This article was downloaded by: [142.150.190.

39] On: 16 April 2017, At: 19:39

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Fifty Years of Vehicle Routing

Gilbert Laporte,

To cite this article:

Gilbert Laporte, (2009) Fifty Years of Vehicle Routing. Transportation Science 43(4):408-416. http://dx.doi.org/10.1287/
trsc.1090.0301

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2009, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 informs ®
Vol. 43, No. 4, November 2009, pp. 408–416
issn 0041-1655
eissn 1526-5447
09
4304
0408 doi 10.1287/trsc.1090.0301
© 2009 INFORMS

Fifty Years of Vehicle Routing

Gilbert Laporte
CIRRELT, Distribution Management, HEC Montréal, 3000, Montréal, Québec H3T 2A7, Canada
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

gilbert@crt.umontreal.ca

T he Vehicle Routing Problem (VRP) was introduced 50 years ago by Dantzig and Ramser under the title “The
Truck Dispatching Problem.” The study of the VRP has given rise to major developments in the fields of
exact algorithms and heuristics. In particular, highly sophisticated exact mathematical programming decompo-
sition algorithms and powerful metaheuristics for the VRP have been put forward in recent years. The purpose
of this article is to provide a brief account of this development.
Key words: vehicle routing problem; traveling salesman problem; exact algorithms; heuristics; metaheuristics;
survey
History: Received: August 2009; revision received: September 2009; accepted: September 2009. Published
online in Articles in Advance October 21, 2009.

1. Introduction set E = i j i j ∈ V  i < j

. In what follows we con-
The year 2009 marks the 50th anniversary of first centrate on the symmetric case, but some formula-
paper published on the vehicle routing problem (VRP), tions associate variables with arcs and some heuristics
under the title “The Truck Dispatching Problem” apply indiscriminately to both the symmetric and the
(Dantzig and Ramser 1959). To commemorate this asymmetric cases. For a review of exact algorithms for
event we highlight the main contributions in the his- the asymmetric VRP, see Toth and Vigo (2002b).
tory of this important problem. There exists a rich scientific literature on the VRP.
The VRP can be simply defined as the problem of For books and survey articles, the reader is referred
designing least-cost delivery routes from a depot to to Toth and Vigo (2002a), Golden, Raghavan, and
a set of geographically scattered customers, subject Wasil (2008), Laporte and Nobert (1987), Fisher (1995),
Laporte and Semet (2002), Gendreau, Laporte, and
to side constraints. This problem is central to dis-
Potvin (2002), Cordeau et al. (2005, 2007), and Laporte
tribution management and must be routinely solved
(2007). The two books by Toth and Vigo (2002a) and
by carriers. In practice several variants of the prob-
the book by Golden, Raghavan, and Wasil (2008) also
lem exist because of the diversity of operating rules
contain a wealth of information on the most common
and constraints encountered in real-life applications. variants of the VRP.
Thus the VRP should perhaps be viewed as a class of The VRP generalizes the well-known traveling sales-
problems. However, it is useful from a methodolog- man problem (TSP) but is much more difficult to solve
ical point of view to work on archetypal versions of in practice. Whereas there exist exact algorithms capa-
the problem. Here we focus on the classical VRP for- ble of routinely solving TSPs containing hundreds or
mally defined as follows. Let G = V  A be a directed thousands of vertices (Applegate et al. 2007), this is not
graph where V = 0     n
is the vertex set and A = the case of the VRP for which the best exact algorithms
i j i j ∈ V  i = j
is the arc set. Vertex 0 represents can only solve instances involving approximately 100
the depot whereas the remaining vertices correspond vertices (Fukasawa et al. 2006; Baldacci, Christofides,
to customers. A fleet of m identical vehicles of capac- and Mingozzi 2008). Because real instances of the VRP
ity Q is based at the depot. The fleet size is given a often exceed this size and solutions must often be
priori or is a decision variable. Each customer i has a determined quickly, most algorithms used in practice
nonnegative demand qi . A cost matrix cij is defined on are heuristics. In recent years, several powerful meta-
A. For simplicity, we consider travel costs, distances heuristics have been developed.
and travel times to be equivalent. The VRP consists In the following three sections of this paper some
of designing m vehicle routes such that each route of the main exact algorithms, classical heuristics, and
starts and ends at the depot, each customer is visited metaheuristics for the VRP will be reviewed. This will
exactly once by a single vehicle, the total demand of be followed by some conclusions.
a route does not exceed Q, and the total length of a
route does not exceed a preset limit L. 2. Exact Algorithms
In the symmetric case, i.e., when cij = cji for all Over the past 40 years exact algorithms for the VRP
i j ∈ A, it is customary to work with an edge have evolved from basic branch-and-bound schemes
408
 Laporte: Fifty Years of Vehicle Routing
Transportation Science 43(4), pp. 408–416, © 2009 INFORMS 409

to highly sophisticated mathematical programming a lower bound on the contribution made by i to the
engineering pieces. In what follows we retrace the routing cost. This bound can be improved through an
main steps of this development. ascent procedure.
Christofides, Mingozzi, and Toth (1981a) have suc-
2.1. Branch-and-Bound Algorithms cessfully solved instances with 10 ≤ n ≤ 25 with these
One of the first known branch-and-bound algorithms two lower bounds. The bound derived from q-routes
for the VRP appeared in “An Algorithm for the is generally better than the k-DCT bound. Using the
Vehicle Dispatching Problem” (Christofides and Eilon
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

same bounds, Hadjiconstantinou, Christofides, and

1969), heralding a slight departure from the original Mingozzi (1995) have developed an improved branch-
Dantzig and Ramser (1959) appellation. In this algo- and-bound algorithm capable of optimally solving
rithm, m is an input parameter. The graph is first instances with n ≤ 50. Later, Fisher (1994) later incor-
extended by adding m − 1 artificial depots to the porated the k-DCT lower bound within a branch-and-
graph and setting the inter-depot arc costs equal to cut algorithm for a restricted version of the VRP in
infinity. An m-TSP is then solved on this graph by which return trips between the depot and a customer
branching on arcs, as in the Little et al. (1963) TSP are disallowed. He was able to successfully solve,
algorithm, except that a shortest spanning tree bound within a small tolerance, several instances containing
is computed at each node, instead of a relaxed assign- up to 134 vertices.
ment problem. The VRP side constraints are handled
through simple fathoming rules. An improvement 2.2. Dynamic Programming
to this algorithm, based on a modification of the Eilon, Watson-Gandy, and Christofides (1971) have
Carpaneto and Toth (1980) TSP algorithm was later formulated the VRP as a dynamic program as fol-
proposed by Laporte, Mercure, and Nobert (1986). lows. Let cS be the optimal cost of a single vehicle
The Christofides (1976) paper seems to be the first to route through the vertices of S ⊆ V \0
. The objec-
have used the name “vehicle routing problem.” The

tive is to minimize m r=1 > cSr  over all feasible parti-
branch-and-bound algorithm it describes branches on tions S1      Sm
of V \0
. Let fk U  be the least cost
routes rather than arcs. The resulting search tree is achievable using k vehicles and delivering to a subset
therefore rather wide but has a depth limited to m. U of V \0
. Then
Neither of these two algorithms was very successful 
in solving anything but rather small or easy instances. 
cU  k = 1
The performance of early branch-and-bound algo- fk U  = (1)
rithms was greatly enhanced by the derivation of 
 ∗ min fk−1 U \U ∗ +cU ∗ 
k > 1
U ⊆U ⊆V \0

two sharp lower bounds based on k-degree centre

trees (k-DCT) and q-routes (Christofides, Mingozzi, The solution cost is fm V \0
 and the optimal parti-
and Toth 1981a). The first computes a tree in which
tion corresponds to the optimizing subsets in (1).
the degree of vertex 0 is equal to k, where k = 2m − y
The state space can be reduced by using fea-
and y ≤ m. The set of all edges belonging to a solution
sibility or dominance criteria. A state-space relax-
can be partitioned into E1  E2  E3
, where E1 is the set
ation algorithm has also been applied by Christofides,
of edges forming a k-DCT, E2 is a set of y edges inci-
Mingozzi, and Toth (1981b). These authors report hav-
dent to 0, and E3 is a set of m − y edges not incident
ing solved instances with 10 ≤ n ≤ 25 using this tech-
to 0. The problem is formulated in terms of three sets
nique, but this avenue of research does not seem to
of variables associated with the three elements of the
have been pursued.
partition, with linking constraints. Lower bounds are
then computed for the subproblem defined by each
2.3. Vehicle Flow Formulations and Algorithms
variable type by incorporating the linking constraints
Two-index vehicle flow formulations for the VRP are
in the objective function in a Lagrangean fashion.
rooted in the work of Laporte and Nobert (1983) and
The second lower bound is based on the concept of
Laporte, Nobert, and Desrochers (1985) and extend
q-route put forward by Houck et al. (1980). Let W be
the classical TSP formulation of Dantzig, Fulkerson,
the ordered set of all possible demands of a potential
and Johnson (1954). Let xij be a 0–1–2 variable equal
vehicle route, starting with the smallest value. Let ql
to the number of times a vehicle travels on edge i j.
be the value of the lth element of W and let l i be
Then the problem is
the value of a least cost route (1) passing through i,
(2) starting and ending at the depot, (3) having no 
(VF) minimize cij xij (2)
two-vertex cycle, and (4) having a demand equal

ij∈E
to ql. Then ni=1 > minl=1 
W
> l iqi /ql
is a
valid lower bound on the cost of an optimal VRP solu- 
n
subject to x0j = 2m (3)
tion. This bound is the sum, over all customers i, of j=1
 Laporte: Fifty Years of Vehicle Routing
410 Transportation Science 43(4), pp. 408–416, © 2009 INFORMS

 
xik + xkj = 2 variables xij are equal to 1 if and only if edge i j is
i<k j>k used in the solution, variables yij represent the vehicle
k ∈ V \0
 (4) load on i j, and variables yji = Q − yij represent the
 empty vehicle space on i j whenever xij = 1.
xij ≤
S
−vS The formulation is
ij∈S

S ⊆ V \0
 (5) (CF) minimize cij xij (8)
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

ij∈E
x0j = 012 j ∈ V \0
 (6) 
subject to yji −yij  = 2qi i ∈ V \0
 (9)
xij = 01 ij ∈ V \0
 (7) j∈V
 
where vS is a lower bound on the number of vehi- y0j = qi  (10)
cles needed to service all vertices of S. Constraints (2) j∈V \0
i∈V \0

and (3) are degree constraints, whereas constraints (4)  

yj0 = mQ− qi  (11)
eliminate subtours and enforce the capacity and route j∈V \0
i∈V \0

length restrictions. In the presence of capacity con- 

yn+1j = mQ (12)
straints, vS is often set equal to  i∈S qi /Q or is
j∈V \0

computed by solving a bin packing problem with item

weights qi i ∈ S and bin size Q. Laporte, Nobert, and
yij +yji = Qxij ij ∈ E (13)
Desrochers (1985) show how to compute vS within a  
branch-and-cut scheme in the presence of route length xik + xkj = 2
i<k j>k
constraints.
The model defined by (VF) can be reinforced k ∈ V \0
 (14)
through the inclusion of valid inequalities. These yij ≥ 0yji ≥ 0
ij ∈ E (15)
include so-called generalized capacity constraints,
frame capacity constraints, VRP comb inequalities, xij = 01
ij ∈ E (16)
inequalities combining bin packing and the TSP, and
some inequalities based on the stable set problem. For Constraints (9)–(12) and (15) define consistent flows
an overview, see Naddef and Rinaldi (2002). These from 1 to n + 1, constraints (13) ensure that the
authors have developed a branch-and-cut algorithm yij and yji variables are feasible, and constraints (14)
enabling them to solve six instances 22 ≤ n ≤ 45 are degree constraints. This problem was solved by
without branching, and nine others 51 ≤ n ≤ 135 branch-and-cut, with valid VRP inequalities expressed
with some branching. For a more recent branch-and- in terms of the xij variables. Several instances with
cut algorithm based on the (VF) formulation, see 16 ≤ n ≤ 135 were solved optimally. The success rate is
Lysgaard, Letchford, and Eglese (2004). very high if 30 ≤ n ≤ 60 and 3 ≤ m ≤ 5.
A number of three-index vehicle flow formulations
have also been put forward. These make use of binary 2.5. Set Partitioning Formulations and Algorithms
variables xijk equal to 1 if and only if arc i j is tra- A straightforward set partitioning formulation of the
versed by vehicle k. They do not seem to have been VRP was first provided by Balinski and Quandt
as successful as two-index formulations. Examples are (1964). Let r denote a route, let air be a binary coeffi-
provided by Golden, Magnanti, and Nguyen (1977) cient equal to 1 if and only if vertex i ∈ V \0
belongs
and by Fisher and Jaikumar (1981). to route r, let c ∗ be the optimal cost of route r, and let
yk be a binary variable equal to 1 if and only if route
2.4. Commodity Flow Formulations and r is used in the optimal solution. The problem is then
Algorithms  ∗
In commodity flow formulations, variables yij (or yijk ) (SP) minimize cr yr (17)
r
define the vehicle load carried on arc i j. An early 
example can be found in Gavish and Graves (1979), subject to air = 1 i ∈ V \0
 (18)
but these authors report no computational result. r

A more recent example is the formulation of Baldacci, yr = m (19)
Hadjiconstantinou, and Mingozzi (2004) which is r
based on the TSP model of Finke, Claus, and Gunn yr = 0 1 all r (20)
(1984). This formulation works on an extended graph
G = V  E,
where V = V ∪n+1
, n+1 is a copy of the Strictly speaking, constraints (19) are not part of
depot, and E = E ∪ i n + 1 i ∈ V
. A vehicle route the standard set partitioning formulation, but they
is defined as a directed path from 0 to n + 1. Binary are used by most authors in the context of the VRP.
 Laporte: Fifty Years of Vehicle Routing
Transportation Science 43(4), pp. 408–416, © 2009 INFORMS 411

A direct application of this formulation is impracti- and generalized multistar inequalities, and hypo-
cal because of the large number of potential routes tour cuts. Embedding this master problem within a
encountered in most nontrivial instances and of the branch-and-cut-and-price algorithm, the authors were
difficulty of computing the cr∗ coefficients which capable of solving instances containing up to 135
requires solving an exponential number of instances vertices.
of an NP-hard problem. The second algorithm, by Baldacci, Christofides,
A number of column generation algorithms were and Mingozzi (2008), works with the (SP) formula-
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

proposed to solve this problem. The first, by Rao and tion, augmented with some inequalities valid for (VF).
Zionts (1968) does not seem to have been numerically This is made possible by using the following identity
tested. A second method, by Foster and Ryan (1976) proved in Baldacci, Hadjiconstantinou, and Mingozzi
generates routes by dynamic programming but was (2004): 
not run to completion. A full column generation algo- xij = aijr yr i j ∈ E (29)
rithm was developed by Agarwal, Mathur, and Salkin r
(1989) who solved instances with 15 ≤ n ≤ 25. where if r is the route 0 j 0, then a0jr = 2, and
Two of the most successful VRP algorithms, devel- aijr = 0 for all i j ∈ E\0 j
; if r contains at least
oped in recent years, make partial use of a set parti- two customer vertices, then aijr = 1 for each edge i j
tioning formulation. The first by Fukasawa et al. (2006)
of route r, and aijr = 0 otherwise.
works with the following set partitioning formulation
The (SP) formulation can be strengthened through
in which the columns correspond to q-routes. Let aijr
the introduction of valid inequalities. For S ⊆ V 0
,
be a binary coefficient equal to 1 if and only if edge
let RS be the set of routes containing at least one
i j appears in q-route r, yr , cr∗ are defined as in (SP),
customer of S. Then the capacity constraints
and xij as in (VF). Then the problem is
 ∗ 
(SP’) minimize cr yr (21) yr ≥ vS S ⊆ V \0
 (30)
r∈RS
r

subject to aijr yr −xij = 0 ij ∈ E (22) are valid. Also, any valid inequality for the VRP
r (Naddef and Rinaldi 2002; Letchford, Eglese, and
and (4) (6) (7) (19) and (20) Lysgaard 2002) of the form

Combining (VF) and (SP’) yields the following $ij xij ≥ % (31)
formulation: i j∈E
 
(VF-SP’) minimize cij aijr yr (23) can be expressed in terms of the yr variables by
r ij∈E
using (29). Finally valid inequalities for the Set Par-
 
subject to aijr yr = 2m (24) titioning Problem (Balas and Padberg 1976; Hoffman
r j∈V \0
and Padberg 1993) are also valid for (SP). Specifically,
  the authors use the clique inequalities. Consider a
  
aikr yr + ajkr yr = 2 graph H whose vertices correspond to vehicle routes;
r i<k j>k two vertices conflict whenever the routes they rep-
k ∈ V \0
 (25) resent share some customers. Then, for any clique C
  of H , the inequality
aijr yr ≥ 2vS
r

i ∈ S j  S yr ≤ 1 (32)
or i  S j ∈ S
r∈C
S ⊆ V \0
 (26)
 is valid.
aijr yr ≤ 1 Baldacci, Christofides, and Mingozzi (2008) use (SP)
r
together with some of constraints (30), (31), and (32).
ij ∈ E\0j j ∈ V \0

 (27) They compute lower bounds on the dual of the lin-

yr = 01 all r (28) ear relaxation of this problem, obtained by applying
three ascent heuristics. The final dual solution is used
A lower bound on the VRP objective is computed by to generate a reduced problem containing columns of
solving the linear relaxation of (VF-SP’) to which are reduced cost between the upper bound and the lower
added the following valid inequalities proposed by bound achieved. This problem is solved by an integer
Letchford, Eglese, and Lysgaard (2002) and Lysgaard, linear programming solver. This algorithm was capa-
Letchford, and Eglese (2004): framed capacity inequal- ble of solving instances with 37 ≤ n ≤ 121 and seems
ities, strengthened combs, multistar, partial multistar, to be slightly better than that of Fukasawa et al. (2006).
 Laporte: Fifty Years of Vehicle Routing
412 Transportation Science 43(4), pp. 408–416, © 2009 INFORMS

3. Classical Heuristics in the original Clarke and Wright algorithm. Solving

The seminal Dantzig and Ramser (1959) paper intro- matching problems is highly time consuming and not
duced the VRP to the research community and pre- worth the effort in comparison with other heuristics.
sented the first heuristic for the problem. Essentially, The last two enhancements are probably of little use
this algorithm iteratively matches vertices, or vertices these days given the state of the art in computer tech-
and partial routes, to form a set of vehicle routes. It nology and improvement heuristics.
does so by solving a sequence of linear programs with
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

variables xij and xji equal to 1 if and only if two points 3.2. Set Partitioning Heuristics
(vertices, partial routes) are matched. Each iteration is A simple heuristic scheme consists of solving the (SP)
called a “stage of aggregation.” The method is illus- formulation with a subset of promising vehicle routes,
trated on a small seven-vertex instance. often called “petals” in this context. This method nor-
Between 1964 and the early 1990s numerous heuris- mally assumes that the vertices are distributed on
tics were put forward. Some are purely constructive a plane. An early example of this methodology is the
but most also include an improvement phase. We call sweep algorithm of Gillett and Miller (1974) in which
these heuristics “classical” because they do not con- nonoverlapping petals are sequentially generated by
tain mechanisms allowing the objective function to rotating a half-line rooted at the depot, as long as the
deteriorate from one iteration to the next. This fea- capacity and route length constraints for a single route
ture is present in “metaheuristics” which have been are satisfied. More sophisticated schemes in which
developed primarily over the past twenty years. intersecting or embedded routes are allowed have
Heuristics are usually tested on two sets of bench- been put forward by Foster and Ryan (1976), Ryan,
mark instances. The first set, called CMT, was pro- Hjorring, and Glover (1993), and Renaud, Boctor, and
posed by Christofides, Mingozzi, and Toth (1979) and Laporte (1996). On the CMT testbed the latter method
contains 14 instances with 51 ≤ n ≤ 199. The second, has produced an average deviation of 2.38%.
called GWKC, was proposed by Golden et al. (1998)
and contains 20 instances with 200 ≤ n ≤ 480. It is cus- 3.3. Cluster-First, Route-Second Heuristics
tomary to compute optimality gaps with respect to The cluster-first, route-second heuristic of Fisher and
best-known solution values. We now outline some of Jaikumar (1981) first locates m seeds and constructs
the most important classical heuristics for the VRP. a cluster for each seed so as to minimize the sum
of customer-to-seed distances, while satisfying the
3.1. The Savings Algorithm capacity constraint. This is achieved by solving a gen-
The savings heuristic put forward by Clarke and eralized assignment problem (GAP). A route is then
Wright (1964) is easy to describe and to implement, determined on each cluster by solving a TSP. Some
and yields reasonably good solutions. This explains procedures for selecting the seeds are described in
its ongoing popularity. It starts with an initial solu- Baker and Sheasby (1999). It is not clear exactly how
tion made up of n back-and-forth routes 0 i 0 the route length constraint is handled within this
i ∈ V \0
. At each iteration, it merges a route end- method but some of the CMT instances solved by
ing with i with another route starting with j, max- Fisher and Jaikumar (1981) do contain a length con-
imizing the saving sij = ci0 + c0j − cij , and provided straint. Exact comparisons with other algorithms are
the merge is feasible. The process stops when it is no difficult to make because the distance rounding con-
longer possible to merge routes. The number of vehi- vention used in the experiments is not specified. This
cles used in the solution is an output of the algorithm. method is natural in that it follows the two phases
This process is normally followed by a three-opt pro- often applied by human dispatchers. It is also inter-
cedure applied to each route. Applying this algorithm esting from a methodological point of view because
to the CMT instances produced deviations of about it can benefit from algorithmic improvements for the
7% within insignificant computing times (Laporte and GAP or for the TSP.
Semet 2002).
Several enhancements to this algorithm have been 3.4. Improvement Heuristics
proposed, namely multiplying cij by a positive Two broad types of methods can be employed to
weight ) (Golden, Magnanti, and Nguyen 1977), opti- postoptimize a VRP solution. Intraroute moves con-
mizing the route merges in a global fashion through sist of improving each route separately by means of a
the use of a matching algorithm (Altinkemer and Gav- TSP algorithm, whereas interroute moves act on sev-
ish 1991, Wark and Holt 1994), accelerating the sav- eral routes simultaneously. It is common to alternate
ings computation (Paessens 1988), and making use of between these two schemes within the same improve-
efficient data structures to speed up the computations ment heuristic.
(Nelson et al. 1985). The first enhancement is useful Intraroute heuristics have been well documented in
in avoiding circumferential routes that tend to occur the TSP literature (Laporte 2009). Interroute methods
 Laporte: Fifty Years of Vehicle Routing
Transportation Science 43(4), pp. 408–416, © 2009 INFORMS 413

typically consist of removing one or several customers In tabu search the solution space is explored by
from a number of routes and relocating them. A large moving from the current solution to the best solution
number of algorithms have been proposed within this in a subset of its neighbourhood. To avoid cycling,
broad framework, most of which are special cases solutions possessing a given attribute of the current
of the b-cyclic, k-transfer scheme of Thompson and solution are not considered for a number of iterations
Psaraftis (1993), where a circular permutation of b (they are declared tabu). An exception is when such
routes is selected and k customers from each route a solution constitutes a new best solution among all
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

are moved to the next route of the permutation. The known solutions possessing the current attribute. This
combinations b = 2 or b variable and k = 1 or 2 seem principle was first formalized by Cordeau, Gendreau,
to produce good results. Kindervater and Savelsbergh and Laporte (1997) and is now known as attribute
(1997) have described efficient ways of handling edge based search (Derigs and Kaiser 2007). In simulated
exchanges, mostly within the context of the VRP with annealing, a solution x is randomly drawn from the
time windows. neighbourhood N xt  of the current solution xt at iter-
ation t. If the objective solution f is to be minimized,
then xt+1 = x whenever f xt+1  ≤ f xt . Otherwise,
4. Metaheuristics the search proceeds to xt+1 = x with probability pt
Metaheuristics can be broadly classified into local and to xt+1 = xt with probability 1 − pt , where pt
search, population search, and learning mechanisms. is a decreasing function of t and of f x − f xt . In
Most VRP metaheuristics are of the first kind but deterministic annealing, x is also randomly selected
there are interesting examples in the other two in N xt  and two different rules can be applied. In
categories. a threshold-accepting algorithm (Dueck and Scheurer
Most metaheuristics can be regarded as improve- 1990), xt+1 = x whenever f x < f xt  + .1 , where .1 is
ment methods. The best ones are rather robust and a positive tolerance; otherwise xt+1 = xt . In record-to-
perform very well even if they are initiated from a record travel (Dueck 1993), a record is the best known
low-quality solution. A general tendency has been to solution x∗ . Then xt+1 = x whenever f x ≤ .2 f x∗ ,
move from algorithms based on a single paradigm where .2 is a positive tolerance; otherwise xt+1 = xt .
to hybrid methods that draw on several principles. Variable neighbourhood search (Mladenović and
Thus, memetic algorithms (Moscato and Cotta 2003) Hansen 1997) works with an ordered list of neigh-
combine population search and local search, resulting bourhoods, which are usually nested. The algorithm
in a search trajectory that is at the same time broad starts with a given neighbourhood and switches to
and deep. the next neighbourhood in the list when it reaches
The number of variants of VRP metaheuristics pub- a local minimum. The search is reinitiated from the
lished in recent years and their level of intricacy make first neighbourhood whenever a new best solution
it rather difficult to provide a comprehensive account is identified or when all neighbourhoods have been
of the relevant literature. Instead, we will explain the explored. An application of variable neighbourhood
basic principles underlying each class of algorithms search to the VRP is described in Kytöjoki et al.
while pointing out the distinguishing features of the (2007). Very large-scale neighbourhood search simul-
most important ones. taneously destroys and reconstructs several parts of
the solution, usually by solving a network flow prob-
4.1. Local Search lem to identify the best neighbour. In a sense, the
Essentially, local search explores the solution space principle underlying this method is related to the
by moving at each iteration from the current solu- destroy and repair mechanism of Shaw (1998). This
tion to another solution in its neighbourhood. Clas- metaheuristic has been applied to the VRP by Ergun,
sical examples include tabu search (Glover 1986), Orlin, and Steele-Feldman (2006). Finally, in adap-
simulated annealing (Kirkpatrick, Gelatt, and Vecchi tive large neighbourhood search (Ropke and Pisinger
1983), deterministic annealing (Dueck and Scheurer 2006, Pisinger and Ropke 2007) several insertion and
1990; Dueck 1993), variable neighbourhood search removal heuristics are defined. At each iteration one
(Mladenović and Hansen 1997), very large neighbour- of these is randomly selected by giving it a weight
hood search (Ergun, Orlin, and Steele-Feldman 2006) proportional to its success rate in previous iterations.
and adaptive large neighbourhood search (Ropke and The local search mechanisms just outlined, and
Pisinger 2006). The main ingredients of local search others that could eventually be proposed, can be
are the rules employed to define the neighbourhood enhanced through various mechanisms. Parallel com-
of a solution and the mechanism put in place to puting, implemented in the tabu search heuristic of
explore it. Typically, local search heuristics perform Taillard (1993), is an obvious example. The granu-
inter-route moves, as described in §3.4, with occa- larity principle (Toth and Vigo 2003) is also of wide
sional intra-route reoptimizations. applicability. It consists of removing long edges from
 Laporte: Fifty Years of Vehicle Routing
414 Transportation Science 43(4), pp. 408–416, © 2009 INFORMS

the data (those that are unlikely to be part of an intelligence. They can learn from experience and
optimal solution) to ease the computational burden. incrementally adjust their weights in an iterative fash-
This idea was successfully applied by Li, Golden, ion. The limited application of this concept to the
and Wasil (2005) in conjunction with record-to-record VRP (Ghaziri 1991; Schumann and Retzko 1995) has
search. It is common to apply a diversification mech- met with mixed success. Ant colony optimization
anism within a local search. A popular rule is to is another form of learning mechanism. It mimics
penalize frequently performed moves in the objec- the behaviour of ants foraging for food and laying
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

tive function (Taillard 1993; Gendreau, Hertz, and pheromone on their paths. With time, pheromone
Laporte 1994; Cordeau, Gendreau, and Laporte 1997; accumulates faster on the shortest paths which are
Cordeau, Laporte, and Mercier 2001). Several local then followed by more ants. In an ant colony opti-
search schemes also allow the exploration of inter- mization algorithm, this idea translates into gradually
mediate infeasible solutions. This is done through giving more weight to the edges appearing frequently
the inclusion in the objective function of penalty in good solutions. An interesting and rather success-
terms weighted by self-adjusting positive parame- ful application of this type of algorithm is provided
ters (Gendreau, Hertz, and Laporte 1994; Cordeau, by Reimann, Doerner, and Hartl (2004).
Gendreau, and Laporte 1997; Cordeau, Laporte, and
Mercier 2001). Finally, limitation strategies can be 4.4. Computational Results
applied to restrict the number of moves consid- On the CMT instances (Christofides, Mingozzi, and
ered within a given neighbourhood structure (Nagata Toth 1979), the most accurate metaheuristics appear
and Bräysy 2008). For example, when applying a to be those of Mester and Bräysy (2005, 2007), Nagata
three-opt algorithm, several potential edge exchanges (2007), Nagata and Bräysy (2008, 2009), Taillard
are unpromising and can thus be disregarded. (1993), and Rochat and Taillard (1995). They all yield
average deviations from the best-known solution in
4.2. Population Search the 0.00% to 0.05% range. On the GWKC instances,
Population search works with a population of solu- these algorithms are also among the best, but the
tions. Genetic algorithms (Holland 1975) are the top-ten list also includes Prins (2009), Pisinger and
best-known example of this paradigm. At each iter- Ropke (2007), Reimann, Doerner, and Hartl (2004),
ation of a genetic algorithm, some parent solu- and Tarantilis (2005). Here the average deviations
tions are extracted from the current population and range from 0.01% to 0.76%. Computation times are
recombined to create offspring which then replace hard to compare because of the variety of computers
the worst elements of the population if this yields used for the experiments and because some heuristics
improvements. It is standard to apply a diversification are based on parallel computing. However, comput-
mechanism, called mutation, to the offspring before ing times are often very modest and much smaller
considering their inclusion in the population. than those of the early metaheuristics put forward 20
As far as we are aware, all known genetic algo- years ago. Many of the recent metaheuristics can solve
rithms applied to the VRP have been applied in con- instances involving more than 100 vertices within
junction with local search. This is usually achieved by 0.1% of the best-known solution value in less than a
improving the offspring through local search. Some minute.
fine examples are provided in Prins (2004, 2009),
Mester and Bräysy (2005, 2007), Nagata (2007), and 5. Conclusions
Nagata and Bräysy (2009). Much progress has been accomplished in the field of
Yet a different way to combine genetic search and vehicle routing in the past 50 years. The VRP is more
local search is to first apply local search and record popular than ever. It has attracted the attention of the
a set of best-known solutions in a memory. After the operations research community; it has been an engine
main phase of the search has ended, the solutions for the phenomenal growth we have witnessed in the
lying in the memory are recombined to create new fields of exact algorithms and heuristics; and some of
partial solutions, local search is reinitiated from these these algorithms have found their way into commer-
partial solutions, and the memory is updated. Rochat cial solvers used by industry.
and Taillard (1995) who have introduced this proce- In the area of exact algorithms, it is apparent that
dure, report excellent results with it. Other applica- the VRP is considerably more difficult to solve than
tions are described in Tarantilis and Kiranoudis (2002) the TSP. We have just about broken the 100 barrier
and in Tarantilis (2005). and it is difficult at this stage to predict how much
further we can go.
4.3. Learning Mechanisms The VRP has also been a major source of motiva-
Learning mechanisms include neural networks which tion for the development of heuristics. For all practical
are derived from concepts borrowed from artificial purposes the classical VRP can be solved adequately
 Laporte: Fifty Years of Vehicle Routing
Transportation Science 43(4), pp. 408–416, © 2009 INFORMS 415

for realistic instance sizes. There is, however, a sense Cordeau, J.-F., G. Laporte, M. W. P. Savelsbergh, D. Vigo. 2007. Vehi-
that several of the most successful metaheuristics are cle routing. C. Barnhart, G. Laporte, eds. Transportation, Hand-
books in Operations Research and Management Science, Vol. 14.
over-engineered and one should now attempt to pro- Elsevier, Amsterdam, 367–428.
duce simple and more flexible algorithms capable of Cordeau, J.-F., M. Gendreau, A. Hertz, G. Laporte, J.-S. Sormany.
handling a larger variety of constraints, even if this 2005. New heuristics for the vehicle routing problem.
were to translate into a small loss in accuracy. On this A. Langevin, D. Riopel, eds. Logistics Systems and Optimization.
Springer, New York, 279–297.
topic, see Cordeau et al. (2002). In addition, time has Cordeau, J.-F., M. Gendreau, G. Laporte, J.-Y. Potvin, F. Semet.
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

probably come to develop algorithms better able to 2002. A guide to vehicle routing heuristics. J. Oper. Res. Soc. 53
incorporate dynamic and stochastic features that are 512–522.
Dantzig, G. B., J. H. Ramser. 1959. The truck dispatching problem.
so common in practice.
Management Sci. 6 80–91.
Dantzig, G. B., D. R. Fulkerson, S. M. Johnson. 1954. Solution of a
Acknowledgments large-scale travelling-salesman problem. Oper. Res. 2 393–410.
This work was partly supported by the Canadian Natu- Derigs, U., R. Kaiser. 2007. Applying the attribute based hill climber
ral Sciences and Engineering Research Council under Grant heuristic to the vehicle routing problem. Eur. J. Oper. Res. 177
719–732.
05-39682; that support is gratefully acknowledged. Thanks
Dueck, G. 1993. New optimization heuristics: The great deluge
are due to a referee who provided valuable comments on algorithm and the record-to-record travel. J. Computational
an earlier version of this paper. Phys. 104 86–92.
Dueck, G., T. Scheurer. 1990. Threshold accepting: A general pur-
pose optimization algorithm. J. Computational Phys. 90 161–175.
References Eilon, S., C. D. T. Watson-Grandy, N. Christofides. 1971. Distribu-
tion Management: Mathematical Modelling and Practical Analysis.
Agarwal, Y., K. Mathur, H. M. Salkin. 1989. A set-partitioning-based Griffin, London.
exact algorithm for the vehicle routing problem. Networks 19 Ergun, O., J. B. Orlin, A. Steele-Feldman. 2006. Creating very large-
731–749. scale neighbourhoods out of smaller ones by compounding
Altinkemer, K., B. Gavish. 1991. Parallel savings-based heuristic for moves. J. Heuristics 12 115–140.
the delivery problem. Oper. Res. 39 456–469. Finke, G. A., A. Claus, E. Gunn. 1984. A two-commodity network
Applegate, D. L., R. E. Bixby, V. Chvátal, W. J. Cook. 2007. The flow approach to the traveling salesman problem. Congressus
Traveling Salesman Problem. A Computational Study. Princeton Numerantium 41 167–178.
University Press, Princeton, NJ. Fisher, M. L. 1994. Optimal solution of vehicle routing problems
Baker, B. M., J. Sheasby. 1999. Extensions to the generalised assign- using minimum K-trees. Oper. Res. 42 626–642.
ment heuristic for vehicle routing. Eur. J. Oper. Res. 119 147–157. Fisher, M. L. 1995. Vehicle routing. M. O. Ball, T. L. Magnanti, C. L.
Balas, E., M. W. Padberg. 1976. Set partitioning: A survey. SIAM Monma, G. L. Nemhauser, eds. Network Routing. Handbooks
Rev. 18 710–760. in Operations Research and Management Science, Vol. 8. North-
Baldacci, R., N. Christofides, A. Mingozzi. 2008. An exact algorithm Holland, Amsterdam, 1–33.
for the vehicle routing problem based on the set partitioning Fisher, M., R. Jaikumar. 1981. A generalized assignment heuristic
formulation with additional cuts. Math. Programming Ser. A 115 for vehicle routing. Networks 11 109–124.
351–385. Foster, B. A., D. M. Ryan. 1976. An integer programming approach
Baldacci, R., E. Hadjiconstantinou, A. Mingozzi. 2004. An exact to the vehicle scheduling problem. Oper. Res. 27 367–384.
algorithm for the capacitated vehicle routing problem based Fukasawa, R., H. Longo, J. Lysgaard, M. Poggi de Aragão, M. Reis,
on a two-commodity network flow formulation. Oper. Res. 52 E. Uchoa, R. F. Werneck. 2006. Robust branch-and-cut-and-
723–738. price for the capacitated vehicle routing problem. Math. Pro-
Balinski, M., R. Quandt. 1964. On an integer program for a delivery gramming Ser. A 106 491–511.
problem. Oper. Res. 12 300–304. Gavish, B., S. C. Graves. 1979. The travelling salesman problem
Carpaneto, G., P. Toth. 1980. Some new branching and bounding and related problems. Working Paper 7905, Graduate School
criteria for the asymmetric travelling salesman problem. Man- of Management, University of Rochester, Rochester, NY.
agement Sci. 26 736–743. Gendreau, M., A. Hertz, G. Laporte. 1994. A tabu search heuristic
Christofides, N. 1976. The vehicle routing problem. RAIRO for the vehicle routing problem. Management Sci. 41 1276–1290.
(Recherche Opérationnelle) 10 55–70. Gendreau, M., G. Laporte, J.-Y. Potvin. 2002. Metaheuristics for the
Christofides, N., S. Eilon. 1969. An algorithm for the vehicle dis- capacitated VRP. P. Toth, D. Vigo, eds. The Vehicle Routing Prob-
patching problem. Oper. Res. Quart. 20 309–318. lem. SIAM Monographs on Discrete Mathematics and Applications.
Christofides, N., A. Mingozzi, P. Toth. 1979. The vehicle routing Society for Industrial and Applied Mathematics, Philadelphia,
problem. N. Christofides, A. Mingozzi, P. Toth, C. Sandi, eds. 129–154.
Combined Optimization. Wiley, Chichester, UK, 315–338. Ghaziri, H. 1991. Solving routing problems by a self-organizing
Christofides, N., A. Mingozzi, P. Toth. 1981a. Exact algorithms for map. T. Kohonen, K. Makisara, O. Simula, J. Kangas, eds. Arti-
the vehicle routing problem, based on spanning tree shortest ficial Neural Networks. North-Holland, Amsterdam, 829–834.
path relaxations. Math. Programming 20 255–282. Gillett, B. E., L. R. Miller. 1974. A heuristic algorithm for the vehicle
Christofides, N., A. Mingozzi, P. Toth. 1981b. Space state relaxation dispatch problem. Oper. Res. 22 340–349.
procedures for the computation of bounds to routing problems. Glover, F. 1986. Future paths for integer programming and links to
Networks 11 145–164. artificial intelligence. Comput. Oper. Res. 13 533–549.
Clarke, G., J. W. Wright. 1964. Scheduling of vehicles from a central Golden, B., T. L. Magnanti, H. Q. Nguyen. 1977. Implementing
depot to a number of delivery points. Oper. Res. 12 568–581. vehicle routing algorithms. Networks 7 113–148.
Cordeau, J.-F., M. Gendreau, G. Laporte. 1997. A tabu search heuris- Golden, B. L., S. Raghavan, E. A. Wasil, eds. 2008. The Vehicle
tic for periodic and multi-depot vehicle routing problems. Net- Routing Problem: Latest Advances and New Challenges. Springer,
works 30 105–119. New York.
Cordeau, J.-F., G. Laporte, A. Mercier. 2001. A unified tabu search Golden, B. L., E. A. Wasil, J. P. Kelly, I.-M. Chao. 1998. Metaheuris-
heuristic for vehicle routing problems with time windows. tics in vehicle routing. T. G. Crainic, G. Laporte, eds. Fleet Man-
J. Oper. Res. Soc. 52 928–936. agement and Logistics. Kluwer, Boston, 33–56.
 Laporte: Fifty Years of Vehicle Routing
416 Transportation Science 43(4), pp. 408–416, © 2009 INFORMS

Hadjiconstantinou, E., N. Christofides, A. Mingozzi. 1995. A new Nagata, Y., O. Bräysy. 2008. Efficient local search limitation
exact algorithm for the vehicle routing problem based on strategies for the vehicle routing problem. Working paper,
q-paths and k-shortest paths relaxations. Ann. Oper. Res. 61 Agora Innoroad Laboratory, University of Juväskylä, Juväskylä,
21–43. Finland.
Hoffman, K., M. W. Padberg. 1993. Solving airline crew scheduling Nagata, Y., O. Bräysy. 2009. Edge assembly based memetic algo-
problems by branch-and-cut. Management Sci. 39 657–682. rithm for the capacitated vehicle routing problem. Working
Holland, J. H. 1975. Adaptation in Natural and Artificial Systems. The paper, Agora Innoroad Laboratory, University of Juväskylä,
University of Michigan Press, Ann Arbor, MI. Juväskylä, Finland.
Houck, D., J.-C. Picard, M. Queyranne, R. Vemuganti. 1980. The Nelson, M. D., K. E. Nygard, J. H. Griffin, W. E. Shreve. 1985. Imple-
Downloaded from informs.org by [142.150.190.39] on 16 April 2017, at 19:39 . For personal use only, all rights reserved.

travelling salesman problem as a constrained shortest path mentation techniques for the vehicle routing problem. Comput.
problem: Theory and computation experience. Opsearch 17 Oper. Res. 12 273–283.
93–109. Paessens, H. 1988. The savings algorithm for the vehicle routing
Kindervater, G. A. P., M. W. P. Savelsbergh. 1997. Vehicle rout- problem. Eur. J. Oper. Res. 34 336–344.
ing: Handling edge exchanges. E. H. L. Aarts, J. K. Lenstra, Pisinger, D., S. Ropke. 2007. A general heuristic for vehicle routing
eds. Local Search in Combinatorial Optimization. Wiley, Chich- problems. Comput. Oper. Res. 34 2403–2435.
ester, UK, 337–360. Prins, C. 2004. A simple and effective evolutionary algorithm
Kirkpatrick, S., C. D. Gelatt, Jr., M. P. Vecchi. 1983. Optimization by for the vehicle routing problem. Comput. Oper. Res. 31
simulated annealing. Science 220 671–680. 1985–2002.
Kytöjoki, J., T. Nuortio, O. Bräysy, M. Gendreau. 2007. An efficient Prins, C. 2009. A GRASP X evolutionary local search hybrid for
variable neighborhood search heuristic for very large scale the vehicle routing problem. F. B. Pereira, J. Tavares, eds. BIO-
vehicle routing problems. Comput. Oper. Res. 34 2743–2757. Inspired Algorithms for the Vehicle Routing Problem. Springer,
Laporte, G. 2007. What you should know about the vehicle routing Berlin-Heidelberg, 35–53.
problem. Naval Res. Logist. 54 811–819. Rao, M. R., S. Zionts. 1968. Allocation of transportation units to
Laporte, G. 2009. A concise guide to the traveling salesman prob- alternative trips—A column generation scheme with out-of-
kilter subproblems. Oper. Res. 16 52–63.
lem. J. Oper. Res. Soc. Forthcoming.
Reimann, M., K. Doerner, R. F. Hartl. 2004. D-Ants: Savings based
Laporte, G., Y. Nobert. 1983. A branch and bound algorithm for
ants divide and conquer the vehicle routing problem. Comput.
the capacitated vehicle routing problem. Oper. Res. Spektrum 5
Oper. Res. 31 563–591.
77–85.
Renaud, J., F. F. Boctor, G. Laporte. 1996. An improved petal
Laporte, G., Y. Nobert. 1987. Exact algorithms for the vehicle rout-
heuristic for the vehicle routing problem. J. Oper. Res. Soc. 47
ing problem. Ann. Discrete Math. 31 147–184.
329–336.
Laporte, G., F. Semet. 2002. Classical heuristics for the capacitated
Rochat, Y., É. D. Taillard. 1995. Probabilistic diversification and
VRP. P. Toth, D. Vigo, eds. The Vehicle Routing Problem. SIAM
intensification in local search for vehicle routing. J. Heuristics 1
Monographs on Discrete Mathematics and Applications, Society for 147–167.
Industrial and Applied Mathematics, Philadelphia, 109–128. Ropke, S., D. Pisinger. 2006. A unified heuristic for a large class of
Laporte, G., H. Mercure, Y. Nobert. 1986. An exact algorithm for vehicle routing problems with backhauls. Eur. J. Oper. Res. 171
the asymmetrical capacited vehicle routing problem. Networks 750–775.
16 33–46. Ryan, D. M., C. Hjorring, F. Glover. 1993. Extensions of the
Laporte, G., Y. Nobert, M. Desrochers. 1985. Optimal routing under petal method for vehicle routing. J. Oper. Res. Soc. 44
capacity and distance restrictions. Oper. Res. 33 1050–1073. 289–296.
Letchford, A. N., R. W. Eglese, J. Lysgaard. 2002. Multistars, partial Schumann, M., R. Retzko. 1995. Self-organizing maps for vehi-
multistars and the capacitated vehicle routing problem. Math. cle routing problems—Minimizing an explicit cost function.
Programming Ser. A 94 21–40. F. Fogelman-Soulie, ed. Proc. Internat. Conf. Artificial Neural Net-
Li, F., B. L. Golden, E. A. Wasil. 2005. Very large-scale vehicle rout- works, Paris, 401–406.
ing: New test problems, algorithms, and results. Comput. Oper. Shaw, P. 1998. Using constraint programming and local search
Res. 32 1165–1179. methods to solve vehicle routing problems. CP98 Fourth Inter-
Little, J. D. C., K. G. Murty, D. W. Sweeney, C. Karel. 1963. An nat. Conf. Principles Practice Constraint Programming, Vol. 1520,
algorithm for the travelling salesman problem. Oper. Res. 11 Lecture Notes in Computer Science. Springer, Berlin, 417–431.
972–989. Taillard, É. D. 1993. Parallel iterative search methods for vehicle
Lysgaard, J., A. N. Letchford, R. W. Eglese. 2004. A new branch- routing problems. Networks 23 661–673.
and-cut algorithm for the capacitated vehicle routing problem. Tarantilis, C.-D. 2005. Solving the vehicle routing problem with
Math. Programming 100 423–445. adaptive memory programming methodology. Comput. Oper.
Mester, D., O. Bräysy. 2005. Active guided evolution strategies for Res. 32 2309–2327.
the large scale vehicle routing problems with time windows. Tarantilis, C.-D., C. T. Kiranoudis. 2002. BoneRoute: An adaptive
Comput. Oper. Res. 32 1593–1614. memory-based method for effective fleet management. Ann.
Mester, D., O. Bräysy. 2007. Active guided evolution strategies for Oper. Res. 115 227–241.
the large scale capacitated vehicle routing problems. Comput. Thompson, P. M., H. M. Psaraftis. 1993. Cyclic transfer algorithms
Oper. Res. 34 2964–2975. for multi-vehicle routing and scheduling problems. Oper. Res.
Mladenović, N., P. Hansen. 1997. Variable neighborhood search. 41 935–946.
Comput. Oper. Res. 24 1097–1100. Toth, P., D. Vigo, eds. 2002a. The Vehicle Routing Problem. SIAM
Moscato, P., C. Cotta. 2003. A gentle introduction to memetic algo- Monographs on Discrete Mathematics and Applications. Society for
rithms. F. Glover, G. A. Kochenberger, eds. Handbook of Meta- Industrial and Applied Mathematics, Philadelphia.
heuristics. Kluwer, Boston, 105–144. Toth, P., D. Vigo. 2002b. Branch-and-bound algorithms for the
Naddef, D., G. Rinaldi. 2002. Branch-and-cut algorithms for the capacitated VRP. P. Toth, D. Vigo, eds. The Vehicle Routing Prob-
capacitated VRP. P. Toth, D. Vigo, eds. The Vehicle Routing Prob- lem. SIAM Monographs on Discrete Mathematics and Applications.
lem. SIAM Monographs on Discrete Mathematics and Applications. Society for Industrial and Applied Mathematics, Philadelphia,
Society for Industrial and Applied Mathematics, Philadelphia, 29–51.
53–81. Toth, P., D. Vigo. 2003. The granular tabu search and its applica-
Nagata, Y. 2007. Edge assembly crossover for the capacitated vehi- tion to the vehicle routing problem. INFORMS J. Comput. 15
cle routing problem. C. Cotta, J. van Hemert, eds. Lecture 333–346.
Notes in Computer Science, Vol. 4446. Springer-Verlag, Berlin Wark, P., J. Holt. 1994. A repeated matching heuristic for the vehicle
Heidelberg, 142–153. routing problem. J. Oper. Res. Soc. 45 1156–1167.