Applied Soft Computing 32 (2015) 154–163

Contents lists available at ScienceDirect

Applied Soft Computing

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

An ACO hybrid metaheuristic for close–open vehicle routing problems

with time windows and fuzzy constraints
J. Brito a,∗ , F.J. Martínez a , J.A. Moreno a , J.L. Verdegay b
a
Group of Intelligent Computing, I.U.D.R. University of La Laguna, E-38271 La Laguna, Spain
b
Department of Computer Science and Artificial Intelligence, University of Granada, E-18071 Granada, Spain

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

Article history: The close–open vehicle routing problem is a realistic variant of the “classical” vehicle routing problem
Received 9 July 2014 where the routes can be opened and closed, i.e. all the vehicles are not required to return to the depot
Received in revised form 14 January 2015 after completing their service. This variant is a planning model that is a standard practice in business
Accepted 12 March 2015
nowadays. Companies are contracting their deliveries to other companies that hire vehicles, and payment
Available online 20 March 2015
is made based on the distance covered by the vehicles. Available information on parameters in real
world situations is also imprecise, and must be included in the optimization model and method. The
Keywords:
aims of this paper are to formulate a model of this novel variant with time windows and imprecise
Close–open vehicle routing problem with
time windows
constraints and to propose a fuzzy optimization approach and a hybrid metaheuristic for its solutions.
Hybrid metaheuristic The full proposal is applied to a real route planning problem with outsourcing, obtaining promising
Ant colony optimization practical results. Customer demands and travel times are imprecise, thus capacity and time windows
Fuzzy constraints constraints are considered flexible and modelled as fuzzy constraints.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction a priority. However, in-depth knowledge of the models on which

they are to be based is required for such systems.
Outsourcing is a current business practice that provides The most widely studied and used model for route planning
increases in efficiency or quality in services and operations. Activ- is the vehicle routing problem (VRP). This model was introduced
ities that are not directly included in strategic business objectives by Dantzig and Ramser in 1959 [7], as a CVRP (capacitated-VRP)
can be outsourced because they do not produce added value or but its current relevance is beyond doubt. In short VRP is the
cost savings. This situation occurs in logistics and transport activi- determination of the optimal set of routes to be performed by a
ties, where many companies opt for collaboration and recruitment fleet of vehicles, in order to satisfy the demand of a given set of
of specialist companies to carry out these activities. Thus, compa- customers. Several variants of this problem have been formulated
nies with their own resources hire the services of other companies, and studied. In particular the vehicle routing problem with time
with complementary resources, to provide for occasional needs or windows (VRPTW) is used for its practical applications. In this
to fulfil part of their business strategy. variant customers must be served in a specific time interval, which
Route planning is among the operational decisions of logistics, are included as time windows constraints [50]. In other problems
transportation and distribution companies that have a high impact vehicles need not return to the depot after delivering goods to
on the efficiency and quality of services. Nowadays these problems customers, and instead go to other locations. This variant of the
are of the utmost importance, and not only because of economic VRP, in which vehicles do not return to the depot is called the
reasons. The possibility of obtaining high quality solutions as soon open vehicle routing problem (OVRP) or the open vehicle routing
as possible with independence of the human operators cannot be problem with time windows (OVRPTW) when time windows
ignored. Human interaction can fail, delay answers or cause unre- constraints are specified. The OVRP was first described in a paper
coverable errors. Hence, the availability of intelligent systems that by Schrage [44] and specified the differences between closed and
effectively implement operative models and suitable solutions is open trips. Bodin et al. [5] proposed the first solution approach
for the OVRP. OVRPTW was first addressed as a whole by Repous-
sis et al. in [39]. These problems are NP-hard [6] which make
∗ Corresponding author. Tel.: +34 922318190; fax: +34 922318170. them intractable with large instance problems, as is usual in real
E-mail addresses: [email protected] (J. Brito), [email protected] (F.J. Martínez), applications, and hence the need of using heuristics and meta-
[email protected] (J.A. Moreno), [email protected] (J.L. Verdegay). heuristics to provide feasible methods to find optimal solutions.

http://dx.doi.org/10.1016/j.asoc.2015.03.026
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163 155

In recent years several researchers have used various heuristics that are coherent with their fuzzy nature. The fuzzy mathemati-
and metaheuristics to solve the OVRP with success. The most cal programming techniques are brief reviewed in [24], which also
common metaheuristics are based on search, such as tabu search includes this approach as a new trend.
[6,9,15,16,23,42], neighbourhood-based search [14,37,43,57], or COVRPTW with fuzzy constraints can be formulated as fuzzy
threshold accepting algorithm [47,48]. Other researchers have optimization problems. In the literature a model with fuzzy
applied bio-inspired and populations based metaheuristics, par- demand for OVRP is proposed in [13]. In order to solve this problem,
ticle swarm optimization [18,35,54,59], ant colony optimization we propose the combination of FLP approach methods and hybrid
[29,30] or genetic and evolutionary computing [36,40,55]. The metaheuristics to obtain solutions, which are feasible in practice.
OVRPTW has also been studied, but with fewer references in These hybrid solution approaches combine ideas from ACO (ant
the literature, [27,28], in these cases applied to the genetic and colony optimization), GRASP (greedy randomized adaptive search
ant colony algorithm, respectively. Other variants of the OVRP procedure) and VNS (variable neighbourhood search).
have also been considered, with special nodes called driver nodes ACO is a population-based swarm intelligence algorithm and
and time deadlines using tabu search metaheuristics [3] or the was proposed by Dorigo and Gambardella [11]. Swarm intelli-
multi-depot OVRP by applying threshold accepting algorithm [46]. gence and Ant Colony procedures have been successfully applied
Some companies may be aware of specific time periods when to some classic optimization problems and transportation prob-
there is an increase in demand. These periods exert additional lems, including VRPTW [1,17,20,49,45,60]. These metaheuristics
service demands on its own fleet and compromise the quality of are applied to solve VRPTW with good results, especially hybrid
service. However, this occasional additional demand does not jus- versions. Zhan et al. (2008) presented a hybrid ACS with dynamic
tify investment in new vehicles. One solution to this situation is sweep algorithms [58]. Marinakis et al. (2010) proposed several
to outsource services from other companies to meet the additional hybrids combining PSO with neighbourhood search strategies,
demand. A specific practical route planning problem is one that GRASP and path relinking (PR) strategy in [32] and also with a
occurs in companies that are willing to collaborate and subcon- genetic algorithm in [33]. Yu et al. (2011) used an ACO to improve
tract. They must decide which demand (customers) will be served diversity with a neighbourhood search and tabu search [56]. Ding
by their own fleet and which demand will be served by another et al. (2012) combined ACO with the saving algorithm and -
operator with an external fleet, and then determine how to organize interchange mechanism [10]. Recent researches have proposed
the different routes. The company has to plan routes and allocate some novel schemes to these population metaheuristics which
customers and vehicles, where their own vehicles must return to would be of interest to solve these problems. For instance, immi-
the depot while external ones do not return. Route planning with grants schemes introduced in [34] for increasing diversity in ACO.
owned vehicles that return to depot and outsourced vehicles that Other approaches include the knowledge extraction process pro-
do not need to return correspond to a mixed OVRP and VRP prob- posed in the cultural algorithm [2], the adaptation of PSO based on
lem, and needs to be modelled. This version of the close–open VRP the law of gravity and mass interactions described in [38] and the
was recently addressed in [31]. In this paper we address and focus use of the proportional fairness ordered proposed in [26].
on the interesting version of the close–open vehicle routing prob- The rest of the paper is organized as follows. The next section
lems with time windows (COVRPTW). describes the COVRPTW formulations when capacity and time win-
Moreover, in practical applications of routing problems, cus- dow constraints are fuzzy. Section 3 explains the fuzzy approach to
tomer demand and travel time are often imprecise, not in a solve these variants of VRPTW with fuzzy constraints. In Section 4
probabilistic sense, but in a linguistic one. The difficulty for a cus- we present the hybrid metaheuristic procedure based on GRASP
tomer to know its demand beforehand is quite common, just as is and VNS combined with an ACO design to find solutions. In Section 5
the establishment of a specific form of quantity demanded. Thus it computational experiments and results of applying our approach
is difficult to measure in specific units an amount that will allow to a real problem are described. Finally, conclusions are included in
some comparison criteria to be established with vehicle capacity. the last section.
In addition, road conditions or traffic congestion cannot allow an
accurate travel time to be set. In the problem here considered, cus-
tomer demand and travel time are linguistically imprecise, thus 2. Fuzzy model for COVRPTW
capacity and time windows constraints can be considered flexible
and modelled as fuzzy constraints. An operative approach to man- We propose a COVRPTW model to approach and solve route
age this kind of uncertainty is required when studying and solving planning problems of distribution companies that have a number of
these problems. own vehicles and can contract external vehicles which must serve
Soft computing is presented as an adequate family of problem specified and dispersed demand points, satisfying a finite set of con-
resolution methods such as probabilistic reasoning, fuzzy logic and straints and minimizing total operating costs. The objective consists
fuzzy sets, neural networks and metaheuristics, in addition to all of in finding a set of routes that minimize total distance travelled. Each
the hybrid models which have been derived from the combination route starts at the depot and ends either at the depot (cycle/close
of these components. A variety of soft computing techniques have route) or at a customer (path/open route). The route type depends
been employed to supply chain management, including the vehi- on whether the vehicles are owned or not. Each customer is vis-
cle routing area [25]. In particular fuzzy sets and systems provide ited once and only once by exactly one vehicle and the demand
a suitable methodological approach for dealing with uncertainty of each customer must be completely fulfilled by a single vehicle
which is a product of the imprecise nature of the information and within flexible and imprecise time windows. Additionally, the total
decisions. Likewise metaheuristics are a good referent because they demand served by each vehicle must not exceed vehicle capacity
offer efficient solutions and strategies that are integrated with other with a certain tolerance.
soft-computing tools to facilitate approximate solutions to more The COVRTW includes a single depot, a known set of n customers
complex real world problems [53]. and a variable set of vehicles. The customers are identified by an
In this paper we use the ideas introduced by Bellman and Zadeh index i, i = 1, 2, . . ., n, and the index 0 denotes the depot. The cost
in 1970 [4], for fuzzy optimization problems and the model devel- of travelling between nodes i and j, where i, j = 0, 1, 2, . . ., n repre-
oped by Verdegay et al. [7,51]. These models provide a fuzzy linear sent nodes, is denoted by cij and tij denotes travel time from node
programming (FLP) approach with a number of methods for solv- i to node j. Each customer has demand amount qi and an unload
ing these problems, in a direct and easy way, obtaining solutions time ui . Each customer i has a time window [ei , li ] within which it
156 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163

must be served. All vehicles leave from the depot to carry out the then we are faced with a fuzzy optimization problem. Discussions
deliveries. Each vehicle is represented by an index k, k = 1, 2, . . ., m, concerning solutions do not focus on their feasibility, nor if they
and has a capacity Q. The first p indices correspond to the owned are optimal solutions or not. We, in turn, have chosen to discuss
vehicles that return to the depot. However, for k > p, vehicle k ends the degree of feasibility and optimality of the solution. In fuzzy
its route at the last visited customer. The model contains two sets optimization problems objectives and constraints can be defined
of decision variables xijk and sik , i, j = 0, 1, . . ., n, k = 1, 2, . . ., m. For in an imprecise way and are characterized using a membership
each arc two indices i and j and each vehicle k, we define xijk = 1 if function such as fuzzy sets.
Fuzzy linear programming constitutes the basis for solving fuzzy
and only if vehicle k travels from i to j, and xijk = 0, otherwise. The
optimization problems and their solution methods have been the
decision variable sik is defined for each customer i and each vehicle subject of many studies in the fuzzy context. Different FLP models
k to denote the time vehicle k starts to service customer i. can be considered according to the elements that contain imprecise
The COVRPTW is formulated as a linear programming problem information that are used as a basis for the classification proposed
as follows: in [51]. These models are: models with fuzzy constraints, mod-
Minimize els with fuzzy goals, models with fuzzy costs and models with

m 
n

n fuzzy coefficients in the technological matrix and resources. In
cij xijk (1) addition a fifth model, the general fuzzy problem, in which all of
k=1 i=0 j=0 the parameters are subject to fuzzy considerations, can be stud-
ied. The corresponding approaches offer solutions to FLP [8] that
Subject to: provide methods for solving COVRPTW with fuzzy terms. Therefore

m 
n this problem can be solved in a direct and simple way, obtaining
xijk = 1 j = 1, 2, . . ., n (2) solutions that are coherent with their fuzzy nature.
k=1 i=0 This FLP model is formulated as:


n Minimize z = cx
k
x0j = 1 k = 1, 2, . . ., m (3) (12)
subject to Ax≤f b [x ≥ 0]
j=1
where c ∈ Rn , b ∈ Rm and A ∈ Rm×n ; i.e., c and b are n-vector and m-

n
vector of real numbers and A is an (m × n)-matrix of real numbers.
k
xi0 = 1 k = 1, 2, . . ., p (4)
This is the case where constraints can be satisfied, and conse-
i=1 quently the feasible region can be defined as a fuzzy set; it should be

n

n defined by means of a membership function  : R → [0, 1]. In such
xijk − xjik ≥ 0 k = 1, 2, . . ., m, j = 1, 2, . . ., n (5) a situation, for each constraint, a desirable upper bound b is consid-
i=0 i=0 ered. However, a greater value is accepted until a maximum b + 
( is referred to as violation tolerance level). The symbol ≤f indicates

n

n
the imprecision of the constraints and where each fuzzy constraint
qj xijk ≤f Q k = 1, 2, . . ., m, (6)
ai x ≤ f bi is specified by a membership function in the form:
i=0 j=1
⎧
⎪ 1 ai x ≤ bi
⎨
sik + ui + tij ≤ sjk + M(1 − xijk ) k = 1, 2, . . ., m, i, j = 1, 2, . . ., n
i (ai x) = fi (ai x) bi ≤ ai x ≤ bi + i (13)
(7) ⎪
⎩
0 ai x ≥ bi + i
ej ≤f sjk k = 1, 2, . . ., m, j = 1, 2, . . ., n (8)
This means that, for each constraint i, given the level of tolerance
 i , to each point (n-dimensional vector) x is associated a number
sjk + uj ≤f lj k = 1, 2, . . ., m, j = 1, 2, . . ., n (9)
i (x) ∈ [0, 1] known as the degree of fulfilment of the constraint
sik ≥ 0 k = 1, 2, . . ., m, j = 1, 2, . . ., n (10) i. the functions fi are assumed to be continuous and monotonous
non-decreasing functions with fi (bi ) = 1 and fi (bi +  i ) = 0, usually the
xijk ∈ {0, 1} k = 1, 2, . . ., m, i, j = 1, 2, . . ., n. (11) linear function f(z) = 1 − (bi − z)/ i .
In particular, Verdegay [52], using the representation theorem
Eq. (1) represents the objective function in terms of costs. Con- for fuzzy sets, proves that the solutions for the case of linear func-
straint (2) guarantees that exactly only one route/vehicle visits each tions fi can be obtained from the auxiliary model:
customer. Constraints (3) and (4) ensure that each vehicle leaves
the depot and each owned vehicle returns to it. Constraint (5) estab- Minimize z = cx
lishes that if a vehicle leaves a customer then it has reached it. (14)
subject to Ax ≤ b + (1 − ˛) [x ≥ 0, ˛ ∈ [0, 1]]
Constraint (6) ensures that the total of customer demand in any
route does not exceed the vehicle capacity Q. Constraint (7) states where  = (1 , 2 , . . ., m ) ∈ Rm is the tolerance level vector.
that a vehicle k cannot arrive at j before sik + tij if it is travelling from Let X(˛) = {x ∈ Rn /Ax ≤ b + (1 − ˛)} denote the ˛-cut of the
i to j and guarantee subtours eliminations (M is a large scalar). Con- feasible solution set, the corresponding ˛-cut of the optimal solution
straints (8) and (9) ensure that time windows are observed. Finally set is S(˛) = {x ∈ Rn /z(x) = cx, x ∈ X(˛)}. Note that, given certain
(10) and (11) establish the conditions of the variables. Note that values of tolerance for the constraints, for each value of ˛ we can
symbols ≤f denote that constraints (6), (8) and (9) are considered obtain a feasible solution set. In particular, for a certain feasible
as fuzzy. solutions set we have a certain optimal solution set where we can
search the optimal solution z* (˛) = min z(x), x ∈ S(˛). In another way,
3. Fuzzy optimization approach for each ˛, we obtain a new classical LP problem from which we
can obtain a new optimal solution. The end result for the deci-
If there is imprecision in some of the formulation components of sion marker is an optimal solution range varying in ˛. The result
the optimization problem and it can be expressed with fuzzy terms, is consistent with the fuzzy nature of the problem.
 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163 157

Fig. 1. Hybrid metaheuristic scheme.

In the COVRPTW model, the constraints which are considered metaheuristic based on a systematic change of the neighbourhood
as fuzzy are linear and are expressed with linear membership func- structure during the search [21].
tions. Thus, we use this approach to obtain an equivalent model to The constructive aspect of ACO algorithms distinguish it from
solve the COVRPTW with fuzzy constraints which is obtained by others (such as those that estimate the distribution or particle
replacing (6), (8) and (9) with the following constraints: swarm optimization). GRASP is a fast constructive metaheuristic
to find good solutions. ACO is a constructive nature metaheuris-

n

n
tic that is perfectly suited to GRASP. Applying ACO improves the
qj xijk ≤ Q + 1k (1 − ˛), k = 1, 2, . . ., m, (15)
performance through the exploration of new solutions. The com-
i=0 j=1
bination enables a constructive search for solutions which provide
j new random solutions by GRASP and the process of historical mem-
sjk ≥ ej − 2 (1 − ˛), j = 1, . . .n, k = 1, . . .m (16)
ory introduced by ACO to converge to the global optimum. VNS is
j
sjk ≤ lj − uj + 3 (1 − ˛), j = 1, . . .n, k = 1, . . .m (17) introduced as systematic process of local search to improve the
solutions. Other bio-inspired evolutive algorithms like PSO (parti-
where 1 ∈ Rm and 2 , 3 ∈ Rn are the tolerance level vectors or the cle swarm optimization) or BCO (bee colony optimization) could be
maximum violations in the fulfilment of capacity and time win- used. However, due to the constructive nature, ACO is used to be
dows constraints provided by the decision maker, and ˛ ∈ [0, 1]. combined with the GRASP.
Applying this model we obtain an optimal solution range varying The resulting ACO hybrid approach shown in Fig. 1 works as
with ˛. follows. Following the ACO metaheuristic, the ants construct solu-
tions in parallel using the GRASP. At each iteration of the ACO, each
4. ACO hybrid metaheuristic solutions ant constructs a complete solution of the problem. Specific char-
acteristics of the procedure applied by the ants improve by using
Solutions to real route planning problems require specific heuristic information accumulated and shared by ACO. GRASP gen-
approaches that can handle both the intrinsic complexity and erates new solutions using this information. The VNS improves
the uncertainty of setting. Metaheuristics are algorithmic tech- the best solution obtained at each iteration by applying a series of
niques whose general purpose is to offer global search strategies movements (chain moves) from nested neighbourhoods. The best
to explore a solution space [19]. The main advantages of meta- solution found in the whole procedure is proposed as the heuristic
heuristics are their adaptability to problem characteristics and solution. The selection of new solutions is based on a probabilistic
their use knowledge of the same problem to find a solution. These rule taking into account of both visibility and pheromone informa-
solution procedures are capable of tackling complex optimization tion. The updating of the pheromone trails is a key element to the
problems, including those with imprecise information, evaluat- adaptive learning technique of ACO and the improvement of future
ing possible alternatives and efficiently determining the preferred solutions. The artificial ants use pheromone laid on trails as an indi-
solution. Metaheuristics have been applied to VRP problems and rect communication medium to guide the search. More pheromone
their variants with success. One of the most important charac- deposits on better routes attract more ants for later search. This
teristics of these methods is seen in their ability to adapt to the effect is called dynamic positive feedback and helps speed the con-
instance and the resultant consistency seen when applying slight vergence of ACO. ACO identifies and uses information on the best
changes in the problem model and parameters. Thus it is common solutions at each iteration.
to adapt metaheuristics to different contexts in real world prob-
lems, with relatively few changes while adapting perfectly to the 4.1. GRASP-construct procedure
characteristics found in route planning problems.
Our proposed metaheuristic uses a ACO-GRASP-VNS hybrid The GRASP solution construction mechanism builds a solution
metaheuristic to solve COVRPTW with fuzzy constraints. ACO step-by-step by adding a new node in each stage. In each iteration
is a constructive metaheuristic inspired by the trail following the added node j to the partial solution is obtained by a random
behaviour of real ant colonies. It is a decentralized distributed coop- process on the restricted candidates list (RCL). The RCL is a bounded
erative strategy, where ants act in a concurrent and independent and ordered list of the best candidates, evaluated by its contribution
way, without direct communication, and where they cooperate in to the objective function by adding it to the partial solution. In our
the exchange of information. ACO incorporates new elements to the case the RCL is constructed with the three best nodes (that make the
solution in a probabilistic way, using the accumulated pheromone solution feasible), evaluated based on a function that calculates the
trail and heuristic information of the problem [12]. GRASP is a two- times to go from one node to another, that is, they are the closest
phase metaheuristic made up of a construction phase, in which candidates in time (see Fig. 2).
a feasible solution is produced, and a post-optimization phase, in The random selection of the node to add is carried out by means
which this feasible solution is improved [41]. VNS is an effective of a probabilistic distribution that assigns to each node in the RCL
158 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163

Fig. 3. VNS improvement phase

Fig. 2. GRASP construction phase. phase. First an initial pheromone is applied to the arcs ( ij (0) = 1).
Each one of the solutions using GRASP is then constructed. Once all
the ants have build their solutions, the best solution is enhanced
an equivalent probability to the pheromone value obtained using
with VND. During the constructive process, when the ant moves
ACO. This procedure guarantees some solution diversification in the
from node i to node j the pheromone value  ij is updated, increasing
search for the best solution. One of the main disadvantages of the
its value by  ij . The procedure also reduces the intensity of the
GRASP construction phase is that it does not use information during
pheromone in an amount  throughout the iterations to avoid a
previous iterations. The use of pheromone trails as a bias in the
very fast convergence to a local optimal.
candidate selection is one of the reasons why it is combined with
The increase of the pheromone in every arc (i, j) is:
the ACO. Once the selected element has been added to the partial

solution, then the obtained improvements are updated according C/L∗ if arc (i, j)belongs to the best solution
to the proposed function. The feasibility of the solution means that ij = (18)
the customer is served within the time window and that the total 0 otherwise
demand served does not exceed vehicle capacity. When the RCL
is empty and there are no more customers to serve and the ant where C is a constant fixed by decision maker, L* is the total distance
ends the route. It then initializes another route to visit other new of the best solution found by the ants. The pheromone level at each
customers. This operation is repeated until all customers are visited. cycle t is then obtained as
The last customer visit then leads to a solution. Note that if the
ij (t) = ij (t − 1) + ij (19)
vehicle is external then the route ends at the last customer served
otherwise the route ends and returns to the depot. where  (0 <  < 1) is the amount of pheromone that disappears by
evaporation in each cycle.
4.2. VNS-improvement procedure Note as described previously in Section 4.1 that in the GRASP
constructive stage, the probability that a node (client) j is chosen
The standard VNS obtains a better solution starting from a solu- from the RCL to form part of the solution depends on the pheromone
tion in its current setting and executing a local monotone search value  ij , where i is the last element added to the solution. The
until finding a local optimum. If there is improvement it substitutes probability used for the selection is described as follows:
the current solution, otherwise it systematically modifies the ⎧
neighbourhood structures. That is, it applies an improvement ⎨  ij (t) if j ∈ RCL
movement in the current solution while possible and, otherwise, pkij (t) =  (t)
j∈RCL ij (20)
it changes neighbourhood structure. The series of neighbourhood
⎩
0 otherwise
structures Nk (x) ; k = 1, . . ., kmax is applied successively ordering
from the simplest to the most complex. The improvement phase A description of the proposed algorithm is shown in Fig. 4.
applied to the best solution constructed by GRASP consists in This approach is applied to several instances of the proposed
a variable neighbourhood descent method (VND) [22]. VND is a COVRPTW model to search for high quality solutions.
deterministic version of VNS. It applies a monotone search by sys-
tematically changing the neighbourhood structures each time a 5. Experimentation and results
local minimum is reached. VND is based on the fact that a local
minimum with respect to an neighbourhood structure is not nec- With the aim of verifying the viability of the proposed focus and
essarily a local minimum with respect to another. The final solution evaluating the behaviour of the implemented hybrid procedure we
provided by this procedure is a local minimum with respect to all of carried out an experimental study with instances obtained with
the kmax movements and consequently the probability of reaching a real data provided by a company that distributes frozen bakery
global minimum is greater than using just one structure. The neigh- products.
bourhood structures used are based on a mechanism that moves a
subset of consecutive elements of the solution and can be described
as follows. Let Nk (x) be the set of solutions obtained from x by a k-
chain move. A k-chain move consists in taking a chain or segment
of the solution with length k and moving it to another part of the
solution. Given that k is fixed, we choose two positions i and j of the
solution vector and insert the k elements which are in positions j,
j + 1, j + 2, . . ., j + k − 1 after position i (Fig. 3).

4.3. ACO procedure

The implemented ACO is an iterative procedure, where in each

stage we use k ants (k < n, where n is the number of customers), each
one constructs a solution that is obtained from the GRASP construct Fig. 4. ACO procedure.
 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163 159

Table 1
Parameter values of the instances.

Parameter Units Values Comments

Real instance
Customers n 91 Customers to serve
Available vehicles m 9 Used between 7 and 9
qi , Customer demand box ∼[6, 22]
Q, Load capacity box 250
ui , Upload time seconds [611, 2414] Delivery operations
 i , Level of tolerance seconds  2 ,  3 = 1200, Time windows
 1 = 20 20% more capacity

Generate instance
Customers n 120 Customers to serve
Available vehicles m 7 Used 7
qi , Customer demand box 10, 20, 30
Q, Load capacity box 250
ui , Upload time seconds 600, 1200, 1800 Delivery operations
 i , Level of tolerance seconds  2 ,  3 = 1200, Time windows
 1 = 20 20% more capacity

The company transports goods to retailers which will eventually Table 2

Crisp solutions for VRP and COVRP with time window constraints.
be thawed, baked, and sold to the public. The data used in the study
comes from the company database. Customer data and depot loca- GRASP GRASP-VNS AGV
tion where are given. The time window of the depot indicates the n = 91 Total demand = 1128
working hours when the distribution task should be performed. The VRP 1363 . 8(9) 893 . 1(8) 862.4 (8)
service time of the depot indicates the time needed to load a vehi- COVRP 1171 . 7(9) 793 . 5(8) 798 . 2(8)
cle with the goods to perform a route. The company receives each n = 120 Total demand = 2510
customer demand by midday the day before and then identifies the VRP 1344 . 0(10) 1308 . 5(11) 1268 . 2(11)
routes for the next day. Customer number and demand to be served COVRP 1026 . 2(11) 976 . 0(11) 953 . 6(12)
varies each day. Market estimates can indicate constant moder-
ate growth from customers and general demand, in the next years, We apply to the model proposed in Section 2 the optimiza-
both requiring an expanded fleet. Nevertheless, the company can- tion approach described in Section 3. The solutions were obtained
not know with precision the equivalent units of the load capacity, applying GRASP, GRASP-VNS and ACO hybrid metaheuristics
or client demand but instead must rely only an approximation. Pre- described in Section 4.
cise travel times to go from one node to another are also unknown, As described in Section 3 the fuzzy constraints are checked
because problems such as traffic congestion, road conditions and by the corresponding crisp constraints (15), (16) and (17) of the
stopping times make these times imprecise. Although the company transformed formulation. We get heuristic solutions for the values
has nine vehicles available in its fleet, it only uses seven vehicles ˛ = 0.0, 0.2, 0.4, 0.6, 0.8. A crisp problem is generated for each value
each day when planning distribution routes. The remaining part of of ˛ and each one has a different value of the optimal solution,
the excess load is distributed using externally contracted vehicles which is in line with the fuzzy nature of the problem.
and currently is not considered in the planning. Each solution of the optimization model consists in a set of
The company is interested in stabilizing its fleet and human routes to be used by the vehicles. As described in Section 2 the set
resources to seven vehicles and also in incorporating in the of variables of the optimization problem consists of those variables
planning all their customers and potential demand, including that determine this set of routes. This set consists in two sets of
distribution by external fleets. This approach guarantees greater decision variables xijk and sik , i, j = 0, 1, . . ., n, k = 1, 2, . . ., m. For each
service control and quality, improved customer assignment and
two indices i and j and each vehicle k, xijk = 1 if and only if vehicle k
subcontracted demand, which in turn generates a reduction in
costs. travels from i to j, and xijk = 0 otherwise. The decision variable sik is
Our experimental study used a real instance with 91 customers, defined for each customer i and each vehicle k to denote the time
corresponding with the data from a specific day (approximate daily vehicle k starts to service customer i. The model variables deter-
average of customers served is 80) and a generated instance with mine the order of the different customer nodes that are visited
120 customers, which simulates a growth scenario for the num- during the routes, so the solutions of the problem represent the
ber of customers and their demand. These instances are obtained visited nodes and the order that they are visited.
based on available data of approximately 250 customers provided Following the proposed procedure in Section 4, a solution for
by the company. In both cases the customers and depot locations the problem is obtained by GRASP, checking the model constraints
correspond to real location data provided by the company. in a constructive way to reach a feasible solution. This procedure
The distance matrix is real and travel times are provided by the ensures compliance with all constraints except capacity (15) and
company based on its expertise. Demand amounts, unload times time windows (16), (17), which must be evaluated. These two con-
and time windows in the first case study were real while the second straints are checked every time a new customer is inserted into the
one was generated randomly, taking into account the current cases route, as described in Section 4.1. The cost function is the total dis-
described by the company. Vehicle capacities and demand amounts tance travelled by the vehicles following the selected routes that is
are expressed in number of standard boxes used by the company to computed by 1. The number of iterations of the ACO procedure in
handle goods. Specific tolerance levels are fixed in the constraints the experimentation is 100.
as well as the time windows and load capacity. The values for the parameters in the metaheuristics are:
Table 1 shows the data of the instances used.
We evaluate the performance of the hybrid algorithm that com- • The number of nodes in the RCL is 3.
bines ACO, GRASP and VNS by considering these two instances. • The kmax value in the VNS is 4.
160 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163

Table 3
VRP with fuzzy time window constraints.

˛= 0.0 0.2 0.4 0.6 0.8

n = 91 Total demand = 1128

GRASP 1337 . 7(8) 1360 . 7(9) 1269 . 5(9) 1344 . 1(9) 1456 . 2(9)
GRASP-VNS 903 . 6(8) 946 . 3(8) 840 . 5(8) 899 . 0(8) 930 . 3(8)
AGV 866 . 4(8) 905 . 1(8) 874 . 0(8) 887 . 7(8) 889 . 8(8)

n = 120 Total demand = 2510

GRASP 1351 . 6(11) 1424 . 5(11) 1416 . 0(11) 1374 . 4(11) 1404 . 1(11)
GRASP-VNS 1324 . 1(11) 1397 . 6(11) 1372 . 4(11) 1242 . 9(11) 1280 . 9(11)
AGV 1339 . 0(11) 1291 . 7(11) 1300 . 3(11) 1280 . 3(11) 1303 . 1(11)

Table 4
Close–open VRP with fuzzy time window constraints.

˛= 0.0 0.2 0.4 0.6 0.8

n = 91 Total demand = 1128

GRASP 1158 . 9(9) 1195 . 1(9) 1104 . 6(9) 1165 . 6(9) 1266 . 7(10)
GRASP-VNS 774 . 5(9) 810 . 3(8) 774 . 4(9) 800 . 3(8) 735 . 4(8)
AGV 776 . 1(8) 808 . 1(9) 720 . 9(8) 797 . 5(8) 752 . 3(8)

n = 120 Total demand = 2510

GRASP 1039 . 7(11) 1044 . 1(12) 1056 . 2(11) 1045 . 5(12) 1060 . 7(11)
GRASP-VNS 974 . 4(12) 1044 . 1(12) 980 . 7(11) 943 . 0(12) 1006 . 7(11)
AGV 1030 . 3(11) 953 . 5(11) 1014 . 8(11) 984 . 0(11) 977 . 4(11)

• The number of ants is equal to the number of nodes (customers). parameters must allow the algorithm to end in reasonable com-
• The constant for the pheromone updating rule is C = 100. putational time.
• The pheromone evaporation parameter is  = 0.5. We compared the solutions obtained by our ACO-GRASP-VNS
hybrid metaheuristic (AGV) with those obtained only with GRASP,
and by the GRASP-VNS hybrid that combines GRASP and VNS. The
We used parameters found in the literature which provided proposed algorithms were coded in C++ and executed on IntelCore2
optimal results for a similar problem. The size of the RCL (restricted Duo processor at 2.66 GHz and 4 Gb of RAM.
candidate list) must be large enough to ensure appropriated diver- Tables show the values of the objective function and the number
sification in the constructed solutions. The value kmax for the of routes (vehicles) for the crisp solutions (Table 2) and for the fuzzy
number of neighbourhoods used in the VNS has to be sufficient solutions for ˛ ∈ {0.0, 0.2, 0.4, 0.6, 0.8} (Tables 3–6). Tables 3 and 5
to allow the improvement of the solutions provided by GRASP. The show the results, respectively, for the VRPTW and COVRPTW when
values for the evaporation process mean a balance between the only time windows constraints are fuzzy. Tables 4 and 6 show the
convergence times of the algorithm and the ability to obtain the results, respectively, for the VRPTW and COVRPTW when both, time
global optimum. The smaller evaporation factor ensures sufficient windows constraints and capacity constraints, are fuzzy.
diversity of the search space and guides following ants to explore Figs. 5 and 6 are used to illustrate the effect of the hybridized
better solutions. A more rapid pheromone evaporation allows a developed optimizing algorithm. These figure show the values of
search to be trapped in local minima more easily. All of these the cost function of the COVRPTW model obtained by applying

Table 5
VRP with fuzzy time window and fuzzy capacity constraints.

˛= 0.0 0.2 0.4 0.6 0.8

n = 91 Total demand = 1128

GRASP 1337 . 7(8) 1360 . 7(9) 1269 . 5(9) 1344 . 1(9) 1456 . 2(9)
GRASP-VNS 903 . 6(8) 946 . 3(8) 840 . 5(8) 899 . 0(8) 930 . 3(8)
AGV 866 . 4(8) 905 . 1(8) 874 . 0(8) 887 . 7(8) 889 . 8(8)

n = 120 Total demand = 2510

GRASP 1325 . 6(10) 1344 . 0(10) 1343 . 1(10) 1287 . 9(10) 1353 . 7(11)
GRASP-VNS 1230 . 0(10) 1224 . 3(10) 1227 . 3(10) 1242 . 3(10) 1278 . 6(11)
AGV 1286 . 7(10) 1203 . 2(10) 1228 . 2(10) 1227 . 9(10) 1306 . 5(11)

Table 6
Close–open VRP with fuzzy time window and fuzzy capacity constraints.

˛= 0.0 0.2 0.4 0.6 0.8

n = 91 Total demand = 1128

GRASP 1158 . 9(9) 1195 . 1(9) 1104 . 6(9) 1165 . 6(9) 1266 . 7(10)
GRASP-VNS 774 . 5(9) 810 . 3(8) 774 . 4(9) 800 . 3(8) 735 . 4(8)
AGV 776 . 1(8) 808 . 1(9) 720 . 9(8) 797 . 5(8) 752 . 3(8)

n = 120 Total demand = 2510

GRASP 1090 . 9(10) 1062 . 4(11) 1040 . 7(11) 1046 . 5(11) 1072 . 8(11)
GRASP-VNS 1068 . 5(11) 1010 . 1(10) 1023 . 0(11) 1026 . 5(10) 1016 . 5(11)
AGV 927 . 5(10) 979 . 6(10) 979 . 6(10) 983 . 6(10) 1005 . 2(12)
 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163 161

Fig. 5. Metaheuristic solutions with 91 customers.

Fig. 6. Metaheuristic solutions with 120 customers.

metaheuristics. Fig. 5 presents the best values obtained for the case feasible solutions associated to the increase in load capacity. On
of 11 vehicles and imprecise constraints (capacity and time win- the other hand, load capacity constraints are not determinant, and
dows) for the instance of 91 customers. In Fig. 6 the values with 10 consequently vehicle capacity does not act as a strong constraint.
and 11 vehicles for instance with 120 customers is shown. CPU times for solutions obtained using the different meta-
Firstly we state that our procedures are able to find good solu- heuristics were given in minutes. We tested the use of CPLEX to
tions from a practical point of view. Solutions are found that solve these real instances. However CPLEX required excessive com-
incorporate imprecision in travel times and load units, providing putation time and produced memory overflows with moderate
flexibility in the affected constraints. Secondly, a solution is offered size instances. Consequently execution time was stopped without
in which included subcontracted route planning. Both offer a wide obtaining any solution. Our metaheuristic got high quality solutions
set of flexible and operative solutions which facilitate decision within a reasonable computation time for this problem, one which
making with a follow up evaluation of the best alternative from the company must solve every day.
an economic perspective. The company is interested in assigning
routes which imply improved profitability of its own fleet and lower 6. Conclusions
subcontracting costs.
Regarding implemented procedures, the data in the table reveals Tackling real world vehicle routing problems requires consid-
that the hybrid metaheuristics generate the highest quality solu- eration of specific parameters and constraints. We consider the
tions. In particular, the ACO hybrid metaheuristic proposed in this ability to plan routes with an owned fleet and outsourced vehicles
work improves the results in most cases, especially when the com- and to use uncertainty of available knowledge. Both scenarios offer
plexity is increased because of the problem size. solutions that are required from real cases found in distribution
Note that in the case of adding flexibility in load capacity with companies.
the initial real problem of 91 customers the solutions obtained are We proposed a new variant of VRPTW with close and open
the same as if they were not flexible. This occurs because the time routes. In addition we consider imprecise capacities and time win-
window constraints are the hardest and cannot provide better dows constraints, which are modelled as fuzzy constraints.
162 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163

Our approach applies fuzzy optimization to address these [19] M. Gendreau, J.-Y. Potvin, Handbook of Metaheuristics, Springer, 2010.
problems modelled as linear programming problems. Hybrid meta- [20] S. Häckel, P. Dippold, The bee colony-inspired algorithm (BCiA): a two-stage
approach for solving the vehicle routing problem with time windows, in:
heuristics are also implemented to solve them. In particular, we use Proceedings of the 11th Annual conference on Genetic and evolutionary com-
a ACO hybrid metaheuristic procedure based on GRASP and VNS to putation, ACM, 2009, pp. 25–32.
search for optimal solutions of the problems. The experiments per- [21] P. Hansen, N. Mladenovic, J. Brimberg, J.A. Moreno, Variable neighbourhood
search, in: M. Gendreau, J.-Y. Potvin (Eds.), Handbook in Metaheuristics,
formed show that our hybrid procedure is effective to find a good Springer, 2010.
enough solution in short running time. These experiments also [22] P. Hansen, N. Mladenovic, J.A. Moreno, Variable neighbourhood search: meth-
show that the ingredients introduced in the hybridization improve ods and applications, Ann. Oper. Res. 175 (1) (2010) 367–407.
[23] F. Huang, C. Liu, An improved tabu search for open vehicle routing problem, in:
performance.
Proceeding 2010 International Conference on Management and Service Science
Some aspects can be improved. For instance, the parameter (MASS), 2010, pp. 1–4.
adjustment of the metaheuristics used can be more efficient, [24] M. Inuiguchi, J. Ramík, Possibilistic linear programming: a brief review of fuzzy
mathematical programming and a comparison with stochastic programming
studying other objective functions that consider economic or envi-
portfolio selection problem, Fuzzy Sets Syst. 111 (2000) 3–28.
ronmental aspects and the balanced load, times and distances in [25] M. Ko, A. Tiwari, J. Mehnen, A review of soft computing applications in supply
the routes. chain management, Appl. Soft Comput. 10 (3) (2010) 661–674.
Finally we feel that the implementation of tools in this study can [26] M. Köppen, K. Yoshida, K. Ohnishi, M. Tsuru, Meta-heuristic approach to pro-
portional fairness, Evol. Intell. 5 (4) (2012) 231–244.
be used to assist in the development of intelligent systems that are [27] G. Li, Research on open vehicle routing problem with time windows based on
able to tackle the uncertainty, variability and complexity of logistics improved genetic algorithm, in: Proceeding 2009 International Conference on
distribution problems, and which can improve solutions provided Computational Intelligence and Software Engineering, 2009, pp. 1–5.
[28] G. Li, An improved ant colony algorithm for open vehicle routing problem with
by route distribution optimization applications used at present. time windows, in: Proceeding 2009 International Conference on Information
Management, Innovation Management and Industrial Engineering, vol. 2, 2009,
Acknowledgements pp. 616–619.
[29] X. Li, P. Tian, An ant colony system for the open vehicle routing problem, in: M.
Dorigo (Ed.), Lecture Notes in Computer Science (ANTS 2006), Springer, Berlin,
This paper has been partially supported by the following 2006, pp. 356–363, 4150.
projects: TIN2012-32608 and TIN2011-27696-C02-01 from the [30] X. Li, P. Tian, S. Leung, An ant colony optimization metaheuristic hybridized
with tabu search for the open vehicle routing problem, J. Oper. Res. Soc. 60
Government of Spain. (2009) 1012–1025.
[31] R. Liu, Z. Jiang, H. Hu, S. Yao, A memetic algorithm for the close–open mixed
Appendix A. Supplementary data vehicle routing problem, in: Proceedings of the 2010 IEEE IEEM, 2010, pp.
728–732.
[32] Y. Marinakis, M. Marinaki, G. Dounias, A hybrid particle swarm optimization
Supplementary data associated with this article can be found, in algorithm for the vehicle routing problem, Eng. Appl. Artif. Intell. 23 (4) (2010)
the online version, at http://dx.doi.org/10.1016/j.asoc.2015.03.026. 463–472.
[33] Y. Marinakis, M. Marinaki, A hybrid genetic-particle swarm optimization
algorithm for the vehicle routing problem, Expert Syst. Appl. 37 (2) (2010)
References 1446–1455.
[34] M. Mavrovouniotis, S. Yang, Ant colony optimization with immigrants schemes
[1] T.-J. Ai, K. Voratas, A particle swarm optimisation for vehicle routing problem for the dynamic travelling salesman problem with traffic factors, Appl. Soft
with time windows, Int. J. Oper. Res. 24 (2009) 519–537. Comput. 13 (10) (2013) 4023–4037.
[2] M.Z. Ali, A. Khalid, T. Yahya, Cultural algorithms: emerging social structures [35] S. MirHassani, N. Abolghasemi, A particle swarm optimization algo-
for the solution of complex optimization problems, Int. J. Artif. Intell. 11 (A13) rithm for open vehicle routing problem, Expert Syst. Appl. 38 (2011)
(2013) 20–42. 11547–11551.
[3] D. Aksen, Z. Ozyurt, N. Aras, Open vehicle routing problem with driver nodes [36] L. Pan, Z. Fu, A clonal selection algorithm for open vehicle routing problem, in:
and time deadlines, J. Oper. Res. Soc. 58 (2007) 1223–1234. Proceeding 2009 Third International Conference on Genetic and Evolutionary
[4] R. Bellman, L. Zadeh, Decision-making in fuzzy environment, Manag. Sci. 17 Computing, 2009, pp. 786–790.
(1970) 141–164. [37] D. Pisinger, S. Ropke, A general heuristic for vehicle routing problems, Comput.
[5] L. Bodin, B. Golden, A. Assad, M. Ball, Routing and scheduling of vehicles and Oper. Res. 34 (8) (2007) 2403–2435.
crews: the state of the art, Comput. Oper. Res. 10 (1983) 63–211. [38] R. Precup, R. David, E.M. Petriu, S. Preitl, M. Radac, Novel adaptive gravitational
[6] J. Brandao, A tabu search heuristic algorithm for open vehicle routing problem, search algorithm for fuzzy controlled servo systems, IEEE Trans. Ind. Inform. 8
Eur. J. Oper. Res. 157 (2004) 552–564. (4) (2012) 791–800.
[7] G.B. Dantzig, R.H. Ramser, The truck dispatching problem, Manag. Sci. 6 (1959) [39] P.P. Repoussis, C.D. Tarantilis, G. Ioannou, The open vehicle routing problem
80–91. with time windows, J. Oper. Res. 58 (2007) 355–367.
[8] M. Delgado, J.L. Verdegay, M.A. Vila, A general model for fuzzy linear program- [40] P.P. Repoussis, C.D. Tarantilis, O. Braysy, G. Ioannou, A hybrid evolution strat-
ming, Fuzzy Sets Syst. 29 (1989) 21–29. egy for the open vehicle routing problem, Comput. Oper. Res. 37 (2010)
[9] U. Derigs, K. Reuter, A simple and efficient tabu search heuristic for solving the 443–455.
open vehicle routing problem, J. Oper. Res. Soc. 60 (2009) 1658–1669. [41] M.G.C. Resende, C.C. Ribeiro, Greedy randomized adaptive search procedures:
[10] Q. Ding, X. Hu, L. Sun, Y. Wang, An improved ant colony optimization and its advances, hybridizations, and applications, in: M. Gendreau, J.-Y. Potvin (Eds.),
application to vehicle routing problem with time windows, Neurocomputing Handbook in Metaheuristics, Springer, 2010.
98 (2012) 101–107. [42] R. Russell, W. Chiang, D. Zepeda, Integrating multi-product production
[11] M. Dorigo, L.M. Gambardella, Ant colony system: a cooperative learning and distribution in newspaper logistics, Comput. Oper. Res. 35 (2008)
approach to the travelling salesman problem, IEEE Trans. Evol. Comput. 1 1576–1588.
(1997) 53–66. [43] M. Salari, P. Toth, A. Tramontani, An ILP improvement procedure for the open
[12] M. Dorigo, T. Stützle, Ant colony optimization: overview and recent advances, vehicle routing problem, Comput. Oper. Res. 37 (2010) 2106–2120.
in: M. Gendreau, J.-Y. Potvin (Eds.), Handbook of Metaheuristics, Springer, 2010. [44] L. Schrage, Formulation and structure of more complex/realistic routing and
[13] C. Erbao, L. Mingyong, The open vehicle routing problem with fuzzy demands, scheduling problems, Networks 11 (1981) 229–232.
Expert Syst. Appl. 37 (2010) 2405–2411. [45] X. Tan, X. Zhuo, J. Zhang, Ant colony system for optimizing vehicle routing
[14] K. Fleszar, I.H. Osman, K.S. Hindi, A variable neighbourhood search algo- problem with time windows (VRPTW). Computational intelligence and bioin-
rithm for the open vehicle routing problem, Eur. J. Oper. Res. 195 (3) (2009) formatics, Lect. Notes Comput. Sci. 4115 (2006) 33–38.
803–809. [46] C.D. Tarantilis, C.T. Kiranoudis, Distribution of fresh meat, J. Food Eng. 51 (2002)
[15] Z. Fu, R. Eglese, L. Li, A new tabu search heuristic for the open vehicle routing 85–91.
problem, J. Oper. Res. Soc. 56 (2005) 267–274. [47] C.D. Tarantilis, G. Ioannou, C.T. Kiranoudis, G.P. Prastacos, A threshold accepting
[16] Z. Fu, R. Eglese, L. Li, Corrigendum to the paper: A new tabu search heuristic for approach to the open vehicle routing problem, RAIRO Oper. Res. 38 (2004)
the open vehicle routing problem, J. Oper. Res. Soc. 57 (2006) 1017–1018. 345–360.
[17] L.M. Gambardella, E. Taillard, G. Agazzi, MACS-VRPTW. A multiple colony [48] C.D. Tarantilis, G. Ioannou, C.T. Kiranoudis, G.P. Prastacos, Solving the open
system for vehicle routing problems with time windows, in: New Ideas in vehicle routing problem via a single parameter meta-heuristic algorithm, J.
Optimization, McGraw-Hill, 1999, pp. 63–76. Oper. Res. 56 (2005) 588–596.
[18] H. Ge, T. Zhen, Y. Jiang, Y. Che, An intelligent solution for open vehicle routing [49] D. Teodorović, Swarm intelligence systems for transportation engineering:
problem in grain logistics, in: Q. Luo (Ed.), in: Advances in Wireless Networks principles and applications, Transp. Res. C 1 (6) (2008) 651–667.
and Information Systems, LNEE 72, Springer-Verlag, Berlin/Heidelberg, 2010, [50] P. Toth, D. Vigo, The vehicle routing problem, Monogr. Discret. Math. Appl. 9
pp. 389–396. (2002).
 J. Brito et al. / Applied Soft Computing 32 (2015) 154–163 163

[51] J.L. Verdegay, Fuzzy mathematical programming, in: M.M. Gupta, E. Sanchez [57] E. Zachariadis, T. Kiranoudis, An open vehicle routing problem metaheuris-
(Eds.), Fuzzy Information and Decision Processes, North-Holland, Amsterdam, tic for examining wide solution neighborhoods, Comput. Oper. Res. 37 (2010)
1982, pp. 231–237. 712–723.
[52] J.L. Verdegay, Fuzzy optimization: models, methods and perspectives, in: [58] Q. Zhang, T. Zhen, Y. Zhu, W. Zhang, Z. Mat, A hybrid intelligent algorithm for the
Proceeding 6th IFSA-95 World Congress, 1995, pp. 39–71. vehicle routing with time windows, advanced intelligent computing theories
[53] J.L. Verdegay, R.R. Yager, P.P. Bonissone, On heuristics as a fundamental con- and applications, with aspects of theoretical and methodological issues, Lect.
stituent of soft computing, Fuzzy Sets Syst. 159 (2008) 846–855. Notes Comput. Sci. 5226 (2008) 47–54.
[54] W. Wang, B. Wu, Y. Zhao, D. Feng, Particle swarm optimization for open vehi- [59] T. Zhen, Y. Zhu, Q. Zhang, A particle swarm optimization algorithm for the
cle routing problem, in: D.-S. Huang, K. Li, G.W. Irwin (Eds.), Proceedings of open vehicle routing problem, in: Proceeding 2009 International Conference
the 2006 International Conference on Intelligent Computing: Part II (ICIC’06), on Environmental Science and Information Application Technology, IEEE, 2009,
Springer-Verlag, Berlin, Heidelberg, 2006, pp. 999–1007. pp. 560–563.
[55] S. Yu, C. Ding, K. Zhu, A hybrid GA-TS algorithm for open routing optimization [60] Q. Zhu, L. Qian, Y. Li, S. Zhu, An improved particle swarm optimization algorithm
of coal mines material, Expert Syst. Appl. 38 (2011) 10568–10573. for vehicle routing problem with time windows, in: Proceedings of the 2006
[56] B. Yu, Z.Z. Yang, B.Z. Yao, A hybrid algorithm for vehicle routing problem with IEEE Congress on Evolutionary Computation, Vancouver, BC, Canada, 2016, pp.
time windows, Expert Syst. Appl. 38 (1) (2011) 435–441. 1386–1390.