Professional Documents
Culture Documents
Heuristic methods for the fleet size and mix vehicle routing problem
with time windows and split deliveries
Patrícia Belfiore a,⇑, Hugo T.Y. Yoshizaki b
a
Department of Management Engineering, Federal University of ABC, Rua Teixeira da Silva, 426, ap. 143, São Paulo, SP, Brazil
b
Department of Production Engineering, University of São Paulo (USP), Av. Prof. Luciano Gualberto, 908, Sala FG-207, 2° andar, Cidade Universitária, São Paulo, SP, Brazil
a r t i c l e i n f o a b s t r a c t
Article history: This paper proposes a scatter-search (SS) approach to solve the Fleet Size and Mixed Vehicle Routing
Received 16 April 2012 Problem with Time Windows and Split Deliveries (FSMVRPTWSD). In the Vehicle Routing Problem with
Received in revised form 8 October 2012 Split Deliveries (VRPSD), each customer can be served by more than one vehicle, as opposed to the clas-
Accepted 14 November 2012
sical VRP in which each customer is served only once. In the FSMVRPTW, the customers must be serviced
Available online 8 December 2012
within their time windows with minimal costs using a heterogeneous fleet. Experimental testing and
benchmark examples are used to assess the merit of our proposed procedure. The results show that
Keywords:
the proposed heuristics are competitive with the best results found in the literature.
Heuristics
Scatter search
Ó 2012 Elsevier Ltd. All rights reserved.
Vehicle routing and scheduling
1. Introduction variable costs that depend on vehicle type, a component that is ne-
glected in the first version. The third version, referred to as the het-
The classical vehicle routing problem (VRP) aims to find a set of erogeneous fixed fleet vehicle routing problem (HFFVRP),
routes with minimal cost (by finding the shortest path, minimizing generalizes the second version by limiting the number of available
the number of vehicles, etc.) and beginning and ending the route at vehicles of each type.
the depot such that the known demand of all nodes is fulfilled. Lenstra and Rinnoy Kan (1981) have analyzed the complexity of
Each node is visited only once by only one vehicle, and each vehicle the vehicle routing problem and have concluded that practically all
has a limited capacity. Selected formulations also include con- vehicle routing problems are NP-hard (including the classical vehi-
straints on the maximum travel time. cle routing problem) because they are not solved in polynomial
The VRPSD is a variation of the classical VRP in which each cus- time. Because the FSMVRPTWSD is a combination of the FSMVRP,
tomer can be served by more than one vehicle. Thus, for the VRPSD, VRPTW and VRPSD, it remains NP-hard (Archetti, Savelsbergh, &
in addition to the delivery routes, the quantity to be delivered to Speranza, 2006; Dror & Trudeau, 1990; Dullaert, Janssens, Soren-
each customer by each vehicle must also be determined. The op- sen, & Vernimmen, 2002; Gendreau, Laporte, Musaraganyi, & Tail-
tion of splitting a given demand makes it possible to service a cus- lard, 1999; Solomon & Desrosiers, 1988; Taillard, 1999). Therefore,
tomer whose demand exceeds the vehicle capacity. Splitting may this observation makes a strong case for the application of heuris-
also result in decreased costs. The vehicle routing problem with tics and meta-heuristics to solve the problem.
time windows and split deliveries (VRPTWSD) presents an exten- This paper develops a scatter-search (SS) algorithm to solve the
sion of the VRPSD in which the time window restraints are added. fleet size and mixed vehicle routing problem with time windows
In the literature, three variants of a VRP with a heterogeneous and split deliveries (FSMVRPTWSD). The algorithm proposed was
fleet have been studied. The first was introduced by Golden, Assad, adapted to solve three varieties of benchmark examples available
Levy, and Gheysens (1984) in which the variable costs were uni- in the literature: benchmark problems by Liu and Shen (1999)
formly spread over all vehicle types with the number of available for the FSMVRP, benchmark problems by Ho and Haugland
vehicles assumed as unlimited for each type. This version is ad- (2004) for the VRPSD, and benchmark problems by Solomon
dressed in this paper and is known as the fleet size and mixed vehi- (1987) for the VRPTW.
cle routing problem (FSMVRP), the vehicle fleet mix (VFM) or the The organization of this article is as follows. Section 2 reviews
fleet size and composition VRP. The second version considers the the literature on VRPSD and its extensions as well as FSMVRP
and its extensions. Section 3 presents the problem definition,
including the mathematical formulation, and Section 4 covers the
⇑ Corresponding author. scatter-search overview. Section 5 describes the constructive heu-
E-mail addresses: pbfavero@usp.br (P. Belfiore), hugo@usp.br (H.T.Y. Yoshizaki). ristics and the scatter-search approach proposed for solving the
0360-8352/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cie.2012.11.007
590 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
model. Section 6 presents the computational results, and conclu- techniques for generating a lower bound and an underestimate
sions are stated in the Section 7. of the optimal solutions. A new savings heuristic for the FSMVRP
based on successive route fusion was presented by Desrochers
and Verhoog (1991). In each iteration, the best fusion is selected
2. Literature review
by solving a weighted matching problem.
A mathematical programming formulation for the FSMVRP was
2.1. Literature review for the VRPSD and its extensions
presented by Gheysens, Golden, and Assad (1984). Gheysens, Gold-
en, and Assad (1986) developed a two-stage general assignment-
The vehicle routing problem with split deliveries (VRPSD) was
based heuristic that uses Golden et al.’s lower bound procedure
introduced in the literature by Dror and Trudeau (1989, 1990),
to determine the fleet composition for use in the second phase of
who presented the mathematical formulation of the problem and
a generalized assignment heuristic.
analyzed the savings that result when a customer is served by
A seven-phase improvement heuristic for the fleet size and
more than one vehicle; this situation creates economy with respect
mixed vehicle routing problem was developed by Salhi and Rand
to both to the number of vehicles and the total distance traveled.
(1993). The aims were to match the total demand of a given route
Dror, Laporte, and Trudeau (1994) have presented an integer
to an appropriate vehicle capacity, to eliminate an entire route by
programming formulation of the VRPSD, developed several fami-
dropping its customers and inserting them into other routes, to
lies of valid inequalities, and established a hierarchy between these
reallocate customers from a given route to other routes such that
types. A constraint relaxation branch-and-bound algorithm for the
a route could be served by a smaller vehicle (if profitable) and to
problem was also described.
either combine routes with a small total demand into fewer routes
Frizzell and Giffin (1992, 1995) have developed construction
with a larger total demand or split larger routes into smaller ones.
and improvement heuristics for the VRPSD with grid network dis-
Rochat and Semet (1994) proposed a tabu search approach for
tances. In their second publication, they also considered the time
delivering pet food and flour in Switzerland. The company serves
window constraints.
its customers out of one depot using a heterogeneous fixed fleet,
Mullaseril, Dror, and Leung (1997) have described a feed distri-
and the customers are characterized by time windows and accessi-
bution problem encountered on a cattle ranch in Arizona. The
bility constraints. The last constraint refers to selected customers
problem is cast as a collection of capacitated rural postman prob-
that cannot be served by the same vehicle.
lems with time windows and split deliveries. They presented an
A tabu search procedure for the multi-trip vehicle routing and
adaptation of the heuristics proposed by Dror and Trudeau (1990).
scheduling problem (MTVRSP) was proposed by Brandão and Mer-
Belenguer, Martinez, and Mota (2000) have proposed a lower
cer (1997) in which each vehicle can make several trips per day.
bound for the VRPSD based on a polyhedral study of the problem
Gendreau, Laporte, Musaraganyi, and Taillard (1999) have pro-
that includes new valid inequalities. The authors developed a cut-
posed a tabu search algorithm for the FSMVRP. The initial solution
ting-plane algorithm to solve small instances, and for larger in-
is obtained through the GENI (Generalized Insertion) algorithm. A
stances, the integer values are obtained via a branch-and-bound
post-optimization phase known as US (Unstringing and Stringing)
approaches.
is implemented after GENI. The tabu search is based on an adaptive
Archetti, Savelsbergh, et al. (2006) have performed a worst-case
memory procedure (AMP) developed by Rochat and Taillard
performance analysis for the vehicle routing problem with split
(1995), and new variants of the tabu search for the FSMVRP are
deliveries. The authors have shown that the cost savings realized
developed by Wassan and Osman (2002). These variants use a
by allowing split deliveries is at most 50%. They also studied the
mix of different components, including reactive tabu search con-
variant of the VRPSD in which the customer demand may be larger
cepts, variable neighborhoods, special data memory structures
than the vehicle capacity but each customer must be visited a min-
and hashing functions.
imum number of times. Archetti, Speranza, and Hertz (2006) have
Several insertion-based savings heuristics for the fleet size and
described a tabu search algorithm for the VRPSD. With each itera-
mixed vehicle routing problem with time window constraints
tion, a neighboring solution is obtained by removing a customer
(FSMVRPTW) were described by Liu and Shen (1999). Their method
from a set of routes in which it is currently visited and subse-
is a generalization of the traditional insertion viewpoint instead of
quently inserting it either into a new route or into an existing route
the combination concept used in the savings-based heuristics of
with sufficient residual capacity. The algorithm also considers the
Clarke and Wright (1964). A certain number of candidate fleet com-
possibility of inserting a customer into a new route without
positions are recorded in the construction phase, followed by appli-
removing it from the old route.
cation of a composite improvement scheme to enhance the solution
Ho and Haugland (2004) have developed a tabu search algo-
quality. The heuristics were tested on 168 sample problems. Dulla-
rithm for the VRPTWSD. The first stage constructs a VRP solution
ert et al. (2002) proposed three insertion-based heuristics for the
using node interchanges, and the second stage improves the VRP
FSMVRPTW. The heuristics are an extension of Solomon’s (1987)
solution by introducing and eliminating splits. A pool of solutions
sequential insertion heuristic I1 with adapted formulations of Gold-
is defined by different move operators, and the best solution in
en et al’s savings concepts. The heuristics were tested on the same
the current pool is always chosen.
problem set as that in Liu and Shen’s work.
Belfiore and Yoshizaki (2009) have proposed a scatter-search
Taillard (1999) has presented a heuristic column generation
algorithm to solve a real-life heterogeneous fleet vehicle routing
method for solving the heterogeneous fixed fleet vehicle routing
problem with time windows and split deliveries that occurred in
problem (HFFVRP). The method was also used to solve the fleet size
a major Brazilian retail group. The results show that the total dis-
and mix vehicle routing problem. A new meta-heuristic known as
tribution cost can be reduced significantly when such methods are
BATA (back-tracking adaptive threshold accepting) was developed
used.
by Tarantilis, Kiranoudis, and Vassiliadis (2004) to solve the
HFFVRP. Other methods were also proposed for the HFFVRP,
2.2. Literature review for the VRP with a heterogeneous fleet and its including a tabu search (Brandão, 2011), a multistart adaptive
extensions memory programming (MAMP) approach and a path re-linking ap-
proach (Li, Tian, & Aneja, 2010).
Golden et al. (1984) developed a savings heuristic to solve the In this paper, we have developed a new scatter-search algo-
fleet size and mixed vehicle routing problem in addition to rithm to solve the Fleet Size and Mixed Vehicle Routing Problem
P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601 591
with Time Windows and Split Deliveries (FSMVRPTWSD). The re- presented below based on the work of Dror and Trudeau (1990),
sults are compared with three varieties of benchmark examples Ho and Haugland (2004) and Golden et al. (1984).
available in the literature: Solomon’s instances (VRPTW), Ho and The objective function can be written as follows:
Haugland’s modified problems (VRPTWSD), and those proposed
X
K X
n X
n X
n X
K
by Liu and Shen (FSMVRPTW). min fk xk0j þ ckij xkij
k¼1 j¼1 i¼0 j¼0 k¼1
In this section, we define the problem under study and the nota- X
n
3.1. Customers
X
n X
n
xkip xkpj ¼ 0 p ¼ 0; . . . ; n; k ¼ 1; . . . ; K ð2Þ
The problem is described by a set of customers C = {1, 2, . . . , n} i¼0 j¼0
residing at n different locations, where node 0 denotes the central
depot, and thus N ¼ C [ f0g becomes the set of all locations. Every X
K
pair of locations (i, j), where i; j 2 N and i–j; is associated with a yki ¼ 1; i ¼ 1; . . . ; n ð3Þ
travel time tij and a distance traveled dij that are symmetrical (tij = - k¼1
plied to hard optimization problems. Glover (1977, 1998) proposed applied to transform a trial solution into one or more enhanced
the first description of this method and a scatter-search template. trial solutions. If no improvement occurs for a given trial solution,
The SS operates on a set of reference solutions to generate new the enhanced solution is considered to be the same as the one sub-
solutions via weighted linear combinations of structured subsets mitted for improvement. Step 3 chooses b solutions from P accord-
of solutions. This approach uses strategies for search diversification ing to their quality or diversity to build the reference set (RefSet),
and intensification that have proven effective in a variety of opti- which is a collection of both high quality solutions and diverse
mization problems. solutions. In step 4, the subset generation method produces a
The following parameters are used in the method discussion: subset of the reference set solutions as a basis for creating com-
bined solutions. The simplest form of the subset generation meth-
PSize = size of the set of diverse solutions generated by the od consists of generating all pairs of reference solutions. In step 5,
Diversification Generation Method. the solution combination method is applied to generate one or
b = size of the reference set (RefSet). more trial solutions. For each combined solution, we apply the
b1 = size of the high-quality subset of RefSet. improvement method (step 6). At this point, the reference set
b2 = size of the diverse subset of RefSet. can be updated by the reference set update method (step 7)
MaxIter = maximum number of iterations. depending on the quality or diversity of the new combined solu-
tion. In this work, dynamic and static RefSet updating is imple-
Fig. 1 shows a sketch of the scatter-search method based on the mented as described in Section 5.3. If the search converges, i.e.,
work of Alegre, Laguna, and Pacheco (2007), Yamashita, Arment- no new solutions are found for inclusion in RefSet, then Step 8 re-
ano, and Laguna (2006) and Martí et al. (2006). The first step builds the reference set. The algorithm stops when a termination
(diversification generation method) generates a set P of diverse criterion, i.e., the maximum number of iterations, MaxIter, is
trial solutions with PSize elements and using one or more arbitrary reached.
trial solutions as an input. The improvement method (step 2) is The scatter-search design is presented in Fig. 2.
P
Diversification Generation Improvement
Method method
Subset
S S S Generation
Method RefSet
5. Solution method Golden et al. (1984) extended the Clarke and Wright (1964) sav-
ings technique for the FSMVRP. Fig. 4 presents the savings algo-
The initial reference set is composed of solutions generated by rithms proposed by Golden et al.
two different construction heuristics. An adaptation of Dullaert We implement only the ROS-c algorithm described by Golden
et al.’s sequential insertion heuristic (2002) is applied to generate et al. with the aim of generating diversity. The ROS-c algorithm
quality solutions. To increase the diversity of the solution pool in uses a route shaping parameter that changes the Clarke and Wright
the reference set, an adaptation of Golden et al.’s heuristic is also savings expression to sij ¼ c0i þ c0j ccij . In our tests, c was al-
used. The solution methods are described below. lowed to vary from 0.0 to PSize=21
10
, by tenths, thus generating a total
of PSize/2 trial solutions.
5.1. Adaptation of Dullaert et al.’s heuristic (ADH) We adapt the ROS-c algorithm for the FSMVRPTW by adding
the time window constraints. First, due to the existence of the time
This heuristic is based on the ideas of Ho and Haugland (2004) windows, we must consider the route orientation. Second, it is nec-
and is an extension of the sequential insertion heuristic of Dullaert essary to verify whether the time window constraints are unal-
et al. (2002). tered by joining customers i and j.
In the Dullaert et al. heuristic, the spare capacity of the vehicle The ROS-c algorithm was also adapted to generate split deliver-
must be at least as large as the customer’s demand to allow the ies. Thus, the capacity constraint is not respected such that the de-
customer to be serviced by this vehicle. However, in the ADH, a mand of a customer is added while capacity exists, similar to the
customer is appended to the route without determining whether ADH.
the vehicle is capable of handling the entire demand, only whether Therefore, the AGH is an adaptation of Golden et al.’s heuristic,
it is fully loaded. introducing the option of split deliveries and adding the time win-
Therefore, the ADH is an extension of the heuristic described in dow constraints.
Dullaert et al. by introducing the option of split deliveries. If a cus-
tomer u is chosen to initialize a route k or to be inserted between 5.3. Scatter-search procedure
customers i and j, and this demand exceeds the spare capacity of
the largest vehicle able to serve the route, u is still appended and This section describes the scatter-search procedure proposed to
the excess demand is offloaded to a different vehicle. The algo- solve the fleet size and mixed vehicle routing problem with time
rithm is presented in Fig. 3. windows and split deliveries.
5.2. Adaptation of Golden et al.’s heuristic (AGH) 5.3.1. Diversification generation method
The initial population must contain a broad set of disperse indi-
The second constructive heuristic proposed is based on the viduals. However, it must also include good quality individuals.
ideas of Ho and Haugland (2004) and is an extension of the Golden Therefore, the individuals for the population are created using a
et al. algorithm. random procedure in the constructive heuristics to achieve a
certain level of diversity. In the next step, an improvement method customers and routes with idle capacity. For each possible route
must be applied to these individuals to obtain better solutions. Rj eliminated (we begin with the longest route, i.e., the one with
Therefore, randomized versions of the ADH and AGH heuristics the maximum length), we aim to combine it with another route
known as ADH-random and AGH-random may be constructed by Rk chosen through a candidate list. This candidate list is based
modifying the initialization criterion and the insertion criterion. on the ideas of Corberán, Fernández, Laguna, and Martí (2002).
In ADH-random and AGH-random, customer j is randomly selected
from a candidate list. The candidate list is created by first defining For each route Rj, we randomly choose a candidate route Rk
r_first and r_last as the first and last customers in the list. If the ini- from the candidate list. We attempt to merge the routes, consider-
tialization criterion considers the customer with the earliest dead- ing Rj first and subsequently Rk. If the merge is feasible, we stop the
line, r_first and r_last are the customers with the earliest and latest process and go to the next route Rj. A feasible merging of routes Rj
deadlines, respectively. If the initialization criterion considers the and Rk is such that the resulting route does not violate the capacity
customer farthest from the depot, r_first and r_last are the custom- and time windows constraints. If the merge is not feasible, we
ers farthest from and nearest to the depot, respectively. In the choose another candidate route Rk from the candidate list, while
insertion criterion, r_first and r_last are the customers with the a candidate route Rk exists.
cheapest and more expensive insertion costs, respectively. A possi-
ble customer i is added to the list if 1. Insertion. This exchange method is based on removing a cus-
tomer from one route and inserting it into another route. The
r i 6 r first þ aðr first r lastÞ
insertion movement is implemented for each route Rj, where
where að0 6 a 6 1Þ is the parameter that controls the amount of all feasible customers will be tested in all positions of another
randomization permitted. If a is set to a value of zero, ADH-random route Rk chosen through a candidate list, similar to the process
and AGH-random become the deterministic procedures described described in the route elimination and combination procedure.
in Sections 5.1 and 5.2. In contrast, the candidate list reaches its We begin with the longest route Rj, which is the route with the
maximum possible cardinality when a is set to a value of one. maximum length, and for each route Rj chosen, we begin with
In the diversification generation method, PSize/2 solutions are the customer i farthest from the depot. We select the best inser-
created using the ADH-random approach and the rest are created tion position of customer i into route Rk. The customer is only
using AGH-random. For the two cases, we consider the first solu- inserted if the total cost is reduced and all constraints are
tion with a = 0 (similar to the solution of the heuristics described respected.
in Sections 5.1 and 5.2, respectively), the second with a = 0, 05, 2. Route addition. Similar to the demand reallocation, we consider
the third with a = 0, 1, etc. until a = 1 or the method generates all customers i whose demand is split into more than one route.
PSize/2 solutions. We begin with customer i located farthest from the depot. For
each customer i, we choose the route Rj that delivers the small-
5.3.2. Improvement method est quantity for customer i. In the next step, we relocate the
To each one of the PSize solutions obtained through the diversi- demand of customer i from route Rj to a new route Rk and cal-
fication generation method, we apply the improvement method culate the reduction or addition to the total cost after this
composed of 5 phases: (1) swap in the same route, (2) demand movement. This process is repeated for the other routes in
reallocation, (3) route elimination and combination, (4) insertion, which customer i is inserted. The demand is added to the actual
and (5) route addition. The first exchange procedure is once again route Rk while space exists and the remaining demand is added
applied at the end of the process. to a new vehicle. In the end, the best combination is chosen in
terms of savings that may exist.
1. Swap in the same route. This exchange procedure has the goal of
reducing the length of a route. The procedure is applied to each 5.3.3. Reference set update method
route Rk for k = 1, . . . , m. From i = 1, . . . , nk - 1 and Solutions are included in the reference set due to either quality
j = i + 1, . . . , nk, the procedure tests the reduction in the length or diversity. The subset of quality solutions (RefSet1) contains the b1
of a route Rk produced by exchanging the positions of rk(i) best solutions, and the subset of diverse solutions (RefSet2) contains
and rk(j). If the length is reduced and the constraints are the b2 diverse solutions. The initial RefSet consists of b1 best solu-
respected, the positions are exchanged. The procedure ends tions that belong to P and b2 elements from P that maximize the
when no additional feasible exchanges are possible that result minimum distance to RefSet. The distance between two solutions
in a shorter route length. is calculated by adding the number of non-common arcs of each
2. Demand reallocation. This procedure is applied for each customer solution that exist before the combination. If an arc belongs to more
i, i = 1, . . . , n, that is split into more than one route. We begin than one route, we add one unit for each non-common arc. We con-
with customer i, the farthest from the depot. For a determined sider arc xi0 or x0i only for routes with a full truckload (0 i 0).
customer i, we initially choose the route Rj that delivers the most Two main aspects must be considered when updating the Ref-
quantity for customer i because customers with superior Set. The first aspect refers to the timing of the update, and two pos-
demands have a larger probability of violating the capacity con- sibilities exist: a static update (S) and a dynamic update (D). The
straint if they are not inserted at the beginning (Salhi & Rand, second aspect addresses choosing the criteria for adding to and
1993). Therefore, we calculate the demand reallocation cost of deleting elements from the RefSet. We consider two types of up-
customer i of route Rj for each one of the other routes Rk where dates, by quality (Q) and by quality and diversity (QD).
customer i is inserted, and we choose the cheapest one. The In the static update (S), the reference set does not change until
exchange is enacted only if all problem constraints are respected all solution combinations for the RefSet are completed. In the dy-
and the total cost is reduced. The procedure is repeated for the namic update (D), the reference set is updated when a new and
other routes in which customer i is inserted. The type of vehicle better solution is found.
designated is always the smallest vehicle able to fulfill the In the quality update (Q), if a new solution is better than the
demand of the route (this rule is used in all phases). worst element of RefSet1, then the worst element of RefSet1 is re-
3. Route elimination and combination. This exchange procedure placed by the new solution. In the quality and diversity update
aims to eliminate the vehicles with small capacity. The elimina- (QD), if a new solution is better than the worst element of RefSet1,
tion criterion can also be applied to routes with n or fewer then the worst element of RefSet1 is replaced by the new solution. If
P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601 595
a new solution is worse than the worst element of RefSet1, but it in- demand is added while capacity exists, and the remaining demand
creases the minimum distance between the solutions in RefSet2, will be delivered through Step 2.2.
then the solution with the minimum distance in RefSet2 is replaced If customer i belongs to more than one combined route, and the
by the new solution. total demand was not fulfilled by the actual vehicles, then the min-
imum between the idle capacity of the largest capable vehicle and
the remaining demand of customer i is delivered for each route;
the route chosen is the one that minimizes the additional fixed
5.3.4. Solution combination method
cost. The procedure is repeated until the total demand of customer
The solution combination method is applied to all pairs of solu-
i has been fulfilled or until all vehicles (the largest capable) in
tions in the current referent set.
which customer i is inserted are loaded to maximum capacity.
This method is divided into two steps. Step 1 seeks to combine
By the end of this step, if the total demand of customer i has not
only the common elements of the combined routes and Step 2 sup-
been fulfilled, we go to Step 2.2. Otherwise, the next furthest cus-
plies the remaining demand of the customers. The combination
tomer is chosen and inserted.
method was based on the work of Corberán et al. (2002) and Rego
and Leão (2000).
Step 2.2. If customer i does not belong to any combined route or
the total demand of the customer i has not been fulfilled
Step 1. Let A be a solution with p routes and B a solution with q
through Step 2.1, then Step 2.2 is carried out. Based on the ini-
routes, where Ai is the ith route for solution A, i = {1, . . . , p}, Bk is
tial solutions A and B (before combination), all routes in which
the kth route for solution B and k = {1, . . . , q}. The solutions A
customer i is inserted are verified as well as all of the arcs in
and B are combined. The combination procedure of Step 1 is
which xij = 1 (customer i is served before customer j) or xji = 1
built from the matrix A B, whose components (Ai, Bk) have
(customer i is served after customer j). Therefore, for each j
the number of common elements (common customers)
belonging to one of the combined routes in step 2.1 (except
between the route i of solution A and the route k of solution
the depot (j = 0)), we calculate the cost for inserting customer
B. First, we define which routes are combined, i.e., which com-
i (addition of fixed cost, routing cost and time) before customer
ponent (Ai, Bk) is chosen. This procedure begins with the compo-
j (when xij = 1) or the cost for inserting customer i after cus-
nent that has a greater number of common elements. If there is
tomer j (when xji = 1). We consider xi0 = 1 or x0i = 1 only for
a tie, the component that minimizes the function f,
P routes with a full truckload (0 i 0).
f ¼ Sj¼1 jyAj;i yBj;k j is chosen, where j = 1, . . . , S is the common
customer of the route i of solution A and of the route k in solu-
For each possible position of customer insertion i, we deliver the
tion B; yAj;i is the quantity delivered to customer j from route i in
minimum between the idle capacity of the largest vehicle capable
solution A; and yBj;i is the quantity delivered to the customer j
and the demand of customer i, and we choose the option with the
from route k in solution B.
minimum cost because the time window constraint has been re-
spected. The procedure is repeated until the total demand of the
The combined route is formed by the common elements of the
customer i is fulfilled or while there is a position in which to insert
routes. The quantity delivered to each element is the smallest
customer i.
quantity between the two combined solutions, and the type of
Once all the routes from solutions A and B in which customer i is
vehicle is the smallest vehicle able to service the route demand.
inserted have been verified, if the total demand has not been ful-
The route for the combined solution is similar to that of the best
filled, we add a new route (the vehicle is the smallest able to serve
solution and if there is a tie, we randomly choose the route from
the remaining demand of customer i). The procedure is repeated
one of the solutions. Each combined route is excluded from the list
until the total demand of customer i is fulfilled.
(we delete the line and column referring to components Ai, Bk). The
The combination method guarantees the feasibility of the final
procedure continues as long as there are routes (which have not
solutions. For each combined solution, we apply the improvement
been combined yet) with common elements.
method.
The algorithm stops when MaxIter = 5, or if the maximum time
Step 2. This step aims to meet the remaining demand of custom-
of 1 h is reached, until the last iteration is finished.
ers who have already been served by a route in step 1 and to
meet the complete demand of customers who still do not
belong to any route. This process is carried out through an 6. Experimental results
insertion procedure. Customer i is the farthest from the depot,
and will be inserted. First, we verify whether it already belongs The algorithm was coded in Delphi 7 and run on an AMD Athlon
to any combined route, and after that, two steps can be taken: computer with a 1-GHz processor and 512 MB of RAM memory.
Step 2.1. If customer i already belongs to at least one combined The algorithm proposed was adapted to solve three varieties of
route (Step 1 of the combination method), the route with larger benchmark examples available in the literature: Solomon (1987)
idle capacity is chosen, and the minimum idle capacity between for the VRPTW, Ho and Haugland (2004) for the VRPSD, and Liu
the actual vehicle and the customer demand is delivered. While and Shen (1999) for the FSMVRP.
the total customer demand i is not yet serviced and a route
exists with idle capacity in which customer i is inserted, this 6.1. Solomon’s problem sets
procedure is repeated (without changing the actual vehicle).
Solomon generated six sets of problems: R1, R2, C1, C2, RC1 and
If customer i belongs to only one combined route, and the total RC2. Their design highlights several factors that affect the behavior
demand has not been served with the actual vehicle, then the vehi- of routing and scheduling algorithms, including the geographical
cle is exchanged for the smallest vehicle able to fulfill the remain- data, the number of customers serviced by a vehicle, and such time
ing demand. If the remaining demand of customer i is larger than window characteristics as percent of time-constrained customers
the idle capacity of the largest vehicle able to serve it, then the and tightness and positioning of the time windows.
596 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
Table 2
Results of T-test comparing the means of Table 1.
Table 3
Demand values. diversity) for different PSize values (10, 20, 30, 40). The results,
for [l, u] = [0.70, 1.00] are shown in Table 5.
l u
In terms of distance and considering different PSize values, the
0.01 0.50
best result obtained was for PSize = 40. However, in terms of vehi-
0.02 1.00
0.50 1.00
cles, the best result obtained was for PSize = 30. For both cases, the
0.70 1.00 best result for the updating aspects was obtained with quality and
diversity (QD) and static (S) updates. Because the results with
PSize = 30 were similar to those with PSize = 40 and the computa-
frequency (S). These parameters were tested for values tional time obtained was lower, we chose the first option. There-
[l, u] = [0.70, 1.00]. The average values for classes R1, C1, RC1, R2, fore, the complete results presented below consider PSize = 30,
C2 and RC2 are listed in Table 4. b1 = 5, b2 = 5, with static (S) and quality and diversity (QD) updates.
According to Table 4, the best result obtained was for b1 = 5 e Table 6 compares the results of the algorithm presented in this
b2 = 5. The shortest computational time was obtained for b1 = 3 paper with the results from Ho and Haugland (2004) for each class
and b2 = 2, but the results were inferior. The computational time of problems and considering different [l, u] values.
(seconds) represents the average of the 56 problems. According to Table 6, we can conclude that, in 6 classes of prob-
Using b1 = 5 and b2 = 5, we tested the first aspect of the update lems, the average from our scatter-search has outperformed the
(static or dynamic) and the second aspect (quality or quality and best results from Ho and Haugland (2004). Moreover, for another
Table 4
Scatter-search results for different values of b1, b2 considering [l, u] = [0.70, 1.00].
Problem b1 = 3, b2 = 2 b1 = 5, b2 = 5 b1 = 7, b2 = 3 b1 = 3, b2 = 7
Distance Vehicles Distance Vehicles Distance Vehicles Distance Vehicles
R1 4485.11 83.42 4464.85 82.75 4471.42 82.75 4470.98 83.42
C1 4961.41 77.50 4950.81 76.88 4954.12 76.88 4952.33 76.88
RC1 6047.80 82.50 6013.92 82.50 6012.74 82.50 6011.45 82.50
R2 4665.96 83.18 4625.87 82.82 4629.30 82.82 4630.81 83.00
C2 5229.33 77.50 5214.79 76.88 5213.96 77.50 5217.20 76.88
RC2 6237.58 83.88 6217.43 83.25 6220.60 83.25 6219.52 83.25
Total 31627.19 487.98 31487.67 485.08 31502.14 485.70 31502.29 485.93
Timea 302 711 715 715
a
Average processing time (seconds).
Table 5
Results for different updates and different PSize values considering [l, u] = [0.70, 1.00].
Table 6
Comparison with Ho and Haugland (2004) results for Solomon’s modified demands.
l, u R1 C1 RC1
0.01, 0.50 Ho and Haugland 18.25/1471.49 12.22/1182.12 20.13/1965.05
Scatter search 18.42/1475.54 12.22/1160.74 21.00/1941.25
0.02, 1.00 Ho and Haugland 35.00/2291.46 22.22/2168.57 40.00/3339.20
Scatter search 35.83/2302.58 24.00/2009.37 41.75/3425.96
0.50, 1.00 Ho and Haugland 67.00/4040.67 61.00/3979.78 70.00/5453.10
Scatter search 69.50/4035.84 60.75/3975.49 73.75/5231.85
0.70, 1.00 Ho and Haugland 79.00/4581.54 77.00/4962.28 81.00/6095.20
Scatter search 82.75/4464.85 76.88/4950.81 82.50/6013.92
R2 C2 RC2
0.01, 0.50 Ho and Haugland 18.00/1430.62 11.13/1174.29 20,00/1946,07
Scatter search 18.00/1425.40 11.75/1180.34 21.00/1941.42
0.02, 1.00 Ho and Haugland 35.00/2318.04 22.00/1995.59 39,00/3419,85
Scatter search 35.82/2314.65 23.13/1993.47 41.50/3410.65
0.50, 1.00 Ho and Haugland 68.00/4059.26 61.00/4268.02 71,00/5546,20
Scatter search 69.27/4055.29 60.88/4259.14 71.00/5498.32
0.70, 1.00 Ho and Haugland 80.00/4574.17 77.00/5246.12 82,00/6155.49
Scatter search 82.82/4625.87 76.88/5214.79 83.25/6217.43
Table 7
Results of the T-test comparing the means of Table 6.
Table 8
Average processing time (seconds) for Ho and Haugland (2004) modified demands.
l, u R1 R2 C1 C2 RC1 RC2
0.01, 0.50 Ho and Haugland 1442 1842 785 2045 1194 1912
Scatter search 807 982 517 1121 678 1030
0.02, 1.00 Ho and Haugland 1618 1903 1475 1595 1634 1666
Scatter search 840 1022 785 888 905 926
0.50, 1.00 Ho and Haugland 1923 1901 982 1595 1901 1666
Scatter search 1025 1014 574 876 1014 885
0.70, 1.00 Ho and Haugland 1502 1589 1638 1794 1705 1807
Scatter search 812 899 755 922 901 957
11 problems, it was possible to reduce the average distance trav- 6.3. Liu and Shen’s problem set (1999)
eled, but the average number of vehicles used was higher. It has
also been verified that the best results were found for The developed method was applied to solve the benchmark
[l, u] = [0.50, 1.00]e[0.70, 1.00] because higher demand values in- examples proposed by Liu and Shen (1999) and derived from the
crease the possibility of split deliveries. vehicle routing problem with time windows (VRPTW) examples
To investigate whether the performances of the scatter-search of Solomon (1987).
meta-heuristic and the Ho and Haugland algorithm are different
with respect to distance and number of vehicles (data from Ta-
ble 6), a T-test was again performed, and the results are shown
in Table 7. In terms of distance, we can conclude that there is no Table 9
difference in the performances of Ho and Haugland’s algorithm Scatter-search results for different values of b1 and b2.
and the scatter-search method proposed in this paper. b1 b2 Timea Valuesa Valuesb
Table 8 shows the average processing time (seconds) for the
3 2 516 3035.65 1314.87
problem classes R1, C1, RC1, R2, C2 and RC2 with consideration 5 5 687 2997.32 1294.17
of different values of [l, u], and comparison of the results with 7 3 687 2983.00 1288.20
those of Ho and Haugland (2004). In all classes of problems, we ob- 3 7 695 3001.25 1295.01
serve that the computational time obtained by the scatter-search a
Average processing time (seconds) of the 168 problems.
algorithm was significantly lower than those produced by Ho and b
Average value considering the total fixed cost of the used vehicles and the total
Haugland (2004). schedule time (excluding the service time of the customers).
P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601 599
Table 10
The results for different updates and different PSize values.
Table 11 Table 12
Comparison with Liu and Shen’s algorithm. Comparison with Dullaert et al.’s algorithm.
Set Liu and Shen (1999) Scatter search Set Dullaert et al. (2002) Scatter search
Value Time (s) Value D (%) Time (s) Value Value D (%)
R1A 4398 4210.36 4.27 531 R1A 1548.53 1430.84 7.60
R1B 2054 1937.52 5.67 586 R1B 1557.38 1432.57 8.01
R1C 1700 1640.64 3.49 605 R1C 1557.85 1432.57 8.04
R1 2717.33 531 2596.17 4.48 574 R1 1554.59 1431.99 7.89
C1A 8007 7421.33 7.31 644 C1A 1166.09 1004.42 13.86
C1B 2485 2394.7 3.63 629 C1B 1126.01 1001.33 11.07
C1C 1705 1670.94 2.00 665 C1C 1155.45 1003.85 13.12
C1 4065.67 435 3828.99 4.32 646 C1 1149.18 1003.20 12.69
RC1A 5184 5130.44 1.03 619 RC1A 1665.04 1568.96 5.77
RC1B 2235 2190.39 2.00 591 RC1B 1680.55 1574.27 6.32
RC1C 1849 1790.12 3.18 621 RC1C 1689.92 1576.48 6.71
RC1 3089.33 470 3036.98 2.07 610 RC1 1678.50 1573.24 6.27
R2A 3809 3708.68 2.63 672 R2A 1426.52 1345.80 5.66
R2B 1797 1727.04 3.89 806 R2B 1431.49 1345.80 5.99
R2C 1513 1439.27 4.87 794 R2C 1419.81 1336.30 5.88
R2 2373 529 2291.66 3.80 757 R2 1425.84 1341.05 5.84
C2A 6717 6355.09 5.39 672 C2A 821.38 751.62 8.49
C2B 1970 1905.82 3.26 686 C2B 821.38 751.62 8.49
C2C 1288 1276.29 0.91 633 C2C 811.16 747.22 7.88
C2 3325.00 444 3179.07 3.19 664 C2 817.97 750.15 8.29
RC2A 5273 4811.2 8.76 834 RC2A 1800.82 1630.84 9.44
RC2B 2324 2189.66 5.78 870 RC2B 1741.97 1632.57 6.28
RC2C 1978 1894.57 4.22 910 RC2C 1741.75 1632.57 6.27
RC2 3191.67 533 2965.14 6.25 871 RC2 1761.55 1631.99 7.33
Average 3074.11 494 2983.00 4.02 687 Average 1406.20 1288.80 8.05
The original VRPTW examples contain 56 problems. Each of Liu and Shen (1999) adapted Solomon’s (1987) problem set to a
these problems contains 100 customers, and the travel time be- heterogeneous fleet, adding the costs for different types of vehicles
tween nodes is equal to the Euclidean distance. The 56 test prob- (see Appendix A). Three different sets of vehicle costs are consid-
lems are grouped into six problem types. The geographical data ered for each problem of Solomon. Therefore, a total of 168 prob-
are either randomly generated according to a uniform distribution lems were solved. For example, the notation R1a, R1b and R1c
(problem sets R1 and R2), clustered in groups (problem sets C1 and refer to the original R1 set with a larger vehicle cost, a medium
C2), or semi-clustered (problem sets RC1 and RC2). The semi-clus- vehicle cost, and a smaller vehicle cost, respectively.
tered problem contains a mix of randomly generated data and clus-
ters. Furthermore, sets R1, C1 and RC1 have a shorter scheduling 6.3.1. Experiments on Liu and Shen’s problem set (1999)
horizon and allow fewer customers per route. Problem sets R2, The results of the SS algorithm are compared with the heuristics
C2 and RC2 have a longer scheduling horizon and allow a larger proposed by Liu and Shen (1999) and Dullaert et al. (2002). The re-
number of customers per route. sults of these papers were presented in two different ways. Liu and
600 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
Table 13
Results of T-test comparing the means of Tables 11 and 12.
Shen’s solution value considers the vehicle fixed costs and the total and Shen (1999) for the FSMVRP, benchmark problems by Ho
schedule time and excludes the service time for customers, which and Haugland (2004) for the VRPSD, and benchmark problems by
is constant. Dullaert et al.’s solution cost considers only the total Solomon (1987) for the VRPTW.
schedule time, excluding the service time of the customers and For the three problems sets, many parameters of the scatter-
the vehicle fixed costs. search meta-heuristic were tested: PSize, b1, b2, the updating crite-
First, we tested different parameters of the SS algorithm with ria and the updating frequency. We conclude that the calibration of
respect to both the quality of the solutions found and the comput- these parameters influences the performance of the algorithm.
ing times for the following: b1, b2 values (3, 2; 5, 5; 7, 3; 3, 7), PSize Considering Solomon’s problem sets, for certain classes, the
values (10, 20, 30, 40), the first aspect of the update (static or dy- scatter-search algorithm has matched the best results found in
namic), and the second aspect of the update (quality or quality the literature. Problems R1, C1 and RC1 have produced better re-
and diversity). The best result was obtained with PSize = 30, sults than classes R2, C2 and RC2. Moreover, class C problems have
b1 = 7, b2 = 3, dynamic update (D) and quality and diversity update presented better results on average than classes R and RC. The
(QD). The results are detailed in Tables 9 and 10. The average value worst result was presented by class R problems. We conclude,
represents the average results over all 168 problems. therefore, that the algorithm adapts better to a short-term horizon,
Table 11 compares our algorithm with that of Liu and Shen routes with fewer customers, small vehicles and geographically
(1999). Liu and Shen (1999) adapted the savings concept of Golden clustered and semi-clustered customers.
et al. (1984), generating three new heuristics: the MCS (modified Computational testing revealed that our algorithm matched a
combined savings), the MOOS (modified optimistic opportunity subset of the best solutions for Solomon’s problem set with aug-
savings) and the MROS (modified realistic opportunity savings). mented demands. It has also been verified that the best results
An improvement phase is applied to the constructive heuristics were found for [l, u] = [0.50, 1.00]e[0.70, 1.00] because the higher
generated. Table 11 presents only the best result (MCS, MOOS or the demand values, the higher the possibility of split delivery. In
MROS with improvement) for each group of instances (i.e., terms of computational time, our algorithm produced values that
R1a, R1b, R1c, . . . , RC2c), giving the average solution value of Liu were significantly lower than those proposed by Ho and Haugland
and Shen’s algorithm and of our scatter-search algorithm as well (2004).
as the percentage improvement we obtained. Table 11 shows that Considering Liu and Shen’s examples, our approach outper-
our scatter-search algorithm outperforms that of Liu and Shen in formed both Liu and Shen’s algorithm and Dullaert et al.’s heuristic
terms of solution quality. The percentage gap is reduced to 4.02% in terms of solution quality.
on average. In future research, other constructive heuristics will be applied
Finally, Table 12 compares our algorithm with that of Dullaert as initial solutions, and other improvement heuristics will also be
et al. (2002); however, the authors do not give computing times tested. Moreover, the scatter-search meta-heuristic developed
for their constructive heuristic. For each group of instances (i.e., can be adapted for other problems, such as VRP with multiple de-
R1a, R1b, R1c, . . . , RC2c), Table 12 compares the average solution pots and pickup and delivery problems, among others. Finally, a
value of Dullaert et al.’s heuristic with that of our scatter-search new set of problems could be created containing all characteristics
approach, including the percentage gap we obtained. Table 12 from the problem studied.
shows that our approach outperforms that of Dullaert et al.
(2002) for each group of instances. The percentage gap is reduced
to 8.05% on average. References
To determine whether the performance of the scatter-search
Alegre, J., Laguna, M., & Pacheco, J. (2007). Optimizing the periodic pick-up of raw
algorithm differs significantly from that of Liu and Shen’s algo- materials for a manufacturer of auto parts. European Journal of Operational
rithm (Table 11) and Dullaert et al.’s algorithm (Table 12), the T- Research, 179, 736–746.
Archetti, C., Savelsbergh, M. W. P., & Speranza, M. G. (2006). Worst-case analysis of
test was applied once again, and the results are shown in Table 13.
split delivery routing problems. Transportation Science, 40, 226–234.
We can conclude that the results achieved by the scatter-search Archetti, C., Speranza, M. G., & Hertz, A. (2006). A tabu search algorithm for the split
algorithm are significantly better than those of Liu and Shen’s delivery vehicle routing problem. Transportation Science, 40, 64–73.
and Dullaert et al.’s algorithms. Belenguer, J. M., Martinez, M. C., & Mota, E. (2000). A lower bound for the split
delivery vehicle routing problem. Operations Research, 48, 801–810.
Belfiore, P., & Yoshizaki, H. T. Y. (2009). Scatter search for a real-life heterogeneous
fleet vehicle routing problem with time windows and split deliveries. European
7. Conclusions and future research Journal of Operational Research, 199, 750–758.
Brandão, J. (2011). A tabu search algorithm for the heterogeneous fixed fleet vehicle
routing problem. Computers & Operations Research, 38, 140–151.
We have developed a scatter-search algorithm for the Fleet Size Brandão, J., & Mercer, A. (1997). A tabu search algorithm for the multi-trip vehicle
and Mixed Vehicle Routing Problem with Time Windows and Split routing and scheduling problem. European Journal of the Operational Research,
100, 180–191.
Deliveries (FSMVRPTWSD). Two constructive heuristics were pro-
Clarke, G., & Wright, J. W. (1964). Scheduling of vehicles from a central depot to a
posed as initial solutions. The scatter-search framework provided number of delivery points. Operations Research, 12, 568–581.
a means to combine solutions as well as to diversify and intensify Corberán, A., Fernández, E., Laguna, M., & Martí, R. (2002). Heuristic solutions to the
the meta-heuristic search process. problem of routing school buses with multiple objectives. Journal of the
Operational Research Society, 53, 427–435.
The model was adapted to solve three varieties of benchmark Desrochers, M., & Verhoog, T. W. (1991). A new heuristic for the fleet size and mix
examples available in the literature: benchmark problems by Liu vehicle routing problem. Computers & Operations Research, 18, 263–274.
P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601 601
Dror, M., Laporte, G., & Trudeau, P. (1994). Vehicle routing with split deliveries. Li, X., Tian, P., & Aneja, Y. P. (2010). An adaptive memory programming
Discrete Applied Mathematics, 50, 229–254. metaheuristic for the heterogeneous fixed fleet vehicle routing problem.
Dror, M., & Trudeau, P. (1989). Savings by split delivery routing. Transportation Transportation Research Part E: Logistics & Transportation Review, 46, 1111–1127.
Science, 23, 141–145. Liu, F. H., & Shen, S. Y. (1999). The fleet size and mix vehicle routing problem with
Dror, M., & Trudeau, P. (1990). Split delivery routing. Naval Research Logistics, 37, time windows. Journal of Operational Research Society, 50, 721–732.
383–402. Martí, R., Laguna, M., & Glover, F. (2006). Principles of scatter search. European
Dullaert, W., Janssens, G. K., Sorensen, K., & Vernimmen, B. (2002). New heuristics Journal of Operational Research, 169, 359–372.
for the fleet size and mix vehicle routing problem with time windows. Journal of Mullaseril, P. A., Dror, M., & Leung, J. (1997). Split-delivery routing heuristics in
Operational Research Society, 53, 1232–1238. livestock feed distribution. Journal of the Operational Research Society, 48,
Frizzell, P. W., & Giffin, J. W. (1992). The bounded split delivery vehicle routing 107–116.
problem with grid network distances. Asia Pacific Journal of Operational Research, Rego, C., & Leão, P. (2000). A scatter search tutorial for graph-based permutation
9, 101–106. problems. Hearin Center for Enterprise, University of Mississippi, HCES-10-00,
Frizzell, P. W., & Giffin, J. W. (1995). The split delivery vehicle scheduling problem USA. In C. Rego & B. Alidaee (Eds.), Metaheuristic optimization via memory and
with time windows and grid network distances. Computers & Operations evolution: Tabu search and scatter search, 2005 (pp. 1–24). Kluwer Academic
Research, 6, 655–667. Publishers.
Gendreau, M., Laporte, G., Musaraganyi, C., & Taillard, E. D. (1999). A tabu search Rochat, Y., & Semet, F. (1994). A tabu search approach for delivering pet food and
heuristic for the heterogeneous fleet vehicle routing problem. Computers & flour in Switzerland. Journal of the Operational Research Society, 45, 1233–1246.
Operations Research, 26, 1153–1173. Rochat, Y., & Taillard, E. D. (1995). Probabilistic diversification and intensification in
Gheysens, F., Golden, B., & Assad, A. (1984). A comparison of techniques for solving local search for vehicle routing. Journal of Heuristics, 1, 147–167.
the fleet size and mix vehicle routing problem. Operations Research Spectrum, 6, Salhi, S., & Rand, G. K. (1993). Incorporating vehicle routing into the vehicle fleet
207–216. composition problem. European Journal of Operational Research, 66, 313–330.
Gheysens, F., Golden, B., & Assad, A. (1986). A new heuristic for determining fleet Solomon, M. M. (1987). Algorithms for the vehicle routing and scheduling problems
size and composition. Mathematical Programming Study, 26, 233–236. with time windows constraints. Operations Research, 35, 254–265.
Glover, F. (1977). Heuristics for integer programming using surrogate constraints. Solomon, M. M., & Desrosiers, J. (1988). Time window constrained routing and
Decision Sciences, 8, 156–166. scheduling problem. Transportation Science, 22, 1–13.
Glover, F. (1998). A template for scatter search and path relinking. In J.-K. Hao, E. Taillard, É. D. (1999). A heuristic column generation method for the heterogeneous
Lutton, E. Ronald, M. Schoenauer, & D. Snyers (Eds.), Artificial evolution. Lecture fleet vehicle routing problem. Rairo, 33, 1–14.
notes in computer science (vol. 1363, pp. 13–54). Springer-Verlag. Tarantilis, C. D., Kiranoudis, C. T., & Vassiliadis, V. S. (2004). A threshold accepting
Golden, B. L., Assad, A., Levy, L., & Gheysens, F. (1984). The fleet size and mix vehicle metaheuristic for the heterogeneous fixed fleet vehicle routing problem.
routing problem. Computers & Operations Research, 11, 49–65. European Journal of Operational Research, 152, 148–158.
Ho, S. C., & Haugland, D. (2004). A tabu search heuristic for the vehicle routing Wassan, N. A., & Osman, I. H. (2002). Tabu search variants for the mix fleet vehicle
problem with time windows and split deliveries. Computers & Operations routing problem. Journal of the Operational Research Society, 53, 768–782.
Research, 31, 1947–1964. Yamashita, D. S., Armentano, V. A., & Laguna, M. (2006). Scatter search for project
Lenstra, J. K., & Rinnoy Kan, A. H. G. (1981). Complexity of Vehicle and Scheduling scheduling with resource availability cost. European Journal of Operational
Problems. Networks, 11, 221–227. Research, 169, 623–637.