Computers & Industrial Engineering 64 (2013) 589–601

Contents lists available at SciVerse ScienceDirect

Computers & Industrial Engineering

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

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
The classical vehicle routing problem (VRP) aims to find a set of
routes with minimal cost (by finding the shortest path, minimizing
the number of vehicles, etc.) and beginning and ending the route at
the depot such that the known demand of all nodes is fulfilled.
Each node is visited only once by only one vehicle, and each vehicle
has a limited capacity. Selected formulations also include con-
straints on the maximum travel time.
The VRPSD is a variation of the classical VRP in which each cus-
tomer can be served by more than one vehicle. Thus, for the VRPSD,
in addition to the delivery routes, the quantity to be delivered to
each customer by each vehicle must also be determined. The op-
tion of splitting a given demand makes it possible to service a cus-
tomer whose demand exceeds the vehicle capacity. Splitting may
also result in decreased costs. The vehicle routing problem with
time windows and split deliveries (VRPTWSD) presents an exten-
sion of the VRPSD in which the time window restraints are added.
In the literature, three variants of a VRP with a heterogeneous
fleet have been studied. The first was introduced by Golden, Assad,
Levy, and Gheysens (1984) in which the variable costs were uni-
formly spread over all vehicle types with the number of available
vehicles assumed as unlimited for each type. This version is ad-
dressed in this paper and is known as the fleet size and mixed vehi-
cle routing problem (FSMVRP), the vehicle fleet mix (VFM) or the
fleet size and composition VRP. The second version considers the
⇑ Corresponding author.
E-mail addresses: pbfavero@usp.br (P. Belfiore), hugo@usp.br (H.T.Y. Yoshizaki).
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-
erogeneous fixed fleet vehicle routing problem (HFFVRP),
generalizes the second version by limiting the number of available
vehicles of each type.
Lenstra and Rinnoy Kan (1981) have analyzed the complexity of
the vehicle routing problem and have concluded that practically all
vehicle routing problems are NP-hard (including the classical vehi-
cle routing problem) because they are not solved in polynomial
time. Because the FSMVRPTWSD is a combination of the FSMVRP,
VRPTW and VRPSD, it remains NP-hard (Archetti, Savelsbergh, &
Speranza, 2006; Dror & Trudeau, 1990; Dullaert, Janssens, Soren-
sen, & Vernimmen, 2002; Gendreau, Laporte, Musaraganyi, & Tail-
lard, 1999; Solomon & Desrosiers, 1988; Taillard, 1999). Therefore,
this observation makes a strong case for the application of heuris-
tics and meta-heuristics to solve the problem.
This paper develops a scatter-search (SS) algorithm to solve the
fleet size and mixed vehicle routing problem with time windows
and split deliveries (FSMVRPTWSD). The algorithm proposed was
adapted to solve three varieties of benchmark examples available
in the literature: benchmark problems by Liu and Shen (1999)
for the FSMVRP, benchmark problems by Ho and Haugland
(2004) for the VRPSD, and benchmark problems by Solomon
(1987) for the VRPTW.
The organization of this article is as follows. Section 2 reviews
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
scatter-search overview. Section 5 describes the constructive heu-
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

3. Problem formulation and definitions

The model constraints are:

In this section, we define the problem under study and the nota- X
n

tions used throughout the paper. xk0j ¼ 1 k ¼ 1; . . . ; K ð1Þ

j¼1

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

tji and dij = dji). The demand at point i is denoted by qi, i = 1, 2, . . . , n.

The demand of each customer is known. X
n
qi yki 6 ak k ¼ 1; . . . ; K ð4Þ
i¼1
3.2. Fleet of vehicles
X
n
The customers are served from one depot with a heterogeneous yki 6 xkji i ¼ 1; . . . ; n; k ¼ 1; . . . ; K ð5Þ
and unlimited fleet. The vehicles leave from and return to the de- j¼0

pot. In the heterogeneous problems, we define a set T = {1, . . . , K}

k k
of different vehicle types. A vehicle of type k 2 T has a capacity bi þ si þ tij  M ij ð1  xkij Þ 6 bj i ¼ 1; . . . ; n;
ak, a fixed cost fk and a per unit-distance variable cost gk. The cost j ¼ 1; . . . ; n; k ¼ 1; . . . ; K ð6Þ
of a vehicle of type k 2 T traversing the pair (i, j) is denoted by ckij ,
which is obtained by multiplying the distance dij and the variable k
cost gk. The number nk of vehicles of type k is assumed to be unlim- ei 6 bi 6 li i ¼ 1; . . . ; n; k ¼ 1; . . . ; K ð7Þ
ited ðnk ¼ 1; k 2 TÞ.
We let Ri = {ri(1), . . . , ri(ni)} denote the route for vehicle i, where yki P 0 i ¼ 1; . . . ; n; k ¼ 1; . . . ; K
ð8Þ
ri(j)) is the index of the jth customer visited and ni is the number of k
bi P0 i ¼ 1; . . . ; n; k ¼ 1; . . . ; K
customers on the route. We assume that every route finishes at the
depot, i.e., ri(ni + 1) = 0.
xkij 2 f0; 1g i ¼ 0; . . . ; n; j ¼ 0; . . . ; n; k ¼ 1; . . . ; K ð9Þ
3.3. Time windows Constraints (1) and (2) state that each vehicle leaves the depot;
after arriving at a customer, the vehicle leaves again, and finally re-
Each customer i 2 C has a time window, i.e., an interval ½ei ; li  turns to the depot.
with ei 6 li , which corresponds to the earliest and latest time to be- Constraint (3) guarantees that the total demand of each cus-
gin service to customer i, respectively. Let si be the service time at tomer will be fulfilled.
customer i. Constraint (4) guarantees that the vehicle capacity will not be
exceeded.
3.4. Split deliveries Constraint (5) guarantees that the demand of each customer
will only be fulfilled if a determined vehicle visits the customer
The demand for a given customer may be fulfilled by more than location. We note that adding up Eq. (5) over all vehicles and com-
one vehicle. This situation occurs in all cases where a certain de- bining it with Eq. (3), we obtain the constraint
mand exceeds the vehicle capacity but may be cost effective in PK Pn k
k¼1 i¼0 xij P 1 J ¼ 0; . . . ; n, which guarantees that each vertex
other cases. will be visited at least once by at least one vehicle.
The decision variables of the model are: Eq. (6) sets a minimum time for beginning service to customer j
8
in a determined route and also guarantees that there will be no
< 1 if a customer j is supplied after customer i
>
sub-tours. The constant Mij is a sufficiently large number, e.g.,
xkij ¼ by a vehicle of type k
>
: Mij = li + tij  ej.
0 otherwise Constraint (7) guarantees that all customers will be served
within their time windows.
k k
bi ¼ moment at which service begins at customer i by a vehicle of type k Eq. (8) guarantees that the decision variables yki and bi are
positive.
yki ¼ fraction of customer’s demand i delivered by a vehicle of type k: Finally, Eq. (9) guarantees that the decision variables xkij will be
binary.
The objective of the model is to minimize the sum of the fixed
vehicle costs and the travel costs of servicing the customers within 4. Scatter-search overview
P
the time window limits. The term nj¼1 xk0j represents the number of
vehicles of type k used such that only the non-empty routes are According to Martí, Laguna, and Glover (2006), the scatter
computed. The mathematical programming formulation is search is an evolutionary method that has been successfully ap-
592 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601

plied to hard optimization problems. Glover (1977, 1998) proposed
the first description of this method and a scatter-search template.
The SS operates on a set of reference solutions to generate new
solutions via weighted linear combinations of structured subsets
of solutions. This approach uses strategies for search diversification
and intensification that have proven effective in a variety of opti-
mization problems.
The following parameters are used in the method discussion:
PSize = size of the set of diverse solutions generated by the
Diversification Generation Method.
b = size of the reference set (RefSet).
b1 = size of the high-quality subset of RefSet.
b2 = size of the diverse subset of RefSet.
MaxIter = maximum number of iterations.
Fig. 1 shows a sketch of the scatter-search method based on the
work of Alegre, Laguna, and Pacheco (2007), Yamashita, Arment-
ano, and Laguna (2006) and Martí et al. (2006). The first step
(diversification generation method) generates a set P of diverse
trial solutions with PSize elements and using one or more arbitrary
trial solutions as an input. The improvement method (step 2) is
applied to transform a trial solution into one or more enhanced
trial solutions. If no improvement occurs for a given trial solution,
the enhanced solution is considered to be the same as the one sub-
mitted for improvement. Step 3 chooses b solutions from P accord-
ing to their quality or diversity to build the reference set (RefSet),
which is a collection of both high quality solutions and diverse
solutions. In step 4, the subset generation method produces a
subset of the reference set solutions as a basis for creating com-
bined solutions. The simplest form of the subset generation meth-
od consists of generating all pairs of reference solutions. In step 5,
the solution combination method is applied to generate one or
more trial solutions. For each combined solution, we apply the
improvement method (step 6). At this point, the reference set
can be updated by the reference set update method (step 7)
depending on the quality or diversity of the new combined solu-
tion. In this work, dynamic and static RefSet updating is imple-
mented as described in Section 5.3. If the search converges, i.e.,
no new solutions are found for inclusion in RefSet, then Step 8 re-
builds the reference set. The algorithm stops when a termination
criterion, i.e., the maximum number of iterations, MaxIter, is
reached.
The scatter-search design is presented in Fig. 2.

Fig. 1. Scatter-search algorithm.

P
Diversification Generation Improvement
Method method

Solution Improvement Reference Set

Combination Method Method Update Method

Subset
S S S Generation
Method RefSet

Stop if no more new solutions

Fig. 2. The scatter-search design.

 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
5. Solution method
The initial reference set is composed of solutions generated by
two different construction heuristics. An adaptation of Dullaert
et al.’s sequential insertion heuristic (2002) is applied to generate
quality solutions. To increase the diversity of the solution pool in
the reference set, an adaptation of Golden et al.’s heuristic is also
used. The solution methods are described below.
5.1. Adaptation of Dullaert et al.’s heuristic (ADH)
This heuristic is based on the ideas of Ho and Haugland (2004)
and is an extension of the sequential insertion heuristic of Dullaert
et al. (2002).
In the Dullaert et al. heuristic, the spare capacity of the vehicle
must be at least as large as the customer’s demand to allow the
customer to be serviced by this vehicle. However, in the ADH, a
customer is appended to the route without determining whether
the vehicle is capable of handling the entire demand, only whether
it is fully loaded.
Therefore, the ADH is an extension of the heuristic described in
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
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
the excess demand is offloaded to a different vehicle. The algo-
rithm is presented in Fig. 3.
5.2. Adaptation of Golden et al.’s heuristic (AGH)
The second constructive heuristic proposed is based on the
ideas of Ho and Haugland (2004) and is an extension of the Golden
et al. algorithm.
Fig. 3. ADH algorithm.
Fig. 4. Summary of the savings algorithm.
593
Golden et al. (1984) extended the Clarke and Wright (1964) sav-
ings technique for the FSMVRP. Fig. 4 presents the savings algo-
rithms proposed by Golden et al.
We implement only the ROS-c algorithm described by Golden
et al. with the aim of generating diversity. The ROS-c algorithm
uses a route shaping parameter that changes the Clarke and Wright
savings expression to sij ¼ c0i þ c0j  ccij . In our tests, c was al-
lowed to vary from 0.0 to PSize=21
10
, by tenths, thus generating a total
of PSize/2 trial solutions.
We adapt the ROS-c algorithm for the FSMVRPTW by adding
the time window constraints. First, due to the existence of the time
windows, we must consider the route orientation. Second, it is nec-
essary to verify whether the time window constraints are unal-
tered by joining customers i and j.
The ROS-c algorithm was also adapted to generate split deliver-
ies. Thus, the capacity constraint is not respected such that the de-
mand of a customer is added while capacity exists, similar to the
ADH.
Therefore, the AGH is an adaptation of Golden et al.’s heuristic,
introducing the option of split deliveries and adding the time win-
dow constraints.
5.3. Scatter-search procedure
This section describes the scatter-search procedure proposed to
solve the fleet size and mixed vehicle routing problem with time
windows and split deliveries.
5.3.1. Diversification generation method
The initial population must contain a broad set of disperse indi-
viduals. However, it must also include good quality individuals.
Therefore, the individuals for the population are created using a
random procedure in the constructive heuristics to achieve a
594 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
certain level of diversity. In the next step, an improvement method
must be applied to these individuals to obtain better solutions.
Therefore, randomized versions of the ADH and AGH heuristics
known as ADH-random and AGH-random may be constructed by
modifying the initialization criterion and the insertion criterion.
In ADH-random and AGH-random, customer j is randomly selected
from a candidate list. The candidate list is created by first defining
r_first and r_last as the first and last customers in the list. If the ini-
tialization criterion considers the customer with the earliest dead-
line, r_first and r_last are the customers with the earliest and latest
deadlines, respectively. If the initialization criterion considers the
customer farthest from the depot, r_first and r_last are the custom-
ers farthest from and nearest to the depot, respectively. In the
insertion criterion, r_first and r_last are the customers with the
cheapest and more expensive insertion costs, respectively. A possi-
ble customer i is added to the list if
r i 6 r first þ aðr first  r lastÞ
where að0 6 a 6 1Þ is the parameter that controls the amount of
randomization permitted. If a is set to a value of zero, ADH-random
and AGH-random become the deterministic procedures described
in Sections 5.1 and 5.2. In contrast, the candidate list reaches its
maximum possible cardinality when a is set to a value of one.
In the diversification generation method, PSize/2 solutions are
created using the ADH-random approach and the rest are created
using AGH-random. For the two cases, we consider the first solu-
tion with a = 0 (similar to the solution of the heuristics described
in Sections 5.1 and 5.2, respectively), the second with a = 0, 05,
the third with a = 0, 1, etc. until a = 1 or the method generates
PSize/2 solutions.
5.3.2. Improvement method
To each one of the PSize solutions obtained through the diversi-
fication generation method, we apply the improvement method
composed of 5 phases: (1) swap in the same route, (2) demand
reallocation, (3) route elimination and combination, (4) insertion,
and (5) route addition. The first exchange procedure is once again
applied at the end of the process.
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
route Rk for k = 1, . . . , m. From i = 1, . . . , nk - 1 and
j = i + 1, . . . , nk, the procedure tests the reduction in the length
of a route Rk produced by exchanging the positions of rk(i)
and rk(j). If the length is reduced and the constraints are
respected, the positions are exchanged. The procedure ends
when no additional feasible exchanges are possible that result
in a shorter route length.
2. Demand reallocation. This procedure is applied for each customer
i, i = 1, . . . , n, that is split into more than one route. We begin
with customer i, the farthest from the depot. For a determined
customer i, we initially choose the route Rj that delivers the most
quantity for customer i because customers with superior
demands have a larger probability of violating the capacity con-
straint if they are not inserted at the beginning (Salhi & Rand,
1993). Therefore, we calculate the demand reallocation cost of
customer i of route Rj for each one of the other routes Rk where
customer i is inserted, and we choose the cheapest one. The
exchange is enacted only if all problem constraints are respected
and the total cost is reduced. The procedure is repeated for the
other routes in which customer i is inserted. The type of vehicle
designated is always the smallest vehicle able to fulfill the
demand of the route (this rule is used in all phases).
3. Route elimination and combination. This exchange procedure
aims to eliminate the vehicles with small capacity. The elimina-
tion criterion can also be applied to routes with n or fewer
customers and routes with idle capacity. For each possible route
Rj eliminated (we begin with the longest route, i.e., the one with
the maximum length), we aim to combine it with another route
Rk chosen through a candidate list. This candidate list is based
on the ideas of Corberán, Fernández, Laguna, and Martí (2002).
For each route Rj, we randomly choose a candidate route Rk
from the candidate list. We attempt to merge the routes, consider-
ing Rj first and subsequently Rk. If the merge is feasible, we stop the
process and go to the next route Rj. A feasible merging of routes Rj
and Rk is such that the resulting route does not violate the capacity
and time windows constraints. If the merge is not feasible, we
choose another candidate route Rk from the candidate list, while
a candidate route Rk exists.
1. Insertion. This exchange method is based on removing a cus-
tomer from one route and inserting it into another route. The
insertion movement is implemented for each route Rj, where
all feasible customers will be tested in all positions of another
route Rk chosen through a candidate list, similar to the process
described in the route elimination and combination procedure.
We begin with the longest route Rj, which is the route with the
maximum length, and for each route Rj chosen, we begin with
the customer i farthest from the depot. We select the best inser-
tion position of customer i into route Rk. The customer is only
inserted if the total cost is reduced and all constraints are
respected.
2. Route addition. Similar to the demand reallocation, we consider
all customers i whose demand is split into more than one route.
We begin with customer i located farthest from the depot. For
each customer i, we choose the route Rj that delivers the small-
est quantity for customer i. In the next step, we relocate the
demand of customer i from route Rj to a new route Rk and cal-
culate the reduction or addition to the total cost after this
movement. This process is repeated for the other routes in
which customer i is inserted. The demand is added to the actual
route Rk while space exists and the remaining demand is added
to a new vehicle. In the end, the best combination is chosen in
terms of savings that may exist.
5.3.3. Reference set update method
Solutions are included in the reference set due to either quality
or diversity. The subset of quality solutions (RefSet1) contains the b1
best solutions, and the subset of diverse solutions (RefSet2) contains
the b2 diverse solutions. The initial RefSet consists of b1 best solu-
tions that belong to P and b2 elements from P that maximize the
minimum distance to RefSet. The distance between two solutions
is calculated by adding the number of non-common arcs of each
solution that exist before the combination. If an arc belongs to more
than one route, we add one unit for each non-common arc. We con-
sider arc xi0 or x0i only for routes with a full truckload (0  i  0).
Two main aspects must be considered when updating the Ref-
Set. The first aspect refers to the timing of the update, and two pos-
sibilities exist: a static update (S) and a dynamic update (D). The
second aspect addresses choosing the criteria for adding to and
deleting elements from the RefSet. We consider two types of up-
dates, by quality (Q) and by quality and diversity (QD).
In the static update (S), the reference set does not change until
all solution combinations for the RefSet are completed. In the dy-
namic update (D), the reference set is updated when a new and
better solution is found.
In the quality update (Q), if a new solution is better than the
worst element of RefSet1, then the worst element of RefSet1 is re-
placed by the new solution. In the quality and diversity update
(QD), if a new solution is better than the worst element of RefSet1,
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

In sets R1 and R2, the customer position is created randomly Table 1

through a uniform distribution. In sets C1 and C2, the customers Results of the scatter search for Solomon (1987). Source: Marius M. Solomon http://
web.cba.neu.edu/~msolomon/problems.htm.
are divided into groups. In sets RC1 and RC2, the customers are
grouped in sub-groups, i.e., a portion of the customers is placed Problem Best result Scatter search
randomly and another portion is placed in groups. Moreover, the Distance Vehicles Distance Vehicles % Difference
R1, C1 and RC1 problems have a short-term planning horizon R101 1607.70 18 1651.84 19 –
and, when combined with lighter capacity vehicles, they allow only R102 1434.00 17 1495.83 17 4.31
certain customers (3–8) in each route. Sets R2, C2 and RC2 have a R103 1292.68 13 1302.41 14 –
long term planning horizon and because they contain higher R104 1007.31 9 989.32 10 –
R105 1377.11 14 1450.27 14 5.31
capacity vehicles, they are able to supply more than 10 customers
R106 1251.98 12 1283.75 12 2.54
per route. R107 1104.66 10 1134.96 10 2.74
Solomon’s problem set consists of 56 Euclidean examples, each R108 960.88 9 993.58 9 3.40
with 100 customers. The travel times between customers are taken R109 1194.73 11 1161.93 12 –
R110 1118.59 10 1172.47 10 4.82
to equal the corresponding distances. Furthermore, a homoge-
R111 1096.72 10 1076.48 11 –
neous fleet is assumed. R112 982.14 9 970.06 10 –
Average 1202.37 11.83 1223.58 12.33 3.85

6.1.1. Experiments with Solomon’s problems C101 827.30 10 828.94 10 0.20

C102 827.30 10 838.90 10 1.40
The proposed meta-heuristic was applied to solve Solomon’s
C103 826.30 10 832.48 10 0.75
problem set (1987) with the aim of evaluating its performance C104 822.90 10 840.65 10 2.16
against other methods. For this set of problems, the fleet is as- C105 827.30 10 841.34 10 1.70
sumed to be homogeneous, and the demand cannot be split. There- C106 827.30 10 839.63 10 1.49
fore, the algorithm was simplified to solve the classical benchmark C107 827.30 10 828.94 10 0.20
C108 827.30 10 828.94 10 0.20
for VRPTW problems. Average 826.63 10 834.98 10 1.01
The best result was obtained with PSize = 30, b1 = 3, b2 = 7, and
RC101 1696.94 14 1640.39 15 –
dynamic (D) and quality (Q) updates. RC102 1554.75 12 1491.12 13 –
Table 1 compares the results of the algorithm presented in this RC103 1261.67 11 1307.88 11 3.66
paper with the best results found in the literature according to the RC104 1135.48 10 1170.81 10 3.11
distance and number of vehicles for each one of the instances of RC105 1629.44 13 1678.41 13 3.01
RC106 1424.73 11 1409.81 12 –
Solomon’s problems (1987).
RC107 1230.48 11 1268.34 11 3.08
According to Table 1, we note that the scatter search has RC108 1139.82 10 1139.82 10 0
achieved the best results found in the literature for five problems. Average 1325.52 11.5 1388.32 11.88 2.57
Moreover, in the other 10 problems, the algorithm has reduced the R201 1252.37 4 1283.41 4 2.48
total distance traveled, but the use of one extra vehicle was re- R202 1191.70 3 1237.12 3 3.81
quired. It has been verified that problems R1, C1 and RC1 have R203 939.54 3 980.62 3 4.37
shown better results than classes R2, C2 and RC2, on average. In R204 825.52 2 855.79 2 3.67
R205 994.42 3 1045.69 3 5.16
addition, the problems from class C have shown better results on R206 906.14 3 958.24 3 5.75
average than the problems from classes R and RC. The worst result R207 890.61 2 937.40 2 5.25
was presented by the problems of class R. We can conclude, there- R208 726.75 2 756.99 2 1.41
fore, that the algorithm appears better adapted to a short-term R209 909.16 3 954.77 3 5.02
R210 939.34 3 971.73 3 3.45
horizon, routes with few customers, small vehicles and customers
R211 892.71 2 833.45 3 –
that are clustered and semi-clustered geographically. Average 951.66 2.73 964.97 2.82 4.31
To investigate whether the performance of the scatter-search
C201 591.56 3 599.37 33 1.32
algorithm and the best results of Solomon’s benchmark problems C202 591.56 3 591.56 3 0
are different with respect to distance and number of vehicles, a C203 591.17 3 591.17 3 0
T-test was performed. The null hypothesis of the T-test states that C204 590.60 3 614.55 3 4.06
there is no difference between the two population means. The re- C205 588.88 3 599.08 3 1.73
C206 588.49 3 597.32 3 1.50
sults of the T-test are shown in Table 2. The null hypothesis is re-
C207 588.29 3 588.29 3 0
jected at a 5% level of significance, concluding that a difference C208 588.32 3 588.32 3 0
exists between the distance and vehicle means. Average 589.86 3 596.13 3 1.08
RC201 1406.91 4 1454.62 4 3.39
RC202 1365.65 3 1370.71 3 0.37
6.2. Solomon’s problems with augmented demands RC203 1049.62 3 1053.96 3 0.41
RC204 798.41 3 833.89 3 4.44
Ho and Haugland (2004) changed the demands from those of RC205 1297.19 4 1337.81 4 3.13
Solomon’s problems, with the aim of allowing more than one deliv- RC206 1146.32 3 1127.44 4 –
RC207 1061.14 3 1045.55 4 –
ery per customer. We consider m as the vehicle capacity and wi as
RC208 828.14 3 886.75 3 7.08
customer i’s demand (i = 1, . . . , n). Each demand is recalculated Average 1119.17 3.25 1138.84 3.5 3.14
within the period [lm, um], where l < u are defined within the per-
iod [0, 1]. Therefore, for every i 2 C, we define a new demand
w0i ¼ lm þ mððu  lÞ=ðw  wÞÞðwi  wÞ, where w ¼ minfwi : i 2 Cg 6.2.1. Experiments on Solomon’s problems with augmented demands
and w  ¼ maxfwi : i 2 Cg. The new demand w0i is rounded to the For this group of problems, the fleet is considered as homoge-
closest integer number. neous. Therefore, the proposed algorithm was simplified to solve
For each problem set, the authors use the following values of the Ho and Haugland problem set.
[l, u] (see Table 3). Initially, many different values for b1 and b2 were tested. The
Using this method, sets with a total of 224 problems were other parameters considered were: PSize = 30, criteria updating
created. through quality and diversity (QD) and use of a static updating
 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601 597

Table 2
Results of T-test comparing the means of Table 1.

Paired samples test Paired differences

95% Confidence interval of the difference
t Sig. (2-tailed)
Pair 1 Solomon_Benchmark_Distance – Solomon_ScatterSearch_Distance 4.016 .000
Pair 2 Solomon_Benchmark_Vehicle – Solomon_ScatterSearch_Vehicle 3.882 .000

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].

Parameters Distanceb Vehiclesc

PSize Timea Update
10 257 QD Static 32238.55 493.01
Dynamic 32100.66 492.27
Q Static 32249.47 493.40
Dynamic 32224.90 492.14
20 487 QD Static 32005.57 490.31
Dynamic 31919.37 490.88
Q Static 32014.78 491.34
Dynamic 31876.33 489.67
30 711 QD Static 31487.67 485.08
Dynamic 31709.63 487.03
Q Static 31641.44 489.21
Dynamic 31530.74 486.12
40 1163 QD Static 31480.66 485.27
Dynamic 31490.85 485.05
Q Static 31601.23 486.11
Dynamic 31597.48 485.20
a
Average processing time (seconds) of Solomon’s 56 problems.
b
Sum of the average values (average distance traveled) of the six classes of Solomon’s problems.
c
Sum of the average values (average of vehicles) of the six classes of Solomon’s problems.
598 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601

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.

Paired samples test Paired differences

95% Confidence interval of the difference
t Sig. (2-tailed)
Pair 1 HoHaugland_ Distance – ScatterSearch_Distance 1.658 .111
Pair 2 HoHaugland_ Vehicle – ScatterSearch_Vehicle 4.708 .000

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.

Parameters Valuesb Valuesc

a
PSize Time Update
10 296 QD Static 3047.90 1323.22
Dynamic 3053.66 1315.84
Q Static 3059.71 1323.69
Dynamic 3050.37 1320.35
20 465 QD Static 3022.98 1311.77
Dynamic 3015.74 1309.62
Q Static 3031.34 1318.43
Dynamic 3023.96 1305.92
30 687 QD Static 2998.57 1299.45
Dynamic 2983.00 1288.20
Q Static 3004.85 1310.51
Dynamic 2995.53 1302.70
40 911 QD Static 2989.41 1297.43
Dynamic 2985.27 1290.76
Q Static 3011.64 1297.63
Dynamic 3018.18 1285.88
a
Average processing time (seconds) of the 168 problems.
b
Average value considering the total fixed cost of the used vehicles and the total schedule time (excluding the service time of the customers).
c
Average value considering only the total schedule time (excluding the total fixed cost of the used vehicles and the service time of the customers).

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
these problems contains 100 customers, and the travel time be-
tween nodes is equal to the Euclidean distance. The 56 test prob-
lems are grouped into six problem types. The geographical data
are either randomly generated according to a uniform distribution
(problem sets R1 and R2), clustered in groups (problem sets C1 and
C2), or semi-clustered (problem sets RC1 and RC2). The semi-clus-
tered problem contains a mix of randomly generated data and clus-
ters. Furthermore, sets R1, C1 and RC1 have a shorter scheduling
horizon and allow fewer customers per route. Problem sets R2,
C2 and RC2 have a longer scheduling horizon and allow a larger
number of customers per route.
Liu and Shen (1999) adapted Solomon’s (1987) problem set to a
heterogeneous fleet, adding the costs for different types of vehicles
(see Appendix A). Three different sets of vehicle costs are consid-
ered for each problem of Solomon. Therefore, a total of 168 prob-
lems were solved. For example, the notation R1a, R1b and R1c
refer to the original R1 set with a larger vehicle cost, a medium
vehicle cost, and a smaller vehicle cost, respectively.
6.3.1. Experiments on Liu and Shen’s problem set (1999)
The results of the SS algorithm are compared with the heuristics
proposed by Liu and Shen (1999) and Dullaert et al. (2002). The re-
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.
Paired samples test
Pair 1 LiuShen_Value – ScatterSearchLS_Value
Pair 2 Dullaertetal_Value – ScatterSearchD_Value
Shen’s solution value considers the vehicle fixed costs and the total
schedule time and excludes the service time for customers, which
is constant. Dullaert et al.’s solution cost considers only the total
schedule time, excluding the service time of the customers and
the vehicle fixed costs.
First, we tested different parameters of the SS algorithm with
respect to both the quality of the solutions found and the comput-
ing times for the following: b1, b2 values (3, 2; 5, 5; 7, 3; 3, 7), PSize
values (10, 20, 30, 40), the first aspect of the update (static or dy-
namic), and the second aspect of the update (quality or quality
and diversity). The best result was obtained with PSize = 30,
b1 = 7, b2 = 3, dynamic update (D) and quality and diversity update
(QD). The results are detailed in Tables 9 and 10. The average value
represents the average results over all 168 problems.
Table 11 compares our algorithm with that of Liu and Shen
(1999). Liu and Shen (1999) adapted the savings concept of Golden
et al. (1984), generating three new heuristics: the MCS (modified
combined savings), the MOOS (modified optimistic opportunity
savings) and the MROS (modified realistic opportunity savings).
An improvement phase is applied to the constructive heuristics
generated. Table 11 presents only the best result (MCS, MOOS or
MROS with improvement) for each group of instances (i.e.,
R1a, R1b, R1c, . . . , RC2c), giving the average solution value of Liu
and Shen’s algorithm and of our scatter-search algorithm as well
as the percentage improvement we obtained. Table 11 shows that
our scatter-search algorithm outperforms that of Liu and Shen in
terms of solution quality. The percentage gap is reduced to 4.02%
on average.
Finally, Table 12 compares our algorithm with that of Dullaert
et al. (2002); however, the authors do not give computing times
for their constructive heuristic. For each group of instances (i.e.,
R1a, R1b, R1c, . . . , RC2c), Table 12 compares the average solution
value of Dullaert et al.’s heuristic with that of our scatter-search
approach, including the percentage gap we obtained. Table 12
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.
To determine whether the performance of the scatter-search
algorithm differs significantly from that of Liu and Shen’s algo-
rithm (Table 11) and Dullaert et al.’s algorithm (Table 12), the T-
test was applied once again, and the results are shown in Table 13.
We can conclude that the results achieved by the scatter-search
algorithm are significantly better than those of Liu and Shen’s
and Dullaert et al.’s algorithms.
7. Conclusions and future research
We have developed a scatter-search algorithm for the Fleet Size
and Mixed Vehicle Routing Problem with Time Windows and Split
Deliveries (FSMVRPTWSD). Two constructive heuristics were pro-
posed as initial solutions. The scatter-search framework provided
a means to combine solutions as well as to diversify and intensify
the meta-heuristic search process.
The model was adapted to solve three varieties of benchmark
examples available in the literature: benchmark problems by Liu
Paired differences
95% Confidence interval of the difference
t Sig. (2-tailed)
3.825 .001
14.882 .000
and Shen (1999) for the FSMVRP, benchmark problems by Ho
and Haugland (2004) for the VRPSD, and benchmark problems by
Solomon (1987) for the VRPTW.
For the three problems sets, many parameters of the scatter-
search meta-heuristic were tested: PSize, b1, b2, the updating crite-
ria and the updating frequency. We conclude that the calibration of
these parameters influences the performance of the algorithm.
Considering Solomon’s problem sets, for certain classes, the
scatter-search algorithm has matched the best results found in
the literature. Problems R1, C1 and RC1 have produced better re-
sults than classes R2, C2 and RC2. Moreover, class C problems have
presented better results on average than classes R and RC. The
worst result was presented by class R problems. We conclude,
therefore, that the algorithm adapts better to a short-term horizon,
routes with fewer customers, small vehicles and geographically
clustered and semi-clustered customers.
Computational testing revealed that our algorithm matched a
subset of the best solutions for Solomon’s problem set with aug-
mented demands. It has also been verified that the best results
were found for [l, u] = [0.50, 1.00]e[0.70, 1.00] because the higher
the demand values, the higher the possibility of split delivery. In
terms of computational time, our algorithm produced values that
were significantly lower than those proposed by Ho and Haugland
(2004).
Considering Liu and Shen’s examples, our approach outper-
formed both Liu and Shen’s algorithm and Dullaert et al.’s heuristic
in terms of solution quality.
In future research, other constructive heuristics will be applied
as initial solutions, and other improvement heuristics will also be
tested. Moreover, the scatter-search meta-heuristic developed
can be adapted for other problems, such as VRP with multiple de-
pots and pickup and delivery problems, among others. Finally, a
new set of problems could be created containing all characteristics
from the problem studied.
References
Alegre, J., Laguna, M., & Pacheco, J. (2007). Optimizing the periodic pick-up of raw
materials for a manufacturer of auto parts. European Journal of Operational
Research, 179, 736–746.
Archetti, C., Savelsbergh, M. W. P., & Speranza, M. G. (2006). Worst-case analysis of
split delivery routing problems. Transportation Science, 40, 226–234.
Archetti, C., Speranza, M. G., & Hertz, A. (2006). A tabu search algorithm for the split
delivery vehicle routing problem. Transportation Science, 40, 64–73.
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
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.
Brandão, J., & Mercer, A. (1997). A tabu search algorithm for the multi-trip vehicle
routing and scheduling problem. European Journal of the Operational Research,
100, 180–191.
Clarke, G., & Wright, J. W. (1964). Scheduling of vehicles from a central depot to a
number of delivery points. Operations Research, 12, 568–581.
Corberán, A., Fernández, E., Laguna, M., & Martí, R. (2002). Heuristic solutions to the
problem of routing school buses with multiple objectives. Journal of the
Operational Research Society, 53, 427–435.
Desrochers, M., & Verhoog, T. W. (1991). A new heuristic for the fleet size and mix
vehicle routing problem. Computers & Operations Research, 18, 263–274.
 P. Belfiore, H.T.Y. Yoshizaki / Computers & Industrial Engineering 64 (2013) 589–601
Dror, M., Laporte, G., & Trudeau, P. (1994). Vehicle routing with split deliveries.
Discrete Applied Mathematics, 50, 229–254.
Dror, M., & Trudeau, P. (1989). Savings by split delivery routing. Transportation
Science, 23, 141–145.
Dror, M., & Trudeau, P. (1990). Split delivery routing. Naval Research Logistics, 37,
383–402.
Dullaert, W., Janssens, G. K., Sorensen, K., & Vernimmen, B. (2002). New heuristics
for the fleet size and mix vehicle routing problem with time windows. Journal of
Operational Research Society, 53, 1232–1238.
Frizzell, P. W., & Giffin, J. W. (1992). The bounded split delivery vehicle routing
problem with grid network distances. Asia Pacific Journal of Operational Research,
9, 101–106.
Frizzell, P. W., & Giffin, J. W. (1995). The split delivery vehicle scheduling problem
with time windows and grid network distances. Computers & Operations
Research, 6, 655–667.
Gendreau, M., Laporte, G., Musaraganyi, C., & Taillard, E. D. (1999). A tabu search
heuristic for the heterogeneous fleet vehicle routing problem. Computers &
Operations Research, 26, 1153–1173.
Gheysens, F., Golden, B., & Assad, A. (1984). A comparison of techniques for solving
the fleet size and mix vehicle routing problem. Operations Research Spectrum, 6,
207–216.
Gheysens, F., Golden, B., & Assad, A. (1986). A new heuristic for determining fleet
size and composition. Mathematical Programming Study, 26, 233–236.
Glover, F. (1977). Heuristics for integer programming using surrogate constraints.
Decision Sciences, 8, 156–166.
Glover, F. (1998). A template for scatter search and path relinking. In J.-K. Hao, E.
Lutton, E. Ronald, M. Schoenauer, & D. Snyers (Eds.), Artificial evolution. Lecture
notes in computer science (vol. 1363, pp. 13–54). Springer-Verlag.
Golden, B. L., Assad, A., Levy, L., & Gheysens, F. (1984). The fleet size and mix vehicle
routing problem. Computers & Operations Research, 11, 49–65.
Ho, S. C., & Haugland, D. (2004). A tabu search heuristic for the vehicle routing
problem with time windows and split deliveries. Computers & Operations
Research, 31, 1947–1964.
Lenstra, J. K., & Rinnoy Kan, A. H. G. (1981). Complexity of Vehicle and Scheduling
Problems. Networks, 11, 221–227.
601
Li, X., Tian, P., & Aneja, Y. P. (2010). An adaptive memory programming
metaheuristic for the heterogeneous fixed fleet vehicle routing problem.
Transportation Research Part E: Logistics & Transportation Review, 46, 1111–1127.
Liu, F. H., & Shen, S. Y. (1999). The fleet size and mix vehicle routing problem with
time windows. Journal of Operational Research Society, 50, 721–732.
Martí, R., Laguna, M., & Glover, F. (2006). Principles of scatter search. European
Journal of Operational Research, 169, 359–372.
Mullaseril, P. A., Dror, M., & Leung, J. (1997). Split-delivery routing heuristics in
livestock feed distribution. Journal of the Operational Research Society, 48,
107–116.
Rego, C., & Leão, P. (2000). A scatter search tutorial for graph-based permutation
problems. Hearin Center for Enterprise, University of Mississippi, HCES-10-00,
USA. In C. Rego & B. Alidaee (Eds.), Metaheuristic optimization via memory and
evolution: Tabu search and scatter search, 2005 (pp. 1–24). Kluwer Academic
Publishers.
Rochat, Y., & Semet, F. (1994). A tabu search approach for delivering pet food and
flour in Switzerland. Journal of the Operational Research Society, 45, 1233–1246.
Rochat, Y., & Taillard, E. D. (1995). Probabilistic diversification and intensification in
local search for vehicle routing. Journal of Heuristics, 1, 147–167.
Salhi, S., & Rand, G. K. (1993). Incorporating vehicle routing into the vehicle fleet
composition problem. European Journal of Operational Research, 66, 313–330.
Solomon, M. M. (1987). Algorithms for the vehicle routing and scheduling problems
with time windows constraints. Operations Research, 35, 254–265.
Solomon, M. M., & Desrosiers, J. (1988). Time window constrained routing and
scheduling problem. Transportation Science, 22, 1–13.
Taillard, É. D. (1999). A heuristic column generation method for the heterogeneous
fleet vehicle routing problem. Rairo, 33, 1–14.
Tarantilis, C. D., Kiranoudis, C. T., & Vassiliadis, V. S. (2004). A threshold accepting
metaheuristic for the heterogeneous fixed fleet vehicle routing problem.
European Journal of Operational Research, 152, 148–158.
Wassan, N. A., & Osman, I. H. (2002). Tabu search variants for the mix fleet vehicle
routing problem. Journal of the Operational Research Society, 53, 768–782.
Yamashita, D. S., Armentano, V. A., & Laguna, M. (2006). Scatter search for project
scheduling with resource availability cost. European Journal of Operational
Research, 169, 623–637.