Genetic Algorithm Based Model for Capacitated

Network Design Problem

Meriem Khelifi Mohand Yazid Saidi Saadi Boudjit

Laboratoire Réseaux et Systèmes L2TI, Institut Galilée L2TI, Institut Galilée
Université Badji Mokhtar, Algérie Université Paris 13, France Université Paris 13, France
Email:[email protected] Email: [email protected] Email: [email protected]

Abstract—Efficient design of networks topologies is • The fixed charge MNDP [3][4] in which the link
challenging, especially with the arrival of the virtualization in capacities are known. Solving MNDP consists to de-
these last years. In this paper, we deal with the Capacitated termine the set of edges that should be opened in the
Network Design Problem (CNDP) with modular link capacities final topology.
to design minimum cost network while satisfying the flow
demands. We propose a two levels Genetic Algorithm (GA) • The capacitated MNDP, where the number of modules
based model that can deal with several variations of CNDP. Our to install on the edges are modeled by integers [1]
proposition defines a new encoding scheme to treat the modular
case. Extensive simulation results on Atlanta, France and • The Network Loading Problem (NLP), where the
Germany network instances show that the proposed algorithm is number of module types is limited, each one with a
much more efficient than the Iterative Local Search algorithm. given unit cost and capacity.

keywords: Network Design Problem; two level optimization; Various heuristics and exact approaches have been developed
Genetic algorithm; Modular capacity; Multicommodity flow for designing capacitated networks. However, the heuristic
problem. approaches are more likely to be trapped in local optima, while
the exact approaches are applied only to small or medium
size problems. Due to the weaknesses of the two approaches
I. I NTRODUCTION and the increasing popularity of metaheuristic approaches, we
To save resources (routers, optical fibers, etc.) networks have witnessed many metaheuristics being applied to network
should be efficiently designed. Diverse networks models were optimization problems. In this paper, we propose a novel meta-
then defined and used to represent a wide range of issues heuristic that is based on Genetic Algorithms (GAs), which
in transportation, telecommunications, logistics, production has extensively been used to solve many difficult combinatorial
and distribution networks. All these models consider a graph optimization problems in industrial engineering and operations
composed of nodes and edges (optical fibers, cables, etc). For a research. Genetic algorithms are one of the most powerful
better use of these resources, networks designers should solve and broadly applicable stochastic search and optimization
the modular Capacitated Network Design Problem (CNDP) techniques and have achieved great advancement in related
which consists of selecting edges and the optimal capacities research fields, such as network optimization, combinatorial
to allocate to route a set of commodities between a source and optimization, multi-objective optimization, and on so on. Our
destination pairs. Each edge of the graph has a potential set contribution consists on an efficient two level evolutionary
of module capacities with their associated costs, a fixed cost algorithm that uses the GA and the Linear Programming (LP)
that is incurred only if the edge is selected, and a routing cost to solve a general model that can deal with diverse variants
which is proportional to the amount of flows along the edge. of capacitated network design problems. We define a Modular
Each commodity is defined by an origin and a destination node Implicit Encoding (IME) to encode individuals which is a very
and the amount to be routed. The objective is to minimize three flexible encoding scheme.
criteria: edge cost, modules and routing. These capacitated The remainder of this paper is structured as follows: related
network design problems are NP-hard and very difficult to work is introduced in Section 2, notations and mathematical
solve in practice. The CNDP is a particular case of the well formulation of the addressed problem are given in Section
known Multicommodity Network Design problem (MNDP), 3. Section 4 describes and explains in details our proposed
in which we distinguish an important number of special cases algorithm. Experimental results are discussed in Section 5
and extensions [1]. The most studied ones are: where we compare our proposition against Iterative Local
Search (ILS) technique. Finally, Section 6 concludes the paper.
• The unsplittable variant where the flow of each com-
modity is required to follow one route between the II. R ELATED WORK
origin and the destination, which increases the diffi-
culty of the problem [2]. Capacitated Network Design Problem is one of the major
research area in network optimization. It is related to two
• The expansion variant, where some edges already issues: Network Design Problem (NDP) and Network Loading
have an existing capacity. Problem (NLP). In the NDP, the goal is to identify the net-
work topology by selecting routers and links that interconnect

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.
them. Thus, the objective function aims to minimize the total framework of general network design problems which combine
constructive cost under some topological constraints. In this location decision and network design decision. In [18] we have
class of problems, the flow is not modeled, consequently it given a brief idea about the use of two level evolutionary
is considered as uncapacitated. In the NLP, it is assumed algorithms for network design optimisation.
that the topology is already established. Thus, solving NLP
consists to search for the set of resources to allocate for III. M ATHEMATICAL F ORMULATION
the network components. These problems are complementary.
Generally, NDP and NLP are solved separately though they Let G = (V, E) be an undirected network where V is the
are combined together in some cases. One can say that most set of vertices and E is the set of undirected edges. Let K be
of network optimization problems can be seen as a kind of the set of commodities. For each one k ∈ K, P k is the set
(1) NDP, (2) NLP or (3) a combination of both where the of paths associated to commodity k, and dk the flow demand
objective and the constraints may differ from one problem to of commodity k. Let fij be the fixed cost of including edge
another: connectivity [5] [6], limited budget [7], hop limit [8] (i, j) in the network, rij the unit variable flow cost on (i, j),
[2], delay [9] [10], reliability [8] [9], and survivability [10]. and pij the pre-installed capacity on the edge (i, j).
The NLP in capacitated or uncapacitated case and with both The formulation of CNDP is shown below:
single or multiple facilities is a special case of the well known
Multicommodity Network Design Problem (MCND). Previous
X X X
M in z(x, y, n) = rij xkp
works on this problem can be classified as:
(i,j)∈E k∈K P ∈P k
• Uncapacitated network design problems where on
X X X
+ fij yij + clij nlij
each link of the network, it is only possible either to (i,j)∈E (i,j)∈E l∈L
open the link with infinite capacity and a given fixed
cost, or the cost is zero and no capacity is available X
[11]. xkp = t(k), ∀k ∈ K . (1a)
p∈P (k)
• Single facility capacitated network loading problem
where, the capacity can be done by installing on each
link an integer unit of a given basic facility [12]. X X X
xkp ≤ pij + ml nlij yij , ∀(i, j) ∈ E (1b)
• Two facilities capacitated network loading problems k∈K p∈P k l∈L
where the capacity can be achieved by means of two X X
types of modules, each capacity has a specific cost xkp ≤ dk yij , ∀(i, j) ∈ E, ∀k ∈ K (1c)
citerefKerivin. k∈K p∈P k

• Multi-type facility capacitated network loading prob- yij ∈ {0, 1}, ∀(i, j) ∈ E (1d)
lems where various types of capacities can be installed
on each link, each facility has a specific cost [7]. xkp ≥ 0, k
∀k ∈ P , k ∈ K (1e)
The early works on capacitated modular network problems nlij ∈Z , +
∀l ∈ L, ∀(i, j) ∈ E (1f)
were focused on the approximation methods. These methods
define residual capacity and cut-set inequalities for single com- This formulation is a mixed integer linear program which
modity and multicommodity cases on directed, undirected and uses three types of variables: the first type is a binary design
bidirected link models [13][14]. Since these works consider variable, which is defined as yij = 1 if (i, j) is included in
that the underling network is established, they focus only on the network and yij = 0 otherwise. The second type is a
the determination of the facilities allowing the accommoda- continuous path flow variable xkp , which represents the amount
tions of flow demand. Their effectiveness depends on the size of flow of commodity k routed on p ∈ P k . The third type is
of the problem instance. an integer allocation module variable, which is defined by nlij .
With the appearance of metaheuristics, both the NLP and It represents the number of module type l allocated on edge
the NDP have attracted some attention. The authors benefit (i, j), where L is the set of potential modules. Each l ∈ L is
from their efficiency to deal with more complex variants characterized by a capacity ml and an installation cost clij . A
with real size instances. In [15], the author compared several positive capacity of edge (i, j) implies that it is used to route
neighborhood structures to solve the uncapacitated facility demands in the two directions: from i to j or from j to i. This
location problem. In [10], the authors proposed an evolutionary formulation corresponds to a general model that can deal with
approach for capacitated network design considering cost, several variants of capacitated network design problems.
performances and survivability. The objective is to minimize
network cost and packet delay. Kleeman et al. [9] used an The objective function correspond to the sum of the
evolutionary algorithm to solve multicommodity capacitated variable flow costs of commodities, the fixed costs of edges
network design problem with an objective function optimiz- and the allocated module costs. These costs are relative to
ing costs, delay, robustness, invulnerability and reliability. A the problem that we deal with and are not all aggregated in
tabu search heuristic algorithm with real costs on facilities some cases. For instance, the fixed charge problem MNDP
is developed in [16]. A firefly algorithm is proposed by includes only the edge costs. The modules and routing costs
Ragheb et al [7], they combined facility location and network on edges are nil. Constraints (1a) consist of flow conservation
design problem with multi-type of capacitated link and limited equations, which represent the fact that the sum of path flows
budget on facilities. Contreras et al. [17] presented a unified of commodity k is equal to the demand. Constraints (1b)

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.
provide the capacity constraints, which prohibit flowing if the function. After a number of generations, highly fit individuals,
edge is excluded, yij = 0, and allow for flow up to the which are analogous to good solutions to a given problem, will
edge capacity if the edge is included, yij = 1. Constraints emerge. Genetic algorithms consist of five components:
(1c) provide the forcing constraints, which prohibit flowing
commodity k if the edge is excluded, and allow for flow up 1. A method for encoding potential solutions into chro-
to the demand if the edge is included. Constraints (1d) and mosomes;
(1e) express respectively the binarity and the non-negativity 2. A means of creating the initial population;
of variables decisions. Constraints (1f) show that the modules 3. An evaluation function that can evaluate the fitness
facility are allocated in a discrete amounts. of chromosomes;
4. Genetic operators that can create the next generation
population;
5. A way to set up control parameters; e.g., population
X X X
M in z(x(ȳ, n̄)) = rij xkp
size, the probability of applying a genetic operator,
(i,j)∈E(ȳ) k∈K P ∈P k
etc.
X
xkp = t(k), ∀k ∈ K (2a) A. Individual representation
p∈P (k)
X X X In the design of genetic algorithms, the encoding is the
xkp ≤ pij + ml n̄lij ȳij , ∀(i, j) ∈ E(ȳ) (2b) most important task. There are some methods to encode each
k∈K p∈P k l∈L individual in a population, such as binary encoding, integer
X X encoding, etc. In this paper, we define a new encoding method
xkp ≤ dk ȳij , ∀(i, j) ∈ E(ȳ), ∀k ∈ K (2c) called IME (Implicit Modular Encoding) that is relative to
k∈K p∈P k our modular case. An individual built by IME is shown in
Figure 1. Each individual I is a matrix In,m , where n and
xkp ≥ 0, ∀k ∈ P k , k ∈ K (2d) m corresponds to the number of modules and to the number
of edges respectively. Hence, I[li ][ej ] gives the number of
e multicommodity flow problem formulation presented in module types li allocated on edge ej . Our encoding represents
the linear programming (2) is obtained by replacing the vectors
(y, n) in the original CNDP formulation (1) by theirs fixed
values (ȳ, n̄). The objective function aims to minimize the
routing of flows. Constraint (2a) is the flow conservation
constraint which ensures that for each demand the flow routed
on paths is equal to the demand flow. The capacity constraint
(2b) enforces the use of the limited capacities on edges without
exceeding it. Constraint (2c) allows to flow up the demands
only on the selected edges. the non-negativity of the flow
decision variables are expressed in the constraint (2d).
A solution to the CNDP can be viewed as a binary Fig. 1: Individual with Implicit Modular Encoding
assignment (ȳ) to each design variable, an integer vector
assignment (n̄) to the allocation module design variables
and the optimal flow of the corresponded multicommodity
minimum cost flow problem x∗ (ȳ, n̄). So the CNDP objective the decision vector n and implicitly the decision vector x.
function value associated to a solution (ȳ, n̄, x∗ (ȳ, n̄)) is the For example, the case of one module type, T [l1 ][ej ] = 2,
sum of the fixed cost of the open edges in (ȳ), the cost of the means that we allocate two modules on the edge ej . Thus,
modules allocated (n̄) and the objective function value of the we implicitly deduce that edge ei exists in the final topology.
CMFP associated to x(ȳ, n̄)∗ . However, edge ej will not be opened in the final network
since T [l1 ][ej ] = 0. When multiple types of modules are
In the next section, we use the genetic algorithms to explore allowed, the edge exists if at least one module is allocated on
different potential solution areas by choosing various vector it, i.e:
values (ȳ, n̄). Then, we apply the CMFP model (the above  Pn
linear program) to solve CNDP. xe = 1 if li=1 T [li ][e] > 0
xe = 0 otherwise
IV. G ENETIC A LGORITHM FOR CNDP
B. Initial population
Genetic algorithms introduced by Holland [19], are based
on the mechanics of natural selection and natural genetics. There are two ways to generate an initial population;
They start with an initial set of random solutions, called random initialization and heuristic initialization. In our case,
a population. Each individual in the population is called a we use the Iterative Local Search (ILS) heuristic [20] to solve
chromosome, representing a solution to the problem. The a CNDP problem. Since the ILS heuristic gives a unique
initial population evolves through successive iterations, called solution, and we need our initial population to be diverse. To
generations. A measure of fitness defines the quality of an create this diversity, we encoded ILS solution according to
individual chromosome. In each generation, chromosomes are IME and we considered it as the initial individual I0 . Then,
evaluated by a fitness function, also called an evaluation we apply some perturbation on I0 to generate other individuals,

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.
 x1 x2
wherein there are some not feasible. For each new generated I1 I2
individual, we used the LP solver (CPLEX optimizer) to check
for the existence of feasible flows. An individual is added to the
initial population if feasible flows are determined. The initial
Inew
population is generated according to the algorithm depicted in
Figure 2.
Fig. 3: Crossover operator
Algorithm 1 InitialPopulation
Inputs: S0 , K
Local variables: P0
P0 ←− ∅
Generate initial individual I0 throughout encoding ILS solution
with IM E encoding
P ←− I0
foreach module {ln ∈ L} do
foreach link {em ∈ E} do
if I0 [ln ][em ] > 0 then
0 Fig. 4: Mutation operator
Generate
0
new individual I ←− I
I [ln ][em ] = I0 [ln ][em ] − 1
0
if (CplexSolver(I , K) == T rue) then
0
Add individual I to P0 and Figure (4), we assumed that the network is composed of
end six edges and with two module types.
end
end E. The genetic algorithm
end After explaining and detailing the basic components of our
Return P0 proposed genetic algorithm, we describe below its operation
(see Figure 5 for instructions). In our algorithm, we first ini-
tialize the population through InitialP opulation() procedure.
Then N successive populations are generated by applying the
Fig. 2: S0 is the ILS solution, K is the set of flow demands. P0 is the initial two-point crossover operator (Crossover()).
population. L is the set of capacity modules. E is the set of links. I0 is the
0
initial individual. I is the new individual. CplexSolver() is the procedure As said previously, only individuals allowing a feasible
0
that solve the M CF P on the network represented by individual I , returns multicommodity flow solution are added to the current popu-
T rue if it find a feasible flow. lation. This is verified by the running of CplexSolver() pro-
cedure. The best solution Gbest is updated at each generation
and returned when the termination condition is satisfied.
CleanP op() procedure allows to switch from one population
C. Fitness function to another by selecting individuals from the first population.
The fitness function corresponds to the objective function It is based on elitist strategy. The algorithm stops its running
of CNDP. It is computed as the sum of the allocated module after a fixed number of generations or when the result is not
costs, the fixed edge costs and the routing costs. Note that the improved after a certain number of generations.
first two costs are deduced from the individual representation
whereas the routing costs are given by solving the CMNP V. R ESULTS
linear program.
In our experiments, we used three real world instances of
D. Genetic operator network topologies including Atlanta, France and Germany50.
All can be downloaded from http : //sndlib.zib.de [21]. We
1) Crossover: Our crossover operator is two point followed the model filter specified in Table I. The population
crossover, We randomly chose two integers in the individual size is 50 and the number of generations is 15.
length interval (0 < x1 ≤ x2 ≤ m) and two individuals (I1
and I2 ) in the current population. Then, we used two-point
crossover operator to generate new individuals (see Figure 3). TABLE I: The model filter
Typically, a new individual Inew is generated by selecting the
modules of edges in [e1 , ex1 [∪[ex2 , em ] from I1 and [ex1 , ex2 ] Demand model Undirected demand (U)
Link model Undirected links (U)
from I2 . Link capacity model Modular link capacities(M)
Fixed-charge model No fixed-charge cost (N)
2) Mutation: One gene is randomly chosen in chromo- Routing model Continuous (C)
some, then the values of the allocated modules on this gene are Admissible path model All paths (A)
Hop limit model No hop-limits (N)
modified by making some permutations between the modules Survivability model No survivability (N)
as shown in Figure (4). In the examples shown in Figure (3)

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.
 corresponds to the best solutions published in [21].
Algorithm 2 Genetic-Algorithm
Inputs: S0 , K
Local variables: i, P , Gbest GAP (A) = {Cost(A) − Cost(BS)}/Cost(BS) ∗ 100 (3)
P ←− InitialP opulation(S0 , K)
Gbest ←− best individual in the population P
T ermination ←− f alse
while ! Termination do TABLE III: The ILS and GA solutions
i ←− size of population P
while i < Max-Size do Instance BS ILS Gap% GA Gap%
(I1 , I2 ) ←− RandomSelection(P) Atlanta 86492550 92904547 7.41 87959303 1.69
France 20200 21400 5.94 20600 1.98
(x1 , x2 ) ←− RandomSelection(P-size) Germany50 645520 719060 11.39 667840 3.45
(x1 , x2 ) ←− Order(x1 , x2 )
Inew ←− Crossover(I1 , I2 , x1 , x2 )
if (CplexSolver(Inew , K) == T rue) then
Add individual Inew to population P As depicted in Table IV, GA is better than ILS since it
Update Gbest determines solutions more close to the optimums than those
i ←− i+1 of ILS. Concretely, the mean gap obtained with ILS is 3.5
end times higher than the mean gap obtained with GA. This can
end be explained by the exploration of multiple solution areas with
P ←− CleanPop (P) GA while ILS determines only a local minimum.
Update(Termination)
Figures 6, 7 and 8 show the allocated capacities and
their usage for Atlanta, France and Germany50 networks
end
respectively. We compared the total installed link capacities,
Return Gbest the total working capacities and the total unused capacities for
BS, ILS and GA solutions. We remark that the total installed
link capacities in ILS and GA are more larger than the BS
Fig. 5: S0 is the initial solution, K is the set of flow demands. Gbest is ones. This justifies the cost gap. On the other hand, ILS
the best value . P is the current population. M axSize is the fixed size of the uses much more working capacities than BS because ILS
population. I is the individual. CplexSolver() is the procedure that solves wastes module resources. Indeed, instead of splitting flows and
0
the CM F P on the network represented by individual I . It returns the value exploring the small unused capacities on links, ILS routes the
T rue if it founds a feasible flow. majority of demands on single shortest paths. With GA, the
CPLEX optimizer try to exploit the residual quantities on the
allocated modules to route flows. This leads to a bifurcation
of demands on multiple paths that could be arbitrary long
Each instance is characterized by the number of nodes |V |, the (although routing costs slightly limit the path lengths).
number of potential links |E| = m and the number of traffic
demands |K|. Table II summarizes the instance specification
details. We classify them into two categories; instances with VI. C ONCLUSION
Single Facility allocation (SF) and instances with Two Facili-
ties allocation (TF). The set of capacity modules L differs from In this paper, we proposed a two level evolutionary ap-
one network instance to another. The allocation cost MCost is proach to solve several special cases and variants of the capac-
variant on links and it is fixed, however, in France instance. itated network design problem. Our algorithm has two levels,
Atlanta instance assumes a Pre-installed capacities pij on their the higher one is the genetic algorithm, that deals with the link
potential links with a unit routing cost rij . See [21], for more selection and the modules allocation decisions. The lower level
details on the filter model and on the setting parameters. is the LP solver (CPLEX optimizer), which fixes the routing
decision by searching for a feasible flow according to the
network configuration made in the higher level. For efficiency,
TABLE II: The Instance Setting Parameters we rigorously defined the main components of the genetic
algorithm. The initial solution is generated by an iterative
Problem Instance |V | |E| |K| N br L M Cost pij rij
local search algorithm, which is combined with an heuristic
Atlanta 15 22 210 TF 1000, 4000 variant yes yes procedure to construct the initial population. To better explore
France 25 45 300 SF 2500 fixed no no
Germany50 50 88 662 SF 40 variant no no
the solution space, we defined a very flexible and meaningful
encoding scheme called IME (Implicit Modular Encoding),
two point crossover operator and an elitist population strategy.
In Table III, ILS and GA corresponds to the solutions ob- The results are very satisfactory. Indeed, the basic idea of
tained by iterative local search algorithm and genetic algorithm combining genetic algorithms and linear programming for
respectively. We examine the quality of a given algorithm A solving the problem in two levels is effective. Simulations
(A could be GA or ILS) by computing its optimality gap (see show that our two level approach is better than the iterative
equality 10) that is defined as the ratio between the difference local search approach since it determines solutions close to the
of the A’s cost and the Best Solution (BS) cost. Note that BS best known ones.

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.
 TABLE IV: Working and Unused capacities
Atlanta France Germany50
Instance BS ILS GA BS ILS GA BS ILS GA

Total installed link capacities 294000 307000 300000 252500 270000 257500 7200 8000 7440
Total working flow 282338.50 281188 284503 246938 237952 240351 7140 7024 7265.83
Total Unused flow 11661.5 25812 15497 5562 32048 17149 60 976 174.17

teruniversitaire de recherche sur les réseaux dâentreprise, la logistique

et le transport, Montreal, 2013.
[5] A. Balakrishnan, M. Banciu, K. Glowacka, and P. Mirchandani, “Hier-
archical approach for survivable network design,” European Journal of
Operational Research, vol. 225, no. 2, pp. 223–235, 2013.
[6] M. Abd-El-Barr, “Topological network design: A survey,” Journal of
Network and Computer Applications, vol. 32, no. 3, pp. 501 – 509,
2009.
[7] R. Rahmaniani and A. Ghaderi, “A combined facility location and
network design problem with multi-type of capacitated links,” Applied
Mathematical Modelling, vol. 37, no. 9, pp. 6400–6414, 2013.
Fig. 6: Atlanta network
[8] T. Liu, W. Yang, and J. Huang, “Reliable network design problem under
node failure with benders decomposition,” Applied Mathematics, vol. 5,
p. 241, 2014.
[9] M. Kleeman, G. Lamont, K. Hopkinson, and S. Graham, “Solving
multicommodity capacitated network design problems using a multiob-
jective evolutionary algorithm,” in IEEE Symposium on Computational
Intelligence in Security and Defense Applications, 2007, pp. 33–41.
[10] A. Konak and A. Smith, “Capacitated network design considering
survivability: an evolutionary approach,” Engineering Optimization,
vol. 36, no. 2, pp. 189–205, 2004.
[11] T. Magnanti, P. Mireault, and R. Wong, Tailoring Benders decomposi-
tion for uncapacitated network design. Springer, 1986.
[12] F. Barahona, “Network design using cut inequalities,” SIAM Journal on
optimization, vol. 6, no. 3, pp. 823–837, 1996.
Fig. 7: France network [13] A. Atamtürk, “On capacitated network design cut–set polyhedra,”
Mathematical Programming, vol. 92, no. 3, pp. 425–437, 2002.
[14] A. Atamtürk and D. Rajan, “On splittable and unsplittable flow capac-
itated network design arc–set polyhedra,” Mathematical Programming,
vol. 92, no. 2, pp. 315–333, 2002.
[15] D. Ghosh, “Neighborhood search heuristics for the uncapacitated facil-
ity location problem,” European Journal of Operational Research, vol.
150, no. 1, pp. 150–162, 2003.
[16] O. Brun, J. Garcia et al., “A tabu search heuristic for capacitated
network design,” in IEEE Symposium on Computers and Communi-
cations,ISCC, 2009, pp. 461–467.
[17] I. Contreras and E. Fernández, “General network design: A unified view
of combined location and network design problems,” European Journal
of Operational Research, vol. 219, no. 3, pp. 680–697, 2012.
Fig. 8: Germany50 network [18] M. Khelifi, S. Boudjit, and M. Y. Saidi, “Two level evolutionary
algorithm for capacitated network design problem,” in 13th IEEE
Annual Consumer Communications & Networking Conference, CCNC
2016, Las Vegas, NV, USA, January 9-12, 2016, 2016, pp. 299–300.
[19] D. E. Goldberg and J. H. Holland, “Genetic algorithms and machine
R EFERENCES learning,” Machine learning, vol. 3, no. 2, pp. 95–99, 1988.
[20] M. Khelifi, M. Babes, S. Ghanemi, M. Y. Saidi, and S. Boudjit,
[1] A. Frangioni and B. Gendron, “0â1 reformulations of the multicommod- “Hybrid heuristic for capacitated network design problem,” J. High
ity capacitated network design problem,” Discrete Applied Mathematics, Speed Networks, vol. 21, no. 4, pp. 313–330, 2015.
vol. 157, no. 6, pp. 1229 – 1241, 2009.
[21] S. Orlowski, R. Wessäly, M. Pióro, and A. Tomaszewski, “Sndlib 1.0
[2] B. Thiongane, J.-F. Cordeau, and B. Gendron, “Formulations for survivable network design library,” Networks, vol. 55, no. 3, pp. 276–
the nonbifurcated hop-constrained multicommodity capacitated fixed- 286, 2010.
charge network design problem,” Computers Operations Research,
vol. 53, no. 0, pp. 1 – 8, 2015.
[3] I. Rodríguez-Martín and J. J. Salazar-González, “A local branching
heuristic for the capacitated fixed-charge network design problem,”
Computers & Operations Research, vol. 37, no. 3, pp. 575–581, 2010.
[4] D. C. Paraskevopoulos, T. Bektaş, T. G. Crainic, and C. N. Potts, A
Cycle-Based Evolutionary Algorithm for the Fixed-Charge Capacitated
Multi-Commodity Network Design Problem. Tech. rep., Centre in-

Authorized licensed use limited to: Beijing Jiaotong University. Downloaded on February 22,2021 at 10:07:16 UTC from IEEE Xplore. Restrictions apply.