J.S. Breese, D. Heckerman, C. Kadie, Empirical analysis

of predictive algorithms for collaborative filtering, in:

Proceedings of the Fourteenth Annual Conference on

Uncertainly in Artificial Intelligence, Madison, 1998.

https://www.google.co.jp/

http://b.hatena.ne.jp/


Comparison of Contemporary Solutions for High Speed

Data Transport on WAN 10 Gbit/s Connections

Dmitry Kachan1, Eduard Siemens1 and Vyacheslav Shuvalov2

1 Department of Electrical, Mechanical and Industrial Engineering, Anhalt University of Applied Sciences, Kothen 06366, Germany

2 Department of Transmission of Discrete Data and Metrology, Siberian State University of Telecommunications and Information

Sciences, Novosibirsk 630102, Russia


Abstract: This work compares commercial fast data transport approaches through 10 Gbit/s WAN (wide area network). Common

solutions, such as FTP (file transport protocol) based on TCP/IP stack, are being increasingly replaced by modern protocols based on

more efficient stacks. To assess the capabilities of current applications for fast data transport, the following commercial solutions

were investigated: Velocity—a data transport application of Bit Speed LLC; TIXstream—a data transport application of Tixel

GmbH; FileCatalyst Direct—a data transport application of Unlimi-Tech Software Inc; Catapult Server—a data transport application

of XDT PTY LTD; ExpeDat—a commercial data transport solution of Data Expedition, Inc. The goal of this work is to test solutions

under equal network conditions and thus compare transmission performance of recent proprietary alternatives for FTP/TCP within

10 Gigabit/s networks where there are high latencies and packet loss in WANs. This research focuses on a comparison of approaches

using intuitive parameters such as data rate and duration of transmission. The comparison has revealed that of all investigated

solutions TIXstream achieves maximum link utilization in presence of lightweight impairments. The most stable results were

achieved using FC Direct. ExpeDat shows the most accurate output.

Key words: High-speed data transport, transport protocol, WAN acceleration, managed file transport.

1. Introduction

The growing demand for the fast exchange of huge

amounts of data between distant locations has led to

the emergence of many new commercial data

transport solutions that promise to transport huge

amounts of data many times faster than conventional

FTP/TCP solutions. Currently, most common

solutions for reliable data transport in IP networks

are based on the TCP protocol, which was developed

in 1970s. A number of papers describe how TCP,

with some tuning, can perform reasonably on LAN

(local area networks) with a high available

bandwidth [1]. However, it is well known that TCP

has a very limited performance when used in long

Corresponding author: Dmitry Kachan, MEng., research

fields: high speed data transport, IP networks. E-mail: [email protected].

distance networks with a high bandwidth—called

―LFN (long fat pipe network)‖ [2]. For example, a

test with Iperf using the topology described in Fig. 1

on an end-to-end 10 Gbit/slink with a 50 ms RTT

(round trip time) delay and in the presence of at least

0.1% packet loss rate shows a data rate of about 40

Mbit/s. Even after increasing socket buffers and

windows sizes to 128 MiBytes, the performance of

TCP and, accordingly, of most of solutions based on

it (SCP, Rsync, FTP), does not reach more than 60

Mbit/s. Comparable measurements of TCP over 10

Gbit/s were also performed by Wu et al. [1]. In their

work, the authors obtained a data rate of less than 1

Gbit/s even in the presence of a loss rate of 0.001%

and an RTT of 120 ms. They show a significantly

decreasing trend in a data rate with growing packet

loss rate. Another example of TCP weaknesses over

Comparison of Contemporary Solutions for High Speed Data Transport on WAN 10 Gbit/s Connections

784

long distances is described by Armbrust et al. in Ref.

[3], where the transmission of 10 TBytes of data

from Berkeley, California to Seattle, Washington via

a common TCP connection takes about 45 days,

whereas transmission of 10 TBytes hard drive takes

less than one day. A similar solution is described by

Armbrust et al. in Ref. [4]. Nevertheless, many

scenarios of remote collaboration (e.g, cloud

computing) demand data transport with maximum

synchronization times for huge data sets from a few

minutes to hours. As a result, many large companies,

for which the exchange of huge amounts of data is

often critical, avoid using legacy TCP-based

transport solutions and either prefer commercial high

speed approaches based on both TCP and UDP

transport protocols [5, 6] or develop their own

solutions based on an open source fast protocol

stacks such as UDT [7] and RBUDP [8].

Sections 2 and 3 contain brief overview of related

work and the motivation to this work; Section 4

describes the testbed that has been used for

experiments; Section 5 contains short information

about each solution under test and results of the tests;

Sections 6 and 7 are dedicated to comparison of

results, summary and conclusion of the work.

2. Related Work

The main goal of our work is to assess the

capabilities of transport solutions in a 10 Gbit/s

network. Of interest is the maximal possible

end-to-end application data rate on such networks in

the presence of impairments such as packet losses and

high round-trip times. Currently, there are a few

different performance measurements that have been

used to assess these impairments in open source and

freeware solutions. For example, Grossman et al. [9]

present the performance evaluation of UDT [7]

through a 10 Gbit/s network. The article shows how

using UDT and in the presence of 116 ms of RTT, this

network has a maximum throughput of 4.5 Gbit/s

within a single data stream. Two parallel streams

achieve together about 5 Gbit/s and within 8 parallel

streams about 6.6 Gbit/s are achieved. Further, a

performance result for data transmission using

RBUDP was presented at the Cine Grid 3rd Annual

International Workshop [10]. While the disk access

speed limited the data transport speed to 3.5 Gbit/s, on

the link from Amsterdam to San Diego only 1.2 Gbit/s

was reached. The distance of that path is about 10,000

km through optic fiber, which corresponds to about

100 ms of RTT.

Most other data transport performance results are

presented for 1 Gbit/s networks e.g. three rate based

transport protocols have been evaluated by Wu et al.

in Ref. [11]: RBUDP, SABUL and GTP. The overall

data rate of applications based on these protocols was

compared for all three protocols and for ―standard

unturned TCP‖. The experiment was performed on a

real network in the presence of 58 ms of RTT and a

loss rate of less than 0.1%. The results showed that all

solutions utilize the 1 Gbit/s link approximately 90%.

These test results show that for open source data

transfer solutions, even those using parallel streams, it

is quite hard to achieve full, or even close to full,

utilization of 10 Gbit/s links.

For commercial closed source solutions, the

situation differs significantly. There are several

published performance results of commercially

available solutions, provided by the manufacturers

themselves: Velocity [12], TIXstream [13], FC Direct

[14] and Catapult Server [15]—all of whom report

perfect results. However these results are mainly

providing commercial information to attract potential

customers and the investigative conditions vary. To

overcome this deficit, the main idea behind the

auhtors’ work is to place all investigated

solutions under equal conditions within the same

environment.

3. Background

For IP networks, packet loss behavior depends on

many factors such as quality of transmission media,


785

CPU performance of intermediate network devices,

presence of cross traffic etc.. It is therefore

impossible to use one universal value of packet

losses for all cases. The best way to assess the

approximate values of packet losses is through

empirical measurements. In Ref. [16], Paxson

discusses the heterogeneity of packet losses and

shows that, even in 1994, the value of packet loss

rate in experiments between 35 sites in 9 countries

was about 2.7%. He also shows that, within one year,

the value of packet losses increased up to 5%.

Probably, such packet loss values are not relevant to

the current Internet; however the author pointed out

that distribution of packet losses is not uniform. Thus,

for some connections, ACK packet loss was not

observed at all. Nevertheless, relative values of all

lost IP packets in both directions for all experiments

were approximately equal. Recent views on the

packet loss ratio are presented by Wang et al. in Ref.

[17]. In this research, tests were made across 6

continents between 147 countries, involving about

10,000 different paths. The authors show that across

all continents for more than 70% of continental

connections, packet loss rate is less than 1%, in

Europe and North America this value is even on

about 90% of connections. The authors also

highlighted that for intercontinental connections,

packet loss value in general is lower than for

intra-continental—across the entire world, packet

loss rate is lower than 1% for about 75% of the

connections.

In Ref. [18], Settlemyer et al. use a hardware

emulator to emulate 10 Gbit/s paths, and they

compare throughput measurement results of the

emulated paths with real ones. The maximal RTT of a

real link used in the research is 171 ms. The authors

have shown that differences between profiles of both

kinds of paths—emulated and real ones—are not

significant, and they conclude that using emulated

infrastructure is a much less expensive way to

generate robust physical throughput.

4. Testbed Topology Description

In this work, the following solutions have been

investigated: Velocity, TIXstream, FileCatalyst Direct,

Catapult Server, ExpeDat. Manufactures of all these

solutions claim that their transport solutions are able

to handle data transmission via WAN in the most

efficient way.

Since all these solutions are commercial and closed

source, it was necessary to get in touch with the

support team of each manufacturer for both obtaining

trial licenses of their products and consulting them

about achieved results. Unfortunately, not all

manufactures were interested in such investigations.

Thus, for example, it would have been interesting to

test Aspera’s solution for fast data transport. However

the authors received no answer from this vendor.

To avoid unexpected inaccuracies, the scheme of

test topology is kept simple. Fig. 1 presents the

typology. The core of the test environment was the

WAN emulator Apposite Netropy 10G [19], which

allows the emulation of WANs under various

conditions such as packet delay, packet loss rate and

jitter in different variations, with an accuracy of about

20 ns. By comparison, software emulators, such as

Net Em, provide an accuracy of about tens of

milliseconds and this value is greatly dependent on the

hardware and operating system [20]. Moreover,

software emulators are very limited in their maximum

achievable data rates. Apposite 10G, for example,

enables a transmission through Ethernet traffic with

an overall throughput of up to 21 Gbit/s on both

copper and fiber optic links.

The testbed topology used here contains two

servers, connected via the 10 Gbit/s Ethernet switch

Extreme Networks Summit x650 and via the WAN

Emulator.

Fig. 1 Logical view of topology.

Internet

Host 1 Host 2

RTT up to 200 Packet loss up to 1%


786

The typology was implemented by means of fiber

optics with a 10 Gbit/s bandwidth in Fig. 2.

There is no background traffic on the path since this

investigation focuses on the pure applications’

performance, not on the fairness aspects of the

protocols. The setup corresponds to the case when a

L2-VPN is used for big data transmission and another

application’s traffic is isolated by means of QoS.

Each server is equipped as follows:

(1) CPU: Intel Xeon X5690 @3.47GHz;

(2) RAM: 42 GiBytes (speed 3,466 MHz);

(3) OS: Linux CentOS 6.3;

(4) NIC: Chelsio Communications Inc T420-CR, 10

Gbit/s.

Operating system socket buffers were extended up

to:

(1) /proc/sys/core/net/wmem_max—64MiBytes

(2) /proc/sys/core/net/rmem_max—64MiBytes

The MTU size of all network devices along the path

was set to 8,900 Bytes.

For sending and receiving data with a rate of 10

Gbit/s, it is necessary to have a storage device that can

read on the sender side and write on the receiver side

with a sustained rate not less than 1,250 MByte/s

(corresponding to 10 Gbit/s). Off-the-shelf hard drives

provide read and write rates of up to 100 MByte/s, so

in the investigated case, data transfer rate would have

been limited by the hard drives. To circumvent this

limitation, storage systems such as RAID arrays with

write/read rates not less than the expected transport

rate must be used.

In the presented experiments, both storage write

and read bottlenecks and inefficient file access

implementations were avoided by using a RAM-based

Fig. 2 Technical view of topology.

file system on both servers. In comparison to common

hard drives, the read rate of RAM disk, as obtained in

several test runs during these investigations, was not

less than 4,500 MiBytes/s; the write rate of RAM disk

was not less than 3,000 MiBytes/s. Therefore, the

servers used for tests were equipped with 42 GiBytes

of RAM onboard, but due to the operating system’s

RAM requirements, it was only possible to use 30

GiBytes of space on RAM disk for test purposes.

Under ideal conditions, a transmission of 30

GiBytes through the network with a bandwidth of 10

Gbit/s without impairments, as explained in Eq. (1),

should take about 26 seconds.

sRi

ST 76.25

1010

10/81024303

63

(1)

where, T—time of transmission; s–size of data,

iR —ideal data rate.

However, this calculation disregards L2-L4 headers

along with some proprietary protocol headers and the

overhead for connection management and

retransmissions.

So, under real conditions, for some packet overhead

and retransmission handling, each solution needs a

certain amount of time for connection initialization

and the releasing of the network path. Besides this, in

high-performance implementations, initialization of

the protocol stacks and the internal buffers often takes

a significant amount of time, which is also

investigated during this research.

5. Experimental Results

The experiment on each data transport solution

under consideration consists of 25 consecutive tests.

Each test comprises the transfer of a 30 GiBytes file

from one server to another through the network

emulator. The RTT latency range is varied from 0 to

200 ms in steps of 50 ms and the packet loss rate takes

the values 0, 0.1%, 0.3%, 0.5% and 1%. Since 1 km of

fiber optics delays the signal by about 5 µs, the

maximum RTT in this test corresponds to 20,000 km

of fiber channel in both the forward and the backward

directions. The RTT is configured symmetrically

Server 1 Server 2Extreme Networks

Summit x650

Apposite


787

across the forward and backward paths of the

emulator; thus 200 ms of RTT would delay data by

100 ms and acknowledgments or other control

information in the backward direction by another 100

ms. The packet losses are randomly injected according

to a normal distribution, whereby the set loss ratio is

applied to both forward and backward direction. Such

packet loss behavior is easier to reproduce, and it is

more complicated for protocols to handle than typical

packet loss behavior on the Internet [16]. An attempt

was made to configure each solution so that the

maximum possible data rate and the minimum

possible overall transmission time were achieved. The

tuning of the operating system and the configuring of

parameters are described below for each solution. All

the tests were repeated 4 times to avoid inaccuracies,

and the best result of each series is presented on the

plots.

The results of each test contain two parameters:

data rate and transfer duration. The first parameter is

average data rate, i.e. the average speed of data

transportation shown by the application during the

experiment. The second parameter is independent of

the solution output and represents the time interval.

This time interval was collected by means of the

operating system and shows the period of time

between the launching of the send command and the

time of completion of this command. This time

interval contains not only the time of actual data

transmission but also the time for service and

retransmission overhead.

5.1 Velocity

This solution was developed in the USA. It is a

TCP-based file transfer application, and according to

the vendor’s web site, it allows the available path

capacity to be fully utilized. Velocity ASC is also

available with on-the-fly data encryption of up to 24

Gbit/s and AES encryption of up to 1,600 Mbit/s. The

supported platforms are Windows, Mac OSX, Linux

and Solaris. According to the user manual, this

solution automatically adapts its parameters to

network conditions and chooses the optimal

parameters for data transmission. Fig. 3 shows the

behavior of the transport rate. The results of tests in

the presence of delays of more than 0.1% are not

shown since the data rate here was lower than 100

Mbit/s.

Increasing latencies do not significantly affect

Velocity’s data rate behavior, slowing it down to only

8 Gbit/s. The solution performs reasonably in the

presence of small packet loss rates without any

emulated delay (back-to-back RTT latency in the

testbed is about 0.15 ms). Thus it achieves a data rate

of 9 Gbit/s in the presence of 0.1% of packet loss, and

this value decreases down to 500 Mbit/s with a packet

loss of 1%. However, this configuration does not

correspond to the situation in Wide Area Networks. In

the presence of 0.1% packet loss and at least 50 ms

RTT, the data rate is reduced to 250 Mbit/s.

Multi-streaming TCP is using by default velocity. It

opens 7 TCP sockets on every single test on each side.

When the number of streams is manually set to 1, the

data rate in presence of 200 ms RTT without packet

loss is about 2.2 Gbit/s. The transfer durations of the

solution are shown in Fig. 4. The numbers on the plot

are obtained for two cases: without latency and with a

latency of 200 ms. A result worth noting was obtained

at a loss rate of 0.1% and an RTT of 0 ms. Under

these conditions, the data rate in the presence of losses

is, with 800 Mbits/s, less than the value without loss

Fig. 3 Behavior of data rate of velocity.


788

Fig. 4 Data transfer duration of velocity.

rate. However, the transfer duration in the latter case

is higher by only 0.1 ms. This behavior was observed

in several experiments.

5.2 TIXstream

This transfer engine has been developed by

TixelGmbH, Germany, which spun off from

Technicolor Corporate Research in 2010. The core of

TIXstream is Tixel’s proprietary RWTP (reliable

WAN transfer protocol) [21], which provides

high-performance data transmission between two

hosts in the network using only one UDP-socket on

each host. The application works under Linux OS.

TIXstream 3.0 (the latest version) provides up to 20

Gbit/s end-to-end performance. It has a

platform—independent web-based user interface for

the management of data transmission between remote

SAN- and NAS systems. TIXstream also provides

on-the-fly AES-256 encryption of data without any

effect on data rate [22]. TIXstream has a peer-to-peer

architecture and uses one TCP socket for control

communication and one UDP socket for data

transmission connection on both sides.

Parameters of application:

(1) RWTP Buffer size—4,362,076,160 Bytes (4

GiBytes);

(2) MSS = 8,800 Bytes;

(3) Receiver buffer size (on both sides) =

1,073,741,824 Bytes (1 GiByte);

(4) Sender buffer size (on both sides) =

1,073,741,824 Bytes (1 GiByte).

The behavior of TIXstream’s data rate as a function

of network impairments is shown in Fig. 5.

There is no visibly decreasing effect on data rate

behavior till 100 ms RTT and till 0.3% of packet loss.

The solution achieves not less than 9.7 Gbit/s (97% of

capacity) with these impairments. With higher delays

in the presence of heavy packet losses, the figure

shows decreasing data rates down to 3,750 Mbit/s, as

on a path with 200ms of RTT and 1% of packet losses.

However, with modest impairments that correspond to

fairly normal WAN links, for example RTT 150 ms

and packet loss rate 0.1%,TIXstream achieves a data

rate of about 8,700 Mbit/s; an 87% utilization of a 10

Gbit/s link. It is worth noting that in the presence of

50 ms of latency, TIXstream performs better than

without any latency for all values of loss rate. This

behavior was found in several experiments.

Fig. 6 shows that the transfer duration quite

accurately corresponds to the behavior of the data

rates. However, the theoretically minimum time of

transmission calculated in Section 0, with a data rate

of 8,700 Mbit/s, is 29.62s versus the 37.25 s that was

measured in the experiment. These 7.63 s were spent

on connection initialization, and the establishing and

releasing of the control channel. Since no packet loss

Fig. 5 Behavior of data rate of TIXstream.

Fig. 6 Data transfer duration of TIXstream.


789

shall occur on this link, time for packet retransmission

shall be neglected.

5.3 FileCatalyst Direct

FileCalatyst Direct was developed by Unlimi-Tech

Software Inc., a Canadian based company. Like

TIXstream, it transmits data via UDP and implements

packet loss management, rate and congestion control

in the user layer. The application obeys a client-server

architecture and the solution operates under Windows,

Mac OSX, Linux and Solaris operating systems. The

data sheet on the vendor’s website shows that this

solution provides data rates of up to 10 Gbit/s [23] and

that there is an option to use AES encryption for

secure transmission. FC Direct provides both,

graphical and command line user interfaces for server

and client applications.

Parameters of application:

(1) Start rate = 9,000,000 (9 Gbit/s);

(2) MSS = 8,800 Bytes;

(3) Buffers = 3,840,000,000 Bytes (3,58 GiBytes);

(4) Number of send sockets = 10;

(5) Number of receiver sockets = 4.

As shown on Fig. 7, FC Direct achieves 90% to

94% link utilization, even under high network

impairments. Data rate behavior is fairly immune to

growing latency and packet loss ratio. The data rate of

FC Direct shows values between 9 Gbit/s and 9.4

Gbit/s for all the tests. During the tests, the Linux

system monitor reveals that each data transmission

opens 10 UDP sockets on the sender side and 4 UDP

sockets on the receiver side and 1 TCP socket on each

side. In this mode, maximal data rates can be achieved.

Data packets from ten sender sockets are not

uniformly distributed over all four receiver sockets,

but according to a special proprietary distribution rule.

The vendor does not call it multi-streaming but ―more

intelligent resource management‖. However, with this

behavior, significant firewall transversal issues are to

be expected.

The distribution of session time durations showed

on Fig. 8 has slightly monotonically increasing

behavior with rising latencies.

5.4 Catapult Server

The Catapult Server is TCP-based and was

developed by XDT PTY LTD, Australia. This

solution follows a client-server architecture and,

according to the vendor’s web site, provides up to 8

Gbit/s on the 10 Gbit/s link. The solution functions

under the Windows, MAC OSX and Linux operating

systems. The vendor positions the solution as a high

data rate transmitting tool for networks with a high

level of latency but without any packet losses. To

prepare the operating system for high speed

transmissions, the vendor suggests using a shell script

to change network parameters. By default, this script

extends the TCP buffers to 32 Mbytes. However, for

the authors’ tests, the 64 Mbytes setting was chosen

since better performance had been reached with this

setting.

The script changes are:

(1) tcp_congestion_control = htcp;

(2) net.ipv4.tcp_rmem = 40,968,738,067,108,864;

Fig. 7 Behavior of send rate of FC direct.

Fig. 8 Data transfer duration of FC direct.


790

(3) net.ipv4.tcp_wmem = 40,966,553,667,108,864;

(4) net.ipv4.tcp_no_metrics_save = 1;

(5) net.core.netdev_max_backlog = 250,000;

(6) net.core.rmem_max = 67,108,864;

(7) net.core.wmem_max = 67,108,864.

To improve the behavior of this solution in the

presence of packet losses, the manufacturer’s support

team also suggested applying the following

configurations:

(1) net.ipv4.tcp_timestamps = 1;

(2) net.ipv4.tcp_sack = 1.

Note that the command line client of XDT-sling

shot copy, which was used for the tests, shows the

data rate as GB/s, probably it means GiByte/s.

Furthermore the value 1.1 GB/s immediately follows

the value 1.0 GB/s, without any intermediate values.

However, the solution shows a transfer duration with an

accuracy of up to milliseconds. Therefore, the data rate

was calculated as shown on Eq. (2):

,T

S

0

XDTR (2)

Where by RXDT—is the data rate of XDT, which is

used for result presentation; S—data size (30 GiByte),

T0—transfer duration from the output of client

application.

Fig. 9 represents the data rate of Catapult Server

dependent on network impairments. The presence of

packet losses on the link makes the transmission

ineffective, so the data rate is reduced to less than 100

Mbit/s. However, in the presence of 150 ms RTT

without packet loss, transmission is about 8,300

Mbit/s

Fig. 10 shows the transfer durations for Catapult

technology.

5.5 ExpeDat

ExpeDat is a UDP-based data transport solution

developed by Data Expedition Inc., USA. The core of

this application comprises the MTP (multipurpose

transaction protocol) [24], developed by the founder

of Data Expedition. ExpeDat supports Windows, Mac

OSX, Linux / Solaris, NetBSD / FreeBSD, AIX and

HP-UX platforms. According to the company’s web

site, ExpeDat allows transmission of data with 100%

utilization of allocated bandwidth and in the presence

of on-the-fly AES encryption [25]. It implements the

transport protocol logics on a UDP channel, and uses

a single UDP socket on each side of the connection

for both data transmission and control information.

Though the product web site [25] claims that the

solution has ―zero-config installation‖, the significant

increase of data rate, namely from 2 Gbit/s up to 9

Gbit/s, even without impairments (RTT = 0 ms,

packet loss = 0%), was obtained only after application

of configuration changes as follows:

MSS—8,192 Bytes;

Environment variable MTP_NOCHECKSUM = 1.

With high values of packet loss on the channel, the

higher results were achieved using the following

option on the command line:

N 25.

Fig. 9 Behavior of data rate of XDT catapult.

Fig. 10 Data transfer duration of XDT catapult.


791

The use of this option shows that heavy packet loss

rate is introduced in a channel.

In Fig. 11, the data rate values of ExpeDat tests are

presented. The plot shows that network latencies lead

to a much higher reduction of the transmission rate

than packet loss.

The distribution of transfer times is presented in

Fig. 12.

6. Comparison of the Solutions

Since not all of the investigated solutions perform

well in the presence of heavy packet losses, the

comparison of data rates was split into two stages. The

first stage is dedicated to a comparison of all

presented solutions on the networks with different

packet latencies and without any packet loss. In the

second stage, the solutions are compared under harder

conditions for terrestrial networks—with packet loss

of 1% and the whole range of investigated RTT. Only

solutions that show a data rate higher than 1% from

maximal channel capacity (100 Mbit/s) have been

considered in the second stage.

A comparison between the distribution of transfer

duration without packet losses and the ideal value

shows which solution spends more time on service

needs such as the initialization and releasing of a

channel.

A further comparison shows the discrepancy

between the actual time of transmission and the

calculated time from the output of all solutions. This

analysis is also split into two stages as described

above and shows how the values from the output of

the particular solution correspond to reality.

Fig. 13 shows a consolidated diagram of

transmission data rates of all tested solutions in the

presence of increasing latency and without any packet

loss on the path. The first set of bars shows how fast a

large set of data can be transmitted in a back-to-back

connection. For this case, the highest result was

achieved by Velocity. However all solutions showed

results of not less than 9.3 Gbit/s. With increased

latencies, Velocity performs worse, and of all the

remaining cases, TIXstream shows the best

performance with up to 9.8 Gbit/s (98% channel

utilization) without any significant decrease at higher

RTTs. FC Direct also shows very stable results. For

all presented impairments, its data rate lies between

9.2 Gbit/s and 9.3 Gbit/s. All other solutions show

decreasing data rates on increasing round-trip-times.

TCP-based solutions obviously can not cope

efficiently with the presence of packet losses on the

path. Although for all solutions except Velocity, the

Fig. 11 Behavior of data rate of ExpeDat.

Fig. 12 Data transfer duration of ExpeDat.

Fig. 13 Comparison of data rates of investigated solutions;

packet loss = 0.


792

support teams of the respective manufacturers were

involved, those solutions did not provide adequate

results in the presence of packet loss. Fig. 14

represents the behavior of solutions of stage two in the

presence of 1% of packet loss.

With an RTT of 0 and 50 ms, TIXstream shows the

best results. However, starting at 100 ms, throughput

is decreased whereas FC Direct shows nearly constant

data rates. The ExpeDat data rate abruptly decreases

down to 5.7 Gbit/s on zero-delay links, and with

increasing latencies ExpeDat’s results are lower than 1

Gbit/s.

As pointed out in Section 0, the theoretical

minimum transfer duration for the transport of 30

GiBytes of data via a 10 Gbit/s WAN is 25.76 s. Fig.

15 compares, for each solution, the ideal value with

the lowest transfer durations obtained during the

experiments.

The minimum transfer duration was achieved by

XDT Catapult server. The time is only 1.4 s longer

than the theoretical minimum. This means that XDT

Catapult initializes the software stack along with

protocol buffers, and establishes and closes the

connection within less than 1.4 s. The time overhead

of Velocity, TIXstream and ExpeDat is slightly higher

but still less than 2 seconds. The worst result was

obtained when using FC Direct: it needed about 3.8 s.

for ramping-up and finishing the application.

Also of interest is the accuracy of the performance

outputs of the transport solutions. During the

experiment, the actual data rate was obtained from the

output of the running application, and program run

time was also measured by means of the operating

system. Transfer durations with a transmission of 30

GiByte with a data rate from output are calculated as

shown on Eq. (3).

][,R

[ B i t s ] S, t

0 sBits

T (3)

where Tt—is the calculated transfer duration; R0—data rate from output; S—data size (30 GiByte).

The differences between calculated transfer

durations and real program run time for the tests

performed without packet loses are presented in Fig.

16. The discrepancy of the values generally has an

increasing trend at higher latencies. Velocity showed

the lowest value of discrepancy along the tests without

Fig. 14 Comparison of data rate of tested solutions;

packet loss = 1%.

Fig. 15 Comparison of transfer durations with theoretical

minimum.

Fig. 16 Difference between calculated and actual transfer

durations; packet loss = 0.

Fig. 17 Difference between calculated and actual transfer

durations; packet loss = 1.


793

packet loss. The second TCP-based

solution—XDTCatapult shows good results on RTTs

below 200 ms. However, with 200 msRTT, this

solution shows the worst of all results. ExpeDat

almost always has the lowest discrepancy values for

all cases. In the presence of RTTs of 100 ms and 150

ms without loss rate, the solution showed negative

result, meaning that the actual transfer duration was

lower than the calculated one. It is evident that the

output of some solutions show the average achieved

data rate including service processes such as

connecting, initializing and releasing the link, and

some of the solutions show the average data rate of

the transmission process only.

Similar to Fig. 16 but with a packet loss of 1%, Fig.

17 shows the differences of calculated transfer

durations and real program run time. A comparison of

these two figures shows that FC Direct has almost the

same discrepancies for all RTT cases except for the

200 ms and 150 cases, where the discrepancy is higher

in the presence of packet loss than without it. The

results of TIXstream have a decreasing trend and

ExpeDat shows again the lowest values—the actual

times of transmission are almost equal to the

calculated ones.

7. Conclusions

This work compares the state of the art of

commercial solutions for reliable fast data transport

via 10 Gbit/s WAN IP networks in the presence of

high delays and varying packet loss rates. The main

problem of such research is that the vendor companies

usually hide the technology used for the accelerated

data transport. The protocol used in ExpeDat solution

—MTP—is covered by some US patents. However

this does not mean that ExpeDat does not use any

algorithms besides the ones described in those patents.

The only independent method to assess these

commercial solutions is to externally observe the

solutions during tests under well-defined conditions.

All investigated solutions position themselves as

reliable high-speed transfer applications designed to

provide alternatives to FTP/TCP and overcoming the

pure TCP performance on 10 Gbit/s-WAN

connections by orders of magnitude. Two of them,

Velocity and XDT Catapult, exploit the TCP stack of

the Linux OS and the rest-FC Direct, TIXstream and

ExpeDat use UDP sockets and implement the protocol

logics in the user-level.

The results obtained show that solutions based on

TCP inherit its native problems on 10 Gbit/s links—a

significant decrease of data rate down to 1% of the

link capacity in the presence of packet loss on the path.

The commercial solutions achieve a higher speed by

increasing TCP window size or by establishing

multiple parallel TCP streams. However, the

experiments show that this solution only works on

links without any packet loss. However, even the

known STCP [26] on WAN networks with a low loss

rate show a reasonable result of about 5 Gbit/s [1].

Although in that paper, the authors tested pure

protocol performance, their results show that it is

possible to achieve good results by only tuning the

TCP on such networks.

UDP-based solutions show a good utilization of a

10 Gbit/s path even under bad network conditions

such as a loss rate of 1% in the presence of RTTs of

up to 200 ms. The best link utilization at the highest

impairment value was achieved by FileCatalyst

Direct—the values were never lower than 93% for all

performed tests. For the loss ratio up to 0.3% and RTT

up to 100 ms, TIXstream shows a better utilization of

about 97 %.

Transmission duration measurements were

primarily intended to prove that the solutions show

accurate data transport numbers in their outputs. The

comparison showed that the lowest transfer duration

of each solution is fairly close to the ideal one and that

the discrepancy of the obtained output values are close

to reality for all solutions.

Each solution uses some time for the allocation of

system resources and the initialization of network


794

resources. This time can not be neglected, at least not

in sessions with up to 30 GiBytes data transport. The

comparison presented in Fig. 15 attempts to assess

this service time. Probably, the time overhead is also

due to the solutions not fully utilizing the bandwidth.

It was found that the data rate of FC Direct is not the

lowest one, but the transfer duration is higher than for

all other solutions under test conditions. This result is

possibly due to known java performance bottlenecks,

because FC Direct is the only solution written purely

in java.

When the solutions work under very light network

impairment conditions (for example back-to-back) and

the data rate achieves maximal value, the CPU usage

is fairly high. For example, the maximum achieved

data rate of ExpeDat seems to be due to CPU

limitations. The system monitor showed 99% CPU

usage in the ExpeDat process, and it also showed that

one core of twelve is used in 99%. Other solutions, e.g.

TIXstream, showed a CPU usage of about 150%, the

usage of two used cores was about 70% and 80%

respectively on the sender side, and on the receiver

side 3 cores were used with a usage of 40%, 90%,

30%. This solution distributes the performance among

several cores to maximally use the available

bandwidth when possible.

One more significant point of resources

management is the socket use. As shown in Section 0,

FC Direct uses different numbers of sockets on the

sender and receiver sides. This use causes no

problems for corporate LANs or simple back-to-back

connections. However, for data transport using more

complex structures, like real Internet connections, this

use could cause problems on such devices such as

firewalls. Such problems are well known even in

simple multi streaming cases. This is an even worse

situation in which each sender socket is sending data

to different destination ports, so at least M × N port

pairs must be tunneled in the firewall. It is very likely

that intrusion detection systems can consider such

behavior as violent traffic.

The analyzed solutions were tested on their abilities

in the presence of high values of latency and packet

losses. However, delay jitter is also a common

network impairment and measurements using

different values and different jitter patterns would also

be of interest.

The present research shows the behavior of the

solutions in an empty path such as VPN. Further

investigations could be made into the behavior of the

solutions in the presence of back-ground traffic.

The testbed topology was simplified to get a first

representation of the presented solutions. An

extension of the experimental topology makes sense

for in-depth investigations.

During the experiments, only the performance of

data transfer for commercial applications was

investigated. To get a deeper understanding of only

the telecommunication part, it would be of interest to

make tests with the technology cores (e.g. protocol

stacks without any wraps such as file system).

The questions of system resource consumptions

were addressed very briefly here. It would also be

interesting to research this topic more extensively.

References

[1] Y. Wu, S. Kumar, S.J. Park, Measurement and

performance issues of transport protocols over 10 Gbps

high-speed optical networks, Computer Networks 54

(2010) 475-488.

[2] H. Kamezawa, M. Nakamura, M. Nakamura, Inter-layer

coordination for parallel TCP streams on long fat pipe

networks, in: Proc. of the 2004 ACM/IEEE conference on

Supercomputing, Pittsburg, PA, USA, 2004.

[3] M. Armbrust, A. Fox, R. Griffith, A.D. Joseph, R.H. Katz,

A. Konwinski, Above the clouds: A berkeley view of

cloud computing, Tec. Rep. UDB/EECS 28 (2009) 19-25.

[4] M. Armbrust, A. Fox, R. Griffith, A.D. Joseph, R.H. Katz,

A. Konwinski, et al., A view of cloud computing,

Communications of the ACM 53 (2010) 50-58.

[5] S. Hohlig, Optimierter dateitranfer über 100 Gigabit/s, in:

100 Gigabit/s Workshop of the DFN, Mannheim, 2011.

[6] Aspera, Custumer Deluxe Digital Studios, [Online],

http://asperasoft.com/customers/customer/view/Customer

/show/deluxe-digital-studios/ (accesseed: Nov., 2012).

[7] Y. Gu, R.L. Grossman, UDP-Based datatransfer for

high-speed wide area networks, Computer Networks 51


795

(2007) 1465-1480.

[8] E. He, J. Leigh, O. Yu, T.A. DeFanti, Reliable blast UDP :

Predictable high performance bulk data transfer, in: Proc.

of IEEE Cluster Computing, Chicago, USA, 2002.

[9] R.L. Grossman, Y. Gu, X. Hong, A. Antony, J. Blom, F.

Dijkstra,Teraflows over Gigabit WANs with UDT,

Journal of Future Computer Systems 21 (2005)

501-513.

[10] L. Herr, M. Kresek, Building a New User Community for

Very High Quality Media Applications on Very High

Speed Networks, [Online], CineGrid,

http://czechlight.cesnet.cz/documents/publications/networ

k-architecture/2008/krsek-cinegrid.pdf. (accessed: Feb.,

2013).

[11] X. Wu, A.A. Chien, Evaluation of rate-based transport

protocols for lambda-grids, in: Proc. of 13th IEEE

International Symposiumon High Performance

Distributed Computing, Honolulu, Hawaii, USA, Jun. 4-6,

2004.

[12] Bitspeed LLC, From Here to There-Much Faster,

Whitepaper,[Online]

http://www.bitspeed.com/wp-content/uploads/2011/10/Bi

tSpeed-White-Paper-From-Here-to-There-Much-Faster.p

df (accessed: Oct., 2012).

[13] Tixel GmbH, Tixstream: Overview, [Online],

http://www.tixeltec.com/ps_tixstream_en.html (accessed:

Oct., 2012).

[14] File Catalyst, Accelerating File Transfers, Whitepaper,

[Online],

http://www.filecatalyst.com/collateral/Accelerating_File_

Transfers.pdf (accessed: Oct., 2012).

[15] XDT PTY LTD, High-Speed WAN and LAN data

transfers, XDT, [Online],

http://www.xdt.com.au/Products/CatapultServer/Features

(accessed: Oct., 2012).

[16] V. Paxson, End-to-End internet packet dynamics,

networking, IEEE/ACM Transactions 7 (1999) 277-292.

[17] Y.A. Wang, C. Huang, J. Li, K.W. Ross, Queen:

Estimating packet loss rate between arbitrary internet

hosts, in: Proc. of the 10th International Conference on

Passive and Active Network Measurement, Seoul, Korea,

2009.

[18] B.W. Settlemyer, N.S.V. Rao, S.W. Poole, S.W. Hodson,

S.E. Hick, P.M. Newman, Experimental analysis of 10

Gbps transfers over physical and emulated dedicated

connections, in: Proc. of Computing, Networking and

Communications (ICNC), Maui, Hawaii, USA, 2012.

[19] Apposite Technologies, [Online],

http://www.apposite-tech.com/index.html (accessed: Oct.,

2012).

[20] A. Jurgelionis, J.P. Laulajainen, M.I. Hirvonen, A.I.

Wang, An Empirical Study of NetEm Network Emulation

Functionalities, ICCCN (2011) 1-6.

[21] Tixel GmbH, White Papers and Reports, Tixel, [Online],

http://www.tixeltec.com/papers_en.html (accessed: Nov.,

2012).

[22] Tixel GmbH, Tixel news. tixel.com, [Online],

http://www.tixeltec.com/news_en.html (accessed: Oct.,

2012).

[23] FileCatalyst, Direct, [Online],

http://www.filecatalyst.com/collateral/FileCatalyst_Direc

t.pdf (accessed: Oct., 2012).

[24] Data Expedition, Difference, [Online],

http://www.dataexpedition.com/downloads/DEI-WP.pdf

(accessed: Oct., 2012).

[25] Data Expedition, [Online],

http://www.dataexpedition.com/expedat/ (accessed: Oct.,

2012).

[26] R. Stewart, Q. Xie, Motorola, K. Morneault, C. Sharp,

Cisco, Stream Control Transmission Protocol, IETF, RFC

2960, [Online], http://www.ietf.org/rfc/rfc2960.txt

(accessed: Jan., 2013).

http://www.bitspeed.com/wp-content/uploads/2011/10/BitSpeed-White-Paper-From-Here-to-There-Much-Faster.pdf



http://www.xdt.com.au/Products/CatapultServer/Features


Awaken the Cyber Dragon: China’s Cyber Strategy and

its Impact on ASEAN

Miguel Alberto N. Gomez

Center for Networking and Information Security, College of Computer Studies, De La Salle University, Manila 1004, Philippin


Abstract: The increase in cyber attacks launched against the ASEAN (Association of South East Asian Nations) regional bloc that have been attributed to the PRC (People’s Republic of China) sets a precedent for the future of low impact cyber conflicts. Conflicting interests between members and the nature of the ASEAN principles suggests that any eventual cyber defense policies can only be framed within the context of individual state interests. This paves the way for low impact cyber attacks that may enable the aggressor to influence the different instruments of national power within the region and serves as a viable tool to project power and influence with minimal risk of escalation. This study aims to discuss how the PRC has adopted this strategy in response to recent disputes with members within ASEAN. The study highlights the low impact approach that the PRC has chosen as a means of exerting its influence within the region. The study discusses the underlying factors that allow the PRC to operate freely by taking advantage of the fundamental weaknesses of ASEAN as a platform for establishing a cyber defense mechanisms within the region and goes on to caution as to such long term repercussions .

Key words: Cyber defense, cyber warfare, cyber strategy, security informatics.

1. Introduction1

The appearance of several high-profile cyber

security incidents the past year has brought the

possibility of cyber war into mainstream

consciousness. Incidents such as last year’s Stuxnet

outbreak and related malware Duqu and Flame, as

well as the recent compromise of the oil firm Saudi

Aramco has prompted commentary regarding the

implications these threats have on global security. So

that in October 2012, United States Defense Secretary

Leon F. Panetta has been quoted saying that the

United States faces an impending cyber-Pearl Harbor

that would be capable of crippling the nation’s critical

infrastructure [1]. While discourse such as this is not

novel, the frequency with which doomsday scenarios

have been proposed is increasing. The question that

faces policy makers, military leaders, and security

1Corresponding author: Miguel Alberto Gomez, lecture, research fields: security informatics, machine learning. E-mail: [email protected].

professionals alike is whether or not these events

signal a future trend in warfare. While acknowledging

the possibility of significant damage caused by attacks

against existing and future cyber infrastructure, a

study of current cyber conflicts demonstrate that the

impact of most have been imperceptible or

significantly milder than what has previously been

claimed; influencing public perception and national

policy rather than causing damage to key

infrastructure. To validate this argument, the study

uses the case of the PRC (People’s Republic of

China)—whose cyber capabilities have been

acknowledged by analysts in the field—and its recent

territorial conflicts with members of the ASEAN

regional bloc, namely the PHL (Republic of the

Philippines).

To support these claims, the study is divided into 4

sections. The first provides a distinction between

cyber war and cyber conflict. The distinguishing

factors between the two serve as a fundamental point

Awaken the Cyber Dragon: China’S Cyber Strategy and its Impact on Asean

797

to explain the PRC’s current approach. The second

and third section explores how cyber conflicts, and

LICC (low impact cyber conflicts) in particular,

contribute to the PRC’s cyber strategy and foreign

policy. Finally, the last section covers ASEAN’s how

inherent weaknesses are exploited when the PRC’s

cyber strategy is put to bear as a means to support

on-going policies and disputes. All discussions within

this study are viewed from the perspective of the

activities that take place between the PRC and

ASEAN. This, however, does not mean that

conclusions derived from this study are not applicable

to other regions or situations. So long as the initial

conditions that lead to LICCs are present, the

proponent believes that the thesis should hold true.

For the succeeding sections of this paper, the term

cyber conflict is used in place of cyber warfare. This

distinction is crucial, as the proponent of this study

does not believe that events that have occurred and

continue to transpire in cyber space are true forms of

warfare as per the definition used by Clausewitz.

Further discussion regarding this distinction is made

in the second section of this study.

2. Cyber War vs. Cyber Conflict

Existing literature and popular media have labeled

recent cyber attacks as being instances of cyber war.

While parallels have been drawn between kinetic

warfare and that of cyber warfare, these two can not

be taken or viewed on equal footing due to inherent

differences between them. This point is emphasized if

one were to take the definition of war as established

by Clausewitz wherein he provides three

characteristics that must first exist, these are its:

violent character, instrumental character, and political

nature. The first refers to the violent characteristics of

warfare, which is crucial as this is viewed to be

present in all forms of conflict that have been labeled

as war and not merely in the metaphorical sense of the

word. Second, war is said to be instrumental, as it

must serve as a means to an end. While war is violent,

this merely acts as a vehicle to reach the goals set by

the aggressor—which is typically to force the other

party to concede to the aggressor’s will or to act in a

manner which it would otherwise avoid. Lastly, all

conflict deemed to be war is political in nature. “War

is a mere continuation of politics by other means”, as

Clausewitz is most often quoted as saying clearly

illustrates this point. While the violence displayed in

war is clear, it is never a single act. Rather, these

actions all serve to further political interests of the

aggressor that would otherwise be unattainable

through other means. While events in cyberspace have

exhibited these traits, none of these have, at present,

been shown to demonstrate all three simultaneously.

Consequently, the claim that cyber warfare is simply

warfare moved to cyber space can not easily be made

as past events that, branded as cyber warfare, have

only displayed two of the three characteristics

highlighted above [2, 3].

If the assumption that cyber warfare is simply

warfare transplanted to cyberspace is maintained,

additional distinction may be made with regards to

whether or not such actions are visible to the general

populace. Two classifications are available in this

regard; that cyber warfare may either be subrosa or

non-subrosa with the former indicating that actions

are, for the most part unknown, while the later is

highly visible. While most incidents discussed in

existing literature are documented or at least known to

the general public, a compelling case exist for discrete

forms of cyber warfare. This being the threat of

retaliation against an aggressor and subsequent

escalation should their identity revealed. It should be

noted that despite the advantages gained in actions

conducted within cyberspace, a level of deterrence

exists between involved parties. Take the hypothetical

case of a disruptive attack launched against a nation’s

power grid. Should a rival launch a cyber attack

against it, and knowledge of such an attack is revealed

to the general public; the threat of escalation may

increase with calls for the affected party to retaliate in


798

kind. Inversely, should knowledge of the attack be

kept or at least attributed to a more begin cause, the

respective governments of the rival parties are given

the opportunity to arrange for a de-escalation of the

hostilities without further escalation. In theory, this

gives the aggressor the ability to launch attacks with a

reduced fear of escalation—assuming that the parties

involved react as expected [4].

The ability to control information, particularly

regarding disruptive events that affect a significant

portion of the population, is difficult at best. In line

with this, revealing crucial information after-the-fact

affects a government’s credibility and would move

them towards full disclosure instead of re-attribution

or deception. Disclosure, while risking escalation, is

perhaps the best option as it provides the affected

party with more liberty to respond with or without

retaliation (legal action, diplomacy, etc.) [4]. While

this sounds counter-intuitive or even reckless, a recent

study of past cyber attacks launched between states

has shown these to be innocuous in nature. From this

it can be assumed that such attacks may actually form

the status-quo in terms of cyber relations between

states. Consequently, so long as an attack does not

lead to massive loss of life or significant damage, the

threat of escalation is minimal and the actions are

tolerated to a certain degree [2].

With these points in mind, the study posits that

cyber warfare does not exist primarily in light of the

fact that parties can not or are not willing to commit to

the first characteristic of warfare, violence, as it risks

escalation and mitigates advantages obtained by

operating in cyberspace to begin with. In its place, the

term cyber conflict is suggested as a means to label

conflicts that regularly exists between states within

cyber space. From this, two further subcategories are

established: LICC and HICC (high-impact cyber

conflicts). The former refers to cases of cyber

conflicts that are aimed towards influencing or

shaping public opinion within the target state while

the later refers to attacks that cause damage to specific

cyber infrastructure of the target state. However, due

to the risk of going beyond the established threshold

of cyber relations, the proponent believes that HICCs

are unlikely increase the likelihood of escalation as

these. Consequently, it can be postulated that current

and future incidents are likely to be in the form of

LICCs.

3. China’s Cyber Strategy

As the previous section has established, incidents

that have taken place in cyber space are not true forms

of warfare but rather instances of what the study has

termed as low-impact cyber conflicts. With this in

mind, the feasibility of such needs to be established

beyond the realm of intellectual discourse and must be

analyzed against the backdrop of current events. To

achieve this, the study takes note of the recent

conflicts between the PRC and ASEAN members

such as Vietnam and the Republic of the Philippines

as a means to demonstrate the existence of low-impact

cyber conflicts as part of a continuing conflict. Prior

to discussing this point further, the concept of cyber

strategy as it relates to the PRC must first be

established. From there, the process of low-impact

cyber conflict may then be presented.

Cyber strategy is defined as “the development and

employment of strategic capabilities to operate in

cyber space, integrated and operated with other

operational domains, to achieve or support the

achievement of objectives across the elements of

national power in support of national security

strategy” [5]. From this it can be deduced that cyber

strategy can be utilized to support different

instruments of power aside from military power. Most

notable amongst these is economic power that can be

gained through the theft of proprietary information

from rivals (not necessarily between states) [6]. In the

case of the PRC, strategy is defined as, “the analytical

judgment of such factors as international conditions,

hostilities in bilateral politics, military economics,

science and technology, and geography as they apply


799

to the preparation and direction of the overall

military/war plan” [5]. Complementary to this, China

views cyber strategy as, “the use of information, a

crucial component of cyber processes, to influence or

control the direction of an opponent’s decision

making activities”.

From these definitions, two crucial points in the

PRC’s cyber strategy may be established. The first is

its willingness to utilize multiple instruments of

national power as a means to determine the most

appropriate military option—if one should be used at

all. Second is the end goal of influencing an

adversary’s decision making activities rather than

confronting them in direct action. To quote the

Chinese strategist Li Bingyan, “How do you get a cat

to eat a hot pepper? You can stuff down his throat.

You can put the pepper in cheese and make him

swallow it. Or you can ground the pepper up and

spread it on his back, which makes the cat lick himself

and receive the satisfaction of cleaning up the pepper”.

From these statements, it can be posited that the

PRC’s cyber strategy aims to influence an adversary’s

decision as a means to minimize escalation while

achieving its desired result. Niu, Li and Xu [7]

emphasize this point when they cited several

stratagems that the PRC may utilize to achieve

information supremacy and to seize the initiative in

the event of conflict. From these, four can be linked to

establish the study’s view of the PRC’s cyber strategy.

These are: thought directing, intimidation through

momentum-building, information-based capability

demonstration, and releasing viruses to muddy the

flow [5].

Thought directing functions by manipulating the

cognitive processes of an adversary such that an

incorrect decision, from their part, is reached. This

may be achieved by releasing factious information

that an adversary may mistake as true. Consequently

disguising the true intention of its initiator.

Intimidation through momentum building is the

process of generating psychological pressure directed

towards an adversary through intimidation.

Information-based capability is another form of

intimidation, though this is achieved through the

perceived unintentional demonstration of one’s

capabilities. These two differ in terms of the perceived

attempt on the part of the initiator. For the former, the

initiator would intentionally perform actions or release

information to cause discord and distress on the part

of the adversary. The latter is performed under the

guise of a routine action that is not intentionally meant

to intimidate but would do so nonetheless. Lastly,

releasing viruses as a means to muddy the flow

pertains to the corruption of information, thus denying

an adversary access to resources to assist in their

decision-making processes. This process would

perhaps be the most straightforward and intrusive of

those previously discussed as it attempts to

intentionally cause the corruption of an adversary’s

information resources [2, 5].

Taken collectively, the evidence that points to the

PRC’s cyber strategy supports the concept of

low-impact cyber conflicts. This is achieved by its

calls for an “influenced-based” approach rather than

that of a direct violent conflict. While the stratagems

provided by Niu, Li and Xu [7] also entertain the

possibility of more violent and damaging actions

recent actions on the part of the PRC and its foreign

policy demonstrates otherwise.

4. China Threat Theory and Cyber Conflict

Recent articles discussing the impending cyber

threat that the global community faces have often

cited the PRC’s liberal use of such tools to further its

political aims. While there is no disputing that

evidence collected in the aftermath of recent incidents

has highlighted the possibility of the PRC’s

involvement, the question is whether or not claims of

the PRC’s cyber capabilities and their intention to use

these pose an actual threat. This requires objective Attribution of cyber attacks does not provide certainty due to the nature of cyberspace.


800

analysis of their foreign policy through the lens of

their cyber strategy that was discussed in the previous

section.

While it would not be possible to divine the PRC’s

true intentions, the perceived threat posed can be

traced to the China Threat Theory. Through this,

claims of the PRC’s hegemonic ambitions are used as

a basis to justify the impending threat posed by an

emerging China in the form of economic, military,

and ideological/political aggression and expansion.

This theory, while appealing in light of its current

events, may be drawn into question on the grounds

that it stands on two fundamental and simplistic

assumptions: (1) the PRC has ambitions to dominate

the region or the world and (2) it has the capacity to

develop capabilities that will allow it to challenge the

United State’s hegemony [7, 8]. While a substantial

amount of literature has been formulated to discuss

the validity of this perspective, little to no consensus

has been made in this regard. So as to be balanced in

terms of providing an explanation on the situation,

defining the PRC as either a status quo or revisionist

power is highly dependent on the environment in

which it exists.

Taking into consideration Chinese history, the shift

between status quo and revisionist has been driven by

external political, economic, and ideological factors.

Focusing on its contemporary history, during its

“opening up” period between 1970’s and early 2000’s,

China has found it necessary to shift its focus from

military, political, and territorial security issues into

that of a more “cooperative” and comprehensive

approach that emphasized the need to maintain

stability and participation in the global political

economy [7]. This need to adhere to the status quo is,

however, dependent on the relatively benign global

conditions that were present during the PRC’s rise to

power. Presently, with the global economic crisis in

both the United States and the European Union and

the PRC’s dependence on the stability of this system

forces it to reassess and act in order to maintain and/or

expand its position both regionally and globally.

Consequently, it can be said that the PRC is

revisionist in this sense, but not to the extent and

capabilities ascribed to it by proponents of the China

Threat Theory. Its peaceful rise is possible, but it only

through careful consideration of its actions [7].

Within the context of this cautious approach, the

existence of LICCs lends themselves well to further

the PRC’s attempts to consolidate its

power—specifically within South East Asia—through

the use of the traditional instruments of power:

political, informational, military, and economic [9].

This is due in part to the pervasiveness of ICT

(information and communication technologies) that

not only comes to support these but have also been a

cornerstone of development within the region [10].

Corollary to this, the nature of low-impact cyber

conflicts minimizes the possibility of escalation if

used as an instrument of power or as a supportive tool.

To further the point, Buzan has mentioned that

China’s rise within the region is dependent on its

ability to assume a benign posture towards its

neighbors. The ambiguity provided by LICCs and the

minimal damaged caused by these would dissuade or

limit retaliatory action against the PRC for fear of

escalation brought about by rash or incorrect action

[6]. Consequently, this places it in an advantageous

position to push for its interests with little fear of

damaging the benign image it continually attempts to

present to the international community.

5. ASEAN Vulnerability to Low-Impact Cyber Conflicts

While instances of the PRC’s cyber capabilities

have been suggested to occur at a global scale,

analysis of its use within the ASEAN regional bloc is

of particular interest as it has coincided with the

on-going territorial dispute in the South China Sea.

Specifically, the cyber attacks launched against the

Philippines are of particular interest.

For the Philippines, cyber attacks were initiated by


801

suspected PRC nationals in April 20, 2012 with the

defacement of the University of the Philippines

website. A day later, on April 21, 2012, Filipino

hackers associate with Anonymous Philippines

retaliated with the defacement of PRC websites. In

response to rising tensions, the Philippine government

calls for a cessation of aggression between the

different parties involved. Despite these public calls

on multiple media, neither side yielded to the call;

suggesting that, at least from the Philippines, the

attacks were not state sponsored. Despite such calls,

cyber attacks between suspected PRC hackers and the

Philippines continued until May 11, 2012. The attacks

launched against the Philippines included DDoS

(distributed denial-of-service) attacks against public

and private organizations as well as the disclosure of

sensitive information [11].

It is important to point out that none of the activities

that have transpired during this period could be

dubbed as cyber warfare as was established in earlier

sections. That is to say, potential targets such as

critical infrastructure which cyber warfare advocates

often cite as crucial targets were not at all affected.

The act of defacing websites (in particular government

websites), denying access to them, and releasing

sensitive information to the public are, for the most

part, a demonstration of capacity and intimidation

rather than outright destruction—this was identified in

previous sections as a component of the PRC’s cyber

strategy. While it would be easy to dismiss this as a

result of a lack of “suitable targets”, it should be

pointed out that ICT serves as a crucial component of

the Philippine’s economic development [12].

A recent report has indicated that the Philippines’

BPO (business process outsourcing) industry—one

that is heavily reliant on ICT—grows annually at a

rate of 20% with its projected value to be 25B PHP by

2016 [13]. This figure does not, as of yet, include the

dependency of MSMEs on ICT to support their

operations—these include retail, financial services,

and telecommunications, etc.. Corollary to this, a

recent survey by the Economist Corporate Network

has shown that 35% of its respondents have indicated

that the Philippines is an attractive service provider

for IT-BPO services [14]. If direct damage to the

Philippines was the desired goal, there would have

been no shortage of targets. With the Philippines still

identified as an emerging economy, it remains to be

highly dependent on the presence and revenue

generated from these investments. Attacks against

these would impact economic activities within the

country and would, in the long run, reduce confidence

in the Philippine’s ability to protect foreign

investments.

While it would be convenient to categorize this

exchange as an isolated case for the Philippines,

similar cyber attacks have been observed against

Vietnam that followed the same form. Interestingly

enough, the triggering factor that initiated cyber

attacks in that case were also territorial disputes

concerning the South China Sea [15]. Besides these, a

study conducted by Dell Secure Works has shown an

increase in the number of cyber attacks that may be

categorized as LICCs within the ASEAN bloc (Vietna,

Brunei and Myanmar). The question though is what

makes ASEAN particularly vulnerable to this form of

cyber attack?

By analyzing the different historical, political and

social conditions within the ASEAN bloc, its

particular vulnerability may be explained through the

Socio-Political Cohesion framework provided by

Buzan as a means to analyze how nation states view

threats to their security relative to their current

socio-economic and military power (refer to Table 1)

[16]. While a through analysis of Buzan’s model

would prove interesting, this is beyond the goals of

this study. To briefly explain the proposed framework,

nation states are categorized based on their

socio-political cohesion and power, which in this case

refers to power. From this, four categories may be

established: Weak P/Weak SP, Weak P/Strong SP,

Strong P/Weak SP, and Strong P/Strong SP. Nation


802

Table 1 Vulnerabilities and types of states Ref. [16].

Socio-Political (SP) Cohesion

Weak Strong

Power (P) Weak Highly vulnerable to most types of threats Particularly vulnerable to military threats

Strong Particularly vulnerable to political threats Relatively invulnerable to most types of threats

Table 2 Mapping to cyber threats Ref. [17].

Socio-Political (SP) Cohesion

Weak Strong

Power (P) Weak

De-stabilizing political actions in cyberspace, attacks on Internet infrastructure, criminal activities

DDoS and other major attacks on critical infrastructure

Strong De-stabilizing political actions in cyberspace Criminal activities in cyberspace

states for each category would view threats in a

significantly different manner from each other. For

example, a state that is viewed as Weak P/String

SP—which is characteristic of the tigers of ASEAN

such as Malaysia—lacks significant military power

but boasts a formidable economy. These would view

direct military action as threats to its security. States

that are identified as being Strong P/Weak SP, such as

the authoritarian regime of Myanmar, would be

viewed as a threat with internal destabilizing factors.

According to an earlier study conducted by Buzan,

most of ASEAN (with the exception of Singapore)

may be labeled as weak. While significant political

and economic developments have been made since,

none of the ASEAN members have reached the 4th

quadrant. This situation leads to a disjointed view of

threats within the region [17].

Table 2 illustrates how the framework may be

mapped to threats presented in cyberspace relative to

the type of state based on the factors explained earlier.

With the 1st and 2nd quadrants being the best

representations for the Philippines and Vietnam,

the attacks that were assumed to have been

initiated by the PRC match those predicted by the

framework.

A second and related factor that contributes to

ASEAN’s vulnerability is the ASEAN Way. This

originates from the region’s collective experience as

colonies of the major powers during the late 19th

century to the mid 20th century. This regional set of

values espouse, among others, non-interference in

internal affairs and consensus-based decision making.

While ideal, these two traits of the regional bloc have

historically proved themselves to be limiting when

addressing security threats within the region [18, 19].

From the perspective of cyber attacks and analyzing

these through the lens of the framework provided

earlier, the ASEAN Way is a major hindrance in

developing a feasible mechanism to collectively

address these threats-regardless of its origin. First, the

need for a consensus-ased decision prior to passing

any resolution is flawed. Given that the region is

composed of members with varying levels of

economic and military power, the heterogeneous

nature of the region would not permit any form of

consensus on security matters. This has been shown at

different points of ASEAN’s history [20].

Second, actions on the part of members that suggest

interference with internal affairs may limit the

efficacy of attempts to determine the true nature of a

cyber attack. The appeal of using cyberspace as

political tool or for criminal activities is the inherent

anonymity that it provides. Part of this anonymity is

achieved through the use of proxies2 to redirect the

source of the attack; consequently disguising the

actual source from the intended target. To mitigate

this, those wishing to identify the origin of the attack

would typically require the cooperation of the proxies

2This is not used in its technical sense rather simply as a reference to mediators of a particular act.


803

involved. For this study, the authors can view these as

simply the states through which the cyber attack

passed through prior to reaching its final target. This

situation is unavoidable due to the interconnected

nature of cyberspace and its most popular

manifestation, the Internet. Unwillingness on the part

of the proxies to cooperate would limit the amount of

information gathered regarding the attack and would

in turn severely limit the decisions to be made due to

lack of information.

A hypothetical scenario that may be used to

illustrate this challenge would be a case in which a

cyber attack from the PRC passes through Cambodia

prior to reaching Vietnam. While the proponent is not

aware of any such case, the current political climate in

which Cambodia appears to be favoring Beijing’s

policies [21] can introduce difficulties in obtaining

their cooperation if such an event were to take place.

A similar issue that has occurred—though outside the

region in this study—attacks against Estonia in 2007.

A postmortem of the events identified the initial

source of attack as being a computer located within

Russia. But without further cooperation from Russia,

and the presumed source of the attacks, the allegations

remained as such—allegations of possible Russian

involvement with the incident [22].

Taken together, the unbalanced power distribution

within South East Asia and the inherent limitations of

the ASEAN regional bloc in the form of

non-interference and consensus-based decision

making restricts the possibility of a unified defense

against cyber attacks aimed at individual member

states or the region as a whole. From the perspective

of the PRC’s assumed cyber strategy, it allows actions

in cyber space—particularly low-impact cyber

attacks—to go unanswered. Considering the changing

political alliances within the region, it is also possible

that such attacks may be tolerated to maintain stability

within the region while at the same time allowing the

PRC to extend its influence through the cyber domain

unchallenged.

6. Conclusions

While this study presents a cross-section of the

current developments within cyberspace concerning

both ASEAN and the PRC, it does not attempt to

predict the future cyber conflict within the region with

completely certain. The conflicts based on the form

and nature of cyber attributed to the PRC that has

taken place within the region, the likelihood of a

“cyber Perl Harbor”, at least within ASEAN, remains

highly unlikely. Even though the analysis presented

has shown the PRC to be a revisionist power at this

point in time, its dependence on international

structures and their stability minimizes the likelihood

of any form of conflict with grave destabilizing

capabilities. On the other hand, activities that limit

damage and demonstrate or imply the capabilities of

the PRC have been observed, as in the case of the

Philippines and Vietnam, to be commonplace. These

forms of low-impact cyber conflicts, while not at the

scale and ferocity that most have claimed, are capable

to extending the PRC’s influence towards issues that

occur outside cyberspace while at the same time

limiting the threat of escalation and retaliation.

For ASEAN, the varying level of socio-economic

and military power between members continues to

challenge attempts at establishing a unified view of

security threats to the region. Rather, threats to

regional security and stability are addressed at a

national level or, at best, through bilateral agreements

among member states with similar socio-economic

and military characteristics. This situation is further

aggravated by the fact that the ASEAN Way may

limit the amount and quality of information gathered

from each other on the basis that this might be

perceived as interference.

It is widely known that ASEAN is developing into

a global economic focal point that is highly dependent

on ICT for its operations. This is best illustrated with

the establishment of the ASEAN ICT Master plan that

lays the groundwork that aims to further enhance the

use and adoption of ICT within the region [10]. At the


804

same time, the PRC has, in the last three decades,

established itself as a formidable entity that is willing

to project and expand its power and influence through

non-traditional mediums such as cyberspace. The

vulnerability of ASEAN as a whole marks itself as the

perfect platform on which the PRC can demonstrate

and practice with efficiency its current cyber strategy.

To mitigate this threat, while taking into consideration

the constraining factors, the best solution would be for

individual states, at present to strengthen or develop

their respective cyber defense policies and practices as

a means to limit the impact of attacks that are

currently being attributed to the PRC.

While this is strictly a stopgap solution, it would

minimize the impact of certain forms of LICCs.

Programs such as strengthening existing cyber crime

legislation, information security awareness, and

private-public partnerships to assist in securing

government assets are all viable ways to lessen the

impact of these forms of conflict. These, however,

would only serve to rebuff attacks rather than provide

a constant deterrence that may only be achieved if a

more open and cooperative environment were to exist

within the region when it comes to addressing cyber

conflicts. As a first step, the respective members of

the ASEAN bloc, particularly those with

underdeveloped or nonexistent cyber legislation,

should take proactive steps to develop the necessary

laws to curb these threats at the home front. Initially,

this would require close collaboration between the

public sector and the private sector which posses a

significant amount of expertise in addressing cyber

threats. For its part, the public sector has the

manpower and resources required to apply this

expertise at a national level [17, 23]. As an example of

this form of cooperation, consultation may be

conducted between national governments, respective

national CERT (computer emergency response teams),

and the telecommunications industry. Once this has

been achieved, steps can then be taken to improve

information exchange and assistance programs

between member states when it comes to cyber attacks

[23-24]. The need for this has already been established

in the ASEAN ICT Masterplan. Unless these

measures, or some form of them are taken, it is the

proponent’s belief that the form of aggression which

is assumed to originate from the PRC will continue so

long as the factors that render it an attractive option

continues to persist within South East Asia.

References

[1] E. Bumiller, T. Shanker, Panetta warns of dire threat of cyberattack on U.S. [Online], 2012, http://www.nytimes.com/2012/10/12/world/panetta-warns-of-dire-threat-of-cyberattack.html?pagewanted=all &_r=0.

[2] B. Valeriano, R. Maness, Cyberwar and rivalry: The dynamic of cyber conflict between antagonists [Online], 2001-2011, http://tigger.uic.edu/~bvaler/Cyberwar%20and%20Rivalry%20Dyanamics%20of%20Cyber%20Conflict%20JPR.pdf.

[3] T. Rid, Cyber war will not take place, Journal of Strategic Studies 35 (2012) 5-32.

[4] M.C. Libicki, Sub rosa cyber war, in: Proceedings 2009 CCDCOE Conference on Cyber Warfare, Estonia, 2009.

[5] T.L. Thomas, Nation-State Cyber Strategies: Examples from China and Russia, Cyberpower and National Security, Potomac Books Inc., Washington, D.C., 2009.

[6] M. Hjortdal, China’s use of cyber warfare: Espionage meets strategic deterrence, Journal of Strategic Security 4 (2011) 1-24.

[7] O. Kwok, China’s foreign policy: Harmonious world, Is it a Mere Strategem, or an Abiding Policy the World can Trust, The Defence Academy Yearbook 2011: A Selection of Commended Essays, Defence Academy of the United Kingdom (2011) 116-138.

[8] K.E.A. Rodhan, A critique of the China threat theory: A systematic analysis, Asian Perspective 31 (2007) 41-66.

[9] D. Tuthill, Reimagining waltz in a digital world: Neorealism in the analysis of cyber security threats and policy, Dissertation, University of Kent, Brussels, 2012.

[10] ASEAN: ASEAN ICT Masterplan 2015, Association of South East Asian Nations (ASEAN), Proposal, ASEAN 2011.

[11] P. Passeri, Philippines and China, on the Edge of a New Cyber Conflict? [Online], 2011, http://hackmageddon.com/2012/05/01/philippines-and-china-on-the-edge-of-a-new-cyber-conflict/.

[12] CICT: The Philippine Digital Strategy Transformation 2.0, Digitally Empowered Nation, Proposal, Republic of


805

the Philippines, Commission on Information and Communication Technology, 2011.

[13] T. Noda, IT-BPO sector eyes $25B target in 2016 [Online], 2012, http://www.philstar.com/breaking-news/795644/it-bpo-sector-eyes-25b-target-2016.

[14] M. Hamlin, Economist: Philippines is Number One [Online], 2012, http://www.mb.com.ph/articles/360743/economist-philippines-is-number-one#.UMROV5NetH8.

[15] Vietnam and China hackers escalate Spratly Islands row [Online], 2011, http://www.bbc.co.uk/news/world-asia-pacific-13707921.

[16] B. Buzan, People, States and Fear: An Agenda for International Security Studies in the Post-Cold War Era, European Consortium for Political Research Press, 2008.

[17] D. Kuehl, From Cyberspace to Cyberpower: Defining the Problem, Cyberpower and National Security, Potomac Books, Inc., Washington D.C., 2009.

[18] R. Sokolsky, A. Rabasa, R. Neu, The Role of Southeast

Asia in U.S. Strategy Towards China, Technical Report, RAND Corporation, 2001.

[19] A. Floristella, Building Security in the South East Asian Region: The Role of ASEAN, Technical Report, The European Consortium for Political Research, 2010.

[20] K. Jong, ASEAN way and its implications and challenges for regional integration in Southeast Asia, Journal of Southeast Asian Studies 12 (2010).

[21] S. Strangio, Cambodia as divide and rule pawn [Online], 2012, http://www.atimes.com/atimes/Southeast_Asia/NG18Ae03.html.

[22] N. Anderson, Massive DDoS attacks target Estonia; Russia accused [Online], 2007, http://arstechnica.com/security/2007/05/massive-ddos-attacks-target-estonia-russia-accused/.

[23] E. Tikk, Ten rules for cyber security, Survival 53(2011)119-132.

[24] Cyber Storm II Final Report, U.S. Department of Homeland Security, Technical Report, 2009.


Inter-MAC Green Path Selection for Heterogeneous

Networks

Olivier Bouchet1, Abdesselem Kortebi2 and Mathieu Boucher3

1. WIDE, WASA, RESA, Orange Labs, France Telecom, Rennes 35510, France

2. MAG, ANA, RESA, Orange Labs, France Telecom, Lannion 22307, France

3. Neo-Soft, Lannion 22307, France


Abstract: Gigabit home networks represent a key technology to make the Future Internet success a reality. Consumers will require networks to be simple to install, without the need of any new wire and with green consideration. To make this network ubiquitous, seamless and robust, a technology independent MAC layer will ensure its global control and provide connectivities to any number of devices the user wishes to connect to it in any room. In order to make this vision comes true, substantial progress is required in the fields of protocol design and in systems architectures. For this goal, we introduce a new convergence layer denoted inter-MAC, which provides a common infrastructure to the home network. Among the inter-MAC functionalities, we focused the interest on the path selection mechanism including the green aspect. To improve quality of service, reduce energy consumption and radiofrequency devices number, we define a new path selection protocol based on several metrics. We present some preliminary results obtained on an experimental test bed.

Key words: Heterogeneous network, convergence layer, quality of service, green path selection, energy and radiofrequency device minimization.

1. Introduction

Home networks are becoming more attractive for

the researchers as the customers are becoming more

demanding, especially concerning the green aspects.

Nowadays, users are interested in services and content

such as HDTV streaming, gaming and high-quality

video conferencing that require high data rates and

high quality of services with a technology

transparency for the user.

To satisfy users’ demands, there was a

technology-independent layer called inter-MAC, from

Omega project [1, 2] which is a convergence layer

located above the different technologies MAC layer.

The inter-MAC realizes the convergence of the

different wired and wireless technologies and permits

Corresponding author: Olivier Bouchet, M.Sc. & MBA, research fields: hybrid network, optical wireless communication. E-mail: [email protected].

a high throughput within the home with high quality

of services requirements, while maintaining a single

IP sub network.

One of the main functionalities of the inter-MAC

layer is the path selection mechanism. This protocol is

capable of forwarding frames from a source A to a

destination B according to the flow identifier, and we

have added some parameters and modified the patch

selection algorithm to reduce energy consumption and

radiofrequency devices number. To do so, we define

several metrics per link which optimizes the patch

selection requested and chose the best outgoing

interfaces.

This paper is organized as follows: Section 2

proposes an overview on existing standards in this

area. Section 3 presents the design and the inter-MAC

architecture. Section 4 presents the metrics and the

green path selection mechanism. Finally, before

DAVID PUBLISHING

D

Inter-MAC Green Path selection for Heterogeneous Networks

807

conclusion, Section 5 presents some objectives and

preliminary results about the implementation process

and the green path selection mechanism use cases.

2. The Current Standardization Initiatives

Various network standards and industry forum’s

specifications could be used to build a local

heterogeneous network.

2.1 IEEE 802

IEEE 802 has many groups for interworking and

convergence. Group 802.1 is managing LAN/MAN

architecture and interworking between 802

technologies. Regarding “routing” protocols, they are

either very simple to manage star topology with single

links, either complex with manual configuration and

resources provisioning. Group 802.21 provides tools

to facilitate seamless handover between a terminal and

different access networks (mobile, WiMax, WiFi).

Finally, the 802.11s is defined for wireless mesh

networks with the path selection based on two

protocols: RM-AODV (radio metric ad hoc on

demand distance vector) and RAOLSR (radio-aware

optimized link state routing). The HWMP (hybrid

wireless mesh protocol) is a hybrid approach. These

protocols use MAC addresses (layer 2 routing) and

also used a radio-aware routing metric for the

calculation of paths. But PLC and MoCA are out of

the scope.

2.2 IEEE P1905

The IEEE P1905.1 [3] working group was created

in November 2010, thanks to both the results obtained

during the OMEGA project [2] and the interest of

manufacturers in “layer 2.5 convergence” or

inter-MAC concepts. The aim of P1905.1 is to

facilitate the usage of new services everywhere in the

home, by defining a unified framework for

multi-interface or multi-technologies devices.

A first version of the IEEE P1905.1 draft standard

is completed in spring 2012. Further enhancements of

this standard could be proposed (P1905.2).

2.3 HGI (Home Gateway Initiative)

The HGI [4] was founded by major BSPs

(broadband service providers) in 2004. It publishes

requirements for digital home building blocks. This

includes home gateways and home network devices.

HGI projects are triggered by the service-oriented

vision of the BSP members and built on the technical

collaboration of all the HGI participants.

At the end of 2011, the main on-going activities in

the HGI are related to the definition of key

performance indicators for evaluating the performance

of home networking technologies. This includes the

definition of requirements for energy saving in home

network infrastructure devices.

2.4 IETF/TRILL

TRILL is a layer 2 solution designed by IETF to

provide redundant/multi-path topology in data centers.

The main goal is to replace the spanning tree

architecture and to bring “link state” routing

intelligence at layer 2. The base protocol specification

was published in July 2011 as RFC6325 and

commercial products are available in 2011 by Cisco

(branded as FabricPath), Broadcom, Blade Networks

(IBM), etc.. A software solution is available in Solaris

Express where the implementation is done in kernel

space. Today, there is no Linux implementation

available.

Technologies supported by TRILL are PPP

(Point-to-Point Protocol) and Ethernet. If Wi-Fi and

PLC need to be supported, they are hidden by an

Ethernet port bridged to this technology. The path

selection protocol used for the control plane is IS-IS

(intermediate system to intermediate system) Protocol

with extra TLVs (Type-Length-Values) to carry

TRILL information (defined in RFC 6326).

3. Inter-MAC Convergence Layer

The OMEGA project [1] has developed the “2.5


808

convergence” concept or inter-MAC network and it is

a potential image of the future home networks. It is a

heterogeneous network (Fig. 1) able to use existing

and future wireless technologies (802.11 technologies,

UWB, wireless optics...) [5] and “without new wire”

[6] technologies such as PLC (power line

communications).

This network is divided into two parts: a backbone

part composed of the extenders (nodes used to extend

the area of connection supposed fixed) and a terminal

part composed by end devices (possibly mobile). Each

inter-MAC node is supposed to have multiple

interfaces.

To manage the different interfaces and the different

technologies used, it was introduced a new

intermediate sub-layer (Fig. 2), named the inter-MAC

convergence layer. This inter-MAC convergence layer

is positioned above the different MAC layer of the

different technologies. The inter-MAC is designed to

manage multiple technologies within the home with a

guarantee of quality of services.

The inter-MAC layer is composed of three

Fig. 1 Inter-MAC network (Omega project).

Fig. 2 Inter-MAC architecture.


809

functional planes: the control plane, the data plane and

the management plane. The control plane is

responsible of short-term decisions related to the

establishment and the release of paths and the

processing of upper layer requests. The control plane

is divided into several engines: the QoS Engine, Path

Selection Engine, the Monitoring Engine, the Link

Setup/Teardown Engine and the inter-MAC Adapter.

The data plane is responsible of frame forwarding.

The inter-MAC layer is designed to be capable of

managing the different MAC layers (located under the

inter-MAC) in the best way to satisfy the upper layers

quality of service demands (e.g., UPnP QoS TSPEC).

The inter-MAC network is also designed to use layer

two forwarding instead of layer three routing. This

choice is basically made to provide seamless

transparency for applications while maintaining a

single IP sub network.

4. The Green Path Selection

After the end of the Omega project, a new project

was created to include the green aspect due to the

focus group results from customer tests. This

development is a part of EconHome project [7]. The

inter-MAC layer has been implemented during the

Omega project as a software module using C/C++

languages under Linux OS. The data plane was

implemented in kernel space for performance reasons,

while the control plane was implemented in user

space.

We extended the existing implementation by adding

green parameters to the path selection protocol as

described in the upcoming sections.

4.1 Green and QoS Metrics

The green path selection protocol takes into account

the following three metrics:

(1) Available bandwidth: For QoS purposes,

network bandwidth is a crucial criterion. In the current

implementation, available bandwidth is computed

form passive measurements relying on interfaces

statistics with respect to received/sent bytes,

periodically. Other methods can obviously be also

applied. Note that bandwidth measurement in hybrid

links (WiFi, PLC) with time variant capacities is not

trivial, as indicated in Ref. [8];

(2) Energy consumption: One of the green

parameters is related to energy consumption on each

network interface due to traffic transmissions; it is an

additive metric. It is worth mentioning that there is no

additional consumption on a given interface when

there is already traffic on it. In other words, when a

link is used by an ongoing flow, actually there is little

extra consumption due to a second flow in that same

link, as suggested by Fig. 3 [7]. So, we need to take into account the state of the interfaces when

Fig. 3 Energy consumption vs. throughput.


810

choosing a path to minimize energy consumption. In

order to minimize the global energy consumption of

the home network, we propose to compute the

difference between energy consumption in “idle” (the

interface is ON, but there is no significant data traffic

on it) and “ON” states. For interface i, we have:

If interface i is idle

ΔEi = Ei(ON) – Ei(idle)

Else

ΔEi = 0

Then, for a path k which is a succession of n

interfaces, we have:

Epath = ∑i = 1, n ΔEi

We assume that MAC layers take into account point

to multi point links, where energy consumption should

include a multiplicative factor depending on the

number of concerned nodes. Indeed, for shared media

such as WiFi and PLC, there are multiple receivers,

even when there is a single MAC destination.

Currently, as it is impossible to obtain dynamic values

in real time, we use static values for energy

consumption according to the interface type (with

watt-hours unit), as indicated in Table 1. These values

were obtained through a series of tests and

measurements conducted in EconHome project [7]. In

the future, it could be more interesting to rely on real

time indications provided by the physical and MAC

layers in order to have more accurate values;

(3) Radio frequency: the second green parameter is

related to electromagnetic pollution in the home. The

idea is to minimize radio frequency emissions due to

wireless and power line interfaces use in the network.

It is not an additive metric since radio values depend

on the considered frequency band (e.g. WiFi at 2.4

Ghz, WiFi at 5 Ghz, etc.). Thus, we propose to

associate a radio frequency cost to each interface,

which is a function of the transmission power,

frequency band and the potential impact on human

body. Then, the cost of the path is the sum of crossed

interfaces radio frequency costs. The definition of the

cost function is beyond the scope of this paper. Note

that, currently, even in the idle state, wireless

interfaces do emission of radio frequency. However,

we assume that future products will be able to enter

sleep mode (without radio frequency emissions) when

needed.

For sake of simplicity, in this paper, we propose to

use the binary values as indicated in Table 1 in our

testbed. Note that PLC radio frequency is much lower

than radio wireless interfaces, so we apply 0 instead of

1. When there is a radio wireless interface in “idle”

state in the considered path, we increment the metric

of radio frequency of the considered path which is

equal to the total number of active wireless interfaces

of the path.

Finally, in addition to these metrics, it is possible to

use the flows service class (based on DSCP marking,

for instance) to decide which criteria to apply for path

selection.

4.2 The Green Path Selection Protocol

The path selection protocol is based on a reactive

approach, where a path is calculated when a new flow

arrives. In fact, when there is no entry in the

forwarding table for an incoming packet, the path

calculation is triggered. Moreover, a new path is

established when there is a link failure or when there

is a better path to the destination. The process is based

on the following 7 control messages:

(1) Route request: it permits to compute a new (or

Table 1 Green metrics and statics values.

Interface type/Metric Energy consumption (W.h), ON state Energy consumption (W.h), Idle state Radio frequency cost

WiFi at 2.4 Ghz 3 1.7 1

PLC 3.2 2.4 0

GigaEthernet 0.6 0.5 0

FastEthernet 0.3 0.2 0


811

better) path. It is sent by the source on all its interfaces

towards the destination. When the packets cross the

intermediate nodes, the metrics information (available

bandwidth, energy consumption and radio frequency)

are updated. These values are added into the route

request message;

(2) Route reply: it is sent by the destination on all

its interfaces in response to a route request message.

The route reply message contains the end to end path

from the source to the destination (including

intermediate nodes). The source stores the obtained

response, in order to choose an appropriate path after

receiving the different answers;

(3) Route activate: it is sent by the source to the

destination, after selecting the appropriate path. The

forwarding tables of the intermediate nodes are then

updated accordingly;

(4) Route activate confirm: it is sent by the

destination when it receives the route activate message

to indicate that the path is computed successfully.

Then, the flow can be transmitted on the

corresponding interface in the source node;

(5) Route delete: it is flooded in the network after a

link failure event. The corresponding entries in the

forwarding tables are erased. When the source of a

flow, which is impacted by the failure, receives this

message, it starts the computation of a new path;

(6) Route check: it is sent periodically by the source

to the destination in order to update the metrics values

(available bandwidth, energy consumption and radio

frequency). At the same time, to be able to detect all

possible paths, a route request message is sent (with a

specific flag). Since the metrics values are updated, a

better path than the initial one might be selected for a

given flow;

(7) Route check reply: it is sent by the destination

in response to a route check message. It contains the

information about the end to end path with the

updated metrics values.

The path selection algorithm applies iteratively the

following criteria to choose the best path for a given

flow. The intent is to choose a path that meets QoS

requirements and ensures minimum energy and radio

frequency consumption:

Among all possible paths between a source and a

destination, we select those that provide an available

bandwidth higher that the required one (we assume

that for some flows such as IPTV, the required rate is

known; otherwise we use a threshold value)

Among the remaining paths, we select the two that

offer lower energy consumption (we assume that there

are at least two possible paths, otherwise, the choice is

limited to the only remaining path).

Among the last 2 paths, we select the one that has

the minimum number of radio interfaces.

It is rather a simple algorithm which can be

implemented easily. Obviously, it is possible to apply

a different combination of these metrics depending on

which criteria we choose to prioritize. In fact, we want

to avoid complex NP hard optimization based

algorithms that could be more “intelligent”, but that

are too complex to implement in home network

environment where cost is a major issue, especially

for infrastructure equipments.

5. Objectives and Preliminary Results

The implementation of the green path selection

protocol on an experimental testbed is undergoing. In

this section, we discuss the objective of the testbed

and present some preliminary results.

5.1 Objectives

Fig. 4 represents a heterogeneous home network

configuration with Ethernet, PLC and WiFi links. The

testbed contains 4 mini PCs which play the role of

extenders, they implement the green path selection

protocol. The end devices (including a PC, NAS, TV,

tablet and mobile phone) are connected to these

extenders. The Home Gateway connects the home

network to the Internet. The idea is to highlight energy

consumption reduction while maintaining satisfactory

QoS for different flows, thanks to the green path


812

Fig. 4 Testbed example.

Fig. 5 Graphical interface.

selection protocol described in section IV.

5.2 Preliminary Results

In this section, we present through an example

some typical results that we can obtain on the test bed.

Fig. 5 shows a graphical interface obtained on the

manager node in a test bed with 3 mini PCs referred

as Omega1, Omega2 and Omega3 nodes. It indicates

the devices where the path selection protocol is

running and their interconnection. We have a Fast

Ethernet link between Omega1 and Omega2, a WiFi

link between Omega1 and Omega3 and a PLC link

between Omega2 and Omega3. Currently, the

graphical interface does not distinguish the types of

the links.

We implemented a wirehark module to decode

inter-MAC frames and analyze them. Fig. 6 shows a

wireshark capture obtained on interface eth0 of

Omega1 when there is a flow from Omega1 to

Omega3. It indicates the path selection protocol

messages described in Section 4.2

Fig. 7 presents a wireshark capture, obtained on

Omega1 eth0 interface, of the route reply message

sent from Omega3 to Omega1. As people can see,


813

Fig. 6 Wireshark capture of control messages.

Fig. 7 Wireshark capture of route reply messages.

the message contains the energy consumption metric

that we implemented (highlighted with a circle), in

addition to available bandwidth on transmitting and

receiving interfaces. The values correspond to the 2

hops path (O1-O2-O3) including Ethernet and PLC

links respectively. Note that available bandwidth on

PLC link is computed based on the maximum 200

Mbps physical capacity, as an example, which is


814

higher than the actual IP throughput. It is possible to

modify this value as explained in Section 4.1 using an

appropriate measurement method. Assuming that both

paths offer sufficient bandwidth and that initially all

interfaces are in idle state, the selected path is the 2

hops one (O1-O2-O3) because is ensures lower

energy consumption (0.9 w.h vs. 1.3 w.h, as delta

energy values ((0.3 – 0.2) + (3.2 – 2.4)) vs. (3 – 1.7)).

This final part of the implementation is close to the

end. Indeed, if we compare the total energy

consumption in the network using the possible paths,

we have:

Direct Path (O1-O3): E = 11.2 w.h (= 2 × 3 (WiFi

ON) + 2 × 0.2 (Eth idle) + 2 × 2.4 (PLC idle));

Two hops path (O1-O2-O3): E = 10.4 w.h (= 2 ×

1.7 (WiFi idle) + 2 × 0.3 (Eth ON) + 2 × 3.2 (PLC

ON)).

6. Conclusions

The inter-MAC network was a good initiative to

reach gigabit throughput and high quality of service

with multi-technologies transparency (Omega project).

The green path selection protocol that we defined (part

of EconHome project) aims to ensure a high quality of

service with energy and radiofrequency minimization.

The preliminary results foreshadow positively the

demonstrator for heterogeneous home network

configuration with Ethernet, PLC and WiFi links.

Acknowledgments

The research leading to these results has received

funding from the French competitivity cluster FUI

project also referred as ECONHOME.

References

[1] ICT OMEGA Project Website, http://www.ict-omega.eu.

[2] V. Suraci, Convergence in Home Gigabit Networks:

Implementation of the inter-MAC layer as a pluggable

kernel module, PIMRC 2010.

[3] IEEE P1905.1, http://grouper.ieee.org/groups/1905/1/.

[4] HGI Website, http://www.homegatewayinitiative.org/.

[5] IEEE P802.11n/D9.0, Draft standard, Wireless LAN

Medium Access Control (MAC) and Physical Layer

(PHY) Specifications, Amendment 5: Enhancements for

Higher Throughput, Mar. 2009.

[6] HomePlug AV Specification, Version 1.1, May 2007

[7] J.P. Javaudin, Orange vision for green home networks, in:

IEEE WCNC 2012, Workshop on Future Green

Communications, Paris, 2012.

[8] O.O. Irigoyen, A. Kortebi, L. Toutain, D. Ros, Available

bandwidth probing in hybrid home networks, in: The 7th

IEEE International Workshop on Heterogeneous,

Multi-Hop, Wireless and Mobile Networks, 2012


Web Block Extraction System Based on Client-Side

Imaging for Clickable Image Map

Hiroyuki Sano, Shun Shiramatsu, Tadachika Ozono and Toramatsu Shintani

Department of Computer Science and Engineering, Graduate School of Engineering, Nagoya Institute of Technology, Aichi

466-8555, Japan

Received: February 07, 2013 / Accepted: March 13, 2013 / Published: June 30, 2013.

Abstract: We propose a new Web information extraction system. The outline of the system and the algorithm to extract information

are explained in this paper. A typical Web page consists of multiple elements with different functionalities, such as main content,

navigation panels, copyright and privacy notices and advertisements. Visitors to Web pages need only a little of the pages. A system to

extract a piece of Web pages is needed. our system enables users to extract Web blocks only by setting clipping areas with their mouse.

Web blocks are clickable image maps. Imaging and detecting hyperlink areas on client-side are used to generate image maps. The

specialty of our system is that Web blocks perfect layouts and hyperlinks on the original Web pages. Users can access and manage their

Web blocks via Evernote, which is a cloud storage system. And HTML snippets for Web blocks enable users to easily reuse Web

contents on their own Web site.

Key words: Web information extraction, web service, HTML5.

1. Introduction

We have implemented a system that enables users

to extract specific contents from Web pages. The

extracted content is called “Web block” in this

research. Of course, there are several different systems

already in place to extract only selected areas from

Web pages, but the specialties of our system is that

Web blocks perfect layouts and hyperlinks on the

original Web pages. Users can access and manage

their Web blocks via Evernote, which is a cloud

storage system. In this paper, we propose the outline

of the system and the algorithm to extract Web blocks

are explained.

A typical Web page consists of multiple elements

with different functionalities, such as main content,

navigation panels, copyright and privacy notices, and

advertisements. Fig. 1 shows a screenshot of a page

Corresponding author: Toramatsu Shintani, professor,

research fields: decision support systems and web intelligence. E-mail: [email protected].

from Reuters.com [1], a news site that brings viewers

the latest news from around the world. This Web page

includes many contents. For example, the content

enclosed by a blue line is a header for the news site,

enclosed by a green line is a news article, enclosed by

an orange line is a navigation panel that contains

many hyperlinks for most popular news in

Reuters.com, and enclosed by a red line is a footer for

the news site. The page is very long, so visitors must

scroll the page down on the screen to seek an intended

content through the page. When printing out the

content, printing the whole page wastes much paper

and ink cartridge for the printer. A system to extract a

piece of Web pages is needed.

The rest of the paper is organized as follows:

Section 2 reviews related works and systems in the

area of Web information extraction; In Section 3, we

propose a Web block extraction system based on

client-side imaging and detecting hyperlink areas for

image maps; experiments and their results are

DAVID PUBLISHING

D

Web Block Extraction System Based on Client-Side Imaging for Clickable Image Map

816

Fig. 1 A page from Reuters.com that contains many contents.

in Section 4; Finally, we conclude the paper in

Section 5.

2. Related Works and Systems

There are several different Web information

extraction approaches already in place.

Most of existing algorithms for Web information

extraction are based on DOM (document object model)

structure of HTML or machine learning [2, 3], for

example, Web wrappers are well known tools for

information extraction from Web pages [4, 5]. They

focus on only main content and extract the main


817

content from a Web page. In Fig. 1, we will extract the

news article. But the main contents for the visitors who

want to know most popular news list is that the part

enclosed by an orange line. So, a system is needed that

enables users to set clipping areas and can extract only

selected areas from Web pages.

FireShot [6], which is a Firefox extension, captures

screenshots of Web pages in Web browsers. Users can

capture an entire page or a selected part, add

annotations, and save the images by using the

extension. Users can save the images not only as image

file formats (JPEG, GIF, PNG and BMP) but also as

PDF by using the extension. But the extension does not

have a function to extract hyperlinks information from

the original Web pages, so PDFs generated by the

extension do not contain hyperlinks.

Evernote Web Clipper [7] also enables to crop Web

pages and can contain hyperlinks in cropped parts. But

users can only specify the cropped area based on DOM

structure of HTML, so they can not specify the area

freely. For example, consider the Web page layout in

Fig. 2. In this example, a user wants to extract the part

“2” in Fig. 2a. But the user can extract the part marked

“1” in Fig. 2a, the part “A” and “C” in addition to the

part “2”, by the Clipper. This is because the part “B”

and “D” are separately placed in DOM showed in

Fig. 2b.

Kwout [8] clears up the problems of FireShot and

Evernote Web Clipper. Kwout enables users to quote a

part of a Web page as an image with hyperlinks. Users

can set clipping areas freely by drag and drop with

mouse. Kwout provides server side imaging processing.

Kwout must query the Web server to collect the

resource of original Web pages via HTTP protocols,

Fig. 2 An example that Evernote Web Clipper can not

crop.

render the web pages to get screenshots of the entire

pages, and finally crop and generate an image map in

the rectangle specified by the user. The cropped parts

layout may be different from the displayed layout on

users Web browser. This is because Web rendering

engines may be different in user’s side and server side.

In addition, server side imaging will be unsuccessful

when users extract from Web application results.

Kwout already returns a screenshot of the Web

application’s initial state.

In this paper, we propose a system based on

client-side imaging and detecting hyperlink areas for

image maps to attack the problems described above.

Client-side processing can perfect layout and

hyperlinks of original web pages and enables users to

extract contents from Web application results.

3. Web Block Extraction System

3.1 Outline of the System

Fig. 3 shows the outline of the system. The system is

based on client-server model. The client system is

written in JavaScript and packaged as an extension for

Google Chrome. The extension has modules for

imaging Web pages, cropping the image, and getting

hyperlink areas in the rectangle specified by users.

A user starts up the extension after loading a Web

page and the extension enable the user to set a clipping

area by drag and drop using mouse. The imaging

module in the extension generates the screenshot of the

entire Web page by using Chrome Extensions API [9],

and the specified area by the user in the Web page is

cropped from the entire image. HTML5 canvas API is

used for cropping. The detail of an imaging process is

described in Section 3.2. The hyperlink analysis

module analyzes the HTML of the Web page in order

to get a list of coordinates of hyperlinks placed in the

cropped area. The detail of generating clickable image

maps is described in Section 3.3.

In Ref. [10], Kondo et al. proposed a system that

converts Web pages for personal computers into Flash

lite programs for mobile phones. The system

A<table>

A B C D

<td> <td> <td> <td>

<tr> <tr>

(a)Web Page Layout

B

C D

1 12

2 2

(b)DOM Tree


818

Fig. 3 The system implemented in this study.

Fig. 4 A Web block as a clickable image map by using the system.

generates a screenshot of a Web page and gets a list of

coordinates of hyperlinks, and finally combines the

screenshot with the list into a Flash lite program. In this

study, we combine them into a clickable image map. A

clickable image map is a static map image that can be

clicked and linked with other HTML sources.

Clickable image maps generated by our system are

called as “Web blocks” in this paper. Fig. 4 shows an

example of a Web block generated by our system. The

cropped image and the list of hyperlinks are sent to the

server, and the combine module makes a clickable

image map by the image and the list of hyperlinks.

The system saves the HTML of the generated Web

blocks in Web server. So all Web blocks have URLs

and other systems can access the blocks via HTTP

protocol. And the server adds the generated Web block

to Evernote, which is a cloud storage system, as a new

note. Users can access and manage their Web blocks

via Evernote. Fig. 5 shows a screenshot of Evernote to

which several Web blocks are added. Fig. 5(1) is the

list of Web blocks user extracted from Web pages. The

selected Web block from the left list is displayed in a

large way in Fig. 5(2). And Fig. 5(3) is the HTML

snippet of the selected Web block. Users can easily

reuse and mashup Web blocks by using their snippets

on users’ own Web site.

Fig. 6 shows an example of Web block mashups. We

generate a new Web page by mashuping three Web

blocks. Fig. 6(1) is the Web block that shows the latest

headlines extracted from Reuters.com. Fig. 6(2) is a

news article block about Boeing 787 trouble in Japan

extracted from BBC News. And Fig. 6(3) is a block for

Web Server

User

stored

（URL is publised)

Evernote

・ Image

・ Hyperlinks

set a clipping area

Client S ide Server S ide

generate a

clickable map

Web

pageWWW

Google Chrome

URL

referenceXMLHttpRequest

add as

a new note


819

top three search results extracted from Google, for

which the search query is “Web intelligence”. To

generate such a Web page showed in Fig. 6, a Web

page creator only copies and pastes the HTML snippets

of the Web blocks into a HTML source code of a new

Web page.

3.2 Client-Side Imaging

The method is captureVisibleTab. The method

returns an image of whole the Web page that is

displayed on a current selected tab. The image must be

cropped in the next step by the rectangle that the user

specified.

The Canvas API in HTML5 is used for cropping.

Canvas API is an API for drawing graphics via

JavaScript in Web browsers. The graphic context of

canvas has drawImage (image, sx, sy, sw, sh, dx, dy, dw,

dh) method to draw an image onto the canvas. The

Fig. 5 The system adds the Web blocks into Evernote and users can access and manage their blocks via Evernote.

Fig. 6 This figure shows an example of Web block mashup. In this example, three Web blocks from different Web pages

are reused.

(1) Web block list

(2) selected block from the left list

(3) HTML snippet for the block

(1) (2)

(3)


820

Fig. 7 An algorithm for link analysis.

image parameter is the image to draw. The sx and sy

parameters determine from where in the source image

to start copying a rectangle of the image onto the

canvas. The sw and sh parameters determine the width

and height of the rectangle. The dx and dy parameters

determine where on the canvas the image is drawn. The

dw and dh parameters determine the width and height

to scale the image to when drawing it.

In the implementation of the system, the image

parameter is set to the image object obtained from

captureVisibleTab method. The sx and sy are set to x

and y coordinates of the upper-left corner of the

rectangle that the user specified. The sw and sh are set

to the width and height of the rectangle that the user

specified. The dx and dy are always set to zero, and the

dw and dh are set to the width and height of the

rectangle that the user specified. These parameters

make the method render only the image within the

clipping area that the user set in the canvas.

The system gets base64 data of the cropped image to

send the image to the server. The toDataURL method is

that canvas returns a dataURL of an image drawn in the

canvas. DataURL scheme is used for embedding some

binary data in text format data such as HTML or

JavaScript. The binary data is encoded by base64

format in dataURL scheme. In the proposed system,

the base64 data obtained from toDataURL method is

posted to the server by using XMLHttpRequest and the

server decodes the data into a PNG format image.

Table 1 The processing time for imaging (t1), getting image

maps (t2) and adding to Evernote (t3).

t1 (msec) t2 (msec) t3 (msec)

255.2 33.7 3,362.8

243.4 12.1 2,533.7

268.1 14.3 2,679.0

253.9 13.7 2,642.3

217.0 11.9 2,580.1

185.7 12.1 2,606.7

235.1 15.5 2,531.3

239.1 10.5 2,644.4

262.8 15.6 2,625.5


821

3.3 Link Analysis in the Rectangle

The advantage of imaging is perfectly keeping up

layout and presentation of the Web pages. But the

disadvantage of imaging is losing the hyperlinks in the

rectangle. The Web blocks provided by the system are

generated as clickable image map to add hyperlinks for

the images.

The JavaScript program was run on the user’s Web

browser to get maps for hyperlinks. Fig. 7 shows the

link analysis algorithm.

First, all <A> nodes are collected and assigned to an

array A in line 1. The coordinates of all <A> nodes are

absolute coordinates, and they are transformed to

relative coordinates in lines 2 to 5. The base point of

the relative coordinates is the left-upper side of the

rectangle that the user specified. Next, in lines 6 to 10,

the <A> nodes that do not meet to the specified

rectangle are deleted from the array A. And in lines 11

to 18, the hyperlinks areas protruded outside the

rectangle are cut off.

4. Experimental Results

We measured the processing time to extract Web

blocks from several Web pages. The processing time

for imaging, link analysis and adding to Evernote were

measured. The Web pages that were used in the

experiment were newest 10 news pages that were

delivered on Yahoo! News in Japan at 0 a.m., on

February 3rd, 2013. We extracted 10 Web blocks

randomly from each Web page by using our proposed

system, and we measured the average processing

time.

The experiment was performed on the iMac (Mid

2010) that has Core i3 3.2GHz CPU and 8GB DDR3

SDRAM (PC3-8500). The operating system running

on the machine was Mac OS X 10.6.8 and the Web

browser was Google Chrome 24.0.1312.52.

Table 1 shows the processing time for imaging,

getting image maps, and adding to Evernote as new

notes. Thus, the experiment shows that the proposed

method can get screenshots and image maps very

quickly, and that the method has potential for practical

use, although the overheads of communication and

Evernote’s WebAPI-call are high.

5. Conclusions

We have implemented a system that enables users to

extract Web blocks from Web pages, and proposed the

outline of the system, and the algorithm to extract Web

blocks are explained in this paper. Imaging and

detecting hyperlink areas on client-side are used to

generate image maps. The specialty of our system is

that Web blocks perfect layouts and hyperlinks on the

original Web pages. Users can access and manage our

Web blocks via Evernote, which is a cloud storage

system. And HTML snippets for Web blocks enable

users to easily reuse Web contents on their own Web

site.

We conducted an experiment to show the

processing speed of the proposed method. The results

showed that the method can extracts Web blocks very

quickly, and that the method has potential for practical

use.

Acknowledgments

This work was supported by KAKENHI (22500128)

and the Hori Sciences & Arts Foundation in Japan.

References

[1] Business & Financial News, Breaking US & International

News, Reuters.com, http://www.reuters.com/.

[2] J. Pastenack, D. Roth, Extracting article text from the Web

with maximum subsequence segmentation, in:

Proceedings of the 18th International Conference on

World Wide Web (WWW’09), New York, 2009, pp.

971-980.

[3] S. Sarawagi, Information extraction, found, Trends

Databases 1 (2008) 261-377.

[4] Y. Yamada, D. Ikeda, H. Sakamoto, H. Arimura, An

information extraction from the Web: Automatic

generation of web wrappers, Journal of Japanese Society

for Artificial Intelligence 19 (2004) 302-310.

[5] Y .Kusumura, Y. Hijikata, S. Nishida, Unsupervised

information extraction by crossover and structure analysis,

The Transactions of the Institute of Electronics,

Information and Communication Engineers D 90 (2007)

2495-2509.


822

[6] Webpage screenshot in firefox: Add-ons for firefox,

https://www.addons.mozilla.org/ja/firefox/addon/fireshot/.

[7] Evernote Web Clipper Lets you save webpage text, Links

and images with a single click, Evernote,

http://www.evernote.com/webclipper/.

[8] Kwout Home Page, http://www.kwout.com/.

[9] Google, Hello There! Google Chrome,

http://www.developer.chrome.com/extensions/docs.html.

[10] K. Kondo, S. Asami, T. Ozono, T. Shintani, On a Web

Contents Browsing Support Environment for Mobile

Devices and its Applications, IPSJ SIG technical report,

2008.


Multiple Chaos Generator by

Neural-Network-Differential-Equation for Intelligent

Fish-Catching

Mamoru Minami, Akira Yanou, Yuya Ito and Takashi Tomono

Graduate School of Natural Science and Technology, Okayama University, Okayama 7008530, Japan

Received: June 18, 2012/ Accepted: July 21, 2012 / Published: June 30, 2013.

Abstract: Continuous catching and releasing experiment of several fishes make the fishes find some escaping strategies such as

staying stationary at corner of the pool. To make fish-catching robot intelligent more than fishes’ adapting and escaping abilities from

chasing net attached at robot’s hand, we thought something that goes beyond the fishes’ adapting intelligence will be required. Here

we propose a chaos-generator comprising NNDE (neural-network-differential-equation) and an evolving mechanism to have the

NNDE generate chaotic trajectories as many as possible. We believe that the fish could not be adaptive enough to escape from

chasing net with chaos motions that have much different chaos, since unpredictable chaotic motions of net may go beyond the fishes’

adapting abilities to the net motions. In this report we introduce the chaos generating system by NNDE, which can produce many

kinds of chaos theoretically, and then analyze the chaos with Lyapunov number, Poincare return map and initial value sensitivity.

Key words: Chaos, neural network, genetic algorithm.

1. Introduction

In recent years, visual tracking and servoing in

which visual information is used to direct the

end-effector of a manipulator toward a target object

has been studied in some researches [1, 2]. A new

trend of machine intelligence [3] that differs from the

classical AI has been applied intensively to the field of

robotics and other research areas like intelligent

control system. Typically, the animal world has been

used conceptually by robotics as a source of

inspiration for machine intelligence. For the purpose

of studying animal behavior and intelligence, the

model of interaction between animals and machines is

proposed in researches like Ref. [4]. In their previous

research, the fish emotional behavior has also been

examined and the robot with adaptive ability to react

to the fish status has been conceived. Another crucial

Corresponding author: Mamoru Minami, Ph.D., professor,

research fields: visual servoing, intelligent robot, mobile manipulator. E-mail: [email protected].

characteristic of machine intelligence is that the robot

should be able to use input information from sensor to

know how to behave in a changing environment and

furthermore can learn from the environment for safety

like avoiding obstacle. As known universally that the

robot intelligence has reached a relatively high level,

still the word “intelligence” is an abstract term, so the

measurement of the intelligence level of a robot has

become necessary. A practical and systematic strategy

for measuring MIQ (machine intelligence quotient) of

human-machine cooperative systems is proposed in

Ref. [5]. In our approach to pursue intelligent robot,

we will evaluate the intelligence degree between

fishes and the robot by Fish-Catching operation. we

think that the system combined with chaos is smarter

than the fish when the robot can beat the fish by

catching it successfully even after the fish finds out

some reached a relatively high level, still the word

“intelligence” is an abstract term, so the measurement

of the intelligence level of a robot has become

mailto:[email protected]

Multiple Chaos Generator by Neural-Network-Differential-Equation for Intelligent Fish-Catching

824

Fig. 1 Fish catching system PA10.

necessary. A practical and systematic strategy for

measuring MIQ (machine intelligence quotient) of

human-machine cooperative systems is proposed in

Ref. [5]. In their approach to pursue intelligent robot,

we will evaluate the intelligence degree between

fishes and the robot by Fish-Catching operation. We

think that the system combined with chaos is smarter

than the fish when the robot can beat the fish by

catching it successfully even after the fish finds out

some escaping strategy. As we did not find the

research about the intelligence comparison between

animal and robot, we mainly dedicate ourselves to

constructing a smart system that is more intelligent

than the fish. We consider that the competitive

relation can be very meaningful as one way to discuss

robotic intelligence. So we not only employ the

inspiration of animal’s behavior for robot

intellectualization, we can also conceive a robot that

can exceed the animal intelligence. By evolutionary

algorithms [6], Visual Servoing and Object

Recognizing based on the input image from a CCD

camera mounted on the manipulator has been studied

in their laboratory (Fig. 1) [7], and we succeeded in

catching a fish by a net attached at the hand of the

manipulator based on the real-time visual tracking

under the method of Gazing GA [8] to enhance the

real-time searching ability.

We have learned that it is not effective for fish

catching to simply pursue the current fish position by

visual servoing with velocity feedback control.

Actually, the consistent tracking is sometimes

impossible because the fish can alter motion pattern

suddenly maybe under some emotional reasons of fear.

Those behaviors are thought to be caused by

emotional factors and they can also be treated as a

kind of innate fish intelligence, even though not in a

high level.

While observing the fishes’ adapting behavior to

escape in the competitive relations with the robot, that

is continuous catching/releasing experiments, We

found that we can define a “FIQ” (fish’s intelligent

quotient) [9] representing decreasing velocity of fish

number caught by the net through continuous

catching/releasing operation. Through this measure we

can compare the innate intelligence of the fish and the

artificial intelligence of the robot.

It has been well known that many chaotic signals

exist in our body, for example, in nerves, in motions

of eye-balls and in heart-beating periods [10, 11].

Therefore we thought that imitating such animal’s

internal dynamics and putting chaos into robots have

something meaningfulness to address fishes’

intelligence. We embed chaos into the robot dynamics

in order to supplement the deficiency of their

Fish-Catching system [12].

Therefore what we have to pay attention to the

fishes’ nature that the fish does conceive always

escaping strategy against new stressing situation. This

means that robot’s intelligence to override the fishes’

thinking ability needs infinite source of idea of

catching motions. To generate such catching motion,

we propose in this report NNDE

(neural-network-differential-equation) that can

produce neural chaos and inherently have a possibility

to be able to generate infinite varieties of chaos,

derived from the neural network’s ability to

approximate any nonlinear function as accurate as

with desirable precision [13-15].

This paper is organized as follows: Section 2

discusses the fish tracking and catching; Section 3

introduces the problem of fish-catching; Section 4

defines fish intelligent quotient; Section 5 verifies the

validity of chaos; Section 6 shows

neural-network-differential-equation; Section 7


825

explains chaos verification methods; Section 8

proposes the chaos generate system; Section 9 gives

the verification results of chaos; Section 10 discusses

sensitivity of neuron’s weight; Section 11 concludes

this paper.

2. Fish Tracking and Catching

The problem of recognition of a fish and detection

of its position/orientation is converted to a searching

problem of Ttytxt )](),([)( r in order to maximize

))(( tF r , where ))(( tF r represents correlation

function of images and fish-shaped matching model.

))(( tF r is used as a fitness function of GA [8]. To

recognize a target in a dynamic image input by video

rate, 33 [fps], the recognition system must have

real-time nature, that is, the searching model must

converge to the fish in the successively input raw

images. An evolutionary recognition process for

dynamic images have been realized by such method

whose model-based matching by evolving process in

GA is applied at least only one time to one raw image

input successively by video rate. We named it as

“1-Step GA” [7]. When the converging speed of the

model to the target in the dynamic images should be

faster than the swimming speed of the fish in the

dynamic images, then the position indicated by the

highest genes represent the fish’s position in the

successively input images in real-time. We have

confirmed that the above time-variant optimization

problem to solve )(tr maximizing ))(( tF r could

be solved by “1-Step GA”. Ttytxt )](),([)( r

represents the fish’s position in Camera Frame whose

center is set at the center of catching net, then )(tr

means position deviation from net to fish, means

)()( tt rr . The desired hand velocity at the i-th

control period di

r is calculated as

)( 1 ii

V

i

Pdi

rrΚrKr (1)

where, ir denotes the servoing position error

detected by 1-Step GA [7]. PΚ and VΚ given are

positive definite matrix to determine PD gain. Now

we add chaos items to Eq. (1) above, and we also need

to redefine the meaning of di

r .

The simple PD servo control method given by

Eq. (1) is modulated to combine a visual servoing and

chaos net motion into the controller as follows

chaosi

fishii kk rrr

21 (2)

Here Tfish

ifish

ifish

i yx ],[ r , is the tracking error of

fish from the center of camera frame, and T

chaosi

chaosi

chaosi yx ],[ r denotes a chaotic oscillation

in yx plane around the center camera frame.

Therefore the hand motion pattern can be determined

by the switch value 1k and 2k . 01k and 02k

indicate visual servoing, and 01k and 12k

indicate the net will track chaotic trajectory made by

NNDE being explained later in this report. The

desired joint variable dq is determined by inverse

kinematics from dr by using the Jacobian matrix

)(qJ and is expressed by equation

ddq rJq )( (3)

where, )(qJ is the pseudo inverse matrix of )(qJ .

The robot used in this experimental system is a 7-Link

manipulator, Mitsubishi Heavy Industries PA-10

robot.

3. Problem of Fish-Catching

In order to check the system reliability in tracking

and catching process, we kept a procedure to catch a

fish and release it immediately continuously for 30

min. We released 5 fishes (length is about 40 mm) in

the pool in advance, and once the fish got caught, it

would be released to the same pool at once. The result of this experiment is shown in Fig. 2, in which

vertical axis represents the number of fishes caught in

successive 5 min and horizontal axis represents the

catching time. We had expected that the capturing

operation would become smoother as time passing on

consideration that the fish may get tired. But to their

astonishment, the number of fishes caught decreased

gradually.

The reason of decreased catching number may lie in

the fish learning ability. For example, the fish can

learn how to run away around the net as shown in


826

Fig. 2 Result of fish catching.

Fig. 3 Fish motion.

Fig. 3a by circular swimming motion with about

constant velocity, having made a steady state position

error that the net can not reach to the chasing fish. Or

the fish can stay in the opposite corner against the net

in the pool shown in Fig. 3b. And also, the fish can

keep staying within the clearance between the edge of

the pool and the net shown in Fig. 3c where the net is

inhibited to enter.

To solve these problems, and to achieve more

intelligent fish catching systems, we thought chaos

behavior of the net with many chaotic varieties can be

a possible method to overcome those fishes’ escaping

intelligence, since huge variety of chaos trajectories

seems to be unpredictable for the fish to adapt them.

This strategy to overcome fishes’ adaptive intelligence

is based on a hypothesis that unpredictability of the

motion of the chasing net produced by plural chaos

can made the fishes’ learning logic confuse, getting

the fish catching robot have made intelligence than the

fishes’. Then we propose Neural-Network-Differential

-Equation to generate chaos as many as possible.

4. Fish Intelligence Quotient

To evaluate numerically how fast the fish can learn

to escape the net, we adapted Linear Least-Square

approximation to the fish-catching decreasing

tendency, resulting in 7.20486.0 ty as shown

in Fig. 2. The decreasing coefficient −0.486 represents

adapting or learning velocity of the fishes as a group

when the fishes’ intelligence is compared with robotic

catching. We named the coefficient as “FIQ” (Fish’s

Intelligence Quotient), since learning velocity has

been thought as one of components of the Intelligence

Quotient. The larger minus value means high

intelligence quotient of the fish, zero does equal and

plus does less intelligent than robot’s. To overcome

the fishes’ intelligence, more intelligent robotic

system needs to track and catch the fish effectively, in

other words it comes to the problem on how to use the

item ichaosr chaos in Eq. (2) effectively to exceed the

fish intelligence.

5. Validity of Chaos

In 1982, some experiments revealed that mollusk

neuron cells and plant cells have irregular excitement

and show chaotic nature if gave them periodic current

stimulation. In addition, also chaotic response for

periodic current stimulation had been clarified in the

axon of the cuttlefish in 1984. From these studies, it

became clear that the chaos is associated with biology.

In the late 1980s, the relationship between chaos and

function of the nervous system has been discussed.

Mpitosos and colleagues examined the pattern of

rhythmic firing of motor neurons of sea cucumber and


827

showed that frequency variation of continuous

discharge relates to the rhythm of the movement with

chaotic behavior. Thus, chaos exists in biological

behavior. It is decided whether the nerve cell of the

organism is excited by a stimulation signal, and this is

because it follows the theory of the chaos. Therefore,

animal behavior and strategies can be estimated from

point of chaos, and maybe apply to catch fish. There

has been presented chaoses with a simplified model of

H-H (hodgkin-huxley) model or BVP (bonhoeffer-van

der pol) model. Using one chaos model to produce

unpredictable motion add to catching-net behaviors

seems to be effective, however we thought single

chaos model is not adequate to overcome fishes’

escaping idea since the fishes change their behavior

continuously.

6. Neural-Network-Differential-Equation

Lorenz and Rossler models renowned as chaos

generation comprise three differential equations,

producing three-dimensional chaotic trajectory in

phase space. Since a N.N. (neural-network) has been

proven to have an ability to approximate any

non-linear functions with arbitrarily high accuracy, we

thought it is straightforward to make a differential

equation including N.N. so that it can generate plural

chaoses by changing N.N.’s coefficients. We define

next nonlinear differential equation including N.N.

function ))(( tpf as

)).(()( tt pfp (4)

Ttptptpt )](),(),([)( 321p is state variable. The

nonlinear function of ))(( tpf in Eq. (4) is constituted

by N.N.’s connections, which is exhibited in left part

of Fig. 4 where the N.N. and integral function of

outputs of N.N. and the feedback of the integrated

value to the inputs of N.N. constitute nonlinear

dynamical Eq. (4). We call it as

Neural-Network-Differential-Equation.

7. Chaos Verification Methods

Since there have been no simple criterion to

determine whether irregular oscillation should be a

chaos or not, we have to apply plural evaluations over

the irregularities of trajectories produced by NNDE.

The followings are criteria being used for judging the

chaotic characters.

7.1 Lyapunov Exponent

As one of criteria to evaluate a chaos’ character of

expansion in time domain, Lyapunov exponent

expressed by the following equation is well known,

1

0

)(log1

limN

i

i

N

xfN

(5)

where, positive value can represent that the irregular

oscillation diverts from a standard trajectory, which

expands like a function of ate (a > 0).

7.2 Poincare Section

The trajectories of the motion made by

Neural-Network-Differential-Equation (4) is

examined by using the Poincare section to verify

further whether the resulted trajectories can be

identified as chaos. Next, the Poincare section is

explained. First of all, we examine a simple closed

curve in three dimensions as shown in Fig. 5. The

plane “A” that intersects with this closed trajectory

pointed by “P” is defined as the Poincare section.

The intersecting points are named as

,,,, 21 nnn PPP and corresponding x-axis position on A

are ,,,, 21 nnn xxx which are all pointed to the

Poincare Return Map as ,, 1nn xx as shown in Fig. 6.

With the Poincare return map of Fig. 6 representing a

Fig. 4 Block diagram of chaos generation.


828

Fig. 5 Poincare section.

Fig. 6 Poincare return map.

shape of “Λ”, the closed curve has the structure of

stretching and folding. This structure is a basic

character of chaos. Looking at the left half of Fig. 6,

we can see the inclination coefficient 1/1 nn dxdx

and right half has 1/1 nn dxdx , representing that left

half has expansion and the other does contraction.

8. Chaos Generate System

Fig. 4 represents the block diagram to find chaos by

using GA and Lyapunov number. This GA is not

1-Step GA, described in Section 2 but used as a

normal GA’s procedure that evolves genes

representing neural network coefficient’s volume. The

trajectory )(tp in time domain obtained from

Neural-Network-Differential-Equation is used for the

calculation of Lyapunov number. Here, T],,[ 321 L

is a Lyapunov number. Using this L for the

evolution of GA, fitness function is defined as

follows,

332211 kkkg (6)

This fitness function incorporated the chaotic

property of the Lyapunov spectrum, which is one of

factors to be essential for generating chaos trajectory.

Here, because we discuss three-dimensional chaotic

attractor in phase space, there are 3 Lyapunov

exponents. The relationship between positive and

negative Lyapunov spectrum is (+, 0, −), which means

resulted time trajectory of Eq. (4) may be thought to

be chaos. Parentheses indicate the sign of the

Lyapunov spectrum. In other words, 1 is positive,

2 is also positive or negative small values, 3 is

negative case, the fitness function of Eq. (6) appears

to have relatively large positive value when

0,0,0 321 . In addition, 21,kk , and 3k

are positive coefficients. The gene of GA is defined as

shown in Fig. 7, with connection weights of N.N.

being T

nqqq ],,,[ 21 q . In this report we adopted a

network of 3 × 6 × 3 as shown in Fig. 4, then the line

number of connections and coefficient are 48, i.e., n =

48. The bit length of q is 16 bits. Because the gene

is expressed in binary, converted to decimal and

normalized into a range from 0 to 1. Then, generating

a trajectory )(tp based on a given gene having been

determined by GA at one previous generation and

calculating Lyapunov number, and evolving new

generation of gene are repeated. This GA’s evolution

can find q to have a highest value of g defined by

Eq. (6), which means possible chaos trajectory.

9. Verification Results of Chaos

So far we have found four chaos patterns with

different neural coefficients explore by GA mentioned

in the previous section. We named them with a serial

number as chaos 01∼chaos 04. The followings are


829

Fig. 7 Gene of GA.

the introduction of those chaoses with each individual

character.

9.1 Chaos 01

9.1.1 Lyapunov Number

Lyapunov numbers are 0.0145851 ,

0.003314- 2 and -0.1653813 . These are

corresponding to the Lyapunov spectrum of chaos, (+,

0, −), and the trajectory is exhibited in Fig. 8.

9.1.2 Sensitivity to Initial Value

Two time-profile of trajectories with minutely

different initial value are shown in Fig. 9. The

trajectories of ))(),(),(( 111 tztytx are the results that

originated from the initial values of

00.1)0(,00.1)0(,00.1)0( 111 zyx and ))(),(),(( 222 tztytx

are from .01.1)0(,01.1)0(,01.1)0( 222 zyx

Trajectories of 1x and 2x are shown in Fig. 9.

We can see from Fig. 9 that the two trajectories

with minute difference of initial values divert often

about 800 s having passed, this means the slight

different initial values make large separation with

each other, indicating sensitivity of initial value,

which is one of the character of chaos. As for y and

z coordinates, they are similar, omitted to spare the

space.

9.1.3 Poincare Return Map

Chaos 01’s poincare return map is shown in Fig. 10.

One dimensional map can be seen in Fig.10, from

which we can understand that the map represents

expanding (left half of the Fig. 10) and contracting

(right half) that are essential character to generate

chaos.

Therefore, the property of chaos 01 has been

confirmed from the viewpoints of Lyapunov number,

the sensitivity of initial value, and the Poincare return

map.

9.2 Chaos 02, 03, 04

We searched second chaos by similar produce like

chaos 01. Up to now we have found other three chaos,

02, 03 and 04 whose trajectories are depicted in Figs.

11-13, where Lyapunov numbers are listed in Table 1,

including chaos 01 also. We confirmed all chaos

trajectories have the Lyapunov spectrum of chaos, (+,

0, −).

Fig. 8 Generated chaos trajectory of Chaos 01.

Fig. 9 Generated trajectory 01 of x (300[s] to 1100[s]).

Fig. 10 Poincare return map of Chaos01.


830

Fig. 11 Generated chaos trajectory 02.



10. Sensitivity of Neuron’s Weight

We have noticed weight coefficient of N.N. that

generated chaos 03 is almost similar to chaos 04’s.

That is, only one weight coefficient is different, that is

“ 1q ” in Fig. 14. We think “ 1q ” is related to the

generation of chaos. So we increased the weight

slightly from “-1” and compare their trajectories. The

range of 1q is 0.10.1 1 q and

1q is increased

from -1.0 by 0.1. In the case of

,4.00.1 1 q 2.01.0 1 q and ,0.18.0 1 q we can

not be considered that the trajectories are

semi-periodic trajectories as shown in Figs. 15-17.

They indicate that the trajectories are expanded to

infinity as shown in the three figures. On the other

hand, the range of 0.03.0 1 q and 7.03.0 1 q

made chaos trajectories as shown in Figs. 18 and 19.

This result indicates that continuous changing of 1q

can make various chaos, stemming from continuity of

real variables. The Lyapunov number of each

trajectory is shown in Table 2.

Table 1 Lyapunov number.

chaos01 chaos02 chaos03 chaos04

1 0.014585 0.01919 0.015934 0.01208

2 -0.003314 0.00733 -0.002172 -0.00143

3 -0.165381 -0.10379 -0.123026 -0.075448

Fig. 14 Neural network for nonlinear function generation.

Fig. 15 Weight = -0.7.

Fig. 16 Weight = 0.2.


831


Fig.18 Weight = 0.0.


Table 2 Lyapunov number.

-0.7 0.0 0.2 0.7 0.9

1 0.007314 0.016477 0.008571 0.004668 0.003386

2 -0.004319 0.002719 -0.038939 0.011399 -0.046857

3 -0.30184 -0.106123 -0.061084 -0.136963 -0.048269

11. Conclusions

This paper proposed chaos generating system

composed of Neural Network and GA’s evolving

ability to change the Neural-Network-Differential-

Equation to be able to generate chaos. This chaos

generating system has exploited the neural network’s

nature of approximation of any nonlinear function

with any desired accuracy. We will utilize this chaos

motion for overcoming fishes’ escaping ability from

chasing net for now.

References

[1] R. Kelly, Robust asymptotically stable visual servoing of

planar robots, IEEE Trans. Robot. Automat. 12 (1996)

759-766.

[2] P.Y. Oh, P.K. Allen, Visual servoing by partitioning

degrees of freedom, IEEE Trans. Robot. Automat. 17

(2001) 1-17.

[3] T. Fukuda, K. Shimojima, Intelligent Control for

Robotics, Computational Intelligence (1995) 202-215.

[4] M. Bohlen, A robot in a cage-exploring interactions

between animals and robots, in: Proc. of Computational

Intelligence in Robot. Automat. (CIRA), 1999.

[5] H.J. Park, B.K. Kim, K.Y. Lim, measuring the MIQ

(machine intelligence quotient) of human-machine

cooperative systems, IEEE Trans. 31 (2001) 89-96.

[6] M. Minami, H. Suzuki, J. Agbanhan, T. Asakura, Visual

servoing to fish and catching using global/local GA

search, in: Int. Conf. on Advanced Intelligent

Mechatronics Proc., Como, Italy, 2001.

[7] M. Minami, J. Agubanhan, T. Asakura, Manipulator

visual servoing and tracking of fish using genetic

algorithm, Int. J. of Industrial Robot 29 (1999) 278-289.

[8] H. Suzuki, M. Minami, J. Agbanhan, Fish catching by

robot using gazing GA visual servoing, Transaction of the

Japan Society of Mechanical Engineers C-68-668 (2002)

1198-1206.

[9] J. Hirao, M. Minami, Intelligence comparison between

fish and robot using chaos and random, in: International

Conference on Advanced Intelligent Mechatronics, China,

2008.

[10] K. Aihara, Chaos in Neural System, Tokyo Denki

University Press, Tokyo, 1993.

[11] R. FitzHugh, Impulses and physiological states in

theoretical models of nerve membrane, Biophy. J. 1

(1961) 445-466.

[12] M. Minami, J. Hirao, Intelligence comparison between

fish and robot using chaos and random, in: Int. Conf. on

Advanced Intelligent Mechatronics Proc., China, 2008.

[13] C.T. Lin, C.S. Lee, Neural Fuzzy Systems, Englewood

Cliffs, Prentice Hall PTR, NJ, 1996.

[14] L.M. Peng, P.Y. Woo, Neural-fuzzy control system for

robotic manipulators, IEEE Control Systems Magazine

22 (2002) 53-63.

R. Endo, J. Hirao, M. Minami, Intelligent chaos

fish-catching based on

neural-network-differential-equation, in: SICE Annual

Conference,Taiwan, 2010.

1q


An Efficient Algorithm for the Evaluate of the

Electromagnetic Field near Several Radio Base Stations

Algenti Lala, Sanije Cela and Bexhet Kamo

Faculty of Information Technology, Polytechnic University, Tirana 10000, Albania


Abstract: This paper is motivated by the increased presence of the radio base stations, and the need to calculate the electromagnetic

field near them. The debate on the effects of the electromagnetic field exposure, in line with the increased success and presence of the

mobile telephony, has attracted the public interest and it has become a concern for the community. The standard procedures in place

for estimation of the electromagnetic field require prior knowledge of the criteria for the field evaluation, be it near field, far field,

presence of one or several base stations, the operating frequencies bands and their combinations. Aiming to have a practical method

for the evaluation, the authors will try to do develop a theoretic model, on which base the authors will simulate the antenna of the

base station and prepare the numeric method that will provide the baseline for the application. They will than compare the

calculations for real situations for which all know the geometrical features, with the ones calculated based on a known theoretical

method also knows as method of the moments MoM, simulated with NEC-2 (numerical electromagnetic code), and further more with

the values measured in the field under the same conditions as the ones for the simulated environments. The results are interpreted in

order to define the efficiency of the proposed method as well as to have an idea on the simplicity, accuracy and computing capacities.

Key words: Evaluation, base stations, far field, FDTD method, algorithm.

1. Introduction

The task to be accomplished in this paper are the

theoretical calculations of the electric and magnetic

field, the density of power caused by the antennas of

the radio base stations, considering that this is caused

by joint presence of several base stations at a given

place. The proposed calculation method is based on

the below requirements:

The need for a simple, and practical method, to

evaluate the field nearby the radio base stations.

The need to define the security zones Iso-Curves

(spaces inside which the values of the electromagnetic

field are higher than the standard recommended ones)

The need for quick evaluations—be it time—wise

as well as space—wise: In Albania are operating 4

cellular operators AMC, VODAFONE, EAGLE and

Corresponding author: Algenti Lala, M.Sc., research fields:

systems in radiofrequency, antennas and measurement in radio frequency. E-mail: [email protected].

PLUS. These operators are obliged to monitor the

levels of the field in the vicinity of the radio base

stations, and to assure that the levels are within the

reference values of safety, this in the framework of the

client care. By monitoring it means the constant

measurement of the levels of electromagnetic field

near the stations. This is translated in relatively high

costs due to man work travel and density of the

measurements in time.

The presence of several operators in the same space

causes the levels of the electromagnetic field to

increase, independently that each individual antenna

contribution is within the allowed limits. Under these

circumstances it is required that the regulatory body

checks frequently the levels in the public spaces and

for the safety of the public. The state authorities are

trying to protect the population and at the same time

they must promote the cellular companies and

encourage new emerging services.

mailto:[email protected]

An Efficient Algorithm for the Evaluate of the Electromagnetic Field near Several Radio Base Stations

833

The literature [1-5] offered many models and

numerical methods for the calculation of the

electromagnetic field near the radio base stations.

The basis of these methods is the conversion of the

integral equations in a linear system, which can be

further computed in a computer. The simulations in

software environments provide accurate results but

they require a lot of processing power and they require

a lot of time to be processed.

In this paper the authors have selected a theoretical

model which implementation is based on the

numerical method FDTD (finite-difference

time-domain) by using the formula of the far field.

They are presenting the theoretical calculations in

real life scenarios in which the geometry of the

antennas in a group of radio base stations is known,

and the results are compared with the values acquired

by the simulator NEC-2 (numerical electromagnetic

code) and further compared to the measured values

acquired in the field by using the NARDA SRM 3000

equipment, for the same conditions of the radio base

stations simulated via the two methods.

The paper is organized as follows: Section 2

describes the proposed theoretical model for the

evaluation of the far field; Section 3 describes the

numerical method and the respective proposals;

Section 4 presents the proposed algorithm; Section 5

presents results and discussions; Section 6 gives

conclusions.

2. The Proposed Theoretical Model for the

Far Field

Modeling of the electromagnetic field near the radio

base stations is a way to evaluate and define the

excluded zones near these stations. The selection of an

appropriate model is important in order to have a good

estimation of the levels of the radiation.

In the literature [1-5] can be found many models for

the definition of the zones of the near field as well as

the far field. Attention must be paid to the fact that in

the zones of the near field, the levels of the radiation

depends not only on the distance from the antenna but

also one the movement along the vertical axis,

whereas in the case of the far field the levels depend

only on the distance not on the movement along the

vertical axes. The models of the far field aim towards

a simple formulation and based on them can be

applied numerical methods which make possible the

estimation of the electromagnetic field in a short time

and with modest computer processing power.

The proposed model is based on the model

―Far-field Gain-based‖ [6] as Eq. (1). This model

provides a simple and efficient method for the

evaluation of the levels of the electromagnetic field

radiated by the antennas of the radio base stations

with uniform groups of cells in the zone of the near

field and the ones of far field. The above is achieved

in two steps:

The first step, electrical intensity of the antenna is

calculated by combining the radiation of the far field

of the antenna elements, and the group factors, by

accepting that the antennas of the radio base stations

are an uniform group of cells. (Fig. 1):

3,),(

),(30),,(

1

)

due

d

GPdE

N

i

ii

j

i

iieini (1)

Second step as Eq. (2):

N

DDGG HeVeM

iie

)()(),(

(2)

i

i

di

2)1( (3)

where, N number of radiating cells, (di, θi, φi)

spherical coordinates of the i-th element up to the N-th

one, Pin total radiated power by a given group, Ge (θi,

φi) amplification of the radiating element, di distance

from the i-th element, u (θi, φi) unit vector of the i-th

element, λ wave length ,GM maximal gain of the

antenna, DVe(θ), DHe (φ) the models of the radiating

element in the vertical and horizontal plan and Φi the

differences of the phases between the coefficients of

the radiating element.

In this paper reference is made to Kathrein antennas

specifically to the models, 80010670, 80010671 and


834

Fig. 1 Modeling the antenna with N source cells.

80010672. The Kathrein model 80010671 [7] is used

for the spectrums 900/1800/2100 MHZ. The intensity

of the electric field calculated for each of the elements

(cells) of the antenna is vectorial and it can be

projected according to the axis x, y, z, and obtain the

respective components for each of the axes. The

electric and magnetic fields in the Cartesian

coordinates are composed of each of the three

components Ex, Ey and Ez for each and every

frequency the same is valid for the magnetic field with

respective Hx, Hy and Hz components [8] as Eq. (4):

222

tanRe ( zyxtezul EEEE (4)

The authors propose for the modeling of the base

antennas the below:

(1) The use of the ―Far-field Gain-based‖ model ins

which the intensity of the electric field is calculated by

the Eq. (1) with the approximation that

),( ii

jue i

=1. This approximation influences the

accuracy of the model ―Far-field Gain-based‖ for the

near fields up to 15 λ. This falls within the safety

distances as defined by the standards.

The acquired equation is as Eq. (5):

N

i i

iiein

d

GPdE

1

) ),(30),,(

(5)

(2) The statistical study considered is ―the worst

case scenario the vectors Ex1 Ex2… Exn in the same

phase as the Ey and Ez‖ This definition will lead to and

overestimation of the electromagnetic field on the

given point. The intensity of the electrical field in a

given point (weight per frequency) near the antenna of

a radio base station when the antenna is Three-Band

(900/1800/2100) and by considering the ―vectors Ex1

Ex2… Exn in the same phase as the Ey and Ez‖ is as per

the Eq. (6).

ERfrekuence-i-th—the electromagnetic field radiated by

the antenna for the i-th frequency on the calculated

point [9] .

2

1

2

1

2

1

(

N

z

N

y

N

xthiRfrekuence EEEE (6)

The intensity of the electrical field in a given point

(weight for the three frequencies) near the antenna of

a radio base station when the antenna is Three-Band

(900/1800/2100) is as Eq. (7).

ERTotale Frekuence —The electromagnetic field radiated

for the three frequencies on the calculated point [9]

23

2

2

2

1( FrekuencaFrekuencaFrekuencakuenceRTotaleFre EEEE (7)


835

In the majority of the real situations are

encountered antennas which operate in different

systems that cover the same fields or have their

radiating diagrams overwritten in their main lobbies.

It comes as a natural need to evaluate the generated

electric field by considering at the same time the

radiation of each and every antenna. By the

assumption that the signal sources are not correlated,

the contribution of the field is added in quadrature

therefore the electrical field in a given point will be as

Eq. (8). [9]

2

1

(

n

iEE (8)

where n is the total number of the antennas which

generate the field and Ei is the contribution from the

i-th antenna.

Eq. (5) is the formula proposed for the far field in

the vicinity of a radio base station. The overall

electromagnetic field is calculated by overlapping the

electrical intensity of the electromagnetic field

calculated for each of the contributing antenna as

Eq. (9):

n

j

N

i i

iiein

Rnd

GPdE

1

2

1

) ),(30),,(

(9)

The approximation of the far field can result in and

overestimation of the measured electromagnetic field.


(weight per frequency) near some radio base stations

where the antennas are three-band (900/1800/2100) as

Eq. (10):

2

111 11 1

2(

N

z

nn N

y

n N

XRn EEEEthiFrekuence

(10)


(the weight for the three frequencies) near some radio

base stations where the antennas are three-band

(900/1800/2100) as Eq. (11):

232

22

1( FrekuencaFrekuencaFrekuencakuenceRTotaleFre EEEE (11)

3. The Proposed Numerical Method

The FDTD (finite-difference time-domain) method

is a numerical one suitable for solving the

electromagnetic problems. The space in this method is

divided in tiny rectangular cells. Modeling the

dielectric materials or the antenna structures in the

stations is very simple by the use of this method [9].

As the result the FDTD is an appropriate nomination

for the evaluation of the electromagnetic radiation

near the radio base stations. As it is now known the

simulation of the full waves requires more and more

computing power and time. In order to increase the

efficiency a simple geometry for the antenna is

proposed, by using an optimization algorithm to meet

the required specifications for the antenna.

The proposal is based on the numerical method of

the finite differences in the time domain FDTD with

the below changes:

(1) discretization of the antenna of the radio base

stations, defining the elementary cell (the cell size

meets the criteria ∆ < λ/10);

(2) the space from the radio base station to the

given point is considered as free space;

(3) discretization of the Maxwell’s equations in the

time domain, explaining scheme as per the proposed

theoretical model;

(4) solving of the discretized equation and the

selection of the transitional step (time step ∆t must

meet the CFL criteria);

(5) interpreting the results.

The antenna model used in this study is presented

schematically (Fig. 2). The physical dimensions of the

antenna (height and width) 2,000 mm × 240 mm. The

cells in the cubical form size is Δx (= Δy = Δz) = 10

mm. The accuracy of the model ―cylindrical-wave‖ is

reduced when the distance from the antenna is

increased. At the same time the model

―spherical-wave‖ accuracy is increased.

The selection of the numerical method FDTD with

the respective proposal (discretizing only the antenna

of the radio base stations and the space from the

antenna to the observation point as free space)

influents by decreasing the execution time as compared


836

Fig. 2 Modeling the antenna according to FDTD.

with the NEC-2 simulation for the same machine.

The proposed model leads to an overestimation of

the electromagnetic field. This should not be

interpreted as a weakness of the model because the

goal is calculation of the field and comparison of the

calculated values with the reference ones. Therefore

the overestimation contributes to the understanding

that the real value is somewhat percent less than the

calculated one.

4. The Proposed Algorithm

This algorithm allows a simple and fast estimation

of the electromagnetic environment starting from the

antenna model. The essence of this method is the

calculation in the so called ―free space‖ of the

electromagnetic field, based on the same model used

for the far field. The 2D model is rebuilt by the

criteria of the projection placed in the algorithm. The

applicability of this method is based on the

assumption that the field radiates independently from

the direction of the observation.

Eq. (5) is the proposed theoretical model for the

antenna of a radio base station.

Eq. (9) is the theoretical model proposed for the

estimation of the electromagnetic field near some

radio base stations. The total field is achieved through

overlapping of electrical intensity of the respective

electrical intensity of the electromagnetic field

calculated for each and every antenna. The

approximation of the far field can result in an

overestimation of the electromagnetic field due to the

fact that the statistical study used on this case does

scalar sum.

Eq. (10) is valid for the electromagnetic field

created by the contribution of n antennas at a given

point for the i-the frequency.

Eq. (11) gives the total field from the n antennas at

the given point for the three frequencies

900/1800/2100MHz. By using the open source code

[10] for the calculation of the intensity of the electric

field for the far field scenario, the proposed algorithm

solves numerically as per the proposed FDTD the Eqs.

(9)-(11).

Afterwards by using the worst case criteria for the

projection of the surfaces in the plans Oxy, Oxz and

Oyz the authors build the iso-curves with the values of

the electric intensity E > 41V/m and define the safety

distances near these radio base stations.

The total value of the intensity of the electric field

for a frequency will be calculated as result of the

quadratic sum of the contribution from each and every

antenna in the particular frequency. On the other hand

the total value of the entire frequency spectrum for the

intensity of the electric field will be evaluated based

on the same procedure.

The proposed algorithm enables the calculation of

the electromagnetic field for up to 12 radio base

stations. For each antenna of the radio base stations is

required prior knowledge of some characteristics

which are presented as the inputs for antennas:

(1) Label of the antenna (not required for

calculation, it is the label in the final graphic);

(2) antenna gain (dBi);

(3) antenna mechanical down tilt (+/-): The

electrical down tilt should be included in the radiation

pattern of the antenna;

(4) horizontal plane rotation angle of the antenna

main beam (with reference to a fixed coordinate

system);

(5) antenna centre X-position;


837

(6) antenna centre Y-position;

(7) antenna centre Z-position;

(8) name of the file containing the horizontal plane

radiation pattern;

(9) name of the file containing the vertical plane

radiation pattern;

(10) input power to the antenna (W).

At the end of the file a routine can require the

punctual calculation of the EM field in some points

identified by their Cartesian coordinates this routine

produces the calculated EM field values to be printed

in the output window.

The program output is represented by the following

data, for each iso-curve level introduced in the input

file:

(1) Iso-curve safety zone on the x-y plane (graphic

window), representing the total contribution of the

antennas and the single safety zone of each antenna;

(2) Iso-curve safety zone on the y-z plane (graphic



(3) Iso-curve safety zone on the x-z plane (graphic



(4) safety distances.

Errors in the input file are indicated by a warning

message, and this leads to the termination of the

program.

5. Results and Discussion

To judge the algorithm for evaluating the intensity

of the electric field radiated by the antenna of the

radio base stations what is proposed is related in the

following three steps:

(1) Through the proposed model and the application

of the FDTD numerical method, the values of the

intensity of the electrical field are defined, depending

on the distance along the axis X, Y, Z and this is

presented in a graphical form;

(2) The simulations are performed for the same

given points along the X, Y and Z axis using the

NEC-2 program implementation of MoM [11];

(3) The intensity of the electrical field is measured

by using the NARDA SRM 3000 [12] in the same

points as the simulated and calculated ones, and the

results are presented graphically in the same graph

with the ones of the above steps.

Through the proposed method the authors are

capable to have a quick evaluation be it for the near

zones, as well as for the far ones, and the calculations

complexity is significantly reduced. The cases the

authors have considered consist in a situation in which

there are three cellular operators; each operator has

three antennas in the GSM 900, GSM 1800 and GSM

2100 (in total 9 antennas, 3 for each band). Reference

is made to Kathrein antennas specifically to the

models 80010670, 80010671 and 80010672.

Fig. 3 shows the three cellular operators (three

antennas each) in cartesian coordinates space for the

middle point values antenna 1(5, 5 ,5), antenna 2(2,8,6)

antenna 3(-4 ,-3 ,7).The distances are in meters m.

The calculations are performed for the three

frequencies 900/1800/2100MHz jointly for the below

cases:

Intensity of the electrical field in the X axis (Y = 0

and Z = 0)

Intensity of the electrical field in the Y axis (X = 0

and Z = 0)

Intensity of the electrical field in the Z axis (X = 0

and Y = 0)

In the case according to the X axis the values are

for points between negative value -4 m to 11 m (30

samples).

Fig. 4 shows the intensity of the electrical field,

magnetic field and the power density for two

calculating methods the NEC-2 (MoM method) and

Emf calculation Alg, as well as the measured values in

the site.

In the case according to Y axis the values are for

points between the negative -6 m up to 10 m (30

samples). The graphs below present the intensity of

the electrical field, magnetic field and the density of


838

Fig. 3 Geometry of the simulation.

power for both methods, and the measured in site

values (Fig. 5):

In the case according to Z axis the values are for

points between negative -1.5 m up to 4 m (12 samples)

The graphs upper present the intensity of the electrical

field, magnetic field and the density of power for both

methods, and the measured in site values (Fig. 6):

Table 1 presents the intensity of the electric field as

per the x axis (m) (y = 0 and z = 0) sampling step

0.5m.

The values of the intensity of the electric field(v/m)

are due for three methods.

Simulation by NEC-2 software [11].

The value of a single sample is calculated for an

average time 80-85 s. The total time for the

calculation of the entire blue graph in the Fig. 4 for 30

samples is 2,550 s.

Simulation by using the proposed algorithm Emf

Calculation Alg. The value for a sample is calculated

for an average time of 20 s. The total time for the

calculation of the entire pink graph in the Fig. 4 for 30

samples is 600 s.

The measurement by using the NARDA SRM 3000

[12]. The value for a sample is achieved by an average

of measurements that last 360 s. The total time for the

yellow graph (Fig. 4) for 30 samples is 10,800 s.

The simulations with the NEC-2 and Emf

calculation Alg are performed on the same machine

with the specifications; Server IBM x3650 M4,

processor Xeon 2.0GHz/1333MHz / RAM 8GB.

6. Conclusions

The algorithm for the evaluation of the radiation of

the field in the presence of several radio base stations

was devoleped by using the proposed theoretical model

for the far field, and the method of FDTD (finite

–differences time-domain) with the respective

assumptions. The proposed algorithm ―Emf Calculation

Alg‖ for the calculation of the electromagnetic field

requires prior knowledge about the antennas, usually

provided by the manufacturer (emitting space in the

horizontal plan and vertical plan).


839

,

Fig. 4 Dependence of E, H and S in x axis (m) (y = 0 and z = 0).


840

Fig. 5 Dependence of E, H and S in y axis (m) (x = 0 and z = 0)


841

Fig. 6 Dependence of E, H and S in z axis (m) (x = 0 and y = 0).


842

Table 1 Comparative result.

Coordinates Values Values Values measured

Samples (X.Y.Z) NEC-2 Emf NARDA

Calculation Alg SRM3000

1 -4.0.0 0.30200514 0.2749 0.258406

3 -3. 0. 0 0.36012108 0.3278 0.308132

5 -2. 0. 0 0.50579544 0.4604 0.432776

7 -1. 0. 0 1.14880602 1.0457 0.982958

9 0 .0 .0 1.3853346 1.261 1.18534

11 1. 0. 0 1.1321073 1.0305 0.96867

13 2. 0. 0 0.74199444 0.6754 0.634876

15 3. 0. 0 0.91304646 0.8311 0.972387

17 4. 0. 0 1.93309656 1.7596 1.654024

19 5. 0. 0 2.043587 2.2457 2.110958

21 6. 0. 0 3.31821144 3.0204 2.839176

23 7. 0. 0 3.18835692 2.9022 2.728068

25 8. 0. 0 3.0563052 2.782 2.61508

27 8. 0. 0 3.0563052 2.782 2.61508

29 9. 0. 0 3.98989548 3.6318 3.413892

31 10. 0. 0 3.27196038 2.9783 2.799602

Analysis and comparison between the calculated

values and the measured ones concludes that the

proposed algorithm provides accurate results for the

field close and far from the radio base stations (up to

12 BTS antennas) in a given urban area. The error

between calculated and measured values is less than

10%.

Analysis and comparison between the values

calculated with the proposed algorithm and the ones

simulated with the NEC-2 software, concludes in

values that converge, thus confirming the accuracy of

the proposed algorithm.

Analysis and comparison between the time required

for running the proposed algorithm, and the NEC-2

simulation confirms the speed of calculations for the

authors’ algorithm, making it suitable for low

computing capacity machines.

For the reviewed radio base stations which operate

at the same time in the frequencies band

900/1800/2100MHz the algorithm ―Emf caclultion

Alg‖ calculates the safety zones in the Iso-kurves (the

space where the E ≥ 41 V/m) in the plans x-y, x-z and

y-z.

In an environment in the presence of several radio

base stations can be calculated the intensity of the

electrical field, magnetic field and the density of

power for different distances from the antenna in a

short period of time providing confident and accurate

results.

References

[1]. M. Barbiroli, C. Carciofi, V.D. Esposti, G. Faciasecca,

Evaluation of exposure levels generated by cellular

systems: methodology and results, IEEE Transactions on

Vehicular Technology 51 (2009) 1322-1329.

[2]. M. Martínez, A.F. Pascual, E. Reyes, V.V. Loock,

Practical procedure for verification of compliance of

digital mobile radio base stations to limitations of

exposure of the general public to electromagnetic fields,

IEEE Proceedings Microwaves, Antennas and

Propagation 149 (2009) 218-228.

[3]. Z. Altman, B. Begasse, C. Dale, A. Karwowski, J. Wiart,

M.F. Wong, et al., Efficient models for base station

antennas for human exposure assessment, IEEE

Transactions on Electromagnetic Compatibility 44 (2002)

588-592.

[4]. L. Correia, C. Fernandes, G. Carpinteiro, C.Oliveira, A

Procedure for Estimation And Measurement of

Electromagnetic Radiation in the Presence of Multiple Base

Stations, Instituto Superior Técnico, Lisbon, Portugal, 2002.

[5]. C. Oliveira, C. Fernandes, C. Reis, G. Carpinteiro, L.

Ferreira, Definition of Exclusion Zones around Typical

Installations of Base Station Antennas, Instituto Superior


843

Técnico, Lisbon, Portugal, 2005.

[6]. M. Bizzi, P Gianola, Electromagnetic fields radiated by

GSM antennas, IEEE Transactions on Electromagnetic

Compatibility, Electronic Letters 35 (1999) 855-857.

[7]. K.W. Kg, Technical Information and New Products 790

-2500 MHz Base Station Antennas for Mobile

Communications, Catalogue Issue, 2004.

[8]. A.W. Scott, Radio Frequency Measurements for Cellular

Phones and Wireless Data Systems, A John Wiley&Sons,

New York, 2008.

[9]. V. Prasad, Engineering Electromagnetic Compatibility

Principles, Measurements, Technologies and Computer

Model, A John Wiley & Sons, New York, 2008.

[10]. I. Laakso, T. Uusitupa, S. Ilvonen, Comparison of SAR

calculation algorithms for finite-difference time-domain

method, Physics in Medicine and Biology 55 (2010)

421-431.

[11]. Numerical Electromagnetics Code NEC2 Unofficial

Home Page, http:// www.nec2.org.

[12]. Available at http://www.narda-sts.de/en/products/emc/.

http://www.nec2.org/

http://www.narda-sts.de/en/products/emc/


Composing Specific Domains for Large Scale Systems

Asmaa Baya and Bouchra EL Asri

Models and Systems Engineering team, Mobile and Embedded Information Systems Laboratory, ENSIAS, Mohammed V Souissi

University, Rabat 10000, Morocco


Abstract: DSM (domain-specific modeling) offers many advantages over general purpose modeling, but this type of modeling is

effective just in narrow domains. The recent MDE (model driven engineering) approaches seek to provide a technology to compose

different specific domains in order to cover large scale systems. In this context, this article proposes a new approach for composing

specific domain models. First, we analyze some related works. On the basis of the key findings and conclusions drawn from the

analysis, we propose a multidimensional approach based on the composition of crosscutting concerns contained in the source domain

models. The approach is illustrated by a composition of service domains.

Key words: Composition, specific domain, separation of concerns, large scale systems, modularization.

1. Introduction

System development process becomes more

complex and time-consuming, because systems are no

more restricted to one specific domain, but describe a

set of related domains. So, it is impracticable to

describe the whole system by one model. To resolve

this problem, designer must combine models coming

from different specific domains. This, however, leads

to resolve model composition issues, which was and

remains one of the mean problem of the model driven

engineering.

In this paper, we are interested in domain model

composition as a practical solution to get the

maximum benefit from domains that have been

already developed and tested. This field of research

was intensified through a wide range of research in

order to extend the use of specific domains (e.g., Ref.

[1-3]). Indeed, the trend is to build libraries of

subdomains meta models that are reusable. These

basic domains (like interconnection of modules,

specialization, finite state machines, Petri networks,

Corresponding author: Asmaa Baya, Ph.D., student,

research fields: MDE, specific domains. E-mail: [email protected].

etc.) are used in the construction of most domains.

Thus, modeler can construct new domains just by

extending and defining rules of composition between

these basic subdomains. This basis arises the need to

define a new domain composition approach which

provides clear and practical mechanisms to extend and

compose domains.

In this paper, we treat domain composition issues.

The paper is structured as follows: Section 2 delimits

the context of work; Section 3 presents some model

composition methods; Section 4 presents some

examples of composition approaches; Section 5

presents our approach to compose specific domains.

We begin this section by presenting the process of the

approach, and then present its different phases. To

illustrate the approach, we present in the last section a

case study when we compose riche service and service

registry domains.

2. Context of Work

Model composition is a very wide field of research.

Especially, if we look for generative solution fit all

contexts, the number of constraints to consider

becomes infinite. That is why we must delimit our

http://www.google.co.ma/url?sa=t&rct=j&q=&esrc=s&source=web&cd=6&cad=rja&ved=0CE0QFjAF&url=http%3A%2F%2Fwww-adele.imag.fr%2Fpublications%2F86&ei=xMNZUMbmDYSn0QWksID4Dw&usg=AFQjCNEwDCJl2LilU4er5oHrmRyRMNagyg&sig2=R3iDq1DUBvYy7NYQZtmmbg


845

context of work; we are interested in two specific

aspects of model composition: specific domain

models and separation of concerns.

Specific domain models: DSM (domain-specific

modeling) helps to reduce the complexity introduced

by the specification of general-purpose modeling

languages, since their concepts are aligned with the

problem domain. Expressive power is gained from

using such notations. DSLs (domain specific

languages) provide graphical representations that

allow describing a system by constructing a visual

model using the terminology and concepts from a

specific domain. Analysis can then be performed on

the model. So, we can make profit of this benefit and

achieve a big part of domain model composition at a

graphical level.

Separation of concerns: Although the importance of

specifying concerns is obvious, scattering them

throughout the model decreases their utility. As

models in large scale systems grow in size and

complexity, it is unmanageable to view the contents of

a model in its entirety. So reusing, modifying and

removing concerns becomes a heavy task. The

apparition of the dominant decomposition according

to only one kind of concern at a time do not resolve

the problem, because many kinds of concerns that do

not align with that modularization end up scattered

across many modules and tangled with one another.

As a result, models with crosscutting concerns are

difficult to compose (highly coupled). That is why we

propose an advanced type of separation

“multidimensional separation of concerns” [4] that

will greatly simplify the modular composition of

models.

In this work, we try to bring the benefits of specific

domains (high level of expressiveness) and separation

of concerns (modularization) to the composition of

models. Recent research works [5] explore this new

orientation under the name of AODSM

(aspect-oriented domain-specific modeling). This new

field of research represents the union of AOP

(aspect-oriented programming) [6] and MIC

(model-integrated computing) [7].

3. Model Composition Methods

In this section, we present some well-known

composition methods in MDE (model driven

engineering): merging, weaving and transforming.

These methods will guide our reflection to propose a

generic approach for specific domains composition.

3.1 Merging

Merging is the action of combining two models,

such that their common elements are included only

once and the other ones are preserved [8], as shown in

Fig. 1. This mechanism is used when elements defined

in different source models have the same name and

represent the same concept. It is most often used to

provide different definitions to the same concept.

Model merging can be decomposed into four

phases [9, 10]: comparison, conformance checking,

merging and reconciliation.

Comparison: Called also “matching” or “mapping”.

In this phase, correspondences between similar

(equivalent) elements of the source models are

identified. Usually, we assume tha entities are the

Fig. 1 Principle of merging.


846

same when they have the same name or id, or when

they support the use of an explicit ontology, or a

thesaurus.

Conformance checking: The matching elements

identified in the previous phase are checked in this

phase. The purpose of this phase is to identify

potential conflicts that would render merging

infeasible.

Merging: The output of merging phase is a model

that retains all non-duplicated information in source

models; it collapses information that match model

declares redundant.

Reconciliation and restructuring: The target model

of merging phase may contain inconsistencies that

need fixing like violation of meta-model-specific

constraint. So, the purpose of this phase is to provide a

consistent merged model.

The main approach of model composition that uses

the merging method is “package merge” [11]. This

approach proposes a directed relationship between

two packages that indicates that the contents of the

two packages are to be combined. Other approaches

were developed in order to propose a more

deterministic process of merging [8, 10]. The

approach “Kompose” [12] is a merging approach that

proposes to merge models by comparing the

signatures of their element. The approach proposed in

Ref. [10] presents a specific algorithm that resolves

merge problem. This approach focuses on resolving

conflicts before and after merging.

3.2 Weaving

A weaving involves two actors: an aspect and a

base model. The aspect is made of two parts, a “point

cut”, which is the pattern to match in the base model,

and an “advice”, which represents the modification

made to the base model during the weaving. The parts

of the base model that match the point cut are called

“join points” (join points are, for example, methods or

variable assignments). During the weaving, each join

point is replaced by the advice [8] (Fig. 2).

Fig. 2 Principle of weaving.

Models and aspects can be woven in two ways [13]:

static and dynamic.

Static weaving: Modify source code of model

entities in order to insert aspect-specific statements at

join points. The mean advantage of this type of

weaving is the optimization of code. However, the

aspect-specific statements introduced become difficult

to identify later.

Dynamic weaving: This type of weaving solves the

problem of aspect identification in woven model. In

fact, it allows the explicit identification of aspects

both at weave-time and at runtime. Therefore, aspects

can be added, removed or updated during runtime.

A lot of approaches (e.g., Ref. [13-15]) adopt

weaving method in order to solve composition issues.

For example, the work in Ref. [14] proposes a

weaving process that exhibit composition properties to

allow multiple aspect weavings. Weaving model is

proposed in Ref. [15]. This model captures the

fine-grained relationships between elements of models

to compose. Other works propose an implemented

solution to weaving process. For example, the

approach “PROSE” (programmable extensions of

services) [16], offers a platform that allows aspects to

be woven, unwoven, and replaced at runtime.


847

3.3 Transforming

“Transformation is the automatic generation of a

target model from a source model, according to a

transformation definition” [17].

A transformation definition is a set of

transformation rules (Fig. 3) that together describe

how a model in the source language can be

transformed into a model in the target language.

A transformation rule is a description of how one or

more constructs in the source language can be

transformed into one or more constructs in the target

language.

Composition can be seen as a particular case of

transformation, although a lot of works [18] studied

the relation between model composition and model

transformation. Transformation can also be a step in a

compositional process and can be supported by other

methods like weaving and merging. In this case,

transforming step ensures automatic generation of

target model, unlike weaving and matching that need

human intervention.

Model transformation has been used in many

compositional approaches [19]. AMW [20] is a model

composition framework that uses model weaving and

Fig. 3 Principle of transforming.

model transformations to produce and execute

composition operations. In fact, a weaving model

captures the links between the input model elements,

and then this weaving model is used to automatically

generate a transformation. So, the transformation

takes two input models and produces the composed

model as output. EML (epsilon merging language) [21]

is also a meta model agnostic language for expressing

model compositions. It includes a model comparison

and model transformation steps.

4. Composition Approaches for Specific

Domains: The State of Art

In this section, we analyze some current

composition approaches for specific domains, as we

draw up a summary of the criteria that guide our

contribution.

4.1 Current Domain Composition Approaches

Several research studies focused on the problem of

composition for specific domains. This contributed to

the emergence of several orientations. In what follows,

each paragraph presents a specific orientation.

Conceptual composition: The approach of Vega [1]

proposes to compose domains at the conceptual level

rather than components infrastructure, an additional

level is inserted between the two levels in order to

ensure their synchronization. The first step of this

approach is to materialize the concepts of the

composite domain by classes of meta model. The

meta model of the composite domain becomes an

extension of the source domain meta models.

Transformation to pivot language: “Multi-modeling

views” approach offers a solution to the composition

of heterogeneous models (called high-level models)

by transforming both high-level models into low level

models that conform to the same existing meta model,

or conform to an extension of it [2]. A correspondence

model (CM high-level) is used to align high-level

models by describing the relationships of

correspondence between the composing elements.

Then high-level models specified in different domain


848

specific languages submit a sequence of

transformation steps in order to translate them into a

common low level language. A CM high-level needs

also to be propagated through the complete

transformation chains in order to automatically derive

CM low-level (correspondence model of low level

models). Finally, this approach proposes to make a

homogeneous model composition.

Composition by adopting template: “Template

instantiation” is a meta model composition approach

[3]. This approach is based on the reuse of common

meta modeling patterns in a single composite meta

model. Those patterns are saved as abstract meta

model templates and will be instantiated in

domain-specific meta models. This approach of

composition does not bring any changes to the source

meta models; however, it automatically creates new

relationships between the pre-existing entity types in a

target meta model to make them play roles in a

common meta modeling pattern.

Extending languages: Many approaches propose to

extend language in order to support domain

composition [22, 23]. For example, the approach of

Ledeczi [23] proposes to extend UML (unified

modeling language) with new operators in order to

combine source meta models. Another approach

proposes to extend the UML meta model with

behavior describing symmetric, signature-based

composition of UML model elements [22].

Coordination: The composition by coordination

was adopted in Ref. [24]. This approach proposes

architecture of coordination where every participant

preserves its autonomy, and the coordinator organizes

the collaboration of all participants. Another work

proposes to compose an inter domain specific

languages coordination [25]. This work introduces an

inter-model traceability and navigation environment

based on the complementary use of mega-modeling

and model weaving.

Graphical composition: Some works solve the

problem of composition at the graphical level. In

Ref. [26], researchers define a layer for graphical

syntax composition. This work provides formally

defined operators to specify what happens to graphical

mappings when their respective meta models are

composed.

4.2 Review of Domain Composition Approaches

The proposed approaches are non-exhaustive. MDE

contains a lot more approaches. However, we choose

these methods because each one illustrates a particular

orientation. The approaches above have several

advantages and disadvantages that we present in what

follows.

In Vega’s approach, the conceptual stage before the

composition of meta models aims at conceptualizing

the concepts of composite domain. Thus, the

composition is expressed at a high level abstraction.

The disadvantage is that the approach remains

complex as long as this high level abstraction is not

supported by simple and practical methods.

The approach “Multi-modeling views composition”

is based on a simple principle which is the

transformation of a high-level language to a low-level

language. However, it follows a long process and goes

through several transformations and intermediate

models before getting the final model. This increases

the complexity of this approach and its margin of

error.

“Template instantiation” approach ensures a

high-level abstraction through the use of abstract meta

model templates. However, these templates have

several limitations such as validity, adaptation and

instantiation, which greatly minimize their context of

use.

Extending languages maximizes reuse and makes

profit of existing languages. However, this orientation

is not generative because of its limitation to specific

languages.

The composition by coordination presents a lot of

advantages like the independence of source domains

and a high level of abstraction. However, it proposes


849

usually many languages and techniques which are not

easy to use by domain experts.

A graphical composition is a limited orientation,

because it solves the problem of composition at the

syntactical level only.

Based on these results, we can conclude that

high-level abstraction approaches often provide a high

level reuse and a good operating range. However, this

advantage is often accompanied by a high level of

complexity. Thus, the challenge to overcome is to

ensure a high level of abstraction while avoiding both

introduction of complex mechanisms and limitation of

the context of use.

5. XCOMP approach

In this section, we present a compositional

approach for specific domain models. The challenge

to overcome is to present a generative approach that

resolves the main composition issues in large scale

systems.

In such context, applying one method is not enough

to resolve the whole composition problems that

designer can meet in a standard context. So we

propose a new approach that combines all presented

methods in previous section: merge, weave and

transforming. This combination allows benefiting

from advantages of these methods. In addition, these

methods are simple to use and familiar for designers.

This new approach is supported with additional

methods that simplify the manipulation of concerns

contained in domain models, as it allows also the

re-modularization on demand; because every designer

organizes the source models according to his point of

view.

5.1 Process of Composition

“XCOMP” (X-Composition) is a composition

approach for specific domains. This approach

proposes to compose the models of source domains.

The output is a new model called “model of

composition”. This model references source models or

its elements as it can contain some emergent

constraints of the composed domains.

To obtain the final model, XCOMP propose to

follow the process presented in Fig. 4.

This process proposes organizational and

transforming steps.

Organizational steps propose to reorganize models in

order to increase design flexibility and allow

obtaining modular models. However, transforming

steps allow the introduction of composition rules and

constraints. The rest of this section explains each step

in the XCOMP process.

5.2 Decomposition

This approach of composition is dedicated to

models that describe large scale systems. So the

models to compose are so diversified and complex

that the navigation through such models, and their

composition is extremely difficult and time

consuming. To compose large scale models, we

propose to reorganize them in order to obtain

simplified models that we can manipulate easily

without losing the necessary information needed for

composition.

For this, we thought to divide the domains to

compose into dimensions. This notion of dimensions

was introduced in the “multidimensional separation of

concerns” [4]. In this context, dimension is a set of

concerns gathered in order to form independent unit of

sense. To identify dimensions, designer can choice

one of the several proposed policies in the domain of

multidimensional separation of concerns, or can just

try to make intuitive separation. In spite of the

followed decomposition policy, defined dimensions

must be weakly coupled. This characterization implies

that changes to one dimension have a limited or no

effect on other dimensions, which allow having a big

flexibility while composing domain models.

Fig. 5 presents a proposition to define dimensions that

will simplify composition later. This figure presents

examples of concerns in every type of dimensions.


850

Fig. 4 XCOMP process.

Fig. 5 Decomposition policy.

The first dimension to define in a domain model is

a business dimension “functional dimension”. It

contains concerns that provide the main functionalities

of systems and deal directly with problem domains.

Defining this part of each systems allows to simplify

the accumulation of the principle functionalities in

composed domain model.

The rest of concerns will be organized in

“non-functional dimensions”. However, we propose to

separate them according to their dependence on

context changes or not. So, non-functional concerns

that will be parameterized depending on environment

of execution are gathered in “context dimensions”.

Once we have defined the dimensions, we

decompose the source domain models following these

dimensions. Each dimension must be represented by a

part of the model (called block). To keep the

consistency of the model, we must trace the

relationships between the blocks.

So, the resulted models from the step of

decomposition, named “model of composition”, will

conform to meta model shown in Fig. 6. Each model

is composed of blocks relied between them, and every

block contains a set of entities and relationships.

5.3 Merging

In this phase, dimensions that have been identified

in previous step are examined for conformance with

each other. The purpose of his phase is to identify

potential conflicts and repetitions in the model of

composition. Fig. 7 presents the process of this step.

This process is repeated for each dimension of the

first domain model.

The first activity of this phase is to compare one

dimension (presented by block) of the first source

model with dimensions of the second model. If this

dimension conforms to dimension in the second

model, we merge blocks that represent these

dimensions. To check conformance, we compare all

entities contained in each block, particularly, their

names, ids and attributes.

To merge blocks, we delete the block from one

model and keep it in the second model. Relations

between the deleted block and the rest of the blocks

will be inserted to the same block of the second, so

those relations will henceforth rely blocks from

different source models.

Fig. 6 Model of composition metamodel.


851

Fig. 7 Merging process.

If this dimension does not conform to any

dimension of the second model, then we move to the

second level which is checking dimension entities. So,

we compare each entity of the block that represents

this dimension with entities of the second model. If

entities conform totally, we merge entities by the

same way that we merge dimensions. If entities

conform partially, we merge entities by adding

attributes and relations that not exist in the first entity

to the second entity, then delete the first entity.

It should be noted that a good decomposition policy

will reduce the number of entities merging, because

repeated elements will appear in the same dimensions.

So, merging dimensions will be sufficient to eliminate

overlaps between source domain models.

5.4 Weaving

The main purpose of composing existing domains

is to obtain new domain that contains all

functionalities of source domains. However, resulting

domain will need additional concepts and

functionalities, directly linked to the new problem

domain, but do not exist in composed domains. To

add these emergent concepts, we propose to weave

concerns with model of composition.

Once we have defined concern to weave, all must

identify the block when the new concern will be

inserted, then identify the join points and point cuts.

5.5 Transforming

After the weaving phase, the model of composition

may contain inconsistencies that need fixing. In the

final step of the process, such inconsistencies are

removed and the model is polished to acquire its final

form.

The composed domain will reuse artifacts of source

domains. That is why source domains must remain

intact. All composition constraints and methods are

inserted in the model of composition. This model

ensures the collaboration of source models and the

introduction of emergent concepts. To achieve this,

we propose to represent new entities and relations

(that exist in resulting model and do not exist in

source models) in additional dimensions instead of

dispersing them throughout the model.

To define which transformations must be made to

obtain a consistent model of composition, we define

model of composition and source model matrices as

shown in Figs. 8 and 9.

Matrices are defined as follows:

(1) Each column and line represents a dimension in

a domain model;

(2) Bi is a block which represent a dimension Di;

(3) Ri,j is a relation between the block “i” and the

block “ j”.

In the same way, we define the matrix of model of

composition (Fig 10). The diagonal of model of

composition matrix contains blocks which represent

all dimensions contained in source models. The Ri, j

represent the relationships between blocks from the

same model. The part X and X’ of this matrix varies

according to composition mechanisms.

The transformations to make is to delete the part X

and X’ of the matrix. New entities inserted while the

composition must be gathered in new dimensions that

reference the existing dimensions. Relations of X and


852

Fig. 8 Matrix of the first domain model.

Fig. 9 Matrix of the second domain model.

Fig. 10 Matrix of the model of composition.

X’ must appear in new dimensions in order to keep the

consistence of the resulting model.

Our approach stop at this level, since all proposals

to homogenize blocks must specify the target

language. However, our goals are to propose a

generative approach applicable to any type of

language. Indeed, to adapt the composite domain

model, there are other alternatives such as keeping

models heterogeneous and proceeding directly to the

code generation, or applying one of the transformation

methods from language to another in order to unify

the language used in the model [27].

6. Case Study

To illustrate what we explain above, we present an

example of composition of rich service domain and

service registry domain. Fig. 11 shows the model of

rich service domain and Fig. 12 shows the model of

service registry domain.

The first step is to identify dimensions, blocks and

relations of domain models. The model of rich service

domain can be organized in five dimensions of

concerns: service (B11), infrastructure (B12), routing

(B13), communication (B14) and application (B15).

Each dimension is represented by a block. Fig. 13

shows the blocks and relations contained in the model

of rich service domain.

The model of service registry domain can be

organized in three dimensions: storage (B21), registry

(B22) and service definition (B23). Blocks and

relations of this model domain are shown in Fig. 14.

The second step is to apply the merging process

(Fig. 7). As we can see through the domain models,

we have no repeated dimensions. However, the entity

“service” appears in B11 and B23. So, we have to

merge the entities “service”. We remove this entity

from the block B11 and link the block B11 with the

entity service of the block B23.

To simplify, we suppose that we have no emergent

concepts. So, we go on the transforming step. We

begin by defining the matrices of source domain

models based on the separation of concerns made in

the decomposition step (Figs. 13 and 14).

The Fig. 15 shows the operation of composing rich

service domain model matrix and service registry

domain model matrix. The resulted matrix of this

composition is shown in Fig. 16.

The part X and X’ (presented in Fig. 10) contain in

this case one relation R1, 2. This relation represents


853

Fig. 11 Rich service domain model.

Fig. 12 Service registry domain model.


854

Fig. 13 Separation of concerns in rich service domain model.

Fig. 14 Separation of concerns in service registry domain model.

Fig. 15 Composition of rich service and service registry domain model matrices.


855

Fig. 16 Resulted matrix of composition.

the inheritance between “rich service” entity and

“service” entity of the block (B11). After removing

“service” entity, this inheritance became between

“rich service” entity of the block (B11) and “service”

entity of the block (B23). So, a new dimension must

be inserted. This new dimension must reference the

blocks (B11) and (B23).

7. Conclusions

The main motivation behind this proposal is to

promote the composition of crosscutting concerns,

and to allow the insertion of new concerns throughout

the domains life cycle. However, this approach uses

several theoretical basis that must be implemented in

order to have a practical approach.

To build this approach, we used the basic principle

of the multidimensional separation of concerns. This

multidimensional separation has several advantages.

The separation is done according to multiple concerns,

which deal with the problem of the dominant

decomposition according to only one kind of concern

at a time. In addition, it allows on-demand

remodularization. Indeed, it is possible to add or omit

concerns throughout the life cycle without having to

change the entire model.

This domain composition approach begins first with

a multidimensional decomposition of concerns, and

then merges dimensions and entities that overlap.

After that, new aspect specific to composed domain

must be weaved to resulting model. Finally, it

proposes to insert new dimensions in order to ensure

the consistence of the composed domain.

References

[1] G. Vega, Développement d'applications à grande echelle

par composition de méta-modèles, Ph.D. thesis,

University Joseph Fourier, 2005.

[2] A. Yie, R. Casallas, D. Deridder, D. Wagelaar, A

practical approach to multi-modeling views composition,

in: Proceedings of the 3rd International Workshop on

Multi-Paradigm Modeling, Denver, Colorado, USA,

2009.

[3] M. Emerson, J. Sztipanovits, Techniques for metamodel

composition, in: OOPSLA—6th Workshop on Domain

Specific Modeling, Portland, 2006.

[4] H. Ossher, P. Tarr, Multi-dimensional separation of

concerns in hyperspace, in: ECOOP’99 Corkshop on

Aspect-Oriented Programming, Lisbon, 1999.

[5] J. Gray, T. Bapty, S. Neema, J. Tuck, Handling

crosscutting constraints in domain-specific modeling,

Communications of the ACM 44 (20) 87-93.

[6] Annual Aspect-Oriented Software Development

http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/y/Yie:Andr=eacute=s.html

http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/c/Casallas:Rubby.html

http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/w/Wagelaar:Dennis.html


856

Conference Home Page, http://www.aosd.net.

[7] J. Sztipanovits, G. Karsai, Model-Integrated Computing,

IEEE Computer 30 (1997) 10-12.

[8] J. Marchand, B. Combemale, B. Baudry, A categorical

model of model merging and weaving, in: MiSe 2012—

4th International Workshop on Modeling in Software

Engineering, Zurich, Switzerland, 2012.

[9] D. Kolovos, R. Paige, F. Polack, Merging Models with

the Epsilon Merging Language (EML), in: MoDELS,

Genova, Italy, Oct. 1-6, 2006.

[10] R. Pottinger, P. Bernstein, Merging models based on

given correspondences, in: Proceedings of 29th

International Conference on Very Large Data Bases

(VLDB’03), Berlin, Germany, 2003.

[11] OMG 2003, UML 2.0 Superstructure Final Adopted

specification, Document-ptc/03- 08-02.

[12] R. France, F. Fleurey, R. Reddy, B. Baudry, S. Ghosh,

Providing support for model composition in meta models,

in: Proceedings of the 11th IEEE International Enterprise

Distributed Object Computing Conference, Washington,

DC, USA, 2007.

[13] K. Bollert, On weaving aspects, in: International

Workshop on Aspect-Oriented Programming, ECOOP99,

Portugal, 1999.

[14] J.M. Jézéquel, Modeling and aspect weaving, in: Methods

for Modeling Software Systems (MMOSS), Germany, 2007.

[15] M.D.D. Fabro, J. Bezivin, F. Jouault, E. Breton, G.

Gueltas, AMW: A generic model weaver, in: IDM-

Ingénierie des Modèles, 1ères Journées sur l'Ingénierie

Dirigée par les Modèles, Paris, 2005.

[16] A. Popovici, T. Gross, G. Alonso, Dynamic weaving for

aspect-oriented programming, in: Proceedings of the 1st

International Conference on Aspect-Oriented Software

Development, Enschede, The Netherlands, Apr., 2002.

[17] A. Kleppe, J. Warmer, W. Bast, MDA Explained, The

Model-Driven Architecture: Practice and Promise,

Addison Wesley Longman Publishing, Boston, 2003.

[18] B. Baudry, F. Fleurey, R. France, R. Reddy, Exploring

the relationship between model composition and model

transformation, in: Proceedings of Aspect Oriented

Modeling Workshop, in conjunction with MoDELS’05,

Montego Bay, Jamaica, 2005.

[19] J. Bezivin, S. Bouzitouna, M.D.D. Fabro, M.P. Gervais, F.

Jouault, D. Kolovos, et al., A canonical scheme for model

composition, in: Proceedings of the 2nd European

Conference on Model Driven Architecture—Foundations

and Applications, Bilbao, Spain, Jul. 10-13, 2006.

[20] Atlas Model Weaver Project Web Page, 2005,

http://www.eclipse.org/gmt/amw/.

[21] D.S. Kolovos, Epsilon Project Page,

http://www.cs.york.ac.uk/~dkolovos.

[22] F. Fleurey, R. Reddy, B. Baudry, S. Ghosh, Providing

support for model composition in meta models,

in: Proceedings of EDOC 2007, Annapolis, MD, USA,

2007.

[23] A. Ledeczi, G. Nordstrom, G. Karsai, P. Volgyesi, M.

Maroti, On metamodel composition, in: Proceedings of

the 2001 IEEE International Conference, Mexico, 2001.

[24] T.L. Anh, Fédération: une architecture logicielle pour la

construction d’applications dirigée par les modèles, Ph.D.

thesis, University Joseph Fourier, 2004.

[25] F. Jouault, B. Vanhooff, H. Bruneliere, G. Doux, Y.

Berbers, J. Bezivin, Inter-DSL coordination support by

combining megamodeling and model weaving, in:

Proceedings of the 2010 ACM Symposium on

Applied Computing, Suisse Sierre, 2010.

[26] L. Pedro, M. Risoldi, D. Buchs, B. Barroca, V. Amaral,

Composing visual syntax for domain specific languages,

in: 13th International Conference, HCI International 2009,

USA, 2009, pp. 889-898.

[27] M. Brambilla, P. Fraternali, M. Tisi, A transformation

framework to bridge domain specific languages to MDA,

in: Models in Software Engineering: Workshops and

Symposia at MODELS 2008, Toulouse, France, 2008.

http://www.aosd.net/

http://hal.inria.fr/index.php?action_todo=search&submit=1&s_type=advanced&search_without_file=YES&f_0=LASTNAME&p_0=is_exactly&f_1=FIRSTNAME&p_1=is_exactly&l_0=and&halsid=u34thfjo2epaec08uhl9q3eqa1&v_0=Marchand&v_1=Jonathan

http://hal.inria.fr/index.php?action_todo=search&submit=1&s_type=advanced&search_without_file=YES&f_0=LASTNAME&p_0=is_exactly&f_1=FIRSTNAME&p_1=is_exactly&l_0=and&halsid=u34thfjo2epaec08uhl9q3eqa1&v_0=Combemale&v_1=Benoit

http://hal.inria.fr/index.php?action_todo=search&submit=1&s_type=advanced&search_without_file=YES&f_0=LASTNAME&p_0=is_exactly&f_1=FIRSTNAME&p_1=is_exactly&l_0=and&halsid=u34thfjo2epaec08uhl9q3eqa1&v_0=Baudry&v_1=Benoit

http://www.informatik.uni-trier.de/~ley/pers/hd/p/Paige:Richard_F=.html

http://www.informatik.uni-trier.de/~ley/pers/hd/p/Polack:Fiona_A=_C=.html

http://www.informatik.uni-trier.de/~ley/db/conf/models/models2006.html#KolovosPP06

http://www.informatik.uni-trier.de/~ley/db/conf/models/models2006.html#KolovosPP06

http://www.cs.york.ac.uk/~dkolovos

http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/b/Brambilla:Marco.html

http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/t/Tisi:Massimo.html


Mobile Station Speed Estimation with Multi-bit Quantizer

in Adaptive Power Control

Hyeon-Cheol Lee

Satellite Technology Research Laboratory, Korea Aerospace Research Institute, Daejeon 305-806, Rep. of Korea


Abstract: The adaptive power control with multi-bit quantizer of CDMA (code division multiple access) systems for communications between multiple MSs (mobile stations) with a link-budget based SIR (signal-to-interference ratio) estimate is applied to four inner loop power control algorithms. The speed estimation performances of these algorithms with their consecutive TPC (transmit-power-control) ratios are compared to each inner loop power control algorithm, and the speed shows full linearity with the consecutive TPC ratio information of CS-CLPC (consecutive TPC ratio step-size closed loop power control), FSPC (fixed step-size power control), and KS-CLPC (Kalman gain step-size closed loop power control). These algorithms show more linearity with increased bit quantization. The result however, indicates consecutive TPC ratio of AS-CLPC (adaptive step-size closed loop power control) is independent of target speed. It is concluded that the speed can be estimated with the consecutive TPC ratio with CS-CLPC, FSPC and KS-CLPC. Key words: Speed estimation, adaptive power control, link-budget, SIR, multi-bit quantizer.

1. Introduction

The communication system of multiple MSs

(mobile stations) requires a mobile wireless network

to share data between MSs. One communication

network protocol that may be used is CDMA (code

division multiple access). CDMA differs from both

FDMA (frequency division multiple access) and

TDMA (time division multiple access) in that it uses

the same frequency for multiple users.

Since all users utilize a single frequency, the signal

from each MS may interfere with other MSs’

receivers [1]. This is referred to as the near-far effect.

To eliminate the near-far effect in CDMA systems, the

transmission signal power from every MS must be the

same as the signal power at the receiver. This

technique of controlling the magnitude of the

transmission power according to the distance between

the MS and the BS (base station) is officially termed

Corresponding author: Hyeon-Cheol Lee, Ph. D., principal researcher, research fields: wireless communications and SAR. E-mail: [email protected].

power control. It equalizes the received power and

eliminates the near-far effect, though it is subject to

such complications as path loss, shadowing,

multi-path fading, etc..

This power control technique is differentiated into

open loop power control and closed loop power

control. The closed loop power control is further

divided into inner loop power control and outer loop

power control. The inner loop power control is

responsible for adjusting the power transmitted to

maintain the received SIR (signal-to-interference ratio)

at the BS at a level equal to that at the SIRtarget. The

outer loop power control is responsible for setting the

SIRtarget based on the BER (bit error rate) or service

requirement.

Conventional SIR estimates [2, 3] consider only the

transmission power and the link-gain, but this paper

takes into account the link-budget, which has more

realistic parameters including distance information

than the link-gain. Using the SIR estimate that reflects

the link-budget, speed estimation [4] is introduced

DAVID PUBLISHING

D

Mobile Station Speed Estimation with Multi-bit Quantizer in Adaptive Power Control

858

based on a CTR (consecutive transmit-power-control

ratio). The proposed speed estimation method is

applied to four algorithms with one, two and three-bit

level quantizers, and the results are compared.

This paper is organized as follows: The literature

related to this work is surveyed in Section 2; the inner

loop power control is described in Section 3; the

concept of the link-budget based SIR estimate is

introduced in Section 4; followed by description of

simulation environments in Section 5; Section 6 gives

details of the speed estimation. The simulation results

are analyzed in Section 7; finally, conclusions are

drawn in Section 8.

2. Related Work

Kim et al. introduced in Ref. [5] the AS-CLPC

(adaptive step-size closed loop power control)

algorithm for a narrowband CDMA system. This

algorithm adapts its power control step-size based on

the optimal factors determined with the mean fade

duration which is inversely proportional to the

maximum Doppler frequency. Nourizadeh et al. [6]

proposed the BA-CLPC (blind adaptive closed loop

power control) in which the power control step-size is

adjusted to cope with the user mobility. Taaghol [7]

introduced the SA-CLPC (speed adaptive closed loop

power control) algorithm in which the power control

step-size is selected based on the user speed

estimation categorized by speed ranges. Lee and Cho

[8] proposed the M-ACLPC (mobility based adaptive

closed loop power control) algorithm in which the

power control step-size is adjusted depending on the

combination of the cumulative information of the

three power control commands and speed estimation.

Patachaianand and Sandrasegaran [3] compared

performances of the AS-CLPC, BA-CLPC, SA-CLPC,

and M-ACLPC in terms of PCE (power control error)

under the same simulation environment. In their

comparisons, the AS-CLPC showed the best

performance when the target speed was lower than 25

km/h, while the SA-CLPC was the best when the

speed was greater than 25 km/h.

Patachaianand and Sandrasegaran [3] presented the

CS-CLPC (CTR step-size closed loop power control)

algorithm whose power control step-size is

determined based on a parameter called CTR. They

measured the moving target speed by CTR, then,

calculated the PCE as the RMS (root mean square) of

the difference between the received SIR and the

SIRtarget. They also suggested in Ref. [4] the mapping

equation and mapping table which can yield accurate

speed estimation using CTR.

3. Inner Loop Power Control

In CDMA, the process of inner loop power control

occurs as follows: In the reverse link direction (from

the MS to the BS), the transmission power

information goes to the BS. At the BS, the SIRtarget

and the received SIR are calculated from the

transmission power, the link-gain and the noise power.

Based on these factors, the BS sends a TPC

(transmit-power-control) command to each MS at rate

of 1,500 Hz (= 0.667 ms) in the forward link direction

(from the BS to the MS). This power equalization

increases the maximum communication number

between MSs and consequently eliminates the near-far

effect. These procedures [2, 3] are represented in Eqs.

(1) and (3).

)()()()1( tiTPCtitiPtiP (1)

where Pi(t) is the transmission power, δi(t) is the

power control step-size, and TPCi(t) is the TPC

command for the ith MS at time t. )()()( ,arg tSIRtSIRtx iietti (2)

Where, SIRtarget,i(t) (= SIRtarget(t) here) is the target SIR

and SIRi(t) is the received SIR for the ith MS at time t.

Non-uniform quantizer [9] used in voice coding is

introduced that (m + 1)-bit TPC is adopted TPCi(t) =

C0 C1···Cm where C0 is the sign bit.

)12()1(

)/))(1log((

))(()(

mi

ii

Rptx

txsigntTPC

(3)

where δ = 2(m+1) – 1 and RP is dynamic range of power

adjustment in Table 1.


859

Table 1 TPC coefficients.

M D multilevel RP

1 0.5 ±1,2 0.1

2 0.25 ±1,2,3,4 0.3

3 0.125 ±1, 2, 3, 4, 5, 6, 7, 8 0.5

4. Link-Budget Based SIR

The conventional SIR estimate of the ith MS in

CDMA is described as follows:

ij

Njij

iii

PtGtP

tGtPtSIR

)()(

)()()(

(4)

where Gi(t) is the link-gain between the ith MS and

the connected BS, and Pi(t) is the transmission power

from the ith MS. Gji(t) is the link-gain between the jth

MS and the BS to which the ith MS connects.

This paper introduces the link-budget based SIR as

ijNjiR

iRi PtP

tPtSIR

)(

)()(

,

, (5)

where PR,i is the received power of the ith MS. PR,ji is

the received power of the jth MS with the BS to which

the ith MS connects. The received power is affected

by factors including the free space loss [10] which has

distance information and gaseous path loss [11] varied

by humidity. The speed is estimated from moving

distance per sample.

The power delivered to the receiver [10] is: ))()(/()( ,,,, iGiFiRiTiTiR DLDLGPGP (6)

where GT,i, PT,I, GR,i, LF(Di) and LG(Di) are the

transmission antenna gain, the transmission power, the

received antenna gain of the ith MS, the free space

loss, and the gaseous path loss, respectively (the

component loss is ignored here.). )(log20)(log204.92))(( 1010 iiF DFdBDL (7)

Di is the distance between the ith MS and the BS in

kilometers, and F is the frequency in gigahertz. The

specific attenuation due to dry air and water vapor

from sea level to an altitude of 5 km can be estimated

by Eq. (8).

iWOiG DDL )()( (8)

P.676-5 of Ref. [11] shows equations of γo

(attenuation for dry air) and γw (attenuation for water

vapor). These attenuations are dependent on σ, the

water vapor density (g/m3) specified in Table 2 from

P.836-3 of Ref. [11]. The bigger the σ is, the larger the

attenuation is.

The noise power [10] is: BTKPN (9)

where K is the Boltzmann constant (1.38×10-23 J/K), T

is the temperature in Kelvin, and B is the equivalent

bandwidth in hertz.

5. Simulation Environments

This section presents a simulation of the speed

estimation using Eq. (5). The frequency, the

temperature, the pressure, the water vapor density and

the bandwidth are set to 2.0 GHz, 288 K, 1013 hPa,

20 (Summer in Coast), and 5 MHz, respectively. In

Fig. 1, five MSs are arranged and Dis are set to 250 m,

500 m, 750 m, 1,000 m and 1,250 m. The antenna

gain of each MS is set to 0 dB, as is the antenna gain

of the BS.

MS1 to MS5 complete their power control by FSPC

so that each transmission power shown in Table 3 is

different. Then, MS1, which is 1,250 m away from the

BS, starts to move outward. It moves at five different

speeds and measures the CTR at each speed. SIRtarget

is set to the transmission power of MS3 on this

simulation.

Table 2 Water vapor density at different seasons and regions.

Jan. April July Oct.

Coast(edge of continent) 5 10 20 10

Inland 5 5 10 5

Ocean 20 20 20 20

Fig. 1 Simulation formation.


860

Table 3 Initial condition of MS1 to MS5.

PT(dB) GT(dB) GR(dB) Di(m)

MS1 +8.048 0.0 0.0 1250

MS2 +4.1314 0.0 0.0 1000

MS3 +0.0000 0.0 0.0 750

MS4 -5.229 0.0 0.0 500

MS5 -13.979 0.0 0.0 250

Table 4 MS1 moving distance for 28.0 s (42,000 × 0.667 ms).

Speed (km/h) Speed (m/s) Moving distance (m)

100 27.7778 777.78

200 55.5556 1555.56

300 83.3333 2333.33

400 111.111 3111.11

500 138.8889 3888.89

600 166.6667 4666.67

6. Speed Estimation

There are several algorithms addressing inner loop

power control, including CS-CLPC [3], AS-CLPC [5],

FSPC and Kalman gain Step-size Closed Loop Power

Control, etc..

This section investigates changes in transmission

power for the above four algorithms with the

link-budget based SIR. MS1 moves outward for

42,000 × 0.667 ms (the number of the sample =

42,000) at the six different speeds listed in Table 4;

100 km/h, 200 km/h, 300 km/h, 400 km/h, 500 km/h,

and 600 km/h. As MS1 moves away, the four inner

loop power control algorithms alter the transmission

power to compensate for the distance between BS

and MS1. Eq. (10) [3] measures the CTR as

follows:

t

mtnm

nTPCnTPCdtCTR

1

)1()()( (10)

where d is a scale factor (see Table 1), m = t if t < w,

and m = w if t ≥ w. w is the maximum size of the

window average.

6.1 CS-CLPC

Patachaianand et al. [3] introduced the CS-CLPC

algorithm, where the step-size is adjusted as shown in

Eq. (11).

),(min1)(

maxCTRtCTRt

(11)

where α, β and CTRmax are constants.

6.2 AS-CLPC

Kim et al. [5] suggested the AS-CLPC algorithm.

This algorithm adapts its step-size based on TPC

history. The step-size is given by Eq. (12).

..,/)(

)1()(,)()(

WOLt

tTPCtTPCKtt

(12)

where K and L are positive real constants with ranges

of 1 < K and 1 < L < 2.

6.3 FSPC

In this simulation, the algorithm uses a fixed

step-size.

6.4 KS-CLPC

The well-known Kalman algorithm [12] for

adaptive step-size is used in this simulation.

)()()1()(

)()1()(

tRtHtPtH

tHtPt

Te

Te

K

(13)

)1())()(()( tPtHtItP eKe (14)

where δK(t) is a Kalman gain vector, H(t) is the an

observation matrix, and Pe(t) is an error covariance

matrix at time of t.

7. Simulation Results

Six different speeds are measured with CTR, and

the relationships are shown in Fig. 2 to Fig. 7

according to two window sizes and three quantizers.

The CS-CLPC, FSPC and KS-CLPC algorithms

show a linear relationship between speed and CTR,

in addition they show more linearity and can

estimate higher speed as quantizing bits are increased.

Therefore, the target speed can be measured

by mapping the CTR with the 3 algorithms. But, the

AS-CLPC algorithm deviates from linearity.

All the results are same with different window

sizes.


861

Fig. 2 CTR vs. MS1 speed at 1-bit Quantizer, window size 3 (Summer, Coast).







862

8. Conclusions

This paper introduced a speed estimation of four

different inner loop power control with consecutive

TPC ratios which use the link-budget based SIR in the

CDMA communication systems between MSs. It was

concluded that linear relationship exists between

speed and Consecutive TPC Ratios, and that MS

speed can be estimated using the Consecutive TPC

Ratios of the CS-CLPC, FSPC and KS-CLPC

algorithms.

References

[1] J.P. Romero, O. Sallent, R. Agusti, M.A.D. Guerra, Radio Resource Management Strategies in UMTS, John Wiley & Sons, New York, 2005.

[2] R. Patachaian, K. Sandrasegaran, Performance Comparison of Adaptive Power Control in UMTS, in: Int. Conf. Wireless Broadband and Ultra Wideband Commun. (AusWireless 2007), Sydney, Australia, 2007.

[3] R. Patachaian, K. Sandrasegaran, Consecutive transmit power control ratio aided adaptive power control for UMTS, Electronics Letters 43 (2007) 55-56.

[4] R. Patachaian, K. Sandrasegaran, User speed estimation

techniques for UMTS, Electronics Letters 43 (2007)

1036-1037.

[5] J.H. Kim, S.J. Lee, Y.W. Kim, Performance of single-bit

adaptive step-size closed-loop power control scheme in

DS-CDMA system, IEICE Trans. Commun. E81-B (1998)

1548-1552.

[6] S. Nourizadeh, P. Taaghol, R. Tafazolli, A Novel

Closed-loop Power Control for UMTS, in: Int. Conf. 3G

Mobile Commun. Technologies (3G 2000), London,

England, 2000.

[7] P. Taaghol, Speed-adaptive power control for CDMA

systems, Bechtel Telecommunications Technical Journal,

2 (2004).

[8] H. Lee, D.H. Cho, A new user mobility based adaptive

power control in CDMA systems, IEICE Trans. Commun.

E86-B (2003) 1702-1705.

[9] W. Li, V.K. Dubey, C.L. Law, A new generic multistep

power control algorithm for the LEO satellite channel

with high dynamics, IEEE Commun. Lett. 5 (2001)

399-401.

[10] D. Roddy, Satellite Communications, Prentice Hall, New

Jersey, 1989.

[11] ITU Recommendations 2004, 2004.

[12] R.G. Brown, P.Y.C. Hwang, Introduction to Random

Signals and Applied Kalman Filtering, John Wiley &

Sons, New York, 1997.


On the Comparison Analysis of Two 4G-WiMAX Base

Stations in an Urban Sub-Saharan African Environment

Eric Tutu Tchao1, Kwasi Diawuo1 and Willie Ofosu2

1. Department of Electrical/Electronic Engineering, Kwame Nkrumah University of Science and Technology, Kumasi, Ghana

2. Department of Electrical Engineering Technology, Penn State Wilkes-Barre, USA

Received: June 04, 2013 / Accepted: June 15, 2013 / Published: June 30, 2013.

Abstract: A growth in the demand for WBA (wireless broadband access) technology has been seen in Ghana in the last few years.. The reason for this growth can be attributed to the emergence of the use of multimedia applications, demands for ubiquitous

high-speed Internet connectivity, the massive growth in the wireless and mobile communications sector and the deregulation in the

telecommunications industry. WiMAX, which is a WBA technology, is currently being deployed in the 2,500-2,690 MHz band to help serve the ever increasing needs of broadband internet subscribers in the country. This paper presents simulation results and field trial measurements for two BS (base stations) in a newly deployed 4G-WiMAX network in a typical dense urban Sub-Saharan African environment. We have used a model which uses the interference to noise ratio (n) parameter to obtain the coverage range of these two BS under evaluation. The simulated cell range for site 1 is 4.90 km and that for site 2 is 3.19 km. The final coverage simulation and the field measurements results have also been presented. Key words: Wireless broadband access, WiMAX, field measurements, interference modeling, Sub-Saharan African environment.

1. Introduction

WiMAX is a wireless digital communications

system which is intended for wireless metropolitan

area networks. WiMAX can provide BWA

(broadband wireless access) up to 30 miles (50 km)

for fixed stations, and 3-10 miles (5-15 km) for

mobile stations [1]. WiMAX allows for more efficient

bandwidth use and is intended to offer higher data

rates over longer distances [2]. Because WiMAX

operates in both licensed and non-licensed frequency

bands, it provides a viable economic model for

wireless carriers which make it a BWA technology of

choice for deployment in many developing

countries [3].

Deployment of WiMAX networks has currently

started in many sub-Saharan African countries. In

Ghana, the 2,500-2,690 MHz band, sometimes

Corresponding author: Eric Tchao, Ph.D., student, research

fields: wireless networks. E-mail: [email protected].

referred to as 2.6 GHz band, which is one of the

various bands defined by the ITU (International

Telecommunication Union), has been auctioned for

WiMAX deployment [4].

Pilot deployments are well underway in several

parts of Accra and Tema. The first successful pilot

deployment which covers 55 km2 in the urban centers

of Accra and Tema has about 535 fixed and mobile

CPE (customer premise equipment). Eleven WiMAX

base stations have been used to provide coverage to

the CPEs in the network using an adaptive 4T4R (four

transmit four receive) MIMO (multi input multi output)

antenna configuration. The final distribution of the

antenna sites in the deployment area is shown in

Fig. 1.

In order to deploy this high grade pilot WiMAX

network, engineering tools and techniques that

allowed rapid system design were used to achieve the

main objective of planning the new network to give

ubiquitous coverage to the user terminals in the

On the Comparison Analysis of Two 4G-WiMAX Base Stations in an Urban Sub-Saharan African Environment

864

Fig. 1 Final distribution of base stations in pilot network.

service area. This paper explores the techniques and

the accuracy of network plan which was used to

successful deploy two WiMAX base stations.

These techniques which we will discuss in this

paper were later extended to the deployment of the

extensive and high grade pilot WiMAX network.

2. Propagation Environment and Interference Modeling

In order to estimate signal parameters accurately in

wireless communication systems, it is necessary to

estimate signal propagation characteristics in different

terrain environments, such as free space, urban,

suburb, country and indoor. To some extent, the

communication quality is influenced mainly by

applied terrain environment [5]. Propagation analysis

provides a good initial estimate of the signal

characteristics.

There are two general types of propagation

modeling: site-specific and site general. Site-specific

modeling requires detailed information on building

layout, furniture, and transceiver locations. It is

performed using ray-tracing methods. For large-scale

static environments, this approach may be viable. For

most Sub-Saharan environments however, the

knowledge of the building layout and materials is

limited and the environment itself can change. Thus,

the site-specific technique is not commonly employed.

Site general models provide gross statistical

predictions of path loss for link design and are useful

tools for performing initial design and layout of

wireless systems especially under Sub-Saharan

African conditions.

The Hata-Okumura model which is best suited for

large cell coverage can be used to model the

propagation environment for this WiMAX network.

Because of its simplicity, it is a widely used model for

most of the signal strength predictions in

macro-cellular environment [6, 7], even though its

frequency band is outside the band of WiMAX.

In order to estimate the coverage range of the two

base stations, mapping of the CPEs which will be

served by the two bases stations was done and the

distribution is as shown in Fig. 2.

The propagation environment can be modeled using

the distribution of CPEs as shown in Fig. 2. It consists

of a fixed station F00 (site 1) and a mobile station M00

attempting to establish a radio link in the presence of

n additional fixed stations niFi 1 . Each fixed

station communicates with additional mobile stations.

Mi is the total number of mobile stations

communicating with the CPE Mij at a power Fij on the

downlink channel. Different propagation models are


865

Fig. 2 Distribution of CPE covered by two base stations.

required for different environments. The simplest

model is the single slope path loss model [8]:

oo

tr ddd

dPP

,

(1)

where Pr is the power received at a distance d (relative

to the reference distance) from a transmitter radiating

at a power Pt. The parameter γ is the path loss

exponent (in free space) and κ is the free space path

loss between the transmission antenna and the

reference distance do: 2

r 4

ot dGG

(2)

where Gt and Gr are the antenna gains of the

transmitter and receiver respectively and λ is the

wavelength of the transmission. The single slope path

loss model is used to describe the mean path loss in

large area environments [9, 10]. Fig. 2 however shows

propagation at close ranges in an urban environment.

Propagation at close ranges however behaves more

like the plane earth model and a dual slope path loss

model is more appropriate [9, 11]:

(3)

where b is the breakpoint distance, γ1 is the path loss

exponent before the breakpoint and γ2 is the path loss

exponent after the breakpoint. The breakpoint is

related to the height above plane earth of the

transmitter antenna ht and receiver antenna hr and is

approximately given by [11]:

rthh

b4

(4)

Eq. (4) gives one theoretical expression for the

breakpoint in the plane earth model, however, the

breakpoint is not well defined due to the oscillatory

nature of the signal envelope in the plane earth

model, and different definitions of where the

breakpoint occurs give slightly different expressions

[12, 13]. Over a region of tens of wavelengths, a

received signal will exhibit variation about the mean

power predicted by the path loss models of Eqs. (1)

and (3). Measurements have consistently indicated

these power variations exhibit lognormal statistics [8,

14]. This phenomenon is called lognormal

shadowing and can be incorporated into either path

loss model as a multiplicative factor to the path loss

PL:

1010

Lr PP (5)

where ζ is a normally distributed dB variable with

dbb

d

d

dkP

bddd

dkP

P

ot

oo

t

r21

1


866

zero mean, and a standard deviation σ typically

between 6 and 12 dB in macrocell systems [8]. From

Eq. (5) it can be seen that shadowing models both

signal attenuation due to obstructions and signal

amplification due to waveguide effects. At a receiver,

a link will be considered successful if the signal to

noise plus interference ratio 1N

S is greater than

or equal to the system protection ratio Z, otherwise an

outage is deemed to occur. The region in which this

threshold is maintained is the region in which radio

communication is considered successful and is called

the “cell”. The extent of the cell is thus a function of

the radio signal and interference statistics [15].

Examining Fig. 1, assuming M00 is communicating

with Fo in the presence of a single interferer Fi which

spills a power Pu into the wanted uplink and Pd into

the wanted downlink. A spatial analysis of link outage

in the presence of an interferer but in the absence of

receiver noise was presented by Cook [16] using the

single slope path loss model. When the effect of

receiver noise is incorporated (assuming that the same

protection ratio applies to noise and interference),

M00’s uplink outage contour is a family of circles

centred on the fixed station F0, but the downlink

outage contour is a higher plane curve [17]. By

introducing a parameter called the interference to

noise ratio (n) the equations for both outage contours

can be written in a simple form. n is the total

interference power at a receiver divided by the

receiver noise power. For a single interferer under the

single slope path loss model, n at the fixed station F0

is given by [17]:

N

rPn iu

u

0, (6)

whiles the DL n at the CPE Moo is given by:

N

rPn id

d

0, (7)

Using this parameter, it can be shown that the

equation for the uplink outage contour can be written

[17]:

1

1

1,,uu

uooiuuooo nn

nrKr (8)

and the DL outage contour equation can be written:

1

1

1,,dd

dooiddooo nn

nrKr (9)

where ZNKPt and the parameter is given by

[14]:

tt

sirtt

WZP

LWGGPK (10)

where Ls is a system loss factor, Pi is the interference

power, and Wi and Wt are the bandwidths of the

interfering and wanted signals respectively.

As the receiver S = [N + I] threshold must be

exceeded on both the uplink and downlink in order for

the duplex link to be successful, the range of the

mobile terminal from in any direction is the minimum

of ru and rd. Thus, the range described by Eq. (8)

represents the cell radius regardless of the downlink

conditions.

3. Coverage Simulation

The coverage prediction for the deployed network

is based on the simulation parameters in Table 1 and a

realistic distribution of BS and CPEs in the network.

Genex-Unet has been used to simulate the average

throughput per sector and the final radio network plan.

The CPE antenna configuration used for the

coverage and capacity simulation was 1T2R.

The capacity simulation results in Table 2 and Fig.

3 show better performance by 4 × 4 Antenna

configuration. The minimum simulated cell edge

uplink rate using a DL/UL ratio of 35:12 which was

used for the final network plan shown in Fig. 4 is 256

Kbps. This value was obtained at 4.90 km and 3.19

km for sites 1 and 2 respectively as shown in Fig. 5.

4. Measurement

The measurements were done at several locations

within the network in Accra which are dense urban

and urban environments. The selection of these


867

Table 1 Simulation parameters.

Parameter Site 1 Site 2

Resource frequency 2.5-2.53 GHz

Channel bandwidth 10 MHz

Average users per sector 10

Fast Fourier Transform (FFT) Size 1,024

Subcarrier spacing 10.93 kHz

Useful symbol time 91.4 μs

Guard time 11.4 μs

OFDMA symbol time 102.8 μs

Modulation QPSK, 16-QAM, 64-QAM

Antenna frequency range 2.3-2.7GHz

VSWR ≤ 1.5

Input impedance 50 Ω

Gain 18 dBi

Horizontal beamwidth (3 dB) 60°

Vertical beamwidth (3 dB) 7°

Electrical downtilt 2°

CPE antenna configuration 1T2R

Maximum power (dBm) 43 43

Antenna height 42 m 38 m

Table 2 Capacity simulation results.

Permutation TDD split ratio

WiMAX carrier average throughput per sector

Minimum simulated cell edge throughput per sector

4T4R adaptive MIMO(Mbps) 4T4R adaptive MIMO

DL UL DL (Mbps) UL (Kbps)

PUSC with all SC 1 × 3 × 3 26:21 11.23 5.18 1.21 104

PUSC with all SC 1 × 3 × 3 29:18 13.00 4.32 1.52 108

PUSC with all SC 1 × 3 × 3 31:15 14.18 3.46 1.76 128

PUSC with all SC 1 × 3 × 3 35:12 16.55 2.59 2.24 256

Fig. 3 SINR simulation for the MIMO antenna configurations.


868

Fig. 4 Final radio simulation of the 2 BS.

Fig. 5 Coverage range of the 2 BS.

locations was based on various criteria, where distance,

elevation and line of sight capabilities were the main

factors. The field trial measurement setup comprised:

(1) GPS antenna;

(2) RF cable;

(3) Dongle XCAL-X;

(4) Laptop with XCAL-X software;

(5) WiMAX PCMCIA CARD.

The measurements were divided into physical

measurements and throughput measurements. The

physical measurements collected RSSI (received

signal strength indication) values in about 10,260 and

7,210 locations in sites 1 and 2, respectively. The

overview of the measurement areas around the base

station and the summary of the measured RSSI values

have been summarized in Figs. (6), (7), (8), and (9).

Throughput measurements were performed by

downloading and later uploading FTP file size of 10

MB form and to a remote server as the subscriber

moves away from the base station until the connection

was dropped. The results of the throughput

measurement have been discussed in the next section.

5. Discussion of Results

From the RSSI measurements taken from the

102,060 locations in sites 1, about 86.55% of the

measured RSSI values were greater than –80 dbm.

From the site 2 measurements taken from 7,210


869

Fig. 6 Summary of RSSI results for site 1.

Fig. 7 RSSI measurement areas for site 1.

Fig. 8 RSSI summary results for site 2.


870

Fig. 9 RSSI measurement areas for site 2.

locations within the cell, 79.39% of the measured

RSSI values were greater than –80 dbm. For site 1,

99.45% of the measured values obtained from up to a

of about 5.30 km away from the base station were

greater than or equal to the simulated cell edge RSSI

value of –90 dBm. For site 2 on the other hand,

95.81% of the measured RSSI values obtained from

up to about 3.5 km away from the base station were

greater than or equal to the simulated cell edge RSSI

value. The results of the throughput measurements

have been summarized in Table 3 and Fig. 10.

The maximum measured throughputs for sites 1 and

2 are 8.62 Mbps and 7.20 Mbps, respectively.

Comparing the field results with the Federal

Communications Commission’s broadband

applications support speed guide [18], the least

measured throughput of 108 kbps and 203 Kpbs for

sites 1 and 2 respectively is enough to support

applications like online job searching, navigating

websites, streaming radio, voice over IP calls and

standard video streaming.

Field measurement trials for the overall pilot

network which have been presented in Ref. [19]

showed that the maximum measured throughput

during the entire pilot network trial measurement was

9.62 Mbps as compared to the simulated throughput

per sector of 16.55 Mbps. The measured throughputs

of 8.62 Mbps and 7.20Mbps for sites 1 and 2 show the

two base stations overall network performance have

been very impressive when compared with results in

Ref. [19]. The two sites will undoubtedly give

ubiquitous network coverage to its CPEs given the

obtained 99.45% and 95.81% RSSI values of -90 dBm

measured for sites 1 and 2 respectively. The

propagation models used for the coverage and

capacity simulation were developed for areas where

the harshness of the Sub-Saharan African terrain has

not been considered explicitly. Analysis of field

measurement results in Ref. [20] supports the

assumption that the correction factor for

Hata-Okumara model which have not been specified

for the Sub-Saharan African environment could have

contributed to the differences between the simulated

throughput per sector and the measured values.

Table 3 Summary of throughput measurements.

Site Measured throughput (DL) Measured throughput (DL)

Maximum (Mbps) Distance (m) Minimum (Kbps) Distance (m)

Site 1 8.62 500 108 5,200

Site 2 7.20 420 203 3,500


871

Fig. 10 Throughput measurements results.

6. Conclusions

Simulations and field trial measurement results for

two WiMAX base stations in a newly deployed pilot

network in the urban centers of Accra have been

presented. In this paper we have made in-depth

analysis of the network parameters based on

measurements performed at locations within the

network. Analytical expressions have been used to

support the simulation methodology for deriving the

cell radius for the two WiMAX base stations based on

the stochastic distribution of the customer premise

equipment.

We have used Genex-Unet to simulate the final

radio plan and throughput per sector of an adaptive 4

× 4 MIMO antenna configuration. The correction

factor for Hata-Okumura Model will modeled in

subsequent papers and it effects on the performance of

the network discussed.

It has been seen through the measurement results

that at distances of 5.20 km, users of the network can

enjoy standard video and radio streaming services.

This goes a long way to validate the claim that

WiMAX technology can deliver last mile broadband

technology to subscribers, as have been seen, even

under the harsh conditions in Sub-Saharan African.

References

[1]. IEEE 802.16: IEEE Standard for Local and Metropolitan Area Networks, part 16: Air interface for fixed broadband wireless access systems, IEEE 2004.

[2]. IEEE Standard 802.16e, part 16: Air interface for fixed and mobile broadband wireless access systems, IEEE 2006.

[3]. J.D. Sandra, C. Lee, WiMAX in Africa: the new frontier; and Comparative and International Law Institute Certificate Thesis, The Catholic University of America, Columbus School of Law, May, 2008.

[4]. Selection and award procedure documentation: “Internet/ Broadband Penetration Statistics Broadband Wireless access License in the 2.50 GHz-2.69 GHz Band in Ghana”, Government White paper, January, 2012, 12-14.

[5]. X.M. Huang, J.J. Luo, Automatic configuration and optimization of WiMAX networks, In WiMAX Network Planning and Optimization, 2009, pp. 44-51.

[6]. M. Hata, Empirical formula for propagation loss in land

mobile radio services, IEEE Transactions on Vehicular

Technology, vol. VT-29:317325, Sep. 1981.

[7]. Y. Okumura, Field strength and its variability in VHF and

UHF land-mobile radio-services, Review of the Electrical

Communications Laboratory, vol. 16, Sept.-Oct. 1968.

[8]. G.C. Hess, Land-Mobile Radio System Engineering,

Artech House, Boston, 1993.

[9]. T.S. Rappaport, S.Y. Seidel, M.J. Feuerstein, K.L.

Blackard, H.H. Xia, Path loss, delay spread and outage

models as functions of antenna height for microcellular

system design, IEEE Transactions on Vehicular

Technology 43 (1994) 487.


872

[10]. L.R. Maciel, A. Lindsay-Stewart, H.H. Xia, H.L. Bertoni,

R. Rowe, Microcellular propagation characteristics for

personal communications in urban and suburban

environments, IEEE Transactions on Vehicular

Technology 43 (1994) 743.

[11]. L.R. Maciel, A. Lindsay-Stewart, H.H. Xia, H.L. Bertoni,

R. Rowe, Radio propagation characteristics for line-of-

sight and personal communications, IEEE Transactions

on Antennas and Propagation 41 (1993) 1439.

[12]. M. Taylor, D. Li, V. Erceg, S. Ghassemzadeh, D.L.

Schilling, Urban/suburban out-of-sight propagation

modeling, IEEE Communications Magazine 30 (1992)

56.

[13]. E. Green, Radio link design for microcellular systems,

British Telecom Technology Journal 8 (1990) 85.

[14]. W.C. Jakes, Microwave Mobile Communications, IEEE

Press, New York, 1974.

[15]. B.C. Jones, D.J. Skellern, An integrated

propagation-mobility interference model for microcell

network coverage prediction, Wireless Personal

Communications 5 (1997) 223-258.

[16]. C.E. Cook, Modeling interference effects for land mobile

and air mobile communications, IEEE Transactions on

Communications 35 (1987) 151-165.

[17]. B.C. Jones, D.J. Skellern, Interference modeling and

outage contours in cellular and microcellular networks, in:

2nd MCRC International Conference on Mobile and

Personal Communications Systems, Adelaide, Australia,

Apr. 10-11,1995.

[18]. Broadband Speed Guide, http://www.fcc.gov. (Accessed

Jan. 7, 2013)

[19]. E.T. Tchao, W.K. Ofosu, K. Diawuo, Radio planning and field trial measurement for a 4G-WiMAX network in a Sub-Saharan Africa environment, in: IEEE Wireless Telecommunications Symposium, Phoenix, Arizona, Apr. 17-19, 2013.

[20]. L. Nissirat, M.A. Nisirat, M. Ismail, S. Al-Khawaldeh, A terrain roughness correction factor for hata path loss model at 900 MHz, Progress in Electromagnetics Research C 22 (2011) 11-22.


Error Analysis for the Three-Dimensional Detection and

Reconstruction of the Road Surface

Youquan He and Jian Wang

Information Science & Engineering Department, Chongqing Jiaotong University, Chongqing 400074, China

Received: December 05, 2012/ Accepted: January 11, 2013 / Published: June 30, 2013.

Abstract: Aiming at the series of errors processing and analyzing produced in three-dimensional testing and reconstruction of the highway pavements, this paper conducts a detailed analysis and computation of these process errors, including calibration of signalized points, centerline calculation of light stripe, accumulated system error, etc.. After the contrast experiment and analysis, it finally introduces gravity method to calculate the camera’s internal parameters, and gray centroid algorithm to extract light stripes’ center. Result shows the system deviation is stabilized at 2 mm, which can better meet the needs of engineering practice. Key words: Road surface detection, three-dimensional transform, error analysis, calibration of signalized points.

1. Introduction

Three-dimensional information collecting of the

road surface is the significant application of computer

vision technology in transportation area. After

establishing the relationship between

two-dimensional images (image side) and

three-dimensional road space (object side) which is

quickly obtained by CCD imaging sensor, we can

work out three-dimensional information of the road

surface. As detection methods, algorithm and other

reasons could cause lots of errors in the whole process,

which directly affect precision of the system, it is

essential to make thorough qualitative and

quantitative analysis of the errors to reduce the

diviation and improve the accuracy of the system [1].

Taking binocular imaging based on three-dimensional

damage detection system for example, this paper

analyzes source and type of errors and discusses

how to improve the accuracy of the detection

system.

Corresponding author: Youquan He, Ph.D., professor, research fields: information processing, data mining. E-mail: [email protected].

2. System Components and Measuring Principle

As shown in Fig. 1, the system uses a LED linear

light as an auxiliary light and makes it ray on road

surface vertically. Then, it uses two CCD sensors, one

is on the left and the other is on the right, to access

pavement image. Especially, their central axis should

be set in 30 to 60 degree angle, not parallel with the

light. As we all know, LED line light could only form

a very narrow light stripe when it rays on road surface.

So, if we take no consideration of camera distortion,

the imaging light from any angle should be a straight

line when the road surface has no damage defects. But,

if pit damage exists, luminous stripe’s shape will

change at the defect. That is to say, every point of

light stripe would in different depth in the pothole.

According to the principle of 3-D projection

transformation, these different depth points would

image a deformation slit of light. Therefore, we can

calculate the actual coordinate value of each point

through transformation between object side and image

side coordinate system, and finally reconstruct the real

cross-section accurately [2, 3].

Error Analysis for the Three-Dimensional Detection and Reconstruction of the Road Surface

874

3. Calibration of the Inside and Outside Camera Parameters

Calibrating camera’s inside and outside parameters

is the first step of binocular vision based camera

calibration. Its major purpose is to establish the

relationship between three-dimensional world

coordinates of tested object and two-dimensional

image coordinates of the computer. This paper employs

RAC two-step method for calibration.

During the calibration, we need to make acquisition

and measurement of several points which are marked in

object plane. Generally, camera plane is not parallel

with object plane, which makes the actual standard

circles distorted in the image plane, such as an ellipse.

In order to work out image side coordinates of the

standard samples’ center, we should resort to the

distortion images. Here, we compared the

performances between value points intersection

method and gravity method.

3.1 Value Points Intersection Method

Taking the top left corner of image plane as the

origin of Cartesian coordinates, horizontal as X-axis,

vertical as Y-axis and pixels as the unit, we could obtain

coordinates of every curve’s point in image plane in

Fig. 2.

After figuring out the biggest X value point A and

minimum point C, we link them into a line, so does the

line between biggest Y value point B and minimum

point D. And then the two lines segments intersect at

point O. Given that: A(x2, y3), B(x3, y2), C(x1, y4）,

D(x4, y1), then

(1)

According to the two equation of straight line AC

and BD, we could figure out the coordinate of point O

Fig. 1 Component diagram of the system.

Fig. 2 Location of the image center.

(x, y), which is also the center point of the image and

sample circle. Suppose that the minimum value x has

multiple corresponding values y, which means a

vertical cantlet in the leftmost pixel image, then we

take the means of these values as the point Y

coordinate, so does the situation of multiple

corresponding values x [4].

Finally, we obtained the center coordinates of the

right target, which are shown in the following Table 1.

3.2 Gravity Method

The method takes the image’s barycenter as target’s

center in image plane. We introduce integral to

calculate image’s center. After taking a unit—a pixel as


875

Table1 Center coordinates of the right target.

Circle number (x1, y1) (x2, y2) (x3, y3) (x4, y4) Center coordinates

A 162, 69 227, 272 185, 258 189, 299.67 180.614, 278.225

B 166, 388 225, 374.1 182, 357.1 187.8, 400.96 181.194, 383.099

C 159, 628.7 230.98, 630 178.8, 600.4 182.3, 652.65 176.071, 628.647

E 387, 307 428.17, 308 416, 288 422.8, 330.7 419.771, 306.132

D 381.46, 616 430.9, 611.9 411.07, 591.88 427.8, 637 418.822, 614.562

, we could obtain the area of the image, which is

also the number of pixels. Then we get the formula for

image center’s calculation:

, (2)

Eventually, we figure out coordinates of the five

target circles’ center in Table 2. Table 3 is results

comparison of the two methods. Value points

intersection method would cause a serious deviation.

Therefore, considering their efficiency, accuracy and

easy realization, this paper uses gravity method to

compute centers of the target points.

4. Error Analysis of Light Stripes’ Extraction

As we known, optical information could reflect

objects’ surface geometric information when LED line

light emits structure light on their surface. LED line

light source is of certain width, but only the points in

center line are really needed in measurement, so the

extraction accuracy of line structured light centerline

directly affects precision of the measurement. Because

of the random error, such as uneven light intensity,

different surface properties of the tested object and

noise produced by CCD camera, etc., it is quite

difficult to improve the precision of extraction. There

are some major algorithms to extract the center of a

light stripe, such as threshold method, gray level

barycenter method, Gauss fitting, etc. [5, 6].

Considering their measurement precision, this system

uses gray barycenter method and Gaussian distributed

curve fitting algorithm and makes a contrast between

their results.

The whole comparative trial is conducted on basis of

the detecting system. Firstly, we collect the light

stripe’s images under the detecting conditions in Fig.3;

then intercept the light stripe part of the image; lastly,

obtain the gray level distribution of the stripe’s section.

As it is show in Fig. 4.

This paper introduces two methods to extract the

light stripes’ center. Each time it selects 50 light stripe

sections, as light stripes center is a straight line when

structured light is projected onto the smooth surface,

we could get the maximum deviation and standard

deviation of the center by making least square linear

fitting to the extracted section [7].

Apparently, the standard deviation could reflect the

discrete degrees of the points. The comparison result

shows that gray barycenter method has better

performance in extracting, and its maximum deviation

and discrete degree are comparatively small. Table 4

gives partial analysis results of numerous extraction

experiments.

Table 2 Coordinates computed by gravity method.

Circle number Center coordinates

A 179.5134, 278.158

B 180.083, 382.029

C 175.072, 627.508

E 418.532, 305.823

D 417.898, 613.376

Table 3 Results comparison of the two methods.

Circle number

Coordinates computed by intersection method

Coordinates computed by gravity method

A 180.614, 278.225 179.5134, 278.158

B 181.194, 383.099 180.083, 382.029

C 176.071, 628.647 175.072, 627.508

E 419.771, 306.132 418.532, 305.823

D 418.822, 614.562 417.898, 613.376

dA

A

xdA

xA

A

ydA

yA


876

Fig. 3 Imaging of the light stripe.

Fig. 4 Gray distribution of light stripe’s section.

Table 4 Comparison results of the two extraction methods.

Light stripes number Gray barycenter method based light stripe thining Gauss distribution based curve fitting method

Maximum deviation (um) Discrete degree Maximum deviation (um) Discrete degree

1 5.11 0.3667 6.02 0.3997

2 4.58 0.3381 4.69 0.3492

3 5.07 0.3594 5.23 0.3822

4 4.77 0.3451 4.98 0.3677

5 5.53 0.3733 5.71 0.3981

5. Conclusions

There are so many factors which affected the

accuracy of the system, and it is very hard to analyze

their interrelations in isolation. So lots of comparison

tests and comprehensive evaluation are needed to

select suitable measurement methods for the system.

This paper improved the precision of binocular vision

measurement system by numbers of experiments. The

results show that the system accuracy was quite good

to meet practical application requirements.

References

[1] Y. Huang, Y. Zhao, Algorithm and realization of three-step camera calibration based on 3D-target, Computer Technology and Development 20 (2010) 139-142.

[2] Y.Q. He, J. Wang, A research of pavement potholes

detection based on three-dimensional projection transformation, in: 2011 4th International Congress on Image and Signal Processing, Shanghai, China, 2011.

[3] X.P. Lou, M.P. Guo, Methods of calibration improvement in the binocular vision system, Journal of Beijing Information Science and Technology University 25 (2010) 316-320.

[4] Y. Rui, J.T. Liu, A video editing system based on the depth information ubi-media computing (u-media), in: 2010 3rd IEEE International Conference, Jinhua, 2010.

[5] Z.M. Liu, W.Y. Deng, Extraction algorithm of light stripes center in the measurement system of structured light, Journal of Beijing Institute of Machinery 24 (2009) 142-146.

[6] X.J. Yang, H.H. Chi, An improve method of extracting structured light strip center, Applied Science and Technology 12 (2009) 41-45.

[7] H. Zhang, J.J. Fang, Machine Vision in Two-Dimensional Images of Three-Dimensional Reconstruction, College of Mechanical and Electrical Engineering, North China University of Technology, Beijing, 2006.

Journal of Communication and Computer（Issue 6，2013）

Documents

Transcript of Journal of Communication and Computer（Issue 6，2013）