Int. J.

Production Economics 141 (2013) 34–44

Contents lists available at SciVerse ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

A multiple ant colony optimization algorithm for the capacitated location

routing problem
Ching-Jung Ting a,n, Chia-Ho Chen b,1
a
Department of Industrial Engineering and Management, Yuan Ze University, 135 Yuan-Tung Road, Chung-Li, Taiwan 32003, ROC
b
Splendor Game Technology Co., LTD. 4F.-5 No. 360 Jianxing Road, North District, Taichung, Taiwan 40459, ROC

a r t i c l e i n f o a b s t r a c t

Article history: The success of a logistics system may depend on the decisions of the depot locations and vehicle
Received 8 November 2011 routings. The location routing problem (LRP) simultaneously tackles both location and routing decisions
Accepted 10 June 2012 to minimize the total system cost. In this paper a multiple ant colony optimization algorithm (MACO) is
Available online 17 June 2012
developed to solve the LRP with capacity constraints (CLRP) on depots and routes. We decompose the
Keywords: CLRP into facility location problem (FLP) and multiple depot vehicle routing problem (MDVRP), where
Logistics system the latter one is treated as a sub problem within the first problem. The MACO algorithm applies a
Location routing problem hierarchical ant colony structure that is designed to optimize different subproblems: location selection,
Multiple ant colony optimization customer assignment, and vehicle routing problem, in which the last two are the decisions for the
Nested method
MDVRP. Cooperation between colonies is performed by exchanging information through pheromone
updating between the location selection and customer assignment. The proposed algorithm is
evaluated on four different sets of benchmark instances and compared with other algorithms from
the literature. The computational results indicate that MACO is competitive with other well-known
algorithms, being able to obtain numerous new best solutions.
& 2012 Elsevier B.V. All rights reserved.

1. Introduction denotes the set of candidate depot locations. Each customer iAI
has a demand di. A capacity Rj and an opening cost fj are
The design of a logistics system is an important issue in today’s associated with each candidate depot site jAJ. Associated to each
competitive environment due to the significant contribution of edge (i, j)AE there is a routing cost cij which denotes the traveling
the distribution cost to the total supply chain cost. This kind of distance or traveling cost between nodes i and j. A set K of
problem is commonly solved in two phases: facility location for a homogeneous vehicles with capacity Q and cost C are available.
long term policy and vehicle routing to satisfying customer Each customer must be served exactly once by only one vehicle.
demands for the operational decisions. These two components Each route must begin and end at the same depot and its total
can be treated separately, but may lead to suboptimal solutions load cannot exceed vehicle capacity. The total load of the vehicles
(Salhi and Rand, 1989). The location routing problem (LRP) assigned to a depot cannot exceed the capacity of that depot. The
integrates facility location problem (FLP), which determines the objective is to find the optimal number and locations of the
depot locations and allocates customers to each selected depot, depots as well as the vehicle routes of each opened depot so as to
and vehicle routing problem (VRP), which constructs the vehicle minimize the sum of the fixed facility costs, transportation costs,
routes of the selected depot. Several real world applications can and vehicle costs.
be found in the literature, for example, bill delivery (Lin et al., The CLRP is very difficult to solve since it encompasses two
2002), parcel delivery (Wasner and Zäpfel, 2004), and mobile NP-hard problems: facility location problem and vehicle routing
network design (Billionnet et al., 2005). The LRP with capacities problem (Garey and Johnson, 1979). In CLRP, the location–
on both depots and routes is called capacitated LRP (CLRP) which allocation decision will influence the total cost of vehicle routes
is the focus of this paper. and the architecture of vehicle routes will affect the location of
The CLRP can be represented by a graph G ¼(V, E), where depots and allocation of customers. Consequently, how to deal
V¼I[J. I¼{1, y, n} is the set of customer nodes and J¼{1, y, m} with the interdependence between these decisions is an impor-
tant issue. In this paper, we solve both location and routing
n
problems simultaneously rather than independently with nested
Corresponding author. Tel.: þ886 3 4638800 2526; fax: þ 886 3 4638907.
E-mail addresses: ietingcj@saturn.yzu.edu.tw (C.-J. Ting),
methods based on the ant colony optimization algorithm.
Chiaho@gosmio.biz (C.-H. Chen). We apply a hierarchical structure, with facility location as the
1
Fax: þ886 4 22361120. main problem and vehicle routing as a subordinate one. To wit,

0925-5273/$ - see front matter & 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijpe.2012.06.011
 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44 35

we decompose the CLRP into facility location and multi-depot LRP into a limited set of MDVRP. The bounds provided by their
vehicle routing problem, while the latter problem is embedded model are very tight, being able to solve instances with up to 199
into the first one. This concept of hierarchy is also emphasized by customers and 14 facilities.
Balakrishnan et al. (1987) and Nagy and Salhi (1996). The As the LRP problem is NP-hard, most of the researches used
proposed multiple ant colony optimization algorithm (MACO) is heuristics to solve the LRP. Nagy and Salhi (2007) classified the
evaluated by four sets of CLRP benchmark instances from the heuristics into four different types as follows: sequential, cluster-
literature and its computational results are compared with state- ing-based, iterative, and hierarchical. Sequential methods solve
of-the-art algorithms. the location problem by minimizing the sum of facility to
The remainders of this paper are organized as follows. Section customer distance and the routing problem based on the selected
2 provides an extensive review of LRP in the literature. The depots sequentially. Clustering-based methods (Srivastava, 1993;
multiple ant colony optimization algorithm to tackle the CLRP is Barreto et al., 2007) partition the customers into clusters and then
described in Section 3. In Section 4, the computational results of find a depot for each cluster. The VRP is then solved for each
four groups of benchmark problems are reported. For each bench- cluster. Iterative methods (Hansen et al., 1994; Wu et al., 2002;
mark set we compare to the best available algorithms. Finally, Prins et al., 2007; Duhamel et al., 2010) decompose the LRP into
conclusions are followed in Section 5. two subproblems. Then, subproblems are solved iteratively by
feeding information from one subproblem to the other. Hierarch-
ical methods (Nagy and Salhi, 1996; Albareda-Sambola et al.,
2. Literature review 2005) consider the location problem as the main problem and the
VRP as a subordinated problem. Nagy and Salhi (2007) believed
The LRP has been studied for decades, there are a few LRP that hierarchical methods may provide better solutions. Based on
surveys in the literature (Laporte, 1988; Min et al., 1998; Nagy their observation, the proposed MACO in this paper is a hier-
and Salhi, 2007). Laporte (1988) reviewed early research on archical method.
location routing problems and summarized different types of Many heuristics that hybrid two different heuristic approaches
formulations, solution algorithms and computational results of are proposed in the literature. Tuzun and Burke (1999) proposed a
research published prior to 1988. Min et al. (1998) synthesized two-phase tabu search (TS) approach for the LRP. One phase seeks
the past evolution of location routing literature and explored a good facility configuration while the other one obtains a good
promising research opportunities in incorporation of more realis- routing for this configuration. Wu et al. (2002) presented a
tic aspects, algorithmic design and model complexity. Recently, combined TS and simulated annealing (SA) decomposition
Nagy and Salhi (2007) surveyed the state of the art in location approach to solve the multi-depot location routing problem with
routing problem. They proposed a classification scheme and multiple fleet types and limited number of vehicles for each
looked at a number of problem variants. They also investigated vehicle type. Lin et al. (2002) developed a meta-heuristic
exact and heuristic algorithms and presented some suggestions approach based on threshold accepting (TA) and SA to assist in
for future research. making decisions of facility location, vehicle routing and loading
Most early work on LRP considers either capacitated routes or decision for bill delivery services in Hong Kong.
capacitated depots, but not both (Laporte et al., 1988; Chien, Albareda-Sambola et al. (2005) proposed another two-phase
1993; Srivastava, 1993; Tuzun and Burke, 1999). Recently, a TS heuristic for the LRP which incurs not capacity constraints on
number of studies have been devoted to the case with capacitated vehicles. Wang et al. (2005) proposed a two-phase hybrid heur-
depots and routes (Wu et al., 2002; Prins et al., 2006a, 2006b; istic which decomposes the LRP into location–allocation problem
Bouhafs et al., 2006; Prins et al., 2007; Barreto et al., 2007; and vehicle routing problem. In the location phase, the TS was
Duhamel et al., 2010; Yu et al., 2010). Our study also considers applied to obtain the configuration of facility locations. For each
both depot and route capacities. selected facility location, a vehicle routing problem was solved by
Several exact methods have been devoted to solve the LRP, but ACO in the routing phase. Bouhafs et al. (2006) proposed a hybrid
optimal solutions are only limited to medium-scale or to basic algorithm which combined the SA and ant colony system (ACS) to
uncapacitated instances. Laporte and Norbert (1981) designed a solve the CLRP. A good configuration of facilities was first found
branch-and-bound algorithm for an LRP with a single open depot, by the SA, and then the ACS was applied to construct the routings
and solved instances with up to 50 customers. In Laporte et al. based on the configuration. These two ACO-related heuristics
(1986), the solution to an LRP with vehicle capacity constraints is construct the routing problem and feed back the information for
obtained by a branch-and-cut method. Subtour elimination con- the facility selection phase.
straints and chain-barring constraints guarantee that each route Prins and his coworkers conducted different heuristic methods
starts and ends at the same facility. Laporte et al. (1988) to the LRP. Prins et al. (2006a) combined greedy randomized
addressed an LRP with asymmetrical costs, in which vehicle adaptive search procedures (GRASP) and path relinking to
capacity is replaced by a maximum route length. They elaborated develop a two phase algorithm for the capacitated location
a branch-and-bound algorithm that is able to solve instances with routing problem. In the first phase, the GRASP and a learning
up to 40 customers, but the number of depots is small (2 or 3) and process were implemented to select depots. The second phase
the number of routes per opened depot is limited to 2. Akca et al. was to generate new solutions using a path relinking. Later, Prins
(2009) presented a set-partitioning based formulation of the LRP et al. (2006b) presented a memetic algorithm with population
and proposed a column generation approach to solve instances management (MA9PM) to solve the same problem. Prins et al.
with up to 40 customers. Belenguer et al. (2011) proposed a (2007) proposed a cooperative approach, which combines the
branch-and-cut algorithm based on a zero-one linear model Lagrangean relaxation and granular tabu search (GTS), to solve
strengthened by new families of valid inequalities for solving the capacitated LRP. The algorithm alternates between a location
the CLRP. They solved instances to optimality with up to 50 subproblem, solved by Lagrangean relaxation, and a multi-depot
customers and five depots. Baldacci et al. (2011) proposed a VRP, solved by the GTS. Duhamel et al. (2010) presented a GRASP
branch-and-cut-and-price algorithm for solving the CLRP based with evolutionary location search (GRASP  ELS) approach for
on a set-partitioning-like formulation of the problem. The lower the CLRP.
bounds produced based on dynamic programming and dual Barreto et al. (2007) integrated several hierarchical and non-
ascent methods, are used by an algorithm that decomposes the hierarchical clustering techniques in a sequential heuristic
36 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44

algorithm. Marinakis and Marinaki (2008) developed a hybrid

algorithm (HybPSO), which combined the particle swarm optimi-
zation (PSO), multiple phase neighborhood search-greedy rando-
mized adaptive search procedure (MPNS-GRASP), the expanding
neighborhood search (ENS) and path relinking, to solve the
location routing problem. More recently, Yu et al. (2010) pro-
posed a simulated annealing algorithm to solve the LRP. Both Yu
et al. (2010) and Duhamel et al. (2010) are, on average, the most
effective algorithms on benchmark instances from the literature.
The LRP is generally considered as a deterministic case in the
literature. A few researches have addressed stochastic versions of
the LRP (Laporte et al., 1989; Chan et al., 2001). Berman et al.
(1995) provided good survey on the stochastic LRP.

3. Multiple ant colony optimization for CLRP

This section introduces the proposed algorithm for solving the

CLRP. The proposed multiple ant colony optimization (MACO)
algorithm is based on the ant colony system (ACS) heuristic. The
ant system (AS) was the first ACO algorithm proposed by Dorigo
et al. (1996). Subsequently, many variants of ACO have been
developed and applied extensively in the fields of the combina-
torial optimization problems. Descriptions of available ACO algo-
rithms and related literature review can be obtained in Dorigo
and Stützle (2004).
The concept of multiple ant colonies was first proposed by
Gambardella et al. (1999) to solve the vehicle routing problem
with time windows (MACS-VRPTW). Two ant colonies were
designed to successively optimize two different objective func- Fig. 1. Pseudo code of the proposed MACO algorithm.
tions: the first colony is used to minimize the number of vehicles,
while the second colony minimizes the traveled distances. Our Location selection in line nine creates an ant with random
MACO adopts a hierarchical ACO structure with different transi- number of locations. Each ant may generate different number of
tion rules for different ant colonies. The upper level is for the locations. Customer assignment in line 10 then creates a new
location selection subproblem, while the lower level is for the assignment based on the selected locations. These two ant
MDVRP subproblem. The lower level is further decomposed into colonies will create a one-to-one mapping solution. The construc-
customer assignment and vehicle routing subproblems. The tion rules for both ant colonies are explained in Sections 3.2 and
coordination between the upper and lower level is achieved 3.3, respectively. VRP in line 11 creates vehicle routes for each
through the global pheromone updating rule. selected facility and customer assignment. This procedure is
explained later in Section 3.4 after introducing the transition
3.1. General structure rules for both location selection and customer assignment. The
local search procedure is explained in Section 3.5. In addition,
A solution of the CLRP consists in defining which depots must both local and global pheromone updates are made clear in
be opened, assigning each customer to one and only one opened Section 3.6. The notations used in MACO are listed in the
depot, and determining vehicle routes to serve customers. Three appendix.
ant colonies are used following this solution characteristic. The
first colony (location selection) is applied to determine the
location set. Then the customers are assigned to each selected 3.2. Location selection
facility location by a second colony (customer assignment). A VRP
for each selected facility location (route construction) is solved by The location selection will affect the customer assignment and
the third colony. These three ACOs are applied iteratively until the the vehicle routing construction significantly in the CLRP. Since
stopping criterion is met. Moreover, three different pheromone we consider the location selection as the main problem for the
matrices and pheromone updating rules are used to record CLRP, each ant h may have different number of facilities to select
pheromone information for each colony, respectively. in our MACO. The idea here is for ants to construct solutions
Fig. 1 illustrates the pseudo code of the proposed MACO diversely. The number of locations which will be selected by an
algorithm. The proposed algorithm consists of three inner loops. ant is calculated by Eq. (1).
The main loop (line 6 to 25) undergoes repetition until the $ P %
n
number of iterations for the CLRP problem is reached. Internal d
phs ¼ Pm i ¼ 1 i þUð1,rÞ ð1Þ
loop (line 8 to 17) undergoes repetition equal to number of ants j ¼ 1 Rj =m
(b). The third loop (line 11) which is implicitly undergoes the ant
colony optimization for vehicle routing. Every ant deemed to find where phs ( rm) is the number of locations of ant h at sth iteration,
the best vehicle routing for the selected locations. When the bxc is the largest integer smaller than or equal to x, U(1,r) is a
iteration best ant is obtained, the local search is applied with two random number following the uniform distribution in [1, r], and r
methods: insertion and move (lines 18 and19). Finally, when all is a pre-specified number. The first term is the number of facilities
ants generate their solutions, the global pheromone is updated in that should be selected based on the average facility capacity.
line 24. Then phs locations are successively chosen from m candidate sites
 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44 37

according to the selection rule, Eqs. (2) and (3).

8 h i
>
< arg max tsj ðZj Þa , if q r q0
j¼ j A Ohs ð2Þ
>
: J, otherwise

 a
tsj Zj
J : P hj ¼ P  a ð3Þ
s
t Zj
j A Ohs j

where Ohs is the set of candidate sites which are not selected yet
by ant h at sth iteration, tsj is the pheromone level of location j at
sth iteration, Zj is the ratio of capacity (Rj) to the fixed charge (fj), a
is the parameter that determines the relative influence of tsj
versus Zsj (a 40), and q is uniformly distributed within [0, 1].
With a prespecified frequency q0, the chosen candidate is the
location that gives the largest utility function tsj (Zsj )a among the
set of not yet selected candidate sites, Ohs (qrq0). Otherwise, it
is taken from Ohs using the probability distribution induced in
Eq. (3).

3.3. Customer assignment

Fig. 2. Pseudo code of the proposed ACO algorithm for the VRP.
After determining the facility location set, the MACO allocates
each customer i to a selected location k (iak) based on the
customer assignment rule by another ant colony. The location Step 3: If the capacity constraint of the location F is met or no
selection and customer assignment is mapped one on one. That is, selected location have spare capacity, go to Step 4. Otherwise,
we only construct one customer assignment solution based on the e¼eþ1, go to Step 2.
construction rule in Eqs. (4) and (5) for every facility location Step 4: If the capacity constraint of the location F is still
solution found in Section 3.2. violated, let phs ¼phs þ1 and go to the location selection phase.
8 h  b i Otherwise, stop.
>
< arg max xsik csik , if q0 rq00
k¼ k A W hs ð4Þ
>
: K, The idea to remove the largest demand customer in step 2 is to
otherwise
repair the infeasible solution into a feasible solution in a quick
 b way. If the capacity constant cannot be fixed, we just add one
xsik csik more location in step 4 and go back to the location selection to
K : P hik ¼ P  b ð5Þ
xs csik
k A W hs ik
generate another ant solution.

where Whs is the set of selected facility locations of ant h at the

location selection colony at sth iteration, xsik is the pheromone 3.4. Vehicle routing
level between customer i and location k at sth iteration, csik is the
reciprocal of Dik, which is given by After all customers are assigned to one of the selected facility
locations, the construction of vehicle routes for each location can
Dik ¼ min cil ð6Þ be regarded as an independent vehicle routing problem. In this
l A Ash
k
research, we apply an ACO to solve the VRP for each selected
where Ash k is the set of nodes including location k and those facility location and those customers assigned to the location.
customers that have been assigned to location k of ant h at Fig. 2 illustrates the pseudo code of the proposed ACO for the VRP.
iteration s (initially only location k itself). The idea here is that the The number of ants for VRP is b0 in line 9. Detailed procedures of
closer a customer to a selected facility, the higher possibility that our ACO for the VRP are described in the following sub-sections.
the customer will be served by the facility. Cil is the distance
between customers i and l. b is the parameter that determines the
relative effect of xsik versus csik (b 4 0), q0 is a random variable 3.4.1. Route construction rule of VRP
uniformly distributed in [0, 1], and the parameter q00 determines In our ACO, when located at node i, ant h0 moves to a node v by
the relative importance of exploitation Eq. (4) versus exploration the following state transition rule.
Eq. (5). 8 h  g i
> s0 s0
< arg max0 ziv fiv , if q00 r q000
After assigning the customers, the capacity restriction for each v Z h
v¼ A is0 ð7Þ
selected facility location FAWhs ¼{1, 2, y, phs } will be checked. If >
: V, otherwise
the capacity restriction of the location F is violated, the customer
assignment for F will be revised by the following procedures.  g
0 zs0iv fs0iv
V : P hiv ¼ P  g ð8Þ
Step 1: Sort the demand of customer assigned to the location F zs0 fs0iv
v A Z h00 iv
is
in descending order. Let [e] be the index of the customer with
0
the eth largest demand. Let e¼1. where Zish0 is the set of nodes which are not visited yet by ant h0 at
0
Step 2: Remove the customer [e] from the current location and node i at iteration s0 , zsiv is the pheromone of edge (i, v) at iteration
s0
re-assign this customer to the nearest selected location with s , jiv is defined as the savings of combining two nodes i and v on
0

spare capacity. one tour as opposed to serving them on two different tours at
38 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44

iteration s0 . The jsiv0 is calculated as follows. where

( 0
js0iv ¼ ci0 þ c0v civ ð9Þ ½ðL0w L0b Þ þ ðL0w L0s0 Þ=L0w if fði,vÞ A T 0b or T s0 g
Dzs0iv ¼ ð12Þ
0 otherwise
where civ is the distance between nodes i and v, and the node
0 denotes the selected facility location. g is the parameter that where L0b and L0s0 denote the tour length of the global-best solution
determines the relative influence of pheromone versus heuristic and the iteration-best solution of VRP, respectively, and L0w is the
information (g 40), q00 is a random variable uniformly distributed tour length of the worst solution of the current iteration. Edges
in [0, 1], and the parameter q000 determines the relative impor- that do not belong to the global-best solution and the iteration-
tance of exploitation Eq. (7) versus exploration Eq. (8). best solution just loose pheromone at the rate r00 , which con-
The starting node of each ant is randomly generated. Each ant stitutes the trail evaporation. This choice is intended to make ants
constructs the route based on the state transition rule in Eqs. searching in the neighborhood of the two best tours instead of the
(7) and (8). When the vehicle capacity is violated, the ant will globally best tour to avoid the algorithm being trapped in a local
return to the facility location, and then start from the facility to optimum without finding very good solutions.
build a new route. This process continues until each node is
visited. 3.5. Local search of MACO

After the initial solutions of the MACO are generated by the

solution construction rules, two local search approaches are
3.4.2. Local search of VRP
applied to improve the solutions. The first one is insertion move
In the typical ACO, after the ants have constructed their
between two opened facility locations. A customer is removed
solutions but before the pheromone is local updated, each ants’
from one route of one location, and then it is inserted into one
solution is improved by applying a local search. However, local
route of another location. The other one is swap move between
search is a time-consuming procedure of ACO. To save the
two opened facility locations. Two customers belonging to two
computation time, we will only apply local search to the best
different locations are exchanged. For these local search
solution built in this iteration. The idea here is that the local
approaches, the best improvement strategy is adopted. All possi-
optima reached from good solutions is expected to be better than
ble moves are evaluated and then the move with largest improve-
those obtained from poor solutions. Moreover, our ACO involves
ments is selected. We chose insertion and swap over other
three local search methods: 2-opt, swap and insertion (line 17 to
heuristics available for the local search based on their trade-off
19 in Fig. 2). In 2-opt, two non-consecutive arcs are removed,
between implementation simplicity, accuracy, and computational
either in the same route or in two different routes, and the two
performance.
paths created are reconnected to restore feasibility. Two custo-
mers are exchanged in swap. Insertion is to move one customer
from its current position to another position, in the same route or 3.6. Pheromone updating rules of MACO
in a different route.
Every time an ant constructs a solution, the quantity of
pheromone associated with the node for the location selection
and the customer assignment is decreased, and the location and
3.4.3. Pheromone updating rule of VRP assignment becomes less attractive. On the other hand, global
The pheromone updating of a typical ACO includes global and updating is used to intensify the search in the neighborhood of
local updating rules. The ants apply a local pheromone update the best solution computed. In our MACO, the pheromone updat-
rule immediately after they crossed an edge (i, v) during the tour ing rules of location selection and customer assignment are
construction. The local pheromone updating rule of our ACO is different. The respective local pheromone updating rules for both
selection and assignment pheromone matrices used in our MACO
zs0ivþ 1 ¼ ð1r00 Þzs0iv þ r00 z0 if fedge ði,vÞ A T h0 g ð10Þ are as follows.

where Th0 denotes the routes constructed by ant h0 , r00 is the tsj þ 1 ¼ ð1rÞtsj þ rt0j if flocation j A T h g ð13Þ
pheromone evaporation parameter in the range of [0, 1] that
regulates the reduction of pheromone on the edges. The z0 is the xsijþ 1 ¼ ð1r0 Þxsij þ r0 x0 if fedge ði, jÞ A T h g ð14Þ
initial value of the pheromone matrix for the route construction
rule, and is set to be 1/njLnn (line 5 in Fig. 2), where nj is the where Th is the solution constructed by ant h, 0r r, r0 r1 are the
number of nodes in the VRP for facility location j and Lnn is the pheromone decay parameters, and t0j ( ¼1/fj) and x0 (a very small
length of routes constructed by the Nearest Neighborhood value) are initial level of pheromone matrices, respectively.
heuristic. Since the location selection and customer assignment is one-
Dorigo et al. (1996) and Bullnheimer et al. (1999) used the to-one mapping, these two ant colonies communicate with each
elitist strategy on the trail updating in ant system. Such a strategy other through the global pheromone updating. Both location and
will direct the search of all the other ants in probability toward a assignment colonies use independent pheromone trails but col-
solution composed by some edges of the best tour itself. In our laborate by sharing the variable global-best solution (Tb) and the
ACO, the best elitist tours, including the global-best tour (T 0b ) and iteration-best solution (Ts) of CLRP. The global pheromone updat-
the iteration-best tour (T 0s0 ) of VRP, are allowed to lay pheromone ing for location selection will take into account the number of
on the edges that belong to them. The idea here is to balance customers assigned to each location. The respective global pher-
between exploitation (through emphasizing the global-best tour) omone updating rules for both selection and assignment pher-
as well as exploration (through the emphasis to the iteration-best omone matrices can be described as follows.
tour). The global updating rule of ACO for VRP is described as
tsj þ 1 ¼ ð1rÞtsj þ rDtsj ð15Þ
follow.

zs0ivþ 1 ¼ ð1r00 Þzs0iv þ r00 Dzs0iv ð11Þ xsijþ 1 ¼ ð1r0 Þxsij þ r0 Dxsij ð16Þ
 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44 39

where from 2 to 15 while the numbers of customers range from 12 to 85.

( The characteristics of Perl’s instances and computational results
½ðLw Lb Þ þ ðLw Ls Þ  nj =Lw if fj A T b or T s g
Dtsj ¼ ð17Þ obtained by MACO are summarized in Table 1. Columns 2–5 in
0 otherwise Table 1 show the number of customers n, the number of
 candidate sites m, vehicle capacity Q, the best-known solutions
½ðLw Lb Þ þ ðLw Ls Þ=Lw if fj A T b or T s and i is assigned to jg (BKS). The best known solution corresponds to the best known
Dxsij ¼
0 otherwise solution as reported by previous authors in the literature. Col-
ð18Þ umns 6–8 in Table 1 present the detail information of the best
where Lb and Ls denote the total cost of the global-best solution solution over 10 runs for each instance obtained by MACO,
and the iteration-best solution of the LRP, respectively. Lw is the including the best solution found in these 10 runs C, the running
total cost of the worst solution at current iteration, and nj time in seconds over the 10 runs (CPU), and the relative percen-
represents the number of customers assigned to location j. tage deviation over the best known solution (Gap), computed as
100%  (C  BKS)/BKS. The set of customers and candidate sites in
instances P3–P9 are the same, but the variable costs and the fixed
4. Computational results costs of these instances are different. The computational results
show that our MACO is very effective to solve the Perl’s instances.
In this section, we present the results of our computational The MACO reaches 1 best-known solution and finds eight new
experiments with the multiple ant colony optimization algorithm best solutions. The largest improvement is by 4.05%. For all
described in previous section. The proposed MACO algorithm is instances, the overall average gap over best known solution is
coded in Borland Cþþ Builder 5.0 and runs on a PC with an Athlon  1.37%.
XP 2500 þ(1.83 Ghz) processor and 512 MB RAM, under the The performance of MACO is compared with other methods
Windows XP operating system. We performed a set of prelimin- published in the literature, including Heuristic-P in Perl (1983),
ary experiments in order to find appropriate parameter settings
that produce overall good results across most instances, even if Table 2
they were not the optimal settings for all instances. We consider a Comparison of results for Perl’s instances.
set of parameters for each phase and then modifying one, while
Instance Heuristic-P Heuristic-H SA TS-ACO MACO
keeping the others fixed. The parameter tested include: rA{1, 2, 3,
4, 5}, a, b, gA{1, 2, 3, 4, 5}, r, r0 , r00 A{0.1, 0.3, 0.5, 0.7, 0.9}, q0, q00, P1 0.00 a
0.00 0.00 0.00 0.00
q000A{0.1, 0.3, 0.5, 0.7, 0.9}, bA{4, 10, 15, 20, 25}, b0 A{nj/10, nj/5, nj/3, P2 4.76 1.54 0.00 3.61  0.45
nj/2, nj}(nj is the number of customers assigned to depot j), SA{15, 25, P3 3.16 0.00 3.04 1.16  1.13
50, 75, 100} (maximum number of iterations of MACO) and S0 A{n/10, P4 7.65 0.00 –b –  1.49
P5 2.51 0.00 – –  2.19
n/5, n/3, n/2, n}(maximum number of iterations of VRP). Among all P6 1.79 0.00 – –  0.99
the preliminary experiments we found that the parameter setting: P7 2.03 0.00 – –  1.12
b¼4, b0 ¼dnj/5e, r¼3, a ¼1, b ¼1, g ¼ 4, r ¼ 0.1, r0 ¼0.1, r00 ¼0.1, P8 4.94 0.00 – –  4.05
q0 ¼0.5, q00 ¼0.1, q000 ¼0.5, S¼25 and S0 ¼dn/5e, provided the best P9 4.51 0.00 – –  0.91
Avg. gap (%) 3.48 0.17 1.01 1.59  1.37
results. We then test our MACO by four different groups of bench-
Avg. timec (sec) 136.26 147.52 4.98 4.12 42.56
mark instances from the literature, containing a total of 94 instances,
with the same parameter setting. The four sets of benchmark a
The best among all compared algorithms is in boldface.
b
instances used were designed by Perl (1983), Barreto (2004), Prins The instance was not tested.
c
et al. (2004) and Tuzun and Burke (1999). For each instance, the The average CPU time in seconds on the computer used by each algorithm.

algorithm is implemented for ten runs. The reported result is the best
found solution over the runs with the average CPU time in seconds
over the 10 runs. The best known solution in each instance was
Table 3
obtained based on the best results among comparing algorithms in
Computational results for Barreto’s instances.
their published version or during their parameter analyses.
Instance n m Q BKS C CPU Gap
4.1. Perl’s instances
B1 21 5 6,000 424.90n 424.90a 5.64 0.00
B2 22 5 4,500 585.10n 585.10 4.69 0.00
The first set of instances, which includes nine instances, was B3 29 5 4,500 512.10n 512.10 8.59 0.00
obtained from Perl (1983). The number of candidate sites ranges B4 32 5 8,000 562.22n 562.22 13.05 0.00
B5 32 5 11,000 504.30n 504.30 9.56 0.00
B6 36 5 250 460.40n 460.40 13.14 0.00
Table 1 B7 50 5 160 565.60n 565.60 29.16 0.00
Computational results for Perl’s instances. B8 75 10 140 844.40 844.88 58.72 0.06
B9 100 10 200 833.40n 836.75 83.92 0.40
Instance n M Q BKS C CPU Gap B10 12 2 140 204.00 204.00 2.09 0.00
B11 55 15 120 1,112.10 1,112.58 28.81 0.04
P1 12 2 140 355.58 355.58a 2.171 0.00 B12 85 7 160 1,622.50 1,623.14 77.86 0.04
P2 55 15 120 5,532.28 5,507.25 19.516  0.45 B13 318 4 25,000 557,275.2 560,210.81 831.34 0.53
P3 85 7 160 7,551.61 7,466.17 46.656  1.13 B14 318 4 8,000 673,297.70 670,118.50 1986.33  0.47
P4 85 7 160 8,374.73 8,249.97 52.843  1.49 B15 27 5 2,500 3,062.00n 3,062.00 8.59 0.00
P5 85 7 240 5,775.08 5,648.54 46.344  2.19 B16 134 8 850 5709.00 5709.00 136.63 0.00
P6 85 7 160 8,223.72 8,142.64 64.594  0.99 B17 88 8 9000,000 355.80 355.80 99.53 0.00
P7 85 7 160 7,257.72 7,176.64 64.578  1.12 B18 150 10 8000,000 43,919.90 44,131.02 166.95 0.48
P8 85 7 160 7,781.47 7,466.17 49.438  4.05 B19 117 14 150 12,290.30 12,355.91 77.45 0.53
P9 85 7 160 10,141.60 10,049.56 36.922  0.91 Avg. 68,639.00 68,641.00 191.69 0.08
Avg. 6,777.09 6673.61 42.56  1.37
n
Proved optimal solution.
a a
The solution is better than or equal to the best known solution is in boldface. The solution is better than or equal to the best known solution is in boldface.
40 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44

Table 4
Comparison of results for Barreto’s instances.

Instance CH SA-ACS GRASP MA9PM LRGTS HybPSO GRASP  ELS SALRP MACO

B1 2.59 1.29 1.11 0.00a 0.00 1.88 0.00 0.00 0.00

B2 1.09 0.27 0.00 4.56 0.39 0.58 0.00 0.00 0.00
B3 0.00 0.00 0.59 0.00 0.00 0.00 0.00 0.00 0.00
B4 1.69 1.26 1.73 1.73 4.48 1.53 0.00 0.00 0.00
B5 1.41 0.36 0.00 6.03 0.10 1.35 0.00 0.00 0.00
B6 2.24 2.17 0.00 5.43 3.50 2.24 0.00 0.00 0.00
B7 3.02 –b 5.92 0.00 3.68 3.02 0.00 0.00 0.00
B8 4.96 – 2.04 2.57 2.26 4.96 0.76 0.43 0.06
B9 6.72 – 3.38 2.00 1.14 6.72 0.00 0.59 0.40
B10 0.00 0.00 – – – 0.00 – 0.00 0.00
B11 2.17 0.57 – – – 2.14 – 0.06 0.04
B12 2.12 1.78 – – – 2.12 – 0.00 0.04
B13 4.20 – – – – 4.20 – 1.12 0.53
B14 11.04 – – – – 11.04 – 1.61  0.47
B15 0.00 0.00 0.00 0.00 0.10 0.00 0.00 0.00 0.00
B16 9.27 8.75 4.49 4.22 1.75 9.13 0.18 0.00 0.00
B17 8.18 – 0.31 0.00 3.63 8.18 0.00 0.00 0.00
B18 6.20 – 1.61 0.21 1.06 6.20 0.10 2.71 0.48
B19 1.50 – – – – 1.50 – 1.17 0.53
Avg. gap (%) 3.60 1.50 1.63 2.06 1.70 3.51 0.08 0.40 0.08
Avg. timec (sec) – – 20.28 35.62 17.58 48.48 187.62 464.07 191.69

a
The best among all compared algorithms is in boldface.
b
The instance was not tested.
c
The average CPU time in seconds on the computer used by each algorithm.

Heuristic-H in Hansen et al. (1994), SA in Wu et al. (2002) and TS- MA9PM in Prins et al. (2006b), the LRGTS in Prins et al. (2007),
ACS in Wang et al. (2005). Table 2 presents the comparison of the HybPSO in Marinakis and Marinaki (2008), the SALRP in Yu
results for Perl’s instances obtained by MACO and other methods. et al. (2010), and the GRASP  ELS in Duhamel et al. (2010).
All the percentage deviation values are computed from the Among the existing heuristics the LRGTS of Prins et al. (2007), the
original publications; a dash (‘-’) indicates that the instance is GRASP  ELS in Duhamel et al. (2010), and the SALRP of Yu et al.
not tested by the algorithm. The last row provides the average (2010) have obtained the best results in the literature. Note that
computation time in seconds. The best value on each instance is some of the instances are not tested by all these methods, direct
indicated in boldface. The results indicate that our MACO outper- comparisons may be in many cases biased. For those 13 instances
forms all compared methods on average for the gap and number that most algorithms tested, our MACO and GRASP  ELS are the
of best solutions found. On the other hand, the larger computa- best two approaches and provide similar solution quality. Four
tional effort of our MACO is due to the multi-start of initial algorithms (CH, HybPSO, SALRP, and MACO) test all 19 instances.
solution of MACO. However, its computational time is reasonable The results show that our MACO approach outperforms the other
for such a problem. three algorithms and provides the lowest largest gap of 0.53%. In
addition, the MACO also yields the best solutions among these
4.2. Barreto’s instances four algorithms in 18 out of 19 problems except instance B12.

The second set of 19 instances was collected from Barreto 4.3. Prins et al.’s instances
(2004) and the complete data sets are available at http://sweet.ua.
pt/ iscf143. Different from the first group of instances, the The third set of 30 instances was generated by Prins et al.
vehicle variable cost is not considered in the second group of (2004) which contains instances with capacitated routes and
instances. These instances are either from the literature or are depots. The complete data sets are available at http://prodhonc.
obtained by adding both capacitated and uncapacitated depots to free.fr/. Different from the first two groups of instances, the
classical VRP instances. The number of candidate sites ranges vehicle fixed cost is considered in this group of instances. The
from 2 to 15 while the numbers of customers range from 12 to number of candidate sites ranges from 5 to 10 whereas the
318. Table 3 presents the basic characteristics of Barreto’s number of customers ranges from 20 to 200. The vehicle capacity
instances and the computational results of MACO. The format of is either 70 or 150. Table 5 presents the information of Prins
the table is the same as that in Table 1. An asterisk on the best et al.’s instances and the computational results of the MACO. The
known solution indicates a proven optimal solution value. As can format of the table is the same as that in Table 1. An asterisk on
be seen from Table 3, the MACO obtains 11 best known solutions the best known solution indicates a proven optimal solution
and updates 1 best solution out of 19 instances. Our solutions are value. From Table 5 we can observe that the MACO reaches all
0.08% above the best known solution on average while the largest 12 best-known solutions for the small size instances. The average
gap is 0.53%. The computational results show that our MACO can gap of all instances is 0.36%, while the largest gap is 1.55%. The
solve the CLRP effectively. This is because the diversely sets of computational results show that our algorithm can effectively
facility locations are adopted in the MACO. It is useful to find a solve the large size LRP instances involving as many as 10
better set of facility locations, when the solutions are very candidate facilities and 200 customers.
sensitive to the choice of the facility locations. Moreover, the comparison of algorithms for Prins et al.’s
The comparison of algorithms for Barreto’s instances is pre- instances is presented in Table 6. The compared algorithms are
sented in Table 4. The compared algorithms are the clustering the MSLS in Prins et al. (2004), the GRASP in Prins et al. (2006a),
based heuristic (CH) in Barreto et al. (2007), the SA-ACS in the MA9PM in Prins et al. (2006b), the LRGTS in Prins et al. (2007),
Bouhafs et al. (2006), the GRASP in Prins et al. (2006a), the the SALRP in Yu et al. (2010), and the GRASP  ELS in Duhamel
 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44 41

Table 5 Table 6
Computational results for Prins et al.’s instances. Comparison of results for Prins et al.’s instances.

Instance n m Q BKS C CPU Gap Instance MSLS GRASP MA9PM LRGTS GRASP  ELS SALRP MACO

Pr1 20 5 70 54,793n 54,793a 4.08 0.00 Pr1 1.85 0.42 0.00a 0.62 0.00 0.00 0.00
Pr2 20 5 150 39,104n 39,104 4.78 0.00 Pr2 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Pr3 20 5 70 48,908n 48,908 3.92 0.00 Pr3 1.55 0.00 0.00 0.00 0.00 0.00 0.00
Pr4 20 5 150 37,542n 37,542 5.45 0.00 Pr4 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Pr5 50 5 70 90,111n 90,111 24.47 0.00 Pr5 8.84 0.58 0.05 0.05 0.00 0.00 0.00
Pr6 50 5 150 63,242n 63,242 21.31 0.00 Pr6 14.10 2.40 0.00 0.02 0.00 0.00 0.00
Pr7 50 5 70 88,298n 88,298 24.19 0.00 Pr7 2.14 0.55 0.00 0.47 0.39 0.00 0.00
Pr8 50 5 150 67,308n 67,308 20.13 0.00 Pr8 2.03 1.09 0.87 0.58 0.00 0.05 0.00
Pr9 50 5 70 84,055n 84,055 24.66 0.00 Pr9 –b 0.00 0.00 0.15 0.00 0.00 0.00
Pr10 50 5 150 51,822n 51,822 16.83 0.00 Pr10 – 0.46 0.00 0.33 0.00 0.00 0.00
Pr11 50 5 70 86,203n 86,203 32.78 0.00 Pr11 12.24 1.37 0.00 0.00 0.00 0.29 0.00
Pr12 50 5 150 61,830n 61,830 25.70 0.00 Pr12 0.02 0.10 0.00 0.00 0.00 1.41 0.00
Pr13 100 5 70 274,814n 276,220 116.59 0.51 Avg. gap 4.28 0.58 0.08 0.19 0.03 0.15 0.00
Pr14 100 5 150 213,615 214,323 134.52 0.33
Pr13 3.57 1.68 2.59 1.14 0.78 0.81 0.51
Pr15 100 5 70 193,671n 194,441 237.28 0.40
Pr14 2.08 1.19 1.42 0.59 1.05 1.12 0.33
Pr16 100 5 150 157,095n 157,222 144.33 0.08
Pr15 2.26 3.02 0.98 1.48 0.31 0.23 0.40
Pr17 100 5 70 200,079n 201,038 178.75 0.48
Pr16 1.73 1.56 0.15 0.44 0.18 0.04 0.08
Pr18 100 5 150 152,441n 152,722 151.58 0.18
Pr17 2.41 1.96 0.83 0.94 0.13 0.08 0.48
Pr19 100 10 70 287,983 291,134 105.28 1.09
Pr18 2.25 1.41 0.58 1.49 0.06 0.02 0.18
Pr20 100 10 150 231,763 235,348 81.81 1.55
Avg. gap 2.38 1.80 1.09 1.01 0.42 0.38 0.33
Pr21 100 10 70 243,590n 245,263 122.47 0.69
Pr22 100 10 150 203,988n 205,524 85.03 0.75 Pr19 12.31 12.22 9.93 1.36 4.67 1.06 1.09
Pr23 100 10 70 250,882 254,302 112.09 1.36 Pr20 19.81 17.14 16.61 1.63 16.32 1.06 1.55
Pr24 100 10 150 204,317 204,786 79.70 0.23 Pr21 19.95 4.31 0.63 1.28 0.08 0.91 0.69
Pr25 200 10 70 477,248 478,843 941.78 0.33 Pr22 20.65 1.26 0.52 0.22 0.00 0.65 0.75
Pr26 200 10 150 378,351 378,865 562.05 0.14 Pr23 2.81 7.95 1.11 3.10 1.05 0.00 1.36
Pr27 200 10 70 449,571 451,457 703.66 0.42 Pr24 2.46 5.80 0.24 0.77 0.38 0.34 0.23
Pr28 200 10 150 374,330 374,972 403.53 0.17 Avg. gap 13.00 8.11 4.84 1.39 3.75 0.67 0.95
Pr29 200 10 70 469,433 475,155 878.83 1.22
Pr25 20.72 2.84 1.31 0.93 1.93 0.79 0.33
Pr30 200 10 150 362,817 365,401 490.92 0.71
Pr26 6.28 10.15 0.45 0.60 1.05 1.38 0.14
Avg. 196,640.13 197,674.40 191.28 0.36
Pr27 22.76 14.04 0.50 0.84 0.45 0.28 0.42
n Pr28 3.91 1.51 0.18 0.81 0.22 0.63 0.17
Proved optimal solution.
a Pr29 18.23 5.81 1.85 1.54 1.91 0.95 1.22
The solution is better than or equal to the best known solution is in boldface.
Pr30 23.83 7.22 0.56 0.67 0.65 0.24 0.71
Avg. gap 15.96 6.93 0.81 0.90 1.10 0.71 0.50
Avg. gap 8.24 3.60 1.38 0.69 1.07 0.41 0.36
et al. (2010). Among those comparing algorithms, SALRP is on (%)
average the most effective on this set of benchmark instances Avg. timec 176.57 96.49 76.69 17.48 258.17 422.36 191.28
from the literature. As shown in Table 6, our MACO obtains (sec)

slightly better results than those by SALRP, and it also needs less a
The best among all compared algorithms is in boldface.
computational time. However, since the algorithms were run on b
The instance was not tested.
c
different machines and different compliers were used, it is The average CPU time in seconds on the computer used by each algorithm.
difficult to compare the run time directly. For all instances, the
gaps over best known solutions obtained by the MACO range from
0.0% to 1.55%. This indicates that our MACO is robust to find a
good solution. It can be observed that our MACO outperforms all of instances does not have facility capacity limitation, the results
the other algorithms on average in small size (n r50) and the are not as good as the other three sets of instances as we
largest size instances (n ¼200). The MACO takes longer computa- expected.
tional time than the compared algorithms but the running time is In Table 8, we present information about the comparison of
quite acceptable for a strategic problem like the CLRP, even on the our MACO with other algorithms. The compared algorithms are
largest instances (less than 16 min). the TS in Tuzun and Burke (1999), the GRASP in Prins et al.
(2006a), the MA9PM in Prins et al. (2006b), and the LRGTS in Prins
4.4. Tuzun and Burke’s instances et al. (2007), the SALRP in Yu et al. (2010), and the GRASP  ELS in
Duhamel et al. (2010). GRASP  ELS is on average the most
The fourth set of instances was first described in Tuzun and effective on this set of benchmark instances from the literature.
Burke (1999). The instances can also be found at http://prod On average, our MACO obtains very competitive average gap that
honc.free.fr/. Different from other three sets of instances, the is slighter better than that of the GRASP  ELS. Moreover, it is
Tuzun and Burke’s instances do not possess capacity restrictions noticeable that our MACO provides better results in the medium
for facilities. Therefore, we set the facility capacity as the total and large size instances (n ¼100, 200) than those by GRASP  ELS,
customer demand in this set of instances. Table 7 provides and the largest gap in these instances is much smaller than that of
characteristics of Tuzun and Burke’s instances and computa- the comparing algorithms. In addition, our MACO is the only
tional results of the MACO. An asterisk in the best known algorithm that updates best solutions in three instances. With
solution indicates that the solution was proven to be optimal. respect to computational times the MACO takes longer than the
The average gap of all instances obtained by the MACO is 0.90%, comparative algorithms but it finds good solutions within
while the gap ranges between 0.15% and 4.21%. Our MACO reasonable times.
reaches 4 best known solutions and updates 3 best solutions. Overall, we solved four groups of CLRP instances to test the
Note that our MACO was designed to solve CLRP based on the performance of the MACO. The computational time grows reasonably
facility capacity information to select the location. Since this set with the problem size. For all 94 tested instances, our algorithm was
42 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44

Table 7 Table 8
Computational results for Tuzun and Burke’s instances. Comparison of results for Tuzun and Burke’s instances.

Instance n m Q BKS C CPU Gap Instance TS GRASP MA9PM LRGTS GRASP  ELS SALRP MACO

T1 100 10 150 1467.68n 1489.68 71.08 1.50 T1 6.06 3.92 1.79 1.58 0.39a 0.65 1.50
T2 100 20 150 1449.20 1453.89 46.33 0.32 T2 5.71 5.36 1.53 1.56 0.00 1.50 0.32
T3 100 10 150 1394.80n 1407.78 60.98 0.93 T3 3.49 2.06 1.72 1.24 0.13 0.99 0.93
T4 100 20 150 1432.29 1433.42 54.05 0.08 T4 5.52 3.49 4.20 0.75 0.00 0.00 0.08
T5 100 10 150 1167.16n 1208.04 79.64 3.50 T5 5.48 2.83 0.52 1.75 0.00 0.86 3.50
T6 100 20 150 1102.24 1102.24a 64.53 0.00 T6 2.70 1.94 1.19 1.24 0.00 0.74 0.00
T7 100 10 150 791.66n 792.90 95.38 0.16 T7 4.23 2.82 0.29 2.73 0.05 0.00 0.16
T8 100 20 150 728.30 728.30 65.14 0.00 T8 1.68 2.68 0.30 2.01 0.00 0.50 0.00
T9 100 10 150 1238.24 1265.27 77.17 2.18 T9 6.36 2.82 1.94 2.40 0.17 0.02 2.18
T10 100 20 150 1245.31 1256.95 50.03 0.93 T10 2.34 2.22 0.48 0.87 0.06 0.16 0.93
T11 100 10 150 902.26n 902.26 60.56 0.00 T11 2.05 1.10 0.17 1.20 0.00 0.00 0.00
T12 100 20 150 1018.29 1018.29 68.56 0.00 T12 2.35 0.41 0.46 0.71 0.00 0.56 0.00
T13 150 10 150 1866.75 1945.43 226.63 4.21 Avg. gap 4.00 2.64 1.22 1.50 0.07 0.50 0.80
T14 150 20 150 1833.95 1853.22 100.88 1.05
T13 7.19 7.50 4.96 4.25 4.17 4.67 4.21
T15 150 10 150 1965.12 1991.44 200.81 1.34
T14 3.21 3.00 2.60 2.28 1.65 3.55 1.05
T16 150 20 150 1801.39 1812.34 140.59 0.61
T15 2.90 3.50 0.97 2.31 1.39 4.70 1.34
T17 150 10 150 1443.33n 1499.05 206.06 3.86
T16 2.97 3.04 2.99 1.03 1.88 0.00 0.61
T18 150 20 150 1441.98 1446.63 162.73 0.32
T17 7.79 4.50 0.34 0.37 0.72 0.69 3.86
T19 150 10 150 1205.09 1204.76 218.19  0.03
T18 2.55 1.03 1.24 3.53 0.15 0.94 0.32
T20 150 20 150 930.99 931.73 149.86 0.08
T19 2.18 2.93 0.19 0.50 1.23 0.10  0.03
T21 150 10 150 1699.92 1724.02 226.11 1.42
T20 1.86 1.05 0.41 0.64 1.59 0.39 0.08
T22 150 20 150 1400.01 1401.05 122.86 0.07
T21 3.68 2.18 1.20 1.73 0.72 1.23 1.42
T23 150 10 150 1199.51 1217.29 240.59 1.48
T22 6.31 1.84 2.09 1.76 0.21 1.13 0.07
T24 150 20 150 1152.18 1158.03 129.94 0.51
T23 5.43 2.02 0.33 1.40 1.28 1.44 1.48
T25 200 10 150 2259.87 2304.67 460.73 1.98
T24 2.61 6.87 0.55 0.87 0.33 0.60 0.51
T26 200 20 150 2185.41 2187.65 230.83 0.10
Avg. gap 4.06 3.29 1.49 1.72 1.28 1.62 1.24
T27 200 10 150 2234.78 2231.46 427.89  0.15
T28 200 20 150 2241.04 2275.70 233.89 1.55 T25 5.29 5.49 1.51 1.62 1.59 2.84 1.98
T29 200 10 150 2089.77 2098.56 569.08 0.42 T26 1.20 4.70 4.21 1.01 0.83 3.33 0.10
T30 200 20 150 1709.56 1711.25 276.17 0.10 T27 2.39 1.72 1.78 1.17 0.52 1.14  0.15
T31 200 10 150 1466.62 1472.93 544.30 0.43 T28 5.12 4.64 6.03 0.82 1.09 3.81 1.55
T32 200 20 150 1084.78 1087.57 316.20 0.26 T29 3.29 2.26 0.79 1.48 0.80 1.09 0.42
T33 200 10 150 1970.44 1978.74 386.44 0.42 T30 4.53 5.72 3.62 1.65 4.06 0.79 0.10
T34 200 20 150 1918.93 1959.71 229.63 2.13 T31 5.67 2.05 0.06 1.50 0.52 0.17 0.43
T35 200 10 150 1771.06 1782.94 405.49 0.67 T32 2.60 1.03 0.30 0.54 0.08 0.36 0.26
T36 200 20 150 1393.16 1392.70 268.45  0.03 T33 4.35 3.77 0.14 0.69 1.72 1.20 0.42
Avg. 1505.64 1520.22 201.88 0.90 T34 4.35 8.96 3.13 3.52 2.37 0.68 2.13
T35 6.00 1.00 0.63 0.89 0.44 0.45 0.67
a
The solution is better than or equal to the best known solution is in boldface. T36 1.56 1.11 0.22 0.57 4.35 0.23  0.03
Avg. gap 3.86 3.54 1.87 1.29 1.53 1.34 0.66

Avg. gap (%) 3.97 3.15 1.53 1.50 0.96 1.15 0.90
able to reach 28 best known solutions and obtain 12 new best Avg. Timeb 11.50 195.60 203.10 21.20 606.64 826.42 201.88
solutions. The computational results show that our MACO algorithm (sec)
can solve diverse instances effectively. Moreover, the performance of a
The best among all compared algorithms is in boldface.
our algorithm is also compared with other algorithms in the b
The average CPU time in seconds on the computer used by each algorithm.
literature. For those six algorithms that test the last three sets of
benchmark instances in Tables 4, 6, and 8, our MACO is the only one
that is able to update the best known solutions. The best two routing problem (MDVRP), while the first one is the main problem
algorithms that tested these three sets of instances from the literature and the latter problem to be a subordinate one. Thus, the MDVRP
are the GRASP  ELS by Duhamel et al. (2010) and SALRP by Yu et al. is solved embedded in the facility location problem. In each
(2010). The results show that our MACO approach is comparable to iteration, two ant colonies (location selection and customer
these two algorithms. It is worth observing that our MACO provides assignment) communicate with each other through the global
the lowest average gap among all algorithms for the last three sets of pheromone updating rule.
85 instances. Furthermore, the average gap by MACO in the largest The computational experiments are carried out on four sets of
size instances with number of customers more than 200 (nZ200) is instances from the literature. Our MACO is able to obtain optimal
much lower than the other algorithms. This indicates that our MACO or near-optimal solutions in a reasonable computation time for
can find good solutions in the problems close to the real and practical most benchmark instances. Our MACO reaches 28 and updates 12
situation. best known solutions, respectively, in 94 instances considered in
this study. The results show that the MACO can obtain good
solutions on various kinds of CLRP instances within reasonable
5. Conclusions computational times. The MACO is especially very effective to
solve the largest size LRP instances (n Z200) and the overall
In this paper we proposed a multiple ant colony optimization performance is competitive with other algorithms in the
(MACO) heuristic, which adopts a nested mechanism with three literature.
hierarchical solution construction rules, to solve the capacitated Our MACO checks the facility capacity constraint after finish-
location routing problem (CLRP). The CLRP is to solve each of the ing the customer assignment, in the future we could check this
three decisions in LRP: location selection, customer assignment during the customer assignment construction to reduce the
and route construction. We decompose the CLRP hierarchically computational time. Currently, each colony has its own control
into facility location problem (FLP) and multiple depot vehicle parameters in the MACO approach. It would be interesting to
 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44 43

study the effect of using the same control parameters in each tsj the pheromone level of location j at sth iteration, t0j is
colony in the future. Another research direction could be the the initial level
study on reliable location routing problem, in which some xsik the pheromone level between customer i and location k
facilities are subject to failure. This could be happened in the at sth iteration, x0 is the initial level
disaster relief network. Other possible perspective could be zs0iv the pheromone of edge (i, v) at iteration s0 , z0 is the
applying the MACO to other combinatorial optimization problems initial level
that contain multiple-level decisions.

References
Acknowledgements

Akca, Z., Berger, R.T., Ralphs, T.K., 2009. A branch-and-price algorithm for
This research was supported by a grant (NSC-100-2221-E-155- combined location and routing problems under capacity restrictions. In:
022) from the National Science Council of Taiwan. The authors Chinneck., J.W., Kristjansson, B., Saltzman, M.J. (Eds.), Operations Research
would also like to thank two anonymous referees for their helpful and Cyber-infrastructure, 47. Springer, pp. 309–330.
Albareda-Sambola, M., Diaz, J.A., Fernandez, E., 2005. A compact model and tight
comments and suggestions on improving the presentation of bounds for a combined location–routing problem. Computers & Operations
this paper. Research 32 (3), 407–428.
Balakrishnan, A., Ward, J.E., Wong, R.T., 1987. Integrated facility location and
vehicle routing models: recent work and future prospects. American Journal of
Mathematical and Management Sciences 7 (1–2), 35–61.
Baldacci, R., Mingozzi, A., Wolfler Calvo, R., 2011. An exact method for the
Appendix
capacitated location–routing problem. Operations Research 59 (5),
1284–1296.
The following notations are used for the MACO algorithm. Barreto, S., Ferreira, C., Paixa~ o, J., Santos, B.S., 2007. Using clustering analysis in
capacitated location–routing problem. European Journal of Operational
Research 179 (3), 968–977.
Ash
k the set of nodes including location k and those custo- Barreto, S.S., 2004. Análise e modelizac- a~ o de problemas de localizac- a~ o-distribuic- a~ o
mers that have been assigned to location k of ant h at sth [Analysis and Modelling of Location–Routing Problems]. Dissertation. Univer-
sity of Aveiro, Aveiro, Portugal.
iteration
Belenguer, J.M., Benavent, E., Prins, C., Prodhon, C., Wolfler Calvo, R., 2011. A
b, b0 the number of ants for the MACO and VRP, respectively banch-and-cut method for the capacitated location–routing problem. Compu-
cil the distance between customers i and l ters & Operations Research 38 (6), 931–941.
Berman, O., Jaillet, P., Simchi-Levi, D., 1995. Location–routing problems with
C the vehicle acquisition cost
uncertainty. In: Drezner, Z. (Ed.), Facility Location: A Survey of Applications
Dik the minimum distance between customer i and all and Methods New York. Springer, pp. 427–452.
customers served by location k Billionnet, A., Elloumi, S., Djerbi, L.G., 2005. Designing radio-mobile access net-
fj fixed charge of location j works based on synchronous digital hierarchy rings. Computers & Operations
Research 32 (2), 379–394.
Lb, Ls the total cost of the global-best solution and the itera- Bouhafs, L., Hajjam, A., Koukam, A., 2006. A combination of simulated annealing
tion-best solution of the LRP, respectively and ant colony system for the capacitated location–routing problem. Lecture
L0b , L0s0 the tour length of the global-best solution and the Notes in Computer Science 4251, 409–416.
Bullnheimer, B., Hartl, R.F., Strauss, C., 1999. A new rank based version of the ant
iteration-best solution of VRP, respectively system—a computational study. Central European Journal of Operations
Lw the total cost of the worst solution at current iteration Research 7 (1), 25–38.
L0w the tour length of the worst solution at current iteration Chan, Y., Carter, W.B., Burnes, M.D., 2001. A multiple-depot, multiple-vehicle,
location–routing problem with stochastically processed demands. Computers
nj the number of nodes in the VRP for facility location j & Operations Research 28 (8), 803–826.
Ohs the set of candidate sites which are not selected yet by Chien, T.W., 1993. Heuristic procedures for practical-sized uncapacitated location-
ant h at sth iteration capacitated routing problems. Decision Sciences 24 (5), 995–1021.
Dorigo, M., Maniezzo, V., Colorni, A., 1996. Ant system: optimization by a colony of
phs the number of locations of ant h at sth iteration cooperating agents. IEEE Transactions on Systems, Man and Cybernetics Part B
q, q0 , q00 random variables uniformly distributed in [0, 1] 26 (1), 29–41.
Q the vehicle capacity Dorigo, M., Stützle, T., 2004. Ant Colony Optimization. Bradford Books. MIT Press,
Cambridge, MA.
Rj the capacity of location j
Duhamel, C., Lacomme, P., Prins, C., Prodhon, C., 2010. A GRASP  ELS approach for
T 0b , T 0s0 the global-best tour and the iteration-best tour of VRP, the capacitated location–routing problem. Computers & Operations Research
respectively 37 (11), 1912–1923.
Th the facility location solution and customer assignment Gambardella, L.M., Taillard, E., Agazzi, G., 1999. MACS-VRPTW: a multiple ant
colony system for vehicle routing problems with time windows. In: Corne, D.,
constructed by ant h Dorigo, F., Glover, F. (Eds.), New Ideas in Optimization. McGraw-Hill, London,
T h0 the VRP solution constructed by ant h0 pp. 63–76.
U(1,r) a random number following the uniform distribution in Garey, M.R., Johnson, D.S., 1979. Computers and Intractability: A Guide to the
Theory of NP-Completeness. W. H. Freeman and Co, New York.
[1, r] Hansen, P.H., Hegedahl, B., Hjortkjær, S., Obel, B., 1994. A heuristic solution to the
W hs the set of selected facility locations of ant h at the warehouse location–routing problem. European Journal of Operational
location selection colony at sth iteration Research 76 (1), 111–127.
Laporte, G., 1988. Location routing problems. In: Golden, B.L., Assad, A.A. (Eds.),
Z h0
is0 the set of nodes which are not visited yet by ant h0 at Vehicle Routing: Methods and Studies. North-Holland, Amsterdam,
node i at iteration s0 pp. 163–197.
a the parameter that determines the relative influence of Laporte, G., Louveaux, F., Mercure, H., 1989. Models and exact solutions for a class
of stochastic location–routing problems. European Journal of Operational
tsj versus Zsj Research 39 (1), 71–78.
b the parameter that determines the relative effect of xsik Laporte, G., Nobert, Y., Taillefer, S., 1988. Solving a family of multi-depot vehicle routing
versus csik and location–routing problems. Transportation Science 22 (3), 161–172.
Laporte, G., Norbert, Y., 1981. An exact algorithm for minimizing routing and
g the parameter that determines the relative influence of
operating costs in depot location. European Journal of Operational Research 6
zs0iv versus js0iv (2), 224–226.
Zj the ratio of capacity and the fixed charge for location j Laporte, G., Norbert, Y., Arpin, D., 1986. An exact algorithm for solving a
r, r0 , r00 the pheromone evaporation parameter capacitated location–routing problem. Annals of Operations Research 6 (9),
293–310.
js0iv the savings of combining two nodes i and v into one tour Lin, C.K.Y., Chow, C.K., Chen, A., 2002. A location–routing-loading problem for bill
csik the reciprocal of Dik delivery services. Computers & Industrial Engineering 43 (1), 5–25.
44 C.-J. Ting, C.-H. Chen / Int. J. Production Economics 141 (2013) 34–44

Marinakis, Y., Marinaki, M., 2008. A particle swarm optimization algorithm with Prins, C., Prodhon, C., Wolfler Calvo, R., 2007. Solving the capacitated location–
path relinking for the location routing problem. Journal of Mathematical routing problem by a cooperative Lagrangean relaxation–granular tabu search
Modelling and Algorithms 7 (1), 59–78. heuristic. Transportation Science 41 (4), 470–483.
Min, H., Jayaraman, V., Srivastava, R., 1998. Combined location–routing problems: Salhi, S., Rand, G.K., 1989. The effect of ignoring routes when locating depots.
a synthesis and future research directions. European Journal of Operational European Journal of Operational Research 39 (2), 150–156.
Research 108 (1), 1–15. Srivastava, R., 1993. Alternate solution procedures for the locating–routing
Nagy, G., Salhi, S., 1996. Nested heuristic methods for the location–routing problem. OMEG 21 (4), 497–506.
problem. Journal of the Operational Research Society 47 (9), 1166–1194. Tuzun, D., Burke, L.I., 1999. A two-phase tabu search approach to the location
Nagy, G., Salhi, S., 2007. Location–routing: issues, models and methods. European routing problem. European Journal of Operational Research 116 (1), 87–99.
Journal of Operational Research 177 (2), 649–672. Wang, X., Sun, X., Fang, Y., 2005. A Two-phase hybrid heuristic search approach to
Perl, J., 1983. A Unified Warehouse Location–Routing Analysis. Ph.D. dissertation, the location–routing problem. 2005 IEEE International Conference on Systems,
Northwestern University, Illinois, USA.
Man and Cybernetics, 4, 3338–3343.
Prins, C., Prodhon, C., Wolfler Calvo, R., 2004. Nouveaux algorithmes pour le
Wasner, M., Zäpfel, G., 2004. An integrated multi-depot hub-location vehicle
proble me de localisation et routage sous contraintes de capacité. In: Dolgui, A.,
routing model for network planning of parcel service. International Journal of
Dauze re-Pére s, S. (Eds.), MOSIM’04, 2, Lavoisier. Ecole des Mines de Nantes,
Production Economics 90 (3), 403–419.
France, pp. 1115–1122.
Wu, T.H., Low, C., Bai, J.W., 2002. Heuristic solutions to multi-depot location–
Prins, C., Prodhon, C., Wolfler Calvo, R., 2006a. Solving the capacitated location–
routing problem by a GRASP complemented by a learning process and a path routing problems. Computers & Operations Research 29 (10), 1393–1415.
relinking. 4OR 4 (3), 221–238. Yu, V.F., Lin, S.W., Lee, W.Y., Ting, C.J., 2010. A simulated annealing heuristic for the
Prins, C., Prodhon, C., Wolfler Calvo, R., 2006b. A memetic algorithm with capacitated location routing problem. Computers & Industrial Engineering 58
population management (MA9PM) for the capacitated location–routing pro- (2), 288–299.
blem. Lecture Notes in Computer Science 3906, 183–194.