Optimizing the ecological connectivity of landscapes with ...
-
Upload
khangminh22 -
Category
Documents
-
view
0 -
download
0
Transcript of Optimizing the ecological connectivity of landscapes with ...
HAL Id: hal-03578088https://hal.archives-ouvertes.fr/hal-03578088
Preprint submitted on 17 Feb 2022
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Optimizing the ecological connectivity of landscapeswith generalized flow models and preprocessing
François Hamonic, Cécile Albert, Basile Couëtoux, yann Vaxès
To cite this version:François Hamonic, Cécile Albert, Basile Couëtoux, yann Vaxès. Optimizing the ecological connectivityof landscapes with generalized flow models and preprocessing. 2022. �hal-03578088�
Optimizing the ecological connectivity of landscapes with generalizedflow models and preprocessing I
Francois Hamonica,b, Cecile Albertb, Basile Couetouxa, Yann Vaxesa
aAix-Marseille Univ, CNRS, Universite de Toulon, LIS, Marseille, FrancebAix-Marseille Univ, CNRS, Univ Avignon, IRD, IMBE, Marseille, France
Abstract
In this paper we consider the problem of optimizing the ecological connectivity of a land-
scape under a budget constraint by improving habitat areas and ecological corridors be-
tween them. We consider a formulation of this problem in terms of graphs in which vertices
represent the habitat areas and arcs represent a probability of connection between two ar-
eas that depend on the quality of the respective corridor. We propose a new generalized
flow model that improves existing models for this problem and an efficient preprocessing
algorithm that reduces the size of the graphs on which generalized flows is computed.
Reported numerical experiments highlight the benefice of these two contributions on com-
putation times and show that larger problems can be solved using them. Our experiments
also show that several variants of greedy algorithms perform relatively well on practical
instances while they return arbitrary bad solutions in the worst case.
Keywords: Environment and climate change, landscape connectivity, generalized
network flow, robust shortest path, linear programming, distances in graphs.
1. Introduction
Habitat loss is a major cause of the rapid decline of biodiversity [5]. Further than
reducing the available resources, it also increases the discontinuities among small habitat
areas called patches, this is a phenomenon known under the name habitat fragmentation [9].
While habitat loss tends to reduce the size of populations of animals and plants, which may
be deleterious to their survival in the long term, habitat fragmentation also makes it harder
for organisms to move around in landscapes. This constrains their access to resources and
reduces gene flow among populations, which can also be deleterious for their persistence
IThe research on this paper was supported by Region SUD, Natural Solutions and the ANR projectDISTANCIA (ANR-17-CE40-0015).
Email addresses: [email protected] (Francois Hamonic), [email protected](Cecile Albert), [email protected] (Basile Couetoux), [email protected] (Yann Vaxes)
Preprint submitted to Networks February 17, 2022
in the long term. Landscape connectivity, defined as the degree to which the landscape
facilitates the movement of organisms between habitat patches [19], then becomes of major
importance for biodiversity and its conservation. Accounting for landscape connectivity
in restoration or conservation plans thus appears as a key solution to maximize the return
on investment of the scarce financial support devoted to biodiversity conservation. For
this purpose, graph-theoretical approaches are useful in modelling habitat connectivity
[20]. In this paper, we consider the landscape connectivity that is distinct from the graph
theoretical notion of connectivity defined in terms of minimum cut.
A landscape can be viewed as a directed graph in which vertices represent the habitat
patches and each arc represents a way for individuals to travel from one patch to another.
Each vertex has a weight that represents the quality of the patch – patch area is often
used as a surrogate for quality – and each arc has a weight that represents the difficulty
for an individual to make the corresponding travel – often approximated by the border to
border distance between the patches. Interestingly, this approach can be used for a variety
of ecological systems like terrestrial (patches of forests in an agricultural area, networks
of lakes or wetlands), riverine (patches are segments of river than can be separated by
human constructions like dams that prevent fishes’ movement) or marine (patches can be
reefs that are connected by flows of larvae transported by currents). With this formalism,
ecologists have developed many connectivity indicators [11, 12, 17] that aim to quantify
the quality of a landscape with respect to the connections between its habitat patches.
Among the proposed indicators, the probability of connection – PC – [17] and its derivative
the equivalent connected area – ECA – [16] seem to perform well; they have received
encouraging empirical support [3, 14, 18]. Definitions and relationship between these
two indicators are described in Section 2. In this paper, our purpose is not to discuss
the merit of ECA with respect to other landscape connectivity indicators but to study the
corresponding combinatorial optimization problem from an algorithmic point of view. One
key question for which landscape connectivity indicators have been used in the last years is
to identify which elements of the landscape (habitat patches, corridors) should be preserved
from destruction or restored in order to maintain a well-connected network of habitat
under a given budget constraint [2]. This means, identifying the set of vertices or arcs that
optimally maintain a good level of connectivity. Many mathematical programming models
have been introduced in the literature to help decision-makers to protect biodiversity, for
a deep and wide review of these models, we refer to the survey paper [4] and the references
2
therein.
Here, we will focus on this problem by measuring ecological connectivity with ECA.
The budget-constrained ECA optimization problem consists in finding the best combina-
tion of graph elements (arcs or vertices) among a set Φ of elements that could be restored
in order to maximize the ECA in the landscape within a limited budget. Interestingly, this
problem addresses both restoration and conservation cases. The restoration case starts
from the current landscape and aims to restore a set of elements among the different
feasible options. The conservation case starts with a landscape in which we search for
the elements to protect among the threatened ones, and before the unprotected ones get
degraded. The conservation problem and the restoration problem are equivalent from an
algorithmic point of view. Indeed, an instance of the conservation problem can be viewed
as an instance of the restoration problem in which the initial landscape is the landscape in
which all threatened elements are degraded and the set Φ of elements than can be restored
in the instance of the restoration problem is the set of elements that can be protected in
the instance of the conservation problem. Until now, ecologists have mostly tackled this
problem by ranking each conservation or restoration option by its independent contribu-
tion to ECA, i.e. the amount by which ECA varies if the option is purchased alone. Such
an approach overlooks the cumulative effects of the decisions made like unnecessary redun-
dancies or potential synergistic effects, e.g. consecutive arcs. This could lead to solutions
that are more expensive or less beneficial to ECA than an optimal solution. Some studies
have tried to overcome this limitation by considering tuples of options [13, 15]. In [15],
the authors show that the brute force approach rapidly becomes impractical for landscape
with more than 20 patches of habitat. However, few studies explore the search for an op-
timal solution, most of which being developed for river dendritic networks. For instance,
in [21], a polynomial time approximation scheme is described for the problem where the
landscape is represented by a tree graph. Recently a mixed integer formulation that is
applicable to our problem has been introduced [22]. This formulation is however expen-
sive to solve. It uses O(|V |2|A|) flow variables and O(|V |3) constraints while O(|Φ|+ |V |2)
integer variables are required to approximate the non-linear objective function. The au-
thors of [22] show that solving exactly this formulation does not scale to landscapes of few
hundreds of patches. Their main contribution was to propose a XOR sampling method
for approximating it. Since it is possible to use the XOR sampling method with our exact
and more compact MIP model, it is better to view our contributions as a complement
3
to those of [22]. Therefore, we did not perform numerical experiments to compare our
methodology against those developed in [22].
Our model is based on a generalized flow formulation (see for instance [1]) instead of
a standard network flow formulation, which has two main advantages. On one side, our
formulation has a linear objective function. We do not need to use a piecewise constant
approximation nor to introduce additional binary variables to handle non-linearity. On the
other side, the new formulation is much more compact because it aggregates into a single
generalized flow the contribution to the connectivity of several source/sink pairs having
the same source. This reduces significantly the number of variables and constraints of the
model. It uses O(|V ||A|) flow variables with O(|V |2) constraints, and |Φ| integer variables
instead of O(|V |2|A|) flow variables with O(|V |3) constraints, and O(|Φ| + |V |2) integer
variables in the model of [22]. Recall that Φ is the set of elements that are threatened or
could be restored. Another contribution of this paper is to propose a preprocessing step
that speeds-up the resolution of the problem by reducing the size of the directed graph
on which a generalized flow to a particular sink t have to be computed. The idea is to
identify, for each sink t, the set of arcs (u, v) that are always on a shortest path from u
to t for all the 2|Φ| possible restoration plans. Such arcs, called t-strong arcs, are always
carrying flow and can be contracted in the generalized flow formulation to t. Conversely,
an arc (u, v) that is never on a shortest path from u to t, called t-useless, can be removed
from the generalized flow formulation to t because it never carries any flow. Once t-strong
arcs have been contracted and t-useless arcs have been removed in each generalized flow
formulation to a particular sink t, it remains to solve the problem on a much smaller
instance containing only the arcs that are not t-strong nor t-useless. Obviously it would
be prohibitive in terms of computation times to solve a shortest path problem for each of
the 2|Φ| possible restoration plans in order to determine whether a given arc is t-strong
(resp. t-useless) or not. In Section 4, we design and analyze a very efficient algorithm, that
computes directly the set of all sinks t such that a given arc is t-strong. This algorithm
takes implicitly into consideration all 2|Φ| possible restoration plans but has the same
complexity as the Dijkstra shortest path algorithm. It improves previous result for the
problem of computing t-strong or t-useless arcs [6]. We anticipate that this algorithmic
result is of independent interest and could be useful in many other contexts, in particular
when shortest path computations are embedded in robust optimization models.
In Section 2 we give more details about the landscape modelization and describe for-
4
mally the problem. Section 3 is devoted to the description of our mixed integer formula-
tion of the problem. In Section 4 we first explain why the removal of t-useless arcs and
the contraction of t-strong arcs does not modify the objective function of any restora-
tion/conservation plan. Then, we design and analyze a O(|A| + |V | log |V |) algorithm
that computes the set of all vertices t for which a given arc (u, v) is t-strong or the set
of all vertices t for which (u, v) is t-useless. Section 5 presents experimental cases for
which we compare our optimization approach with simple greedy algorithms in terms of
running times and quality of the solutions found. These experiments also shows that the
preprocessing is very effective on some real instances.
2. Budget Constrained ECA Optimization problem
Our optimization problem takes as input a landscape represented by a directed graph
G = (V,A). Each patch s ∈ V has a weight ws that represents its size. Each ordered
pair of patches (s, t) such that an individual can move from patch s to patch t without
crossing another patch is joined by an arc of G. For each arc (u, v), we suppose that an
estimate of the probability π(u,v) that an individual successfully moves from u to v. This
value will be determined by the landscape and by the species studied. The input of our
optimization problem also includes a list of arcs that can be improved to make them easier
or less dangerous to cross and the costs of these improvements. As objective function and
measure of the quality of the landscape, we use the Equivalent Connected Area defined
by the following formula [16]:
ECA(G,w, l) =
√∑s∈V
∑t∈V
(wswtΠst)
where Πst is the probability of connection from the patch s to the patch t. This probability
is defined as the probability of the most probable (reliable) path from s to t where the
probability of a path is the product of the probabilities of its arcs. The probability of an
arc a is often defined as πa := exp(−αla) where α is a parameter that depends on the
dispersal ability of the species and la is the euclidean distance between the centers of the
two patches joined by a. This dispersal model has been used in many applications. Its
ecological justification, discussed for instance in [7, 8], is out of the scope of this paper. By
monotonicity of the exponential function, a path P is a most reliable path between s and
t if and only if it is a shortest path with respect to arc lengths l and Πst = exp(−αd(s, t))
where d(s, t) is the length of a shortest st-path with respect to l. The equivalent connected
5
area of a landscape is the area of a single patch equivalent to it. This equivalence comes
from the indicator PC on which ECA is based [16]. Indeed, suppose that the landscape is
a rectangle of area W containing all patches. We consider a stochastic process consisting
of choosing two points p and q uniformly at random in the rectangle. The indicator PC
is the expected value of a random variable equal to 0 if either p or q does not belong to
a patch and Πst if p belongs to s and q belongs to t (recall that Πst = 1 if s = t). Since
the probability that p belongs to a patch u is wu/W, and the events p ∈ s and q ∈ t are
independent, by linearity of expectation, PC can be expressed as follows:
PC(G,w, l) =
∑s,t∈V wswtΠst
W2=ECA(G,w, l)2
W2
where W, ws and wt are respectively the area of the rectangle, patch s and patch t. The
equivalent connected area of a landscape is the area of a single patch whose PC value is
equal the PC value of the original landscape. If the area of the patches and the landscape
are normalized to makeW equal to 1 then PC is the square of ECA. Therefore, optimizing
PC and optimizing ECA are equivalent problems. ECA is often considered by researchers
interested in landscape connectivity because it represents an area, a more concrete quantity
than the expected value of a random variable.
An instance I = (V,A,Φ, w, l, l−, c, B) of the Budget Constrained ECA Optimization
Problems – BC-ECA-Opt – is defined by a graph G = (V,A) with a weight wu for each
vertex u ∈ V and a length la for each arc a ∈ A, a subset Φ ⊆ A of arcs that can be
improved, for each arc a ∈ Φ, the cost ca of reducing its length to l−a and the total budget
B for improving the arcs. An optimal solution of the instance I is a subset of arcs S ⊆ Φ
of total cost c(S) :=∑
a∈S ca at most B that maximizes ECA(G, l′, w) where l′a = l−a if
a ∈ S and la otherwise. For the sake of clarity, we first consider only arcs improvements,
but, as explained in Section 3.1, our model can be extended to the case where it is also
possible to improve the landscape by increasing the weight of the vertices. In this case, a
subset of vertices W is given, as well as the cost cu of increasing the weight of u to w+u for
each vertex u ∈ W, and we want to find a subset of arcs S ⊆ Φ and a subset of vertices
T ⊆ W of total cost c(S) + c(T ) ≤ B that maximizes ECA(G, l′, w′) where l′a = l−a if
a ∈ S and la otherwise and w′u = w+u if u ∈ T and wu otherwise.
6
Remark 1. Our optimization problem is stated in terms of arc length reduction but it
would be equivalent to state it in terms of arc probability augmentation.
Remark 2. In our model, it is possible to add a new connection between two existing
patchs u and v. It suffices to place (u, v) in the set of improvable arcs Φ with an initial
length l(u,v) =∞ (i.e. π(u,v) = 0) and a finite improved length l−(u,v) <∞.
3. An improved MIP formulation for BC-ECA-Opt
We first show how to compute f t =∑
s∈V ws ·Πst as the maximum quantity of gener-
alized flow that can be sent to t across the network if, for every patch s ∈ V , ws units of
flow are available at s and appropriate multipliers are chosen for each arc of the network.
Recall that a generalized flow differs from a standard flow by the fact that each arc a has
a multiplier πa such that the quantity of flow leaving arc a is equal to the quantity of flow
φa entering in a multiplied by πa. In our case πa = exp(−αla). For each vertex u ∈ V, let
δoutu be the set of arcs leaving u and δin
u the set of arcs entering u. As explained below,
the linear program (P) has f t as optimum value.
(P)
max z
s.t.∑
a∈δoutu
φa −∑b∈Γ−u
πb · φb ≤ wu u ∈ V \ {t} (A1)
∑a∈δoutt
φa −∑b∈δint
πb · φb = wt − z (A2)
φa ≥ 0 a ∈ A (A3)
Constraints (A1) require that the total quantity of flow leaving u is at most wu plus the
total quantify of flow entering u, i.e. the quantity of flow available in vertex u. Constraint
(A2) requires that z is equal to the total quantity of flow entering t plus wt minus the total
quantity of flow leaving t. Finally, constraints (A3) state that each arc carries a positive
quantity of flow from its source to its sink.
Lemma 1. Any optimal solution of (P) is obtained by sending, for each every vertex
s ∈ V − {t}, ws units of flow from s to t along most reliable st-paths.
Proof. First notice that the quantity of flow arriving at t when ws units flow are sent
from s to t along an st-path P is ws times the probability of path P. Let φ′ be an optimal
solution of (P) maximizing the quantity of flow routed along a path which is not a shortest
7
path. Suppose by contradiction that φ′ send ε > 0 units of flow along an st-path P ′ whose
probability is smaller than the probability of a most reliable st-path P. Let φ be the flow
obtained from φ′ by decreasing the flow sent on path P ′ by ε and increasing the flow sent
on path P by ε. Since the probability of P is larger than the probability of P ′, φ send
more flow to t than φ′, a contradiction.
Corollary 1. The optimal value of (P) is f t =∑
s∈V ws ·Πst.
Proof. Since the objective is to maximize z, no flow leaves t in any optimal solution of
(P), i.e.∑
a∈δoutt
φa = 0. Constraint (A2) ensures that z is the quantity of flow received
by t plus wt = wt · Πtt. By Lemma 1, there exits an optimal solution φ such that every
vertex s distinct from t send ws units of flow on a most reliable st-path. Hence, for every
vertex s distinct from t, the flow received by t from s is ws ·Πst and thus the value of z is∑s∈V ws ·Πst.
With this linear programming definition of f t, we can now present our MIP program
of BC-ECA-Opt. For each arc a = (u, v) ∈ Φ, we add another arc a′ = (u, v) of length
la′ = l−a that can be viewed as an improved copy of arc a. This improved copy can be
used only if the improvement of arc a is purchased. For each arc a ∈ Φ, that could be
improved, the variable xa is equal to one if the improvement of arc a is purchased and
zero otherwise. We denote by Ψ the set of improved copies of arcs in Φ. In the following
MIP program, δoutu and δin
u are defined with respect to the set of arcs A′ = A ∪Ψ.
BC-ECA-Opt
max∑t∈V
wt · f t
s.t.∑
a∈δoutu
φta −∑b∈δinu
πb · φtb ≤ wu t ∈ V , u ∈ V \ {t} (B1)
∑a∈δoutt
φta −∑b∈δint
πb · φtb = wt − ft t ∈ V (B2)
φta′ ≤ xa ·Ma t ∈ V , a ∈ Φ (B3)∑a∈F
ca · xa ≤ B (B4)
xa ∈ {0, 1} a ∈ Φ (B5)
φta ≥ 0 t ∈ V , a ∈ Φ (B6)
Constraints of (B1) and (B2) are simply constraints of (A1) and (A2) for all possible
target vertex t. For each arc a ∈ Φ, the big-M constraints (B3) ensure that if the im-
8
provement of arc a is not purchased, i.e. xa = 0, then the flow on arc a′ is null. Ma is
an upper bound of the flow value on arc a. We could simply take Ma =∑
u∈V wu for all
a ∈ Φ but smaller estimations are important to guaranty a good linear relaxation and thus
a faster resolution of the MIP program. Constraint (B4) ensures that the total cost of the
improvements is at most B. This MIP program has O(|V |(|A| + |Φ|)) flow variables, |Φ|
binary variables and O(|V |2 + |V ||Φ|) constraints.
3.1. Extension to patch improvements
To extend our model to the version with patch improvements where it is possible to
increase the weight of a vertex u from wu to w+u at cost cu, we process all the occurrences
of wu as follows. Let yu be a binary variable equal to 1 if the vertex u is improved and 0
otherwise. We add a term yucu in the budget constraint for every vertex u in the set W
of vertices that can be improved. When wu appears as an additive constant, we simply
replace it by wu+yu(w+u −wu). Note that wu only appears as a coefficient in the objective
function in the form wuft. In this case, to avoid a quadratic term, we use a standard
McCormick linearization [10]. We replace the product wuft by wuft + (w+u −wu)f ′t where
f ′t is a new variable that is equal to ft if yu = 1 and 0 otherwise. To achieve this values of
f ′t we add the constraints f ′t ≤ ft and f ′t ≤ yuM where M is larger than any values of ft.
As it is a maximization program and f ′t appears with a positive coefficient in the objective
function, the first constraint guaranties that f ′t = ft if yu = 1 in any optimal solution and
the second constraint guaranties that f ′t = 0 if yu = 0.
4. Preprocessing
The size of the mixed integer programming formulation of BC-ECA-Opt given in Sec-
tion 3 grows quadratically with the size of the graph that represents the landscape. In
this section, we describe a preprocessing step that reduces the size of this graph. For that,
we introduce a notion of strongness of an arc with respect to a target vertex t ∈ V . We
will call a solution x ∈ {0, 1}Φ of BC-ECA-Opt a scenario and say that the distances are
computed under scenario x when the length of every a ∈ A is l−a if xa = 1 and la otherwise.
We denote dx(s, t) the distance between s and t when the arc lengths are set according
to the scenario x. An arc (u, v) is said to be t-strong if, for every scenario x ∈ {0, 1}Φ,
(u, v) belongs to a shortest path from u to t, i.e. dx(u, t) = dx(u, v) + dx(v, t) for every
x ∈ {0, 1}Φ. An arc (u, v) is said to be t-useless if, for every scenario x ∈ {0, 1}Φ, arc
9
(u, v) does not belong to any shortest ut-path when arc lengths are set according to the
scenario x. We denote by S(t) the set of arcs (u, v) such that (u, v) is t-strong and by
W (t) the set of arcs (u, v) such that (u, v) is t-useless.
Of course, the naive algorithm that tests whether (u, v) belongs to a shortest ut-path
for every scenario x ∈ {0, 1}Φ is very inefficient. In Section 4.2, we describe and analyse
an algorithm that computes the set S(t) and W (t) for every sink t in a much more efficient
way. But first, we explain how the knowledge of these sets can be used to define a smaller
equivalent instance of the problem.
4.1. Operations to reduce the size of the graph
Removal of an arc from W (t). Let (u, v) ∈ W (t), since for all scenarios x ∈ {0, 1}Φ,
(u, v) does not belong to any shortest path from u to t, its removal does not affect the
distance from any vertex to t. It is clear, from the definition of ECA, that the removal of
(u, v) does not affect the contribution of t to ECA. Let fxt (G) :=∑
s∈V wswtΠxst be the
contribution to ECA of all the pairs having sink t in the graph G when the probability of
connection Πxst is computed under the scenario x.
Lemma 2. Let (u, v) ∈W (t) and let G′ be a graph obtained from G by removing arc (u, v)
then, for all scenario x ∈ {0, 1}Φ, fxt (G) = fxt (G′).
Proof. For every scenario x ∈ {0, 1}Φ, (u, v) does not belong to any shortest path from u
to t. Therefore, the removal of (u, v) cannot affect the probability of connection Πut for
any vertex t and fxt (G) = fxt (G′).
Contraction of an arc in S(t). Now, assume that (u, v) /∈ Φ, (u, v) ∈ S(t). The
contraction of (u, v) consists in replacing every arc (w, u) ∈ δinu by an arc (w, v) of length
l′wv = lwu + luv and by removing the vertex u and all its outgoing arcs. The weight of
u in G is moved to the weight of v in G′. Namely, the weight of v in the new graph is
w′v = wv+wu exp(−αluv). Let G′ be the graph obtained from G by contracting (u, v). The
next lemma establishes that the contribution of t to ECA in G is equal to its contribution
in G′.
Lemma 3. Let (u, v) ∈ S(t) and let G′ be the graph obtained from G by contracting
arc (u, v) and modifying accordingly the weight of wv. For every scenario x ∈ {0, 1}Φ,
fxt (G) = fxt (G′).
10
w
u v t...
lwuluv
wu wv
Γ−u
(a)
w
v t...
lwu + luv
w′v
Γ−u
(b)
Figure 1: (a) represents a graph G before the contraction of an arc (u, v). (b) represents the graph G′
obtained from G by contracting arc (u, v). The weight w′v of v in G′ is equal to wv + wu exp(−αluv).
Proof. Let s be a vertex of G and x be a scenario in {0, 1}Φ. We denote fxst(G) the
contribution to ECA of the pair st with scenario x ∈ {0, 1}Φ in G. If s /∈ {u, v}, it is easy
to check that, for any x ∈ {0, 1}Φ the length of the shortest path from s to t in G is equal
to the length of the shortest path from s to t in G′. Moreover, the weight of s is the same
in G and in G′. Therefore, by the definition of ECA, fxst(G) = fxst(G′), for any x ∈ {0, 1}Φ.
On the other hand, the contributions of the pairs ut and vt in G sum to the contribution
of v in G′. Indeed,
fxut(G) + fxvt(G) = wu exp(−αdx(u, t)) + wv exp(−αdx(v, t))
= wu exp(−α(luv + dx(v, t))) + wv exp(−αdx(v, t))
= (wv + wu exp(−αluv)) exp(−αdx(v, t))
= w′(v) exp(−αdx(v, t))
= fxvt(G′)
We conclude that, for any x ∈ {0, 1}Φ, fxt (G) =∑
s∈V fxst(G) = fxut(G) + fxvt(G) +∑
s∈V−{u,v} fxst(G) = fxvt(G
′) +∑
s∈V−{u,v} fxst(G
′) = fxt (G′).
Graph reduction. We obtain Gt by deleting every arcs of W (t) and contracting every
arcs of S(t). By induction the contribution of t is preserved. The experiments of Section
5 show that, when the number of arcs that can be protected is much smaller than the
total number of arcs, replacing G by Gt in the construction of the mixed integer program
reduces significantly the size of the MIP formulation and the running times.
The problem of identifying S(t) and W (t) has been studied as a preprocessing step
for other combinatorial optimization problems having as input a graph whose arc lengths
may change [6]. In [6], given an arc (u, v) and a vertex t, the authors identify a sufficient
(but not necessary) condition for (u, v) ∈ S(t) and use it to design a O(|A| + |V | log |V |)
11
algorithm that can, in most cases, identify if (u, v) ∈ S(t). The authors also propose a
O(|A|+ |V | log |V |) algorithm to test whether a given arc (u, v) belongs to W (s) or not.
Given an arc (u, v) of G, we will show how to adapt the Dijkstra’s algorithm to compute
in O(|A|+|V | log |V |) the set of all vertices t such that (u, v) ∈ S(t) or the set of all vertices
t such that (u, v) ∈ W (t). This improves the results of [6] by providing a necessary and
sufficient condition for (u, v) to be t-strong, i.e. we compute the entire set S(t) while
the algorithm of [6] computes a subset of S(t). It also reduces the time complexity for
computing S(t) and W (t) for all t by a factor |V | as we only need to run the algorithms
on each arc instead of each pair of arc and vertex.
4.2. Computing S(t) for all t
Given a scenario x ∈ {0, 1}Φ, the fiber Fx(u, v) of arc (u, v) is the set of vertex t such
that (u, v) belongs to a shortest path from u to t when arc lengths are set according to x,
i.e.
Fx(u, v) = {t ∈ V : dx(u, t) = lx(u, v) + dx(v, t)}
Let F (u, v) be the intersection of the fibers of (u, v) over all possible scenarios x ∈
{0, 1}Φ, i.e. F (u, v) :=⋂x Fx(u, v). By definition, an arc (u, v) is t-strong if t belongs to
Fx(u, v) for every scenario x i.e.
S(t) = {(u, v) : t ∈ F (u, v)}
In order to compute every S(t), we first compute F (u, v) for every arc (u, v) and then
transpose the representation to get S(t) for every vertex t ∈ V . Let y ∈ {0, 1}Φ be the
following scenario :
ywt =
0 w ∈ F (u, v) or (w, t) = (u, v)
1 w /∈ F (u, v)(1)
Lemma 4. The intersection F (u, v) of the fibers of (u, v) over all possible scenarios is the
fiber of (u, v) under the scenario y, i.e. F (u, v) = Fy(u, v).
Proof. Since the inclusion F (u, v) ⊆ Fy(u, v) is obvious, it suffices to prove that Fy(u, v) ⊆
Fx(u, v) for any scenario x ∈ {0, 1}Φ. By way of contradiction, suppose that x is a scenario
such that Fy(u, v)\Fx(u, v) contains a vertex t at minimum distance from u in the scenario
y. Let P be a shortest ut-path containing (u, v) under the scenario y. For every vertex
12
w of P, dy(u,w) ≤ dy(u, t) and w ∈ Fy(u, v). Hence, by the choice of t, w belongs to
Fx(u, v) for every scenario x. Therefore w belongs to F (u, v) and the length of every
arc of P is set to its upper bound in the scenario y, i.e. ly(P ) = l(P ). Now, let Q
be a shortest ut-path in the scenario x. If the path Q contains a vertex z ∈ F (u, v)
then dx(u, t) = dx(u, z) + dx(z, t) = lx(u, v) + dx(v, z) + dx(z, t) = lx(u, v) + dx(v, t), a
contradiction with t /∈ Fx(u, v). Therefore, the vertices of Q do not belong to F (u, v) and
thus their lengths are set to their lower bound in y, i.e. l(Q) = l−(Q). We deduce that
lx(P ) ≤ ly(P ) ≤ ly(Q) ≤ lx(Q) a contradiction with t /∈ Fx(u, v).
We describe a O(|A| + |V | log |V |) time algorithm that, given an arc (u, v), computes
simultaneously the scenario y defined by (1) and the fiber Fy(u, v) of arc (u, v) with respect
to this scenario. The algorithm is an adaptation of the Dijkstra’s shortest path algorithm
that assigns colors to vertices. We prove that it colors a vertex w in blue if w belongs to
Fy(u, v) and in red otherwise. At each step, we consider a subset of vertices S ⊆ V whose
colors have been already computed. Before the first iteration, S = {u} and the length
of every arc leaving u is set to its lower bound except the length of (u, v) which is set to
its upper bound according to scenario y. At each step, since the color of every vertex of
S is known, the length, under scenario y, of every arc (w, t) with w ∈ S is also known.
Therefore, it is possible to find a vertex t ∈ V − S at minimum distance from u, under
scenario y, in the subgraph G[S ∪ {t}] induced by S ∪ {t}. Following Dijkstra’s algorithm
analysis, we know that the distance under scenario y from u to t in G[S ∪ {t}] is in fact
the distance between u and t in G. This allows us to determine if there exists a shortest
path from u to t in the scenario y passing via (u, v) and to color the vertex t accordingly.
We end-up the iteration with a new vertex t whose color is known and that can be added
to S before starting the next iteration. The algorithm terminates when S = V.
For every vertex w ∈ V − S, the estimated distance d(w) is the length of a uw-path
under the scenario y in the subgraph G[S ∪{w}]. An arc a is blue if a = uv or if its origin
is blue, the other arcs are red, i.e. a is blue if ya = 0 and red if ya = 1. The estimated
color γ(w) of w is blue if there exists a blue uw-path of length d(w) in G[S ∪{w}] and red
otherwise. The correctness of Algorithm 1 follows from the following Lemma.
Lemma 5. For every vertex t ∈ S, γ(t) is blue if and only if t ∈ Fy(u, v).
Proof. We proceed by induction on the number of vertices of S. When S = {u}, the
property is verified. Now, suppose the property true before the insertion in S of the
13
Algorithm 1: Computes the set S of vertices t such that (u, v) is t-strong
Input : G = (V,A, l, l−), (u, v) ∈ A
Output: {t ∈ V : (u, v) is t-strong}
foreach (u,w) ∈ δoutu \ {(u, v)} dod(w)← l−uw
γ(w)← red
d(v)← luv; γ(v)← blue
S ← {u}; γ(u)← red
while S 6= V do
Pick t ∈ V − S with smallest d(t) breaking tie by choosing a vertex t such that
γ(t) = blue if it exists
if γ(t) is blue then
foreach (t, w) ∈ δoutt such that d(w) ≥ d(t) + ltw do
d(w)← d(t) + ltw
γ(w)← blue
else
foreach (t, w) ∈ δoutt such that d(w) > d(t) + l−tw do
d(w)← d(t) + l−tw
γ(w)← red
S ← S ∪ {t}return {t ∈ V : γ(t) = blue}
vertex t such that d(t) is minimum. By induction hypothesis, the lengths of arcs having
their source in S are set according to y. Therefore, following Dijkstra’s algorithm analysis,
we deduce that d(t) is the length of the shortest path from u to t in the graph G under
scenario y. If t has been colored blue then there exists a blue vertex w ∈ S such that
d(t) = d(w) + lwt. By induction hypothesis w ∈ Fy(u, v) and there exists a shortest ut-
path under scenario y containing (u, v), i.e. t ∈ Fy(u, v). Now, suppose that t has been
colored in red. By contradiction, assume there exists a shortest ut-path under scenario
y that contains (u, v). This path cannot contain a vertex outside S except t because
otherwise, since arc lengths are non-negative, its length according to y would be greater
than d(t) by the choice of t. Therefore, the predecessor w of t in this path belongs to S.
Since w belongs to a shortest ut-path passing via (u, v), by induction, it was colored blue.
But in this case, there exists a blue vertex w such that d(w) + l(w, t) = d(t), and t was
colored blue as well, a contradiction.
14
We are now ready to state the main result of this section.
Proposition 1. Given a graph G = (V,A), two arc-length functions l− and l such that 0 ≤
l−a ≤ la for every arc a ∈ A, and an arc (u, v), Algorithm 1 computes in O(|A|+|V | log |V |)
the set of vertex t such that (u, v) is t-strong.
Proof. The correctness of Algorithm 1 is given by Lemma 4 and Lemma 5. Analogously
to the Dijkstra’s algorithm, it can be implemented to run in O(|A| + |V | log |V |) using a
Fibonacci heap as priority queue.
4.3. Computing W (t) for all t
The algorithm that computes, for every arc (u, v), the set of vertices t such that
(u, v) is t-useless is obtained from Algorithm 1 by applying two small changes. Before
describing these changes, we explain how the two problems are related. For that, we first
introduced a strengthening of the notion of strongness. We will say that an arc (u, v)
is strictly t-strong if it belongs to all shortest ut-paths for every scenario x ∈ {0, 1}Φ.
Recall that t-strongness requires only the existence of a shortest ut-path passing via (u, v)
for every scenario x ∈ {0, 1}Φ. Algorithm 1 can be easily adapted to compute for every
arc (u, v) the set of vertex t such that (u, v) is strictly t-strong. It suffices to change
the way, the algorithms breaks tie between a red and a blue path and the choice of t
in case of tie. Namely, in the first internal loop, the condition for coloring w in blue
becomes d(w) > d(t) + ltw while the condition for coloring w in red in the second internal
loop becomes d(w) ≥ d(t) + l−tw. Moreover, when we choose t such that d(t) is minimal,
we break tie by choosing a red vertex if it exists. Clearly, these small changes exclude
the existence of a red path of length d(w) between u and a blue vertex w. Therefore,
(u, v) belongs to every path of length d(w) and (u, v) is strictly w-strong whenever w is
blue. We call the resulting algorithm the strict version of Algorithm 1. The next step
is to extend the notion of strict strongness to a subset of arcs having the same source.
For any vertex u ∈ V, a subset Γ ⊆ Γ+u of arcs is strictly t-strong if, for every scenario
x ∈ {0, 1}Φ, all shortest ut-paths intersect Γ. By definition, an arc (u, v) is t-useless if and
only if Γ+u \ {(u, v)} is strictly t-strong. Indeed, every ut-path avoiding (u, v) intersects
Γ+u \{(u, v)} and, conversely, (u, v) belongs to every ut-path avoiding Γ+
u \{(u, v)}. Hence,
computing the set of vertex t such that (u, v) is t-useless amounts to compute the set
of vertex t such that Γ+u \ {(u, v)} is strictly t-strong. An algorithm that computes this
15
set of vertices can be obtained from the strict version of Algorithm 1 by modifying only
the initialization step: arc (u, v) is colored in red and its length is set to l−uv while arcs
of δoutu \ {(u, v)} are colored in blue and their lengths are set to their upper bound. A
correctness proof very similar to the one of Algorithm 1 (and that we will not repeat)
shows that a vertex is colored blue by Algorithm 2 if and only if (u, v) is t-useless. Since
the two algorithms have clearly the same time complexity, we conclude this section with
the following result.
Algorithm 2: Computes the set of vertex t such that (u, v) is t-useless
Input : G = (V,A, l, l−), (u, v) ∈ A
Output: {t ∈ V : (u, v) is t-useless}
d(v)← l−uv; γ(v)← red
foreach (u,w) ∈ δoutu \ {(u, v)} dod(w)← luw
γ(w)← blue
S ← {u}; γ(u)← blue
while S 6= V do
Pick t ∈ V − S with smallest d(t) breaking tie by choosing a vertex t such that
γ(t) = red if it exists
if γ(t) is blue then
foreach (t, w) ∈ δoutt such that d(w) > d(t) + ltw do
d(w)← d(t) + ltw
γ(w)← blue
else
foreach (t, w) ∈ δoutt such that d(w) ≥ d(t) + l−tw do
d(w)← d(t) + l−tw
γ(w)← red
S ← S ∪ {t}return {t ∈ V : γ(t) = blue}
Proposition 2. Given a graph G = (V,A), two arc-length functions l− and l such that 0 ≤
l−a ≤ la for every arc a ∈ A, and an arc (u, v), Algorithm 2 computes in O(|A|+|V | log |V |)
the set of vertex t such that (u, v) is t-useless.
16
5. Numerical experiments
In this section, we report our computational experiments in order to demonstrate the
added benefit of our MIP formulation and preprocessing step. The numerical experiments
were performed on a desktop computer equipped with an Intel(R) Core(TM) i7-8700k 4.8
gigahertz and a memory of 16 gigabytes. Our model was implemented in Gurobi 9.0.1
with default parameters and the other algorithms were coded in multithreaded C++ using
the LEMON Graph Library.
5.1. Quality of the solutions
In the lack of an efficient method to compute an optimal solution of BC-ECA-Opt,
ecologists often use greedy algorithms to compute sub-optimal solutions. In this subsec-
tion we will compare an optimal solution obtained with our MIP model to the solutions
obtained with different greedy algorithms. The incremental greedy algorithm (IG) starts
from the graph with no improved element, at each step i, the algorithm selects the ele-
ment e with the greatest ratio δie/ce until no more element fits in the budget. Here, δie
denotes the difference between the value of ECA with and without the improvement of
the element e at the step i and ce is the cost of improving the element e (that can be
either an arc or a vertex). The decremental greedy (DG) starts from the graph with all
improvements performed and iteratively removes the improvement of the element e with
the smallest ratio δie/ce. DG finishes with incremental steps to ensure there is no free bud-
get left. These algorithms perform at most |Φ| steps and at each step i need to compute
δie for each element e. It is easy to implement the IG algorithm in O(|V |3 + |Φ|2 · |V |2)
by using an all pair shortest path algorithm to compute in O(|V |3) the initial distance
matrix and then by performing |Φ| steps in which the computation of δie for each arc
e ∈ Φ takes O(|V |2). Indeed, when we decrease the length of an edge we can update the
distance matrix in O(|V |2). For the implementation of DG, when we increase the length
of an edge we cannot update the distance matrix as easily as for the IG algorithm. We
can recompute in |V |3 the whole distance matrix for computing each δie and the algorithm
runs in O(|Φ|2 · |V |3) (we did not try to improve the implementation). These complexities
are already too large for the practical instances handled by ecologists which can have few
thousands of patches. Most of the studies using the PC or ECA indicators use simpler
algorithms that we call static increasing (SI) and static decreasing (SD). These algorithms
are simpler variants of the greedy algorithms which do not recompute the ratio δe/ce of
17
each element e at each step and thus are faster but do not account for cumulative effects
nor redundancies.
In the next section, we provide instances on which IG and DG performs poorly com-
pared to an optimal solution. On these instances, it is easy to check that the solutions
returned by algorithms SI and SD are not better than the solutions returned by algorithms
IG and DG.
5.1.1. Limits of the greedy algorithms
In the instances presented below, all arcs have a probability 1 if improved and 0
otherwise and have unitary costs. Recall that a spider is a tree (sorry for ecologists)
consisting of several paths glued together on a central vertex.
Case A: The graph is a spider with 2k branches: k long branches with two edges, an
intermediate node of weight 0 and a leaf node of weight 1, and k short branches consisting
of a single edge with a leaf of very small weight ε > 0, see Fig.1 (a). All branches are
connected to a central node of weight 1. IG performs poorly on this instance. Indeed,
IG is tricked into purchasing short branches with very small ECA improvement because
purchasing an edge of a long branch alone does not increase ECA at all. An optimal
solution results in larger value of ECA by improving pairs of arcs of long branches.
In this case, IG does not perform well while DG finds an optimal solution computed by
the MIP solver except when the budget is 1. In this case, the reverse occurs: DG performs
badly while IG is optimal. Indeed, DG realizes that the budget is not sufficient to improve
two arcs of a long branch only after removing the improvements of all short branches.
Case B: The graph is obtained from a star with k + 1 branches by replacing one branch
by a path of length k. The central node and all leaves except the leaf of the path have
weight 1. The leaf of the path has weight 1+ε. The internal nodes of the path have weight
0. DG performs poorly on this instance because it removes one by one the branches of
the star for which δe/ce = 1 before removing an edge of the path for which δe/ce = 1 + ε.
When the budget is at least 2, an optimal solution removes all the edges of the path before
removing an edge of another branch.
Case C: The graph is a spider with k+ 1 branches. All branches except one are paths of
length 2 with an internal node of weight 0 and a leaf of weight 1. The last branch is a path
18
8
7
6
5
4
3
2
10
(a)
0 2 4 6 8 10 12budget
0.0
0.2
0.4
0.6
0.8
1.0
opt
rati
o
incremental greedydecremental greedyoptimum (MIP)
(b)
Figure 2: An instance on which Incremental Greedy fails. (a) the graph of case A with k = 4, (b) ratio of
the increase in ECA between the solutions returned by IG and DG and an optimal solution for several
budgets.
����
65
4
3 2
10
(a)
0 2 4 6 8 10budget
0.0
0.2
0.4
0.6
0.8
1.0
opt
rati
o
incremental greedydecremental greedyoptimum (MIP)
(b)
Figure 3: An instance on which Decremental Greedy fails. (a) the graph of case B with k = 5, (b) ratio
of the increase in ECA between the solutions returned by IG and DG and an optimal solution for several
budgets.
of length 2k with internal nodes of weight ε > 0 and a leaf of weight 1 + ε. All branches
intersect in a central node of weight 1. In this case, both Incremental and Decremental
Greedy fail. On one hand, IG selects the edges of the path of length 2k one by one and
does not realize that by taking two edges of a short branch it could improve much more
ECA. On the other hand, DG removes first the edges of the short branch because the
weight of leaf of a long branche is 1 + ε while the weight of the leaf of a short branch is 1.
Hence, DG and IG return the same law quality solution.
19
151413121110987
65
4
3 2
10
(a)
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0budget
0.0
0.2
0.4
0.6
0.8
1.0
opt
rati
o
incremental greedydecremental greedyoptimum (MIP)
(b)
Figure 4: An instance on which both Incremental and Decremental Greedy algorithms fail. (a) the graph
of case C with k = 5, (b) ratio of the increase in ECA between the solutions returned by IG and DG and
an optimal solution for several budgets.
Case C illustrates the fact that the greedy algorithms IG and DG do not provide any
constant approximation guarantee (even for trees), i.e. for any constant 0 < α < 1 there
exists an instance of BC-ECA-Opt such that ALG < αOPT where ALG is the value of
ECA for the best solution among those returned by IG and DG and OPT is the value of
ECA for an optimal solution.
5.1.2. In Practice
In this subsection we investigate on the performances of the greedy algorithms we
discuss in the previous subsection on a real instance. We perform these numerical experi-
ments on a model of the landscape around Montreal from [2]. The landscape model of [2]
is a Gabriel graph – a subset of the Delaunay triangulation – of 8733 patches. For each
patch, we have an estimation of its area that could be lost by 2050. In order to obtain
20
several examples with a reasonable computing time, we extract 72 subgraphs of sizes be-
tween 150 and 400 nodes from this large instance. Then, we consider that protecting a
patch corresponds to buying the portion of it that would otherwise be lost and has a cost
proportional to the purchased area. To model the protection of a patch u, we set its de-
fault weight, wu, to the area it would have in 2050 if unprotected and its improved weight,
w+u , to its current area. If a patch is threatened of total loss, i.e. all its area would be lost
by 2050, we prevent it to contribute to shortests paths if it is not purchased as follows.
We add a node uout, replace each outgoing arc (u, v) by an arc (uout, v) and connect u and
uout with an arc (u, uout) of default probability 0 and enhanced probability 1. In this way,
if a patch threatened of total loss is not protected then it does not contribute to ECA by
its quality nor by its use as stepping stone for other pairs of nodes.
82468245 8244824382428241 8240823982388237
8236
8235
82348233
82328231 8230 82298228822782268225 822482238222 8221
8220
82198218
8217821682158214 821382128211 8210820982088207
8206
8205
8204
8203
8202
820182008199
8198 8197
8195
8194
81938192
81918190
8189
8188
8187
8186
8185
8184
81838182
8181 8180817981788177
81768175
8174
8173 8172
8171
8170
8169
81688167
8166
8164
8163
8162
8161
8160
81598158
81578156
8155 8154
81538152
8151
81508149
8148
8147
81468145
8144
8143
8142
81418140
8139
8138
8137
8136 8135
8134
8133
81328131
8130
8129
8128
8127
8126
8125
8124
8123
8122
81218120
81198118
81178116
8115
8114
8113 8112
8111
8110
8109
8108
8107
81068105
8104
8103
8102
8101
8100
80998098
80968095
8094
8093
8092 8091
8090
8089
8088
80878086
8085
8084
8083 8082808180808079
8078
8077
8076
8074
80738072
8071
8070
8069
8067
8066
8065
8064
8063
8062
8061
8060
8059 8058
8057
805680558054
8053
8052
8051
8049
8048
8047
8046
8045
8044
8043
8042
8041
8040
8039
80388037
8036
8035
8034
8033
8032
8031
8030
8029802880278026
8025
80248023
8022
8021
8018
8017
8016
80158014
8013
8012
8011
8010 8009
8008
8007
8006
8005
8004
80038002
8001
8000
7999
7998
7997
7996
7994
7993 7992
7991
7990
7989
7988
7987
7986
7985 79847983
7982
7981
7980797979787976
7975
7974
7973
7972
7971
79707969
7968
7967
79667965
7964
7963
7962
7960
7958 79577956
7955
7954
7953
79527951
7950
7949
7948
7947 7945
7944
7943
7942
7940
7939
79387937
7936
7935 7934
7933
7932
7931
7930
7929
7928
7927
79267925
7924
79237922
7921
7920
7919
7918
7917
7916
79157914 7913
7912
7911
7910
7909790879057904
7903
7902
7901
79007899
7898
7897
7896
7895
7894
7893
7892
7891
7890
7889
7888 7886
7885
7884
7883
7882
7881
7880
7879
7878 7876
7875
7874
7873
7872
78717870
7869
7868
7867
7866
7864
7863
7862
7861
7860
7859
7858
7857
7856
7855
7854
7853
7852
7851
7850
7849
7848
7846
7845
7843
7842
7841
7839
78387837 7836
7834
7833
7832
7831
7830
7829
7828
7827
78267825
78247823
7822
7821
7820
7819
7818
7817
7816
7815
7814
7813
7812
7811
7810
7809
7808 7807
7806
7805
7804
7803
7802
7801
7800
7799
7798
7797
7796
7795
7794
7793
77927791
7790 778977887787
7786
77857784
7783
7782
7780
7779
7778
77777776
7775
7774
7773
77717770
776877677766
7765
7764
77637762 7761
77607759
7758
7757
7756
775577547753
7752
7751 77507749
7748
7747
7746
774577447743
77427741
7740
7739
7738
7737
7736 77357734
7733
7732
7731
7730
77297728
7726
7725
7724
77237722
77217720
7719
7718
7717
7716
7715
7714
77137712
7711
7710
7709
7708
7707
7706
7705
7704
7703
7702
7701
7700 7699
7698
7697
7696
7695
7694
7693
7692
7690
7689
7688
7687
7685
7684
7683
7682
7681
7680
7679
76787677
76767675
7674
7673
7672
7671
7669
7667
7666
7665
7664
7663
7662
7661
7660
76597658
7657
7656
76557654
7653
7652
7651 7650
7649
7648
7647
7646
7645
7644
7643
7642
7641
76407639 76387637
7636
7635
7634
7632
7631
7630
7629
7628
762776267625
76247623
7621
7620
7619
7618
7617
7616
7615
7614
7613
7612
7611
76107609
7608
7607
7606
7605
7604
7603
7602
7601
7600
7599
7598
7597
7596
7595
75947593
75927591
7588
7587 7586
75857583 7582
7581 7580
757975787577 7576
7575
7574
7573
7572
75717570
7569
7568
7567
7566
7565
7564
7563 7562
7561
7560
7559
7558
7557
7556
75557554
7553
755275517550
7549
7548
75477545
7544
7543
7542 7541
7540
7539
7538
7537
75367535
7534
75337532
7531
7530
7529
7528
7527
7526
7525
7524
7523
75227521
7520
7519
7518 7517
7516
7515
7514
7513 7512
7511
7510
7509 7508
7507
75067505
7504
7503 7502 7501
7500
7499
7498
7497
7496 7495
7494
7493
7492
7491
7490
7489
7488
7487
7486
7485 7484
7483
7482
7481
74807479
7478
747774767475 7474
7473
7472
7471
7470
7469
74687467
74667465
7464
7463
7462
7461
7460
7459
7458
7457
7456
7455 7453
7452
7451
74507449
7448
7447
7446 74457444 7443
7442
7441
74407439
74387436
7435
7434
7433
7432
7431
7430
7429
7428
7427
7426
7425
7424 74237422
7421
7420
7419
7418 7417
7416
7415
7414
7413
7412
7411 74107409
7408
7407
74067405
7404
7403
7402
7398
7397
7396739573937392
7391
7390
7389 7388
73877386
7384
7383
7380
7379
7378
7376
7375
7374
7373
7372
7371
7369
7368
7367
7366
7365
7364
7363
7361 7360
7359
7358
7357
7356
7355
7354
7352 73517350
7348
7347
7346
7345
7344
73437342
7341
734073397338
7336
7334
7333
7332
7331
7330
7329
7328
7327
73267325
7323
7322
7321
7320
7319
7318
7317
7316
7315
7314
73137312 7311
7309
7308
7307
7306
7305
7304
7303 7302
7301
7300
72997298
7296
7295
7294
7293
7292
7291
7290
7289
7288
7287
7286
7285 7283
7282
7281
7280 7279
7278
7276 7275727472737272
7271
7270
7269
7267
7266
7265
7264
7263
7262
7261
7260
7259
7257
7256
72557253
7252
7250
7249
7248
7247
7246
7245
7244
7243
7242
7241
7240
7239
7238
7237
7236
7235
72347233 7232
7231
72307229
7228
7227
7226
7225
7224
7223
7222
7221
7220
72197218
7217
7216
7215
7214
7213
72107209
7208
7207
7206
7205 7204
7203 7202
7201
7199
7198
7197
7196
7195 7194
7190
7189
7188
7187
7186 7185
7182 7181
7180
7179
7178 7177
7176
7175
7173
7171
7169
7168
7167
7166
7165
7164
7163 7162
71617160
7159
71587157
7156 71557154
7153
7152
7150
7149
7148
7147
7145 7144
7143
71427141
7140
7139
7138
7136
7135
7134
7132
7131
7130
7129
7128
71277126
7125
7123
7121
7120
7119
7118
7117
7116
7115
7114
7113
7112
7111
7110
7109
7108
7107
71067104 7103
7102
71017100
7098
7097
70967095
7094
7093
7092
7091
7090
7089
7088
7087
7086
7085
7084
7083
7082
7081
7079
7078
70777076 7074
70737072
7071 70707069
7068
7067
7066
7065
7064
7062
706170607059
7058
7057
70567055
70547053
7052
70517050 7049
7048
7047
7046
7045
7044
7043
7042
7041
7040
7039
7038
7037
7036
7035
7034
7032
7031
7030
70297028
70277026
7025
70247023
7022
7021
7020
7019
7018
7017
7016
7015
7014 7013
7012
701070097008 7007
7006
7005
7004
7002
7001
7000
69996997
69966995
6994
6993
6991 6990
69896988
6987
6986
69856984
6983
6982
6981
6980
6979
6978
6976
6975
6974
6973
6971
6970
69696968
6967
6966
6963
6962
6961
6960
6959
69586957 6956
6955
6954
6953
6952
6951
6950
6949
6948
6947 6946 6945
69446943
6942
6941
6940
69396938
6937
6936
6935
6934
6933
6932
6930
6929
6927
69256924
6923
6922
6921 6920
6919
6918
6917 6916
6915 6913
6912 6911
6910 6909
6908 6907
6906
6905
6904
6903
6901
69006899
6898
6897
6896
689368926891
6890
6889
6888
6886 6885
68846883
6882
6881
6880
6879
6878
6877
6876
6875
6874
6872
68716870
6869
6868
6867
6866
6865686368626861 6860
6859
6858
6857
6856
6855
6854
6853
6852
6851
68506849
68486847
68466845
6844
6843
6842
6841
6840
6839 6838
68376836
6835
6834
6833
68326831 6830
6829
6827
6826
6824
6823
6822
6821
68206819
6818
6817
6816
6815
6814 6813
6812
6811 6810
6809
68086807
6806
6805
6804
68036802 6801
6800
6799
6798
6797
6796
67956794
6793
6792
6791
6790
6789
6787
6786
6785
6784
6783
6781
6780 6779
6778
6777
6776
6775
6774
6773
6770
6769
6768
6767
6766
67656764
6762
6761
6760
6759
6757
6756
6755
67546753
6752
6750
6749
6748
6747
6746 6745
6744
6743
6742
6741
6740
6739
6738
6737
67366735 6732
6731
67306729
6728
6727
6725 6724
67236722
672167206719
6718
6717 6716
6715
6714
6713
6712
6711
6710
6709
6708 6707
67066705
67046703
6702
6701
6700
66996698 66976696
6695
6694
6693
6692
6691
6690
6689
6688
66876686
6684
6683
6682
6681
6679
6677
6675 6673 6672
6670 6669
6668
6667
66666665
6664
6663
66626661 66606659
6658
6657
6656
6655
6654
6653
66516650
6649
6648 66476646 6644
6643
6642 6641
6640
6639
6638
6637
6636
6635
6634
6633
6632
6631
6630
6629
6628 66276626
6625
6624
66236622
6621
6620
66196618 6617
6614
6613 6612
6611
6610
6609
6607
66066605
6604
6603
6602
66016600
65986597
6596
6595
6594
65936592 6591
6590
6589
6588
6587
6586
6585
6584
6583
6582
6581
6580
65796578
6577 6576
6575
6574
6573
6572
6571 6570
6569
6568
6566
6565
6564
6563
6562
6561
6559
6557
6556
6555
6554
6553
6552
6551
65506549
6548
6547 6546
6545
6544
6543
6542
6541
6540
6538
6537
6536
65356534 6533
65326531
6530
6529
6528
652665256523
6522
6521
6520
6519
65176516
6515 6514
6513
6512
6511
6510
6509
6508
6507
6506
6505
6504
6502 6501
6500
6499 6498
6497
6496 6495
6493
6492
6491
6490
6489
6488
6487
6485
6484
6483
64826481
6480
6479
6478 647664756474
6473
6472
6471
6469
6468
6467
6466
6465
6464
6463
6461
6460 6459
6458
6457
64566454
6453
6451
6450
6449
6448
6447
6446 6445
6444
64436442
6441
6440
6439
6438
6437
6436 6435
6434
6433
64326431 6430
6429
64286427
6426
6425
6424
6423
6422
6421
6420
6419
6417
6416 6415
6414
6413
6412 6411
64106409
6408
6407
6406
64056404
6403
6402
64016400 6399
6398
6397
6396
6395
6394
6393
63926391
6390
6389
6388
6387
6386
6385
6384
63836382
6380
6379
6378
6377
63766375 6374
6373
6372
6371
63706369
63686367 6365
6362
6361
6360
6359
6358
6357
6356 6355
6354
6353
6352 6351
63506349
6348
6346
6345
6344 63436341
6340
6339
6338
6337
6335
6334 6333
63326331
6330
6329
6328
6327
6326
6325
6324
6323
6321
6320
6319
6318
6317
6316
6315
6314
6313
6312
6311 6310
63096308
6307
6306
6305 6304
6303
6302
6301
6300
6299 6298 6297
6296
6295
6294
6293
6291
6290 628962886287
6286
6285
6284
6283
6282
6281
6280
6278
6277
6276
627562746273
6272
6271 6270
6269
6268
6266
6265
6264
6263
6262
6261
6260
6259
62586257
6256
6255
6254
6253
6252
6251
62506249
6248
62476246
62456244
6243
6242
6241
6240
6239
6238
6236
6235
6234
6233
6231
6230 6229
6228
6227
62266225
6224
6223 6222
62216220
6219
6218
62176216
6215
6214
6213
6212
6210
6209
62086207
6205
6204
6203
62026201
6200
61996198
6197
6196
61956193
61926191
61906189
61886187
6185
6182
6181
6180
6179
61786176 6175
6174
6173
6172
6171
6170 6169
61686167
6166
61656164
616361616159
6158
6157
6156
6155
6154
6153
6152
61516150 6149
6148
6147
6146
6145
6143
6142 6141
61406139
61386137
6136
6135
6134
6133
6132
6131
6130
6128 61276126
6125
6124
61236121
6120
6119
611861176116
6115
6114
6113
6112
6111
6109
6108
6107
6106
61056104
6103 6101
6100
6099
6098
6097
6096
6095
6094
6093
6092
60916090
6089
6088
6087
6086
6084
6083
6081
6080
6079
60786077
6076
6075
6074
6073 6072
6071
6069
6068 6067
6066
6065 6064
6062
6061
6060
6059
6058
6057
6056
6055
6054
6053
6052
6051
6050
6049
6047
604560446043
6041
6039
60386037
60366035
6034
6033 6032 6030
6029
6028
6027
6026
6025
6024
6023 6022
6021
6020
6019
6018
6017
6016
6015
6014
6013
6012
6011
6010
6009
6008
60066005
6004
6002
6001 6000
5999
5998
5997 5996
5995
5993
5992 5991
5990
5988 5987
5986
5985
59845983
5982
5981
5980
5979
5978
5977
5976
5975
5974
5973
5972
5971
5970
5968
59675966
5965 5964
5963
5962
5961
59605959
5958
5957
59565955
5954
5953
5952
59505949 5948
5947
5945
5944
5943
5941
5940
5939
5938 5936
5934
5932
5930
59295928
5927
5925
5924
5922
59215920
5919
59185917
5916
5915
5914
5913
5912
5911
5909
5908
5906
5905
5903
5902 5901
5900
5899
5898
5897
5896
5895
5894
5893 5892
5891
5890
5889
5888
5887
58865884
5883
5882
5881
5880
58795878
58775876
5875
5874 5873
5872
5871
5870
5869
58685867 58665865 58645863
5862
5861
5860
5858
5857 58555854
5853
5852
5851
5850
5849
5848
5846
5843
5842
58405839
5838
5837
5836
5835
5834
5833
5832 5831
5830
5829
5828 5826
5825
5824 5823
5822
58215820 5819
5818
5817
5815
5814
5813
58125811
5809
5808 5807
5806
5804
5802
5800
5798
5797
5796
5795 5794 5793 57925790
5789 5787
5786
5785
5784
5783 5782
5781
5780
5779
5777
5776
5775
5774
57735772
5771
5770
57695768
5766
5765
5764
5763
5761
5759
5758
5757
5756
5755
5754
57535752
5751
5750
5749
5748
5747
5746
5745 5744
5743
5742
5741
5740
5739
5738
5737
5736
5735
5734
5733
5732
5731
5730
5729
5728
5727
5726
5725
5724
5723
5722
5721
5720
5719
5718
57165715
5714
5713
5712
5710
57095708
5707
5706
5704
5703 5702 5701
5700
5699
5698
5697
5696
5695
5694
5693
5692
5691
5690
5689 5688
5687
56865685
5684
5683
5682
56815680
5679
5678
5677
5676
5675
5674
5673
5672
5671
5670 5669
56685667
5666
5664
5663
5662
5661
56605659
56575656 5654
5653
5652
5650
5649
5648
56475646
5645
5644
5643
5642
5641
5640
5639 56385637
5636
5635
5634 56335632
5631
56305629
5628
56275626
5625
5624
5623
56225621
5620
5619
5617 5616
5615
5614
56125611
5610 5609
5608
5607
5606
5605 5604
5603
5602
5601
5600
5599 5598
5597
5596
5595
5594
5593
5592
5591
5590
5589
5588 5587
5586
55855584
5583
5582
5581
5580
5579
55785577 5576
5575
5574
5573
5572
5571
5570
5569
5568
55675566
5565
5564
55635562
5561
5560
5559
5558
5557
5556
55555554
5553
5552
55515550
5549
5548
55475546
55455544
5543
5542
5541
5540
5539
5538
5537
55365535
55345533 5532
5531
5530
5529
5528
5527
5526
5525
5524
5523
5522
55215520
5519
55185517
5516
5515
5514
5513
5512
5511
5510
5509
5508 5507
5506
5505
5504
5503
5502
5501
5500
5499
5498
5497
5496
5495
5494
5493
5492
5491
5490
5487
5486
5485
5484 5483
5482
5481
5480
54795478
5477
5476
5475
5474
5473
5472
5471
5470
5469
5468
5467
5466
5465
54645463
5462
5461
5460
5459
5458
5457
5456
5455
5454
54535452
5451
5450
5449
5448
5447
5446
5445
5444
5443
5442
5441
54405439
5438
5437
5436
5435
5434
54335432
5431
5430
54295428
5427
5426
5425
5424
5423
5422
5421
5420
54195418
5417
5416
5415
5414
5413
54125411
5410
5409
5408
5407
5406
5404
5403
5402
5401 5399
5398
5397 5396
5395
5394
5393
5392
53915390
5389
5388
5387
5386
5385
53845383
5381
5380
5378
537753765375
5374
5372
5371
5370
5369
5368
5367 5366
5365
5364
5363
5362
5361
5360
5359
5358
53575356 5355
53525351
5350
53495348
5347
5346
5345
5344 53435342
5341
5340
5339
53385337 53365334533353325331 5330
5329
5328 53265325
5324
5323
5322
5321
5320
5319
5318
5317
5315
5314
5313
5312
5311
53105309
5308
5307
5306
53055304
5303
53025301
5300
52995298
52975295
5293
5292
5291 5290 5289
52885287
5286
5285
5284
5283
52825281 5280 527952785277
5276
5275
5274
5273
5272
5271
5270
5269
5268
5267
5266
5265
5264
5263
5262
5261 5260
5259
5258
5257
5256
5254 5252
5251
5250
5249
5248
5247 5246
5245
5244
52435242
5240
5239
5238
5237
52365235 52345233
5232
5231
5230522952285227
5226
52255224 5223
52225221
5220
5219
5218
52175216
5215 5214
5213
5212
5211
5210 5209
5208
5207
5206
5205
5204
5203
5202
5201
5200
5199
5198
5197
51965195
51945193 5192
5190
5189
5188
5187
5186
5185
5184
51835182
5181
5180
5179
5178
5177
5176
51755174
5173
5172
5171
5170
5169
5168
5167 51665165
5164
5163 5162
5161 51605159 51585157
51565155
5154
5153
5152
5151
51505149
5148
5147
5146
5145
5144
514351425141
5140
5139
5138
5136
5135
5134
5133
5132
5131
5130
512951285127
5124
5123
5122
5121
5120
5119
5118
5117
5116 5115
5114
5113
5112
5111
51105109
5108
51075106
5105
51045103
5102 51015100
50995098
5097
5096
5095
5094
5093
5092
5091
5090 5089 5088
5087
5086
5085
5084
5083
5082
508150805079
5078
5077
5076
5075
5074
5073
5072
5071
5070
50695068
5067
5066
5065
5064
5063
5062
5061
5060
5059
50585057
5056
5054
50535052
5051
5049
5048
5047
5046
5045
50445043
5042
5041
5040
5039
5038
5037
5036
5034
5033
5032
5031
5030
50295028
5027
5026
5025
5024
5023 50225021
502050195018 5017
5016
5015
5014
5013
5012
5011
5010
5009
5008
5007
5006
5005 5004
5003
5002
5001
5000
4999
4998 499749964995
4994
4993
4992
4991
4990
4989
4988
4987
4986
4985
4984 4983
4982
4981 4980
4979
4978
4977
4976
4975
49744973
4972
4971 4970
4969
49684967
4966
49654964
4963
4962
4961
49604959
4958
4957
4956
4955
49544953
4952 4951
4950
4949
4948
4947
4946
4945
4944 4943
4942
4941
49404939
49384937
4936
4935
4934
4933
4932 4931
4930 4929
4928
4927
4926
4925
492449234922
4921
4920
4919
49184917
4916
4915
4914
4913
4912
4911
4910
4909
49084907
4906
4905
49044903
4902
4901
4900
4899
4897
4896
48954894
4893
4892
4891
4890
48894888
4887
4886
4885
4884
4883
4881 4880
4879
4878
4877
4876
4874
4873
48724871
4870
486948684867
4866
4865
48644863
4862
4861
4860
485948584857
4856
4855
4854
4853
4852
4851
48504849
4848
4847
4846
4845
4844
4843
4842
48414839
4838
4837
48364835
4834 4832
4831
4830
4829
4828
4827
4825
4824
4823
4822
4820
48194818
4817
4816
4815
48144813
4812
48114809
4808
4807
4806
4805
4804
4803 4802
48014799
4798
4797
4796
4795
4794
4793 4792 4791
4790
47894788
4787
4786
4784
478347824781
4780
4779 4778
4777 47764775
4774
477347724771
4770
4769
476747664765
4764
4763
4762
4761
4760
4759
4758
4757
4756 4755
4754
47534752
4751 475047494748
4747
4746 474547444743
4742
4741
4739
4738 4737
4736
473547344733
4732
47304729
4728
4727
4726
4725
4724
4723
4722 4721
4720
4719
4718
47154714
4713
4712
4711
4710 4709
4708
4707
4706
4705
4704
4703 4701
4700
4699
4698
4697
4696
4695
4694
46934692
4691 4690
4689
4688
4687
4686
4685
4684
46834682
4681
4680
4679
4678
46774676
4675 4674
4673
4672
4671
46704669
4668
4667
4666
4665
4664
46624661 4660
4659 4658
4657
4656 46554654
4653
4651
4650
4649
46484647
4646
4645
4644 4642
4641
4640
4639
4638
4637
4636
4635
4634
46334632 46314630
4629
4628
4627
46264625
4624
4623
4622
4621
4620
4619
4618
4617
46164615
4614
4613
46124611
46104609
4608
4606 4605
4604
4603
4602
46014600 4599
4597 4596
4595
4594
4593
4592
4591
45904589
4588
4587
4586
4585
4584
4583
4582
45814580
45794578
45774576 4574
4573
4572
4571
4570
4568 45674566
45654564
45634562
4561 456045594558
4557
4556
4555
45544552
4551 4550
4549
4548 4547
4546
4545
4544
4543
4542
4541
4540
4539
4538
4537
4536
45354534
4533
4532
45314530
4529
4528
4527 4526 4525
4524
45234522
4521
4520
4519
4518 4517
4516
4515
4514
4513
4512
4511
4510
4509
4508
4507
4506
4505
4504
4503
4502
4501 4500
44994498
4497
4496
4495
44944493
4491
4490
4489
4488
4487
44854484
44834482
44814480
4479
4478
4477
4476
4475
4474
4473
447244714470 44694468
4467
4466
4465
4464
4463
4462
4461
4460
4459 4458
4457
4456
4455
4454 44534452 445144504449
4448
4447
4446
4445
4444
4443
4442
443944384437
4436
4435
4434
4433
4432
4431
44304429
4428
4427
4426
4425
4424
44234421
4420
4419 4418
4417
4416
4415
4414
4413
4412
4411
4410
4409
4408
4406
4405
4404
4403 4402
4401
4400 4398
4397 4395
4394 43934392
43914390
4389
4388
43874385 43844383
4382
4381
4380
4379 4378
4377
43764375
4374
4373 4372
4371
4370
4369
4368
4367
4366
4365
4364
4363
43624361
4360
4359
4358
43574356
4355
4354
4353
4350 4348
4347 4346
4345
4344
43434342
4341
4340
4339
4338 4337
4336
4335
4334
433343324331
4330
43294328
4327
43264325 4324
4323
4322
4321
4319
4318
43174316
4315 4314
4313
4312 4311 4310
4308
43074306
4305 4304
4302
4301
43004299
4298
42974296
4295
4294
4293
4292
4291
4290
4289 4288 428742854284
4283
4282
4281
4280
4279
42784277 4276
4275
4274
4272
4271
42704269
4268
4267
42664265
4264 4263
4262
4260
4259 4258
4257
4256
4255 42544253
4252
42514250 4249
4248
4247 4246
4244
42434242
4241 4240
4239
4238
4237
4236
423542344233
4232
4231
4230
4228
42274226
4225
4224
4223
4222
4221
4220
4219
4218
4217
4216
4215
4214
4213
4212
4211
4210
4209
4208
4207
42054204
4203
4202
42014200
4199
41984197
4196 4195 4194
41924191
4190
4189
4188
4187
4186
4185 4184
418341824181
41804179 41784177 4176
4175
4174
4173
4172
41704169
4168
4167
4166
4165
4164
4163 4162
4161
4160
4159
4158
4157
4156 4155
4154
4153
4151 4150
4149
4147
4146
4145
4144
4143
4142
4141
4140
4139
41384137
4136
41354134 4133
4132
4131 4130
41294128 41274126 4125
4124
4123 4122
41214120
41194118
4116
4115
411441134112
4110
4109
4108
4107
4106
41054104 4103
4102
410141004099
4098
4097
4096
4095
4093
40924091
4090 4089
4087
4086
4085 4084
4083
4082
4080
4078
4077
4076
4075
4074
4072
4071
4070
4069
40684067
4066
40654064 406340624061
4060 4059
4058
4057
4056 4055
4054
4053 4052
4051
4050
4049
4048
4047
40464045
4044
4043 4042
4041 404040394038
4037 4036
4035
40344033
40324031
4030
40294028
4027
4026
4025
4024
4023
4022
4021
4020
40194018401740164015
4014
4013
4012
4011 401040094008
4007 40064005
4004
4003
4002 40013999
3997
3996
39953994
3993 39923991
3990
398939883987
3986
3985
3984
3983
3982
39813980
3979
3978
3977
39763975
3974
3972
3971 3970
3968
3967
3966
3965
39643963
3962
3961
39603959
3958
3957
3956
3955
3954
3953
3952395139503949
3948
39473946
3945
3943
3942
3941
3940
3939
3938
3937
3936
39353934 3932
3931
39303929
3928
3927
3926 3925
3923
3922
39213920
39183917
3916
39153914
3913
3912 3911
3910
3909 3908
3906
39053904
39033902
3901
3900
3899
38983897
3896
3895 3894
3893
3892
3891
3890
3889
3888
3887
3886 3885
3884
3883
3881
3880
38793878
38773876 3875
3874
3873
3872
3871
3870
3868
38673866
3865
3864
3863
3862
3861 3860
3859
38583857
3856
3855
3854
3853
3852
3851 3850
3849
3848
3847
3846
3845
3843
38423841
3840
3839
3838
3837
3836
3835
3834
3833
3832
3831
3830
38293828
3827 3826
3825
38243823
3822
3821
3820
3819
3818
3817
38163815
3814381338123811
3809 3808
3807
3806
3805
38043803
3802
3801 3800
379937983797 3796
379537943793
3792
3791
3790
37893788
3787
3786
3785
3784
3783
37823781
3780
37793778 3777
3776
3775
3774
3773
3772
3771
3770
3769
3768
3767
3766
3765
3764
3763
3762
37613760
3759
3758
3757
37563755
37543753
3752
3751
3750
3749
3748
3747
3745
3744
3743 3742
3741
3740
3739
3738
3737
3736
3735
3734
3733
3732
3731
37303729
3728
3727
3726
3725
3724
3723
3722
3721
3720
3719
3718
3717
3716
3715
3714
3713
3712
3711
3710
3709
3708
37073706
3705
3704
3703
37023701 3700
3699369836973696
3695
3694
3693
3692
3691
3690
3687 3686
3685
3684
3683
3682
3681
36803679
367836773676
3675
3674
3673
36723671
3670
3669
3668
3667
3666
3665
3664
3663 3662
3661
3660
3659
3658
3657
3656
3655
3654 3653
3652
3651
3650
3649
3648
3645
3644
36433642
3641
3640
3639
3638
36373636 363536343633
3632
3631
36303629
3628
3626
3625
3624
3623
3622
3621
3620
3619
3618
36173616
3615
3614
3613
3612
36113610
360936073606
3604
3603 3602
3601
3600
3599
3598
3597
3596
3593
3592
3591
3590
3589
3588
35863585 35843583
3581
3580
3579
3578
3577
3576
357535743573
3572
3571
3570
3569 3568
3567
35663565
35643563
3562
3561
35603559
3558
3557
3556
3555 3554
3553
3552
3551
3550
3549
3548
3547
3546
35453544
3543
3542
3541
3540
3539
3538
3537
3536
3535
3534
3533
3532
3531
3530
3529
3528
3527 3526
3525
3524 3520
3519 3518
3517
3514
3513
35123511
3510
3509 3507
35063504 3503
3502
3501
350034993498
3497 3496
3495
3494
3493
3492
3491
3490
3489
34883487
3486
3485
3484
3483
3482
3481
3480
3479
3478
3477
3476 34753473
3472
3471
3470
3469
34683467
3466
3465
3464 3463
3462
3461
3460
3459
3458
3457
34563455
34543453
3452
34513450 3449
3448
344634453444
3443
344234413440 3438
3437
3436
3435
34343433
3432
3431
3430
34293428
3427
3426
3425
3424
342334223421
3420
3419
3418
3417
3416
3414
3413
3412
3411
3410
3409
3408
3407
3406
3405
3404
3403
3401 3400
3399
3398
3397
3396
3395
3394
3393
3392
3391
3390
3389
3388
3387
3386
3385
3384
33833382
3381
338033793378
3377 3376
3375
3374
3373
3372
3371
3370
3369
33683367
3366
3365
33643363
3362
3361
3360
3359 3358
3357
3356
3355
3353
3352
3351
3349
3348
3347
3346
3345 33443343
33423341
3339
3338
3337
3336 3335
3334
3332
3331
33303328 3327
3326
3325
3324
3323
3322
3321
3320
3319
3318
3317
3316 33153314
3313
3312 3311
3309
33083306
3305
3304
3303
3302
3301 3300
3299
3298 3297
3296
3295
3294
3293
32923291
3290
3289
3288
3287
3286
3285
3284
3283
3282
32813280 3279
3278
3277
3276
3275
3274
32733272
3271
3270
3269
3267
3266
3265
3263
3262
3261
3260 3259
3258
3256
3255
3254
32533252325132503249
3248
3247
3246
3244
3243
32423241
3240
3239
3238
3237
3236
32353234 3233
3232
3231
3230
3229 3228 3227
3226
3225
3224
3223
322232213220
3219
3218
3217
3216
32153214
3213
32123211
3210
3209
3208 3207
3206
3205
3204
3203 3202
3201
32003199
3198
3197
3196
3195
31943193
3192 3191
3190
3189
3188
3187
31863185
3184
3182 31813180
3179
3178
31773176
3174
3173
31723171
3170
3169
3168
3167
3166
3165
3164
3163
3162
3161 3160
3159
315831573156
3154
3152 3151
3150
31493148
3147
31463145
3144
3143
3142
3141 3140
3139
3138 31363135
3133 3132
3129
3127
3126
31253124
3123
3122
3121
3120
31193118
3116
3115
3114
3113
3112
3111
3110
3109
3108
3107
3106
31053104
3102
3101
3100
3099
3097
3096
3095
3094 3093
30923091
3090
3088
3087
3086
3085
30843083
3082
3081
3080
3079
307830773076 3074 30733072
3071
3070 3069
306830673066
30653064
3063
3062
3061
30603059
3058
3057
3056
3055
3054
30533052
3051
3050
3049
3048
30473046
30453044
3043
3042
3041
3040
3039
3038
3037
3036
3035
3033
3032
3031
3030
3029
30283027
30263025
302430233022 3021 3020
30193018 3017
3016
3015
3014
30133012
3011
3010
3009
3008
3006
3005
3004
3003
3002
3001
3000 2999 2998
2997
2996
2995
2994
2993 2992
29912990
2989
29882987
29862985
2984
2983
29822981
2980
2979
2978
2977
2976
2975
2973
29722971
2970
2969
2968
2967
2966
2965
2964
2963
2962
2961
2960
2959
2958
2957
29562955 2954
29532952 29512950
2949
2948
2947
2946
2945
29442943
2942
2941
2940
2939 293829372936
2935
2934 29332932 293129302929
2927
2926
2925
2923
2922
2921
2920
2919
2918
2917
2916 2915
2914
2913
2912 2911
2909
2908 2907
2906
2905
29042903
2902
2901 2900
2899
2898
28972896
2895
2894
2893
2892
2891
2890
2889
2888
2887 2886
2885
2884
2883
2882
2881
2880
2878
2877
2876
2875
2874
2873
28722871
2870
2869 2868 2867
2866
2865
2864
2863
2862
2861
2860
2857
28562855
2854
2853
2852 2851
2850
2849
2848
28472846 2845
2844
28432842
2840
28392838
2837
2836 2835
2833
2832 2831
2830
28292828
2826
2825
2824
2823
2822
2821
2820
2819
2818
2817
2816
2815
2814
2812
2811
2810
2809
2807
2806
2805
2804
2803
2802
28012799
2798
2797
2796
2795 27942793
2792
2791
2790
2788
2787
2786
2785
2784
2783
2782
2781
2780
2779
2778
2777
2776
2775
2774
2773
2772
27712770
2769
2768
2767
2765 2764
2763
27622761
2760
2759
2758
2756
27552754
2753
2752
2751
2750
2749
2748 2747
2746
2745
2744
27432742
27412740
2739 2738
2737
27362735
2734
2733
2732
2731
2730
2729
2728
27272726
2725
2724
2723
27222721
2720
2719
2718
2717
2716
27152714
2713
2712
2710
2709
2708
2707
2706
2705
2704
2703
2702
2701
2700
2699
2698 2697
2696
2695
2694
2693
2692
26912690
2689
2687
2686
2684 26832682
2681
2680
2679
2678
26772676
2675
2674
2673
2672
2671
26702669
2668
2667
2666
2665
2664
2663
2662
266126602659
2658
2657
2656
26552654
2653
2652 2651
2650
2649
2648
2647
2646
2645
2644
26432642
2641
2640
2639
2638 26372636
2634
2633
2632
2631
2630
2629
2628
26272626
26252624
2623
26222621
2620 2619 2618
2617
2616
2615
2614
2613
2612
2611
26102609
2608
2607
2606
2605
26042603
2602
26012600
2599
2598
2597
2596
2595 2594
2593
2592
2591
2590
2589
2588
2587
2585
25842583
2582
2581
2579
2578 25772576
2575 2574
2573
2572
2571
2570
2569
2568
2567
2566
2565 25642563
2562
25612560
2559
2558
2557
2556
2555
2554
2553
2552 2551
2550
2549
2548
25472546 2545
2544
2543
2542
254125402539
2538
25372536
2535
2534
2533
2532
2531
2530
2529 25282527
2526
2525
2524
2523
2522 25212520
251925182517
2515 2514
25132512
2511
2509
2508
2507 2506
2505
2504
2503
2502
2501
2500
2499
24982497
2496
2495
2494
2493
2492
2491
2490
2489
2488
2487
2486
2485
2484
2482
2481
2480
2478 247724762475 2474
2473 2472
2471
2470
2468
2467
2466
2465
2464 2462
2461
2460
2459
2458
2457
2456
24552454
24532452 24512450
2449
2448
2447
244624452444
244324422441
2440 2439
2438
24372436
2435
24342432
2431
2429
2428
2427
24252424 2422
2421
2420
2419
24182417 2416 2415
2413
2412
2411 2410
2409
2408
2407
2406
2405240424032402
2401
2400 2399
2398
2397
2396
2395
23942393
2392
2391
2390
2389
2388
2387
2386
2384
2383 23822381
2380
2379
23782377
2376
2374 23732372
2371
2370 23682367
2366
2365 23642363
2362
2361
23602359
2358
2357
2356
23552354
2353 2351
2350
2349
2348
2347
2346
2345
2344
2343
2342
2341
2340
2339
2338
2337
2336
2335
23342333
2332
2331
2330
2329 2328
2327
2326
2325
23242323
2322
23212320 2319
2318
23172316
2315
2314
2313
2312
231123102309 2308
2307
2306
2305
2304
2303
2302
2300
2299
2298
2297
2296
2295
229422932292
2291
2290
2289
2288
2287 2286
2285
2284
2283
2282
2281
2279
2278227722762275
2274
2273 2272
2271
2270
2269
2268
2267
2266
2265
2264
2263
22622261
2260
2259
2258 2257
2256
2255
2254
2253
2252
2251
2250
2249
2248
2246
2244
2243
2242
2241
2240
2239
2238
2237
2236
2235
2234
2233
22322231
2230
2229
2228
2227
2226
22252224
2223
2222
2221
2220
2219
2218
2217
2216
2215
2214 2213
2212
2211
22102209
2208
2207 2206
2205
2204
2203
2202
2201
22002199
2198
2197 2196
2195
2194
2193
21922191
2190
2189
2188
2187
2185
2184
2183 2182
2181
21802179
2178
2177
2176 2174
2173
2172 2171
2170
2168
2167
2166
2165
216421632162
2161
2160
2159
2158
2157
2156
21552153 21522151 2150
2149
2148
2147
2146
2145 2144
2143
21422141
21402139
21382137
2136
2135
2134
2133
2132
2131
2130
2129
2128
2126
2125
212421232122
2121
2119
2118
2117
2116
2115
2114
2113
2112
2111
2110
2109
2107
2106
21052104
2103
2102 2101
2100
2099
2098
2097
2096
2095
2094
2093
2091
209020892088
2087
2086
2085
2084
2083
2082
2081
2080
2079
2078
2077
2076
2075
2074
2073
2072
2071
2070
2069
206820672066
2064
2063
2062
2061
20602059
2058
2057
2056
2055
2054
2053
2052
2051
2050
2049
2048
2047
2046
2045
2044
2043
2041
2040
2039
2038
2037
2036
2035
2034
2033 2032
20312030
2029 2028
2027
2026
20252024
2023
2022
2021
2020201920182017
2016
2015
2014
2013
2012
2011 2010
2009
2008
2007
2006
2005
2004
2003
2002
2001
20001999
1998 1997
19961995
1994
1993
1992
1991
19901989
19881987 19861985
19841983
1982
1981
1980 197919781977
1976
1975
1974
1973
1972
1971
1970
1969 1968
19671966
1965
1964
1963
1962
1961
1960
1959
1958
19571956 1955
1954
1953
1952
1951
1950
1949
1948
1947
1945
19441943
1942
1941
1940
19391938
1937
1936
1935
1934
1933
1932 1931
19301929 1928
1927
1926
19251924
1923
1922
1921
1920
1919
1918
1917
1916
1915
1914
1912
1911
1910
1909
1907
1906
1904
1903
1902
1901 1900 1899
1898
1896
1895
1894 1893
1892
1891
1890
1889
1888
18871886
188518841883
1882
1881
1880
1879
18781877
1876
1875
1874
1873
1872
1871
1870
1869
1868
1867
186618651864
1863
1862
1861 1860
1859
1858
1857
1856
18551854
1853
1852
1851
1850
1849
18481847
1846
1845
1844
1843 1842
1841 1839
1838
1837
1836
1835
1834
1832
1831
1830
1829
1828
1827
1826
1825
1824
1823
1822
1821 1820
1819
1818
1817
18161815
1813
1812
1810
1809 1808 1807
1806
180518041803
1802
18011800
1799
1798
1797
1796
17941793
1791 17901789
17881787
1786
1785 1784
1783
1782
1781
1780
1779
1778
1777
17761775
1774
1773 1772
1771
1770
1769
1768
1766
1765
17641763
1762
1761 1760
1759
1758
1757
1756 1755
1754
1753
1752
1751
1750
1749
1748
1747
1746
17451743
1742
1741
1740
1739
1738
1737
1736
1735
1734
1733
1732
1731 1730
17291728
1727
1726
1725
1724
1723
17221721
1720
1719
1718
1717
1716
1715
1714
1713
1712
17111710
1709
1708
1707
1706
1705
17041703
1702
1701
1700
1699 1698
1697
169616951694
1693
1692
1691
1690
1689
1688
1687 1686
1685
16841683 1682
1681
1680
16791678
1677
1676
1675
1674
16731672
1671
1670
1669
1668
166616651664
1663
1662
1661
16601659
16581657
1656
1655
16541653
1652 1651
1650
1649 16481647
1646
1645
1644
1643
16421641
1640
16391638
1637
1636
1635
1634 16331632
1631
1630
1629
1628
1627
1626
1625
1624
1623
1622
1621
1620
16191618
1617
1616
1615
1612
1611
1610
1609
1608
1607
1606
16051604
1603
1602
1601
1600
1599
1598
1597
1596
15941593 1592
1591
15901589
1588
1587
158615851584
1583
15821581
1580
1578
1577
1576
1575
1574
1573 1572
15711570
1569
1568
1567
1566
1565 1563 15621561
156015591558
1557
15561555
1554
1553
1552
15511550 1549
1547
15461545
1544
1543
1542
1541
15401539
1538
1537
1536
15351534
1533
1532
1531
1530
1529
1528
1527
1526
15251524
15231522
1521
1520
1519
1518 1517
1516
1515
1514
1513
1512
1511
1510 150815071506 1505
1504
1503
15021500
1499
1498 1497
1496
1495 14941493
1492
1491
1490
1489
1488
148714861485
1484
1483
1482
1481
1480
1479
1478
1477
1476
1475
1474
1473
1472
1471
1470
1469
1468 14671466 1465
1464
146314621461
1459
1458
1457
1456
1455
1453
1452
1451
1450 1449
1448
1447
1446
1445
1444
1443
1442
1441
1440
14391438
1437
1436
1435
1434
1433
1432
1430 14291428
1427
1425 1424
1423
1422
1421
1420
1419
14181417
1415
1414
1413
14121411 1410 14091408
1407
14061405 1404
1403
1402
1401
1400
13991398
13971396
1395
1394
1393 1392
1391
1390
13891388
1387
1386
13851384 1383
1382
1381
13801379
137813771376
1375
1374
1373
1372
1371
1370
13691368
1367
1366 1365
1364
1363
1362
1361
1360
1359
1358
1357
1356
1355
1354
1353
1352
1350
1349
1348
1347
1346
1345
1344
1342
1341
1340
1339
13381337
1336
1335
1334
1333
1332
13311330
1329 1328
1327 1326
1325
13241323
1322
1321
1320
1319
1318
1317
1316
1315
1314
1313
1312
1311
1310
1309 1308
1307
1306 130513041303 13021301
1300
1299
1298 1297
1296
12951294
1293
1292
1291
1290 1289
1288
1287
1286
1285 1284
1283
1282
1281
1280
1279
1278 127712761275
1274
1273
1272
1270
1269
1268
1267
1266
1264
1263
1262
12611260
1259
1258
1257
125612551254 1253
1252
1251
1250
1249
1248 1247 1246
1245
1244
1242
12411240
1239
1238
1237
12361235
1234
1233
1232
1231
1230
1229
1228
12271226
1225
1224
1223
1222
1221
1220
12191218
1217
1216 1215
1214
1213
1212
12111210
1209
1207 1206
1205 12041203
12021201
1200
1199
11981197
11961195
1194
1193
11921191
1190
1189
11871186
1185
1184
11831182
1181
1180
1179
1178
1176
1175 1174
11731172 1171
1170
1169
1168
1167
1166
1165
1164
1163
1162
1161
1160
1159
1158
1157
1155
1154
1153
1152
11511150
1149
1148
1147
1146
1145
1144
1143
1142
1141
1140
1139
1138
1137
1136
1135
1134
11331132 1131
1130
1129
1128
1127
1126
1125
11231122 1121
1120
11191118
1116
11151114
1113
1112
11111110
1109
1108
11071106 1105
1104
1103 1102
1101
1100
1099
1098
10971096
1095
1094
1093
10921091
10901089
108810871086
1085
10841083
10821081
1080
1079
1078
1077
1076
1075
1074 1073
1072
1071
1070
1069
1068
1067
1066
1065
1064 1063
1062
1061 1060
1059
10581057 1056
105510541053
1052
10511050
1049
10481047 1046
1045
1044
1043
1042
1041
1040
1039
1038
1037
1036
1035
1034
1033
1032
10311030
102910281027
10261025
1024
1023 1022
10211020
1019
1018
1017
1016
1015
1014
1013
1012
1011
1010
1009
1008
1007
1006
1005
1004
100310021001
1000
999998
997
996995
994993
992
991
990
989
988
987
986
985 984
983
982
981980 979
978
977
976
975
974
973
972
971
970
969
968967 966
965964 963962
961
960
959
958
957
956 955
954
953
952951
950
949
948947 946945
944
943942 941
940
939
938
937 936 935934 933
932
931
930
929928927926
925
924
923
922
921
920919 917
916
915
914 913
912911
910 909
907 906
905
904
903902901
900
899898
897
896 895
894
893
892
890
889
888
887
886
885
884
883
882
880879878 877
876
875
874873
872
871
870
868
867
866
865
864
863 862
861 860
859
858857
856
855
854
853
852
851
850
849
848
847846
845
844
843
842
841
840
839
838
837
836
835
834833
832
831
830 829
828
827
826
825
824823
822
821 820
819
818
816
815
814
813
812
811
810
809
808807
806
805803 802
801 800
799
798
797
796
795
794
793
792
791 790
789
788
787
786
784
783
782 781780
779
778
777776
775
774
773
772
771
770
769
768766 765 764
763762
761
760
759
758
757
756
755
754
753
752
751
750749
748
747
746
745
744 743
742 741740739
738
737
736
735
734
733
732
731
730
729
728
727
726
725
724 723722721
720719
718
717
716
715
714 713712
711
710709
708
707706
705
704
703
702701
700
699698
697
696
695
694
693
692
691
690
689688
687686685684683
682
681
680679
678
677
676
675674
673
672
671670669
668
667
666665
664
663
662
661660
659
658
657
656
655
654
653652
651
650
649
648 647
646
645
644
642 641 640
639
637636
635634
633
632
631630629
628
627626
625624
623 621
620
619
618 617
616
615
614
613
612611
610609
608
607 606
605
604
603
602
601
600
599
598
597
596
594
593
592
591
590
589588 587
586
585
584
583
582
581580
579
578577
575
574
573
572
571
570 569
568
567566
565
564
563
562
561
560
559
558
557556555
554
553
552
551
550
549
548547
546
545
544
543542
541
540
539 538
537
536535 534533
532
531
530
529
528 527
526
525
524523
522
521 520
519
518
517
516
515
514
513
512
511510
508
507
506
505504
503
502
501
500
499
498
497
496
495
494
493
492
491
490489
488
487 486
485484
483
482481
480 479
478
477
476
475
474
473
472
471
470469
468 467
466
465
464
463
462
461460459 458
457
456
455
454
453
452451 450449
448
447
446
445
444
443
442
441
440
439
438
437436
435
434
433
432
431
430
429 428
427
426
425
424
423
422
421
420
419 418
417 416415
414413412
411
410
409
408
407
406
405
404
403
402
401
400
399
398
397
396
395
394
393
392391
390
389
388
387
386
385
384
383
382
381380
378
377
376
375374
373
372
371370 369
368
366
365
364
363
362
361
360 359
358 357
356
355 354
353
352
351350
349
348347
346
345
344343 342341 340
339
338337
336
335334
333
332
331330
329328
327
326
325324 323322 321
320
319
318
317
316 315314 313
312
311310 309
308
307
306
305
304
303
302
301
300
299
298
297
296
295
294
293
291
290
289
288
287
286
285
284283
282
281
280
279
278
277
276
275
274
273
272
271270269268
267
266
265
264
263
262
261260258
257
256 255
254
253
252
251
250
249
248
247246
245
244 243
241
240
239
238
237
236
235
234
232
231230
229
228227
226
225
224
223
222
221
220 219
218
217
216215
214
213
212
211
210
209208207
206205204
203
202
201
200
199
198
197
196
195194
193192
191190
189
188
187
186
184
183
182181
180
179178
177
176
175174 173
172
171
170169
168167
166
165
164
163
162
161160
159 157
156155154153
152
151 150149
148
147
146145
144
143
142
141
140139
138
137
136 135
134
133
132
131
130
129128
127
126
125
124123
122
121
119118
117
116
115
114
113
112
111110
108107
106
105
104
102
101
100
999897
9695
9493
92
919089 88
87
86
85 84
83
82
81
80
79 7776
75
74
73
72
717069 6867
66
65
64
63
62
61
6059 5857 56
55
54
53
52
5150
49
48
47
46
45
44
43
42
41
40
393837
36
35
34
33
32
31
30
29
28
27
26
24
23
22
21
20
19
18
17
16
15
14
13
12
11
109
8
7
6
5
4
3
2
1
0
Figure 5: Quebec landscape modelization from [2]
Contrastingly to what we have seen in the previous section, greedy algorithms perform
surprisingly well in the experiments we perform on this real case study. As shown in Fig
6, greedy algorithms lead on average to gains that reach 95% of the one obtained with
optimal resolution with respect to the gain expected with a random solution.
5.2. Scalability and benefits of the preprocessing
In this subsection we address the scalability of our approach and the added benefits of
our preprocessing step. For this purpose we build a plausible instance from which we can
easily vary the number of changing elements and integer variables. We take the case of
the city of Marseille which is surrounded by 3 big forest massifs and contains 54 parks of
21
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0budget in %
0.95
0.96
0.97
0.98
0.99
1.00av
erag
e op
tim
um r
atio
incremental greedydecremental greedyoptimum (MIP)
Figure 6: Average value of returned solutions normalized between the average random solution and an
optimal one
#wastelandMIP preprocessed MIP greedy
#var #const time #var #const time time
20 345775 18410 12246 10716 7197 316 766
50 717901 36410 142306 34613 21485 1964 7217
80 1306597 59810 1914478 76281 42039 13141 29479
110 - - - 132225 66759 62268 87948
140 - - - 207999 96600 200558 208179
170 - - - 308355 132701 1701684 416751
Table 1: Comparison of the MIP, the preprocessed MIP and the greedy (algorithm DG) according to the
number of variables (#var), the number of constraints (#const) and the time (in ms) it takes to execute
variable sizes. The idea is to consider a certain number of wastelands throughout the city
that can be protected to preserve the stepping stones between massifs and parks. The base
graph is composed of 96 nodes of which 42 models the frontier of massifs, 43 represents
small parks and 11 big parks. The graph is complete and each arc is weighted with the
border to border distance between the two patches it connects. We model each wasteland
as a patch which is totally threatened (see Subsection 5.1.2). A non purchased wasteland
has then a null quality and does not contribute to any shortest path. We perform our
experiments by adding various number of wastelands, randomly generated across the city,
and looking for variations of the MIP formulation size and computation time.
We see in Table 1 that the computation time of the MIP without preprocessing in-
22
creases very quickly. It takes more than 30 minutes on average for instances with 80+
wastelands whereas the preprocessed one can be solved in less than 30 minutes with up to
170 wastelands. This is due to the preprocessing step that significantly reduces the time
required to solve the linear relaxation by reducing the number of variables, constraints
and non-zero entries of the mixed integer program. The preprocessed MIP is faster than
the greedy algorithm for instances with at most 140 wastelands (the preprocessing step
was not used for greedy algorithms).
Finally, we run our preprocessing on 400 instances randomly generated from the Que-
bec’s graph (Fig. 5) in order to study the impact of the preprocessing on the MIP for-
mulation size. For building these instances, we take 20 connected subgraphs of 500 nodes
and for each graph we create 20 instances by randomly picking a percentage of arcs whose
length could be improved by half. Figure 7 shows the percentage of reduction of the num-
ber of constraints, variables and non-zero entries of the MIP formulation with respect to
the percentage of arcs that could be restored. The preprocessing removes almost all the
elements of the MIP formulation when the number of restorable arcs arrives close to zero.
When 20% of the arcs could be restored, the preprocessing removes more than 75% of the
model’s variables and non-zero entries and around 60% of the constraints.
0 20 40 60 80 100percentage of restorable arcs
0
20
40
60
80
100
perc
enta
ge o
f ele
men
tsre
mov
ed b
y th
e pr
epro
cess
ing constraintsvariablesnon-zero entries
Figure 7: Average percentage of constraints, variables and non-zero entries that the preprocessing
removes with respect to the percentage of arcs that could be restored
6. Conclusion
In this paper we introduce a new MIP formulation for BC-ECA-Opt and we show that
this formulation allows solving optimally instances with around 150 habitat patches while
23
previous formulations [22] are limited to instances of about 30 patches. In order to make
our formulation scale to even larger instances, we present a preprocessing step that reduces
significantly the size of the graphs on which a generalized flow has to be computed. We
show that this preprocessing step allows to greatly reduce the MIP formulation size and
that its benefits increase when the proportion of arcs whose lengths can change decrease.
Thus allowing us to treat instances of up to 300 patches.
This new method allows us to solve optimally larger instances of BC-ECA-Opt and to
compare optimum solutions with the solutions returned by greedy algorithms. Interest-
ingly, we found that greedy algorithms perform well in practice despite the arbitrary bad
cases we spotted. Our next goals will be to experiment our approach on other practical
instances of the problem arising from different ecological contexts. On the theoretical side,
we would like to investigate the problem from an approximation algorithm point of view.
For instance, is it possible to find reasonable assumptions under which greedy algorithms
are guaranteed to return a solution whose ECA value is at least a constant fraction of the
optimal ECA ? If these assumptions are fulfilled by the real instances that we considered,
this would explain our experimental observations. Moreover, since a polynomial time ap-
proximation scheme has been given in the case of trees [21], it could also be interesting
to know for which larger classes of graphs constant factor approximation algorithms for
this problem exist. Another interesting question is to determine whether good solutions
could be obtained by decomposing geographically the problem, by solving independently a
subproblem for each region and then by reassembling the solutions. In this case, a notion
of fairness could help to allocate the budget among the regions so that each region can
enhance its own internal connectivity keeping a part of the budget to enhance the connec-
tivity between the regions. Since ECA is based on the equivalence between a landscape
and a patch, such a multilevel optimization approach looks promising.
[1] Ahuja, R. K., Magnanti, T. L., Orlin, J. B., 1993. Network Flows: Theory, Algo-
rithms, and Applications. Prentice-Hall, Inc., USA.
[2] Albert, C. H., Rayfield, B., Dumitru, M., Gonzalez, A., 2017. Applying network
theory to prioritize multispecies habitat networks that are robust to climate and
land-use change. Conservation Biology 31, 1383–1396.
[3] Awade, M., Boscolo, D., Metzger, J. P., 2012. Using binary and probabilistic habitat
24
availability indices derived from graph theory to model bird occurrence in fragmented
forests. Landscape Ecology 27, 185–198.
[4] Billionnet, A., 2013. Mathematical optimization ideas for biodiversity conservation.
European Journal of Operational Research 231 (3), 514–534.
URL https://www.sciencedirect.com/science/article/pii/S0377221713002531
[5] Brondizio, E. S., Settele, J., Dıaz, S., Ngo, H. T., 2019. Global assessment report on
biodiversity and ecosystem services of the Intergovernmental Science- Policy Platform
on Biodiversity and Ecosystem Services. IPBES, Bonn, Germany.
[6] Catanzaro, D., Labbe, M., Salazar-Neumann, M., 2011. Reduction approaches for
robust shortest path problems. Computers & OR 38, 1610–1619.
[7] de Toda, S. S. M., Agosto 2015. Quantifying dispersal paths in probabilistic habitat
networks: a reply to hock and mumby (2015) and an overview of recent developments
and applications. Artıculo de trabajo 10.13140/RG.2.1.2978.2883, ETSI Mon fos.
URL http://oa.upm.es/37351/
[8] Hock, K., Mumby, P., 02 2015. Quantifying the reliability of dispersal paths in con-
nectivity networks. Journal of The Royal Society Interface 12, 20150013.
[9] Jaeger, J. A., Soukup, T., Schwick, C., Madrinan, L. F., Kienast, F., 2011. Landscape
fragmentation in Europe. Publications office of the European Environmental Agency,
Luxembourg.
[10] McCormick, G. P., 1976. Computability of global solutions to factorable nonconvex
programs: Part I - convex underestimating problems. Math. Program. 10 (1), 147–
175.
URL https://doi.org/10.1007/BF01580665
[11] McRae, B. H., Dickson, B. G., Keitt, T. H., Shah, V. B., 2008. Using circuit theory
to model connectivity in ecology, evolution, and conservation. Ecology 89, 2712–2724.
[12] Pascual-Hortal, L., Saura, S., 2006. Comparison and development of new graph-
based landscape connectivity indices: towards the priorization of habitat patches and
corridors for conservation. Landscape ecology 21, 959–967.
25
[13] Pereira, J., Saura, S., Jordan, F., 2017. Single-node vs. multi-node centrality in land-
scape graph analysis: key habitat patches and their protection for 20 bird species in
ne spain. Methods in Ecology and Evolution 8, 1458–1467.
[14] Pereira, M., Segurado, P., Neves, N., 2011. Using spatial network structure in land-
scape management and planning: A case study with pond turtles. Landscape and
Urban Planning 100, 67 – 76.
[15] Rubio, L., Bodin, O., Brotons, L., Saura, S., 2015. Connectivity conservation priorities
for individual patches evaluated in the present landscape: how durable and effective
are they in the long term? Ecography 38, 782–791.
[16] Saura, S., Estreguil, C., Mouton, C., Rodrıguez-Freire, M., 2011. Network analysis
to assess landscape connectivity trends: Application to european forests (1990-2000).
Ecological Indicators 11, 407–416.
[17] Saura, S., Pascual-Hortal, L., 2007. A new habitat availability index to integrate
connectivity in landscape conservation planning: Comparison with existing indices
and application to a case study. Landscape and Urban Planning 83, 91 – 103.
[18] Saura, S., Torne, J., 2009. Conefor sensinode 2.2: a software package for quanti-
fying the importance of habitat patches for landscape connectivity. Environmental
modelling & software 24, 135–139.
[19] Taylor, P. D., Fahrig, L., Henein, K., Merriam, G., 1993. Connectivity is a vital
element of landscape structure. Oikos 68, 571–573.
[20] Urban, D., Keitt, T., 2001. Landscape connectivity: A graph-theoretic perspective.
Ecology 82, 1205–1218.
[21] Wu, X., Sheldon, D. R., Zilberstein, S., 2014. Stochastic network design in bidirected
trees. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. Q.
(Eds.), Advances in Neural Information Processing Systems. Vol. 27. Curran Asso-
ciates, Inc., pp. 882–890.
[22] Xue, Y., Wu, X., Morin, D., Dilkina, B., Fuller, A., Royle, J. A., Gomes, C. P.,
2017. Dynamic optimization of landscape connectivity embedding spatial-capture-
recapture information. In: 31st AAAI Conference on Artificial Intelligence. Vol. 31.
pp. 4552–4558.
26