See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/233843551

Vehicle Routing

Chapter · January 2007

CITATIONS READS
171 10,870

4 authors, including:

Martin Savelsbergh Daniele Vigo

Georgia Institute of Technology University of Bologna
274 PUBLICATIONS 23,243 CITATIONS 209 PUBLICATIONS 16,557 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Daniele Vigo on 20 May 2014.

The user has requested enhancement of the downloaded file.

 Vehicle Routing

Jean-François Cordeau∗
Gilbert Laporte∗
Martin W.P. Savelsbergh†
Daniele Vigo‡

November 23, 2005

∗
Canada Research Chair in Distribution Management, HEC Montréal, 3000 chemin de la Côte-Sainte-
Catherine, Montréal, Canada, H3T 2A7.
†
School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332-
0205, U.S.A.
‡
Dipartimento di Elettronica, Informatica e Sistemistica, University of Bologna, Viale Risorgimento 2,
40136 Bologna, Italy
1 Introduction
The Vehicle Routing Problem lies at the heart of distribution management. It is faced each
day by thousands of companies and organizations engaged in the delivery and collection
of goods or people. Because conditions vary from one setting to the next, the objectives
and constraints encountered in practice are highly variable. Most algorithmic research and
software development in this area focus on a limited number of prototype problems. By
building enough flexibility in optimization systems one can adapt these to various practical
contexts.
Much progress has been made since the publication of the first article on the “truck
dispatching” problem by Dantzig and Ramser (1959). Several variants of the basic problem
have been put forward. Strong formulations have been proposed, together with polyhedral
studies and exact decomposition algorithms. Numerous heuristics have also been developed
for vehicle routing problems. In particular the study of this class of problems has stimulated
the emergence and the growth of several metaheuristics whose performance is constantly
improving.
This chapter focuses on some of the most important vehicle routing problem types. A
number of other variants have been treated in recent articles and book chapters (see, e.g.,
Toth and Vigo (2002a)). The Pickup and Delivery Vehicle Routing Problem, which has
also been extensively studied, is covered in the “Transportation on Demand” chapter.
The remainder of this chapter is organized as follows. Section 2 is devoted to the
classical vehicle routing problem (simply referred to as VRP), defined with a single depot
and only capacity and route length constraints. Problems with time windows are surveyed
in Section 3. Section 4 is devoted to inventory routing problems which combine routing and
customer replenishment decisions. Finally, Section 5 covers the field of stochastic vehicle
routing in which some of the problem data are random variables.

2 The Classical Vehicle Routing Problem

The Classical Vehicle Routing Problem (VRP) is one of the most popular problems in
combinatorial optimization, and its study has given rise to several exact and heuristic
solution techniques of general applicability. It generalizes the Traveling Salesman Problem
(TSP) and is therefore NP-hard. A recent survey of the VRP can be found in the first
six chapters of the book edited by Toth and Vigo (2002a). The aim of this section is to
provide a comprehensive overview of the available exact and heuristic algorithms for the
VRP, most of which have also been adapted to solve other variants, as will be shown in
the remaining sections.
The VRP is often defined under capacity and route length restrictions. When only
capacity constraints are present the problem is denoted as CVRP. Most exact algorithms

2
have been developed with capacity constraints in mind but several apply mutatis mutandis
to distance constrained problems. In contrast, most heuristics explicitly consider both
types of constraint.

2.1 Formulations
The symmetric VRP is defined on a complete undirected graph G = (V, E). The set
V = {0, . . . , n} is a vertex set. Each vertex i ∈ V \{0} represents a customer having a
non-negative demand qi , while vertex 0 corresponds to a depot. To each edge e ∈ E =
{(i, j) : i, j ∈ V, i < j} is associated a travel cost ce or cij . A fixed fleet of m identical
vehicles, each of capacity Q, is available at the depot. The symmetric VRP calls for the
determination of a set of m routes whose total travel cost is minimized and such that: 1)
each customer is visited exactly once by one route; 2) each route starts and ends at the
depot, 3) the total demand of the customers served by a route does not exceed the vehicle
capacity Q, and 4) the length of each route does not exceed a preset limit L. (It is common
to assume constant speed so that distances, travel times and travel costs are considered
as synonymous.) A solution can be viewed as a set of m cycles sharing a common vertex
at the depot. The asymmetric VRP is similarly defined on a directed graph G = (V, A),
where A = {(i, j) : i, j ∈ V, i 6= j} is an arc set. In this case a circuit (directed cycle) is
associated with a vehicle route. Most results of Sections 2.1 and 2.2 apply to the symmetric
CVRP.
An integer linear programming formulation of the CVRP follows, where for each edge
e ∈ E the integer variable xe indicates the number of times edge e is traversed in the
solution. Let r(S) denote the minimum number of vehicles needed to serve the customers
of a subset S of customers. The value of r(S) may be determined by solving an associated
Bin Packing Problem (BPP) with item set S and bins of capacity Q. Finally, for S ⊂ V ,
let δ(S) = {(i, j) : i ∈ S, j ∈/ S or i ∈/ S, j ∈ S}. If S = {i}, then we simply write δ(i)
instead of δ({i}). The CVRP formulation proposed by Laporte, Nobert and Desrochers
(1985) is then:
X
(CVRP1) Minimize ce xe (1)
e∈E
subject to
X
xe = 2 (i ∈ V \ {0}) (2)
e∈δ(i)
X
xe = 2m (3)
e∈δ(0)
X
xe ≥ 2r(S) (S ⊆ V \ {0}, S 6= ∅) (4)
e∈δ(S)

3
 xe ∈ {0, 1} (e 6∈ δ(0)) (5)
xe ∈ {0, 1, 2} (e ∈ δ(0)). (6)

The degree constraints (2) state that each customer is visited exactly once, whereas the
depot degree constraint (3) means that m routes are created. Capacity constraints (4)
impose both the connectivity of the solution and the vehicle capacity requirements by
forcing a sufficient number of edges to enter each subset of vertices. We note that since the
BPP is NP-hard in the strong sense, r(S) may be approximated from below by any BPP
P
lower bound, such as ⌈ i∈S qi /Q⌉. Finally, constraints (5) and (6) impose that each edge
between two customers is traversed at most once and each edge incident to the depot is
traversed at most twice. In this latter case, the vehicle performs a route visiting a single
customer.
A widely used alternative formulation is based on the set partitioning or set covering
models. The formulation was originally proposed by Balinski and Quandt (1964) and
contains a potentially exponential number of binary variables. Let R = {R1 , . . . , Rs }
denote the collection of all feasible routes, with s = |R|. Each route Rj has an associated
cost γj , and aij is a binary coefficient equal to 1 if and only if vertex i is visited (i.e.,
covered) by route Rj . The binary variable xj , j = 1, . . . , s, is equal to 1 if and only if route
Rj is selected in the solution. The model is:
s
X
(CVRP2) Minimize γj xj (7)
j=1

subject to

s
X
aij xj = 1 (i ∈ V \ {0}) (8)
j=1
Xs
xj = m (9)
j=1
xj ∈ {0, 1} (j = 1, . . . , s). (10)

Constraints (8) impose that each customer i is covered by exactly one route, and (9)
requires that m routes be selected. Because route feasibility is implicitly considered in the
definition of R, this is a very general model which may easily take additional constraints
into account. Moreover, when the cost matrix satisfies the triangle inequality (i.e., cij ≤
cik + ckj for all i, j, k ∈ V ), the set partitioning model CVRP2 may be transformed into
an equivalent set covering model CVRP2’ by replacing the equality sign with “≥” in
(8). Any feasible solution to CVRP2 is clearly feasible for CVRP2’, and any feasible
solution to CVRP2’ may be transformed into a feasible CVRP2 solution of smaller or

4
equal cost. Indeed, if the CVRP2’ solution is infeasible for CVRP2, then one or more
customers are visited more than once. These customers may therefore be removed from
their route by applying shortcuts which will not increase the solution cost because of the
triangle inequality. The main advantage of using CVRP2’ is that only inclusion-maximal
feasible routes, among those with the same cost, need be considered in the definition of R.
This significantly reduces the number of variables. In addition, when using CVRP2’ the
dual solution space is considerably reduced since dual variables are restricted to be non-
negative. One of the main drawbacks of models CVRP2 and CVRP2’ lies in its very large
number of variables, which in loosely constrained medium size instances may easily run
into the billions. Thus, one has to resort to a column generation algorithm to solve these
problems. The linear programming relaxation of these models tends to be very tight, as
shown by Bramel and Simchi-Levi (1997). Further details on these formulations and their
extensions, as well as additional formulations for the symmetric and asymmetric cases, can
be found in Laporte and Nobert (1987) and in Toth and Vigo (2002b,d).

2.2 Exact algorithms for the CVRP

We now review the main exact approaches presented in the last two decades for the solution
of the CVRP. For a thorough review of previous exact methods, see Laporte and Nobert
(1987). We first describe the algorithms based on branch-and-bound, including those that
make use of the set partitioning formulation and column generation schemes, and we then
examine the algorithms based on branch-and-cut. In practice, the CVRP turns out to be
significantly harder to solve than the TSP. The best CVRP algorithms can rarely tackle
instances involving more than 100 vertices, while TSP instances with hundreds and even
thousands of vertices are now routinely solved to optimality.

2.2.1 Branch-and-bound and set partitioning based algorithms

Several branch-and-bound algorithms are available for the solution of the CVRP. Until the
late 1980s, the most effective exact methods were mainly branch-and-bound algorithms
based on elementary combinatorial relaxations. Recently, more sophisticated bounds have
been proposed, namely those based on Lagrangean relaxations or on the additive bounding
procedure, which have substantially increased the size of the problems that can be solved
to optimality. We now describe some branch-and-bound algorithms with an emphasis on
lower bound computations which constitute the most critical component of methods of
this type. More details on the structure of branch-and-bound algorithm strategies and
dominance rules may be found in Toth and Vigo (1998, 2002c,d). We also review in this
section exact set partitioning based algorithms for the CVRP.
Many different elementary combinatorial relaxations were used in early branch-and-
bound algorithms, including those based on the Assignment Problem (AP), on the degree-

5
constrained Shortest Spanning Tree, and on state-space relaxation. Here we outline the two
families of relaxations used as a basis for the more recent branch-and-bound algorithms
for the symmetric and asymmetric CVRP. A first relaxation is obtained from the integer
programming formulations of these problems by dropping the connectivity and capacity
constraints. In the symmetric case the resulting problem is a b-Matching Problem (b-MP),
i.e., the problem of determining a minimum cost set of cycles covering all vertices and such
that the degree of each vertex i is equal to bi , where bi = 2 for all the customer vertices,
and b0 = 2m for the depot vertex. It is easy to see that by adding m − 1 copies of the depot
to G the relaxation becomes a 2-MP. In the asymmetric case the relaxed problem is the
well-known Transportation Problem which may be transformed into an AP by introducing
copies of the depot. Also in this case, the AP may be seen as the problem of determining
a set of circuits covering all vertices and such that each vertex has one entering and one
leaving arc. The solution of these relaxed problems may be infeasible for the CVRP since
the demand associated with a cycle or circuit may exceed the vehicle capacity, and some
of these may be disconnected from the depot. The relaxed problems may then be solved
in polynomial time (see, e.g., Miller and Pekny (1995) for the b-MP and Dell’Amico and
Toth (2000) for the AP). However, the quality of the lower bounds obtained with these
relaxations is generally very poor and not sufficient to solve instances with more than 15 or
20 customers. Toth and Vigo (2002c) report average gaps in excess of 20% with respect to
the optimal solution value on benchmark CVRP instances. The situation is slightly better
for the AP relaxation of the asymmetric CVRP that yields average gaps of about 10% or
less. Laporte, Mercure and Nobert (1986) have proposed a branch-and-bound algorithm for
asymmetric CVRP, based on the AP relaxation and capable of solving randomly generated
problems involving tens of customers and between two and four vehicles.
The second family of elementary relaxations used in recent branch-and-bound algo-
rithms is based on degree-constrained spanning trees. These relaxations extend the well-
known 1-tree relaxation proposed by Held and Karp (1971) for the TSP. The earliest
branch-and-bound algorithm based on this relaxation, proposed by Christofides, Mingozzi
and Toth (1981a), could only solve relatively small instances. More recently, Fisher (1994)
has presented another tree based relaxation requiring the determination of a so-called m-
tree, defined as a minimum cost set of n + m edges spanning the graph. The approach used
by Fisher is based on CVRP1 with the additional assumption that single-customer routes
are not allowed. Fisher modeled the CVRP as the problem of determining an m-tree with
degree equal to 2m at the depot vertex, with additional constraints on vehicle capacity
and a degree of 2 for each customer vertex. The determination of an m-tree with degree
2m at the depot requires O(n3 ) time. The degree-constrained m-tree relaxation is easily
obtained from CVRP1 by removing the degree constraints (2) for customer vertices and
weakening the capacity constraints (4) into connectivity constraints, i.e., by replacing their
right-hand side with 1. The m-tree solution is not always feasible for the CVRP since some
vertices may have degree different from 2 and the demand associated with the subtrees

6
incident to the depot may exceed the vehicle capacity.
For the asymmetric CVRP, similar relaxations may be derived from directed trees, also
called arborescences, spanning the graph and having outdegree equal to m at the depot
vertex. To obtain the final bound a minimum cost set of m vertex-disjoint arcs entering
the depot are added to the constrained arborescence. In this case, the relaxed subproblem
may be solved in polynomial time, but again the quality of the resulting lower bound is
very poor. Toth and Vigo (2002c) report that on benchmark asymmetric instances, the
average gap of these relaxations with respect to the optimal solution value is larger than
25%.
Different improved bounding techniques were later developed to narrow the gap between
the lower bound and the optimal solution value of the CVRP. These include two bounding
procedures based on Lagrangean relaxation proposed by Fisher (1994) and Miller (1995).
These are strengthenings of the basic CVRP relaxations obtained by dualizing some of the
relaxed constraints in a Lagrangean fashion. In particular, they both include in the objec-
tive function a suitable subset of the capacity constraints (4), whereas the Fisher relaxation
also incorporates degree constraints (2) which were relaxed in the m-tree relaxation. As in
related problems, good values for the Lagrangean multipliers associated with the relaxed
constraints are determined by using a subgradient optimization procedure (see, e.g., Held
and Karp (1971) and Held, Wolfe and Crowder (1974)). The main difficulty associated
with these relaxations lies in the exponential cardinality of the set of relaxed constraints
which does not allow for their complete inclusion in the objective function. Both authors
include a limited family F of capacity constraints and iteratively generate the constraints
violated by the current solution of the Lagrangean problem. The process terminates when
no violated constraint is detected (hence the Lagrangean solution is feasible) or a preset
number of subgradient iterations have been executed. Redundant constraints are period-
ically removed from F. The relax-and-cut algorithm of Martinhon, Lucena and Maculan
(2000) generalizes these Lagrangean-based approaches by also considering comb and mul-
tistar inequalities, and moderately improves the quality of the Lagrangean bound.
Some exact algorithms for the CVRP are based on the set partitioning formulation
CVRP2. The first of these is due to Agarwal, Mathur and Salkin (1989) who considered a
relaxation of model CVRP2 not including constraints (9) and solved the resulting model
through column generation. Agarwal, Mathur and Salkin used their algorithm to solve
seven Euclidean CVRP instances with up to 25 customers. Hadjiconstantinou, Christofides
and Mingozzi (1995) proposed a branch-and-bound algorithm in which the lower bound
was obtained by considering the dual of the linear relaxation of model CVRP2, following
the approach introduced by Mingozzi, Christofides and Hadjiconstantinou (1994). By
linear programming duality, any feasible solution to this dual problem yields a valid lower
bound. Hadjiconstantinou, Christofides and Mingozzi (1995) determined the heuristic dual
solutions by combining two relaxations of the original problem: the q-path relaxation of

7
Christofides, Mingozzi and Toth (1981a) and the m-shortest path relaxation of Christofides
and Mingozzi (1989). The algorithm was able to solve randomly generated Euclidean
instances with up to 30 vertices and benchmark instances with up to 50 vertices. Further
details on set partitioning-based algorithms for the CVRP are provided in Bramel and
Simchi-Levi (2002).
Fischetti, Toth and Vigo (1994) have improved the AP relaxation of the asymmetric
CVRP by combining into an additive bounding procedure two new lower bounds based on
disjunctions on infeasible arc subsets and on minimum cost flows. The additive approach
was proposed by Fischetti and Toth (1989) and allows for the combination of different lower
bounding procedures, each exploiting a different substructure of the problem under consid-
eration. The resulting branch-and-bound approach was able to solve randomly generated
instances containing up to 300 vertices and four vehicles. Other bounds for the asymmetric
CVRP may be derived by generalizing the methods proposed for the symmetric case. For
example, Fisher (1994) proposed a way of extending to the asymmetric CVRP the La-
grangean bound based on m-trees. In this extension the Lagrangean problem calls for the
determination of an undirected m-tree on the undirected graph obtained by replacing each
pair of arcs (i, j) and (j, i) with a single edge (i, j) of cost c′ij = min{cij , cji }. No compu-
tational testing for this bound was presented by Fisher (1994). Potentially better bounds
may be obtained by explicitly considering the asymmetry of the problem, i.e., by using
m-arborescences rather than m-trees and by strengthening the bound in a Lagrangean
fashion as proposed by Toth and Vigo (1995, 1997) for the Capacitated Shortest Spanning
Arborescence Problem and for the VRP with Backhauls.

2.2.2 Branch-and-cut algorithms

Branch-and-cut algorithms currently constitute the best available exact approach for the
solution of the CVRP. Research in this area has been strongly motivated by the emergence
and the success of polyhedral combinatorics as a framework for the solution of hard com-
binatorial problems, particularly the TSP. However, in a recent survey on branch-and-cut
approaches for the CVRP, Naddef and Rinaldi (2002) state: “ ... the amount of research
effort spent to solve CVRP by this method is not comparable with what has been dedicated
to the TSP [... the research in this field] is still quite limited and most of it is not pub-
lished yet”. In the following we summarize the main available branch-and-cut approaches
for the CVRP. The reader is referred to Naddef and Rinaldi (2002) for a more detailed
presentation.
The use of branch-and-cut for the CVRP is rooted in the exact algorithm of Laporte,
Nobert and Desrochers (1985). This algorithm uses the linear programming (LP) relaxation
of model CVRP1 without capacity constraints (4) as a basis for the solution of the VRP
with capacity and maximum distance restrictions. This initial relaxation is iteratively
strengthened by adding violated capacity constraints which are heuristically separated by

8
considering the connected components induced by the set of non-zero variables in the
current LP solution. Gomory cuts are also introduced at the root node of the branch-and-
cut tree. The algorithm was capable of solving randomly generated loosely constrained
Euclidean and non-Euclidean instances with two or three vehicles and up to 60 customers.
The first polyhedral study of the CVRP was presented by Cornuéjols and Harche (1993).
The presence of equalities (2) and (3) makes the CVRP non-fully-dimensional. Therefore,
as in the TSP, Cornuéjols and Harche first considered the full-dimensional polyhedron,
containing the CVRP polyhedron as a face, associated with the so-called Graphical VRP
(GVRP) where customers may be visited more than once. The basic properties of the
GVRP polyhedron were also investigated. Conditions under which the non-negativity, de-
gree and capacity constraints define facets of the GVRP and CVRP polyhedra were also
determined. Cornuéjols and Harche have extended to the GVRP and the CVRP several
other families of valid inequalities proposed for the TSP and the Graphical TSP. In partic-
ular, comb, path, wheelbarrow and bicycle inequalities were extended to the capacitated
case and again, sufficient conditions under which these inequalities define facets of the
GVRP and CVRP polyhedra were derived. These inequalities were used by Cornuéjols
and Harche as cutting planes to solve two instances of CVRP with 18 and 50 customers,
within a branch-and-cut algorithm. The detection of violated inequalities was performed
manually, starting from the current optimal LP solution.
Augerat et al. (1995) have developed the first complete branch-and-cut approach for
the CVRP. They described several heuristic separation procedures for the classes of valid
inequalities proposed by Cornuéjols and Harche, as well as four new classes of valid in-
equalities. Separation procedures were further investigated by Augerat et al. (1999). The
resulting approach was able to solve several CVRP instances containing up to 134 cus-
tomers. Ralphs et al. (2003) have presented a branch-and-cut algorithm for the CVRP
in which an exact separation of valid m-TSP inequalities is used in addition to heuristic
separation of capacity inequalities. The resulting algorithm was implemented within the
SYMPHONY parallel branch-and-cut-and-price framework and was able to solve several
instances involving fewer than 100 vertices. Lysgaard, Letchford and Eglese (2004) have
developed new separation procedures for most of the families of valid inequalities proposed
so far (see also Letchford, Eglese and Lysgaard, 2002). Their overall branch-and-cut ap-
proach, which is further enhanced by the use of Gomory cuts, was able to solve within
moderate computing times previously solved instances and three new medium size ones.
Baldacci, Hadjiconstantinou and Mingozzi (2004) have put forward a branch-and-cut
algorithm based on a two-commodity network flow formulation of the CVRP and requiring
a polynomial number of integer variables. It seems to provide an interesting alternative to
other classical formulations (see also Gouveia (1995) for a single-commodity formulation).
The overall algorithm strengthens the LP relaxation by adding violated capacity inequali-
ties and implements various variable reduction and branching rules. The results obtained

9
with this approach are comparable with those of the other branch-and-cut algorithms just
described.
Finally, Fukusawa et al. (2003) have proposed a successful branch-and-cut-and-price
algorithm combining branch-and-cut with the q-routes relaxation of Christofides, Mingozzi
and Toth (1981a), used here in a column generation fashion. This method produces tighter
bounds than other branch-and-cut algorithms and is capable of solving several previously
unsolved instances with up to 75 customers. Baldacci, Bodin and Mingozzi (2004) have
used their set partitioning algorithm, previously developed for a rollon-rolloff VRP, to solve
difficult CVRP instances. Their approach yields bounds whose quality is comparable to
those of Fukusawa et al. (2003), but seems much quicker.
Other branch-and-cut algorithms are described in Achuthan, Caccetta and Hill (1996,
2003) and Blasum and Hochstättler (2000). We also mention that the polyhedral structure
of the special case of CVRP where all the customers have a unit demand was studied by
Campos, Corberán and Mota (1991) and by Araque, Hall and Magnanti (1990). Branch-
and-cut algorithms for this problem are presented by Araque et al. (1994) and by Ghiani,
Laporte and Semet (2004).

2.3 Heuristics for the VRP

An impressive number of heuristics have been proposed for the VRP. Initially these were
mainly standard route construction algorithms, whereas more recently powerful meta-
heuristic approaches have been developed. In the following we separately review these two
families of algorithms. Almost all of these methods were developed, described and tested
for the symmetric VRP. In addition, since finding a feasible solution with exactly m ve-
hicles is itself an NP-complete problem, almost all methods assume an unlimited number
of available vehicles. However, it should be observed that many of the proposed methods
may be quite easily adapted to take into account additional practical constraints, although
these may affect their overall performance (see, e.g., Vigo (1996) for an extension of some
classical heuristics to the asymmetric case).

2.3.1 Classical heuristics

Using the classification proposed by Laporte and Semet (2002), we describe classical VRP
heuristics under these headings: route construction methods, two-phase methods, and
route improvement methods.

Route construction heuristics

Route construction methods were among the first heuristics for the CVRP and still form the
core of many software implementations for various routing applications. These algorithms

10
typically start from an empty solution and iteratively build routes by inserting one or
more customers at each iteration, until all customers are routed. Construction algorithms
are further subdivided into sequential and parallel, depending on the number of eligible
routes for the insertion of a customer. Sequential methods expand only one route at a time,
whereas parallel methods consider more than one route simultaneously. Route construction
algorithms are fully specified by their three main ingredients, namely an initialization
criterion, a selection criterion specifying which customers are chosen for insertion at the
current iteration, and an insertion criterion to decide where to locate the chosen customers
into the current routes.
The first and most famous heuristic of this group was proposed by Clarke and Wright
(1964) and is based on the concept of saving, an estimate of the cost reduction obtained by
serving two customers sequentially in the same route, rather than in two separate ones. If
i is the last customer of a route and j is the first customer of another route, the associated
saving is defined as sij = ci0 + c0j − cij . If sij is positive, then serving i and j consecutively
in a route is profitable. The Clarke and Wright algorithm considers all customer pairs and
sorts the savings in non-increasing order. Starting with a solution in which each customer
appears separately in a route, the customer pair list is examined and two routes are merged
whenever this is feasible. Generally, a route merge is accepted only if the associated saving
is non-negative but, if the number of vehicles is to be minimized, then negative saving
merges may also be considered. The Clarke and Wright algorithm is inherently parallel
since more than one route is active at any time. However, it may easily be implemented in a
sequential fashion. The resulting algorithm is quite fast but may have a poor performance
(see, e.g., Laporte and Semet, 2002). Golden, Magnanti and Nguyen (1977), Paessens
(1988), and Nelson et al. (1985) have proposed various enhancement strategies of the
savings approach aimed at improving either its effectiveness or its computational efficiency
by means of better data structures. Other attempts to improve the effectiveness of the
savings method were made by Desrochers and Verhoog (1989), Altinkemer and Gavish
(1991), and by Wark and Holt (1994) who proposed to implement route merges by using a
matching algorithm, together with a more sophisticated estimate of actual merge savings.
The results obtained with these algorithms are in general better than those of previous
savings methods, but matching-based algorithms require much larger computing times.
Another classical route construction heuristic is the sequential insertion algorithm of
Mole and Jameson (1976). The algorithm uses as selection and insertion criterion the eval-
uation of the extra distance resulting from the insertion of an unrouted customer k between
two consecutive customers i and j of the current route, namely α(i, k, j) = cik + ckj − λcij ,
where λ is a user-controlled parameter. Variations of this criterion taking into account
other factors, such as the distance of the customer from the depot, were also considered.
After each insertion, the current route is possibly improved by using a 3-opt procedure.
A more general and effective two-step insertion heuristic was proposed by Christofides,

11
Mingozzi and Toth (1979). In the first step, a sequential insertion algorithm is used to
determine a set of feasible routes. The second step is a parallel insertion approach. For
each route determined in the first step, a representative customer is selected and a set of
single-customer routes is initialized with these customers. The remaining unrouted cus-
tomers are then inserted by using a regret criterion, where the difference between the best
and the second-best insertion cost is taken into account, and partial routes are improved
by means of a 3-opt procedure. The resulting algorithm is superior to that of Mole and
Jameson and represents a good compromise between effectiveness and efficiency.

Two-phase heuristics

Two-phase methods are based on the decomposition of the VRP solution process into
the two separate subproblems: 1) clustering: determine a partition of the customers into
subsets, each corresponding to a route, and 2) routing: determine the sequence of customers
on each route. In a cluster-first-route-second method, customers are first grouped into
clusters and the routes are then determined by suitably sequencing the customers within
each cluster. Different techniques have been proposed for the clustering phase, while the
routing phase amounts to solving a TSP.
The sweep algorithm, due to Wren (1971), Wren and Holliday (1972), and Gillett and
Miller (1974), is often referred to as the first example of cluster-first-route-second approach.
The algorithm applies to planar VRP instances. The algorithm starts with an arbitrary
customer and then sequentially assigns the remaining customers to the current vehicle by
considering them in order of increasing polar angle with respect to the depot and the initial
customer. As soon as the current customer cannot be feasibly assigned to the current
vehicle, a new route is initialized with it. Once all customers are assigned to vehicles,
each route is separately defined by solving a TSP. Another early two-phase method is
the truncated branch-and-bound method of Christofides, Mingozzi and Toth (1979) in
which the set of routes is determined through an adaptation of an exact branch-and-bound
algorithm that uses a branching-on-routes strategy. The decision tree contains as many
levels as the number of available vehicles, and at each level of the decision tree a given node
corresponds to a partial solution made up of some complete routes. The descendant nodes
correspond to all possible routes including a subset of the unrouted customers. The running
time of the algorithm is controlled by limiting to one the number of routes generated at
each level.
The Fisher and Jaikumar (1981) algorithm solves the clustering step by means of an
appropriately defined Generalized Assignment Problem (GAP) which calls for the deter-
mination of a minimum cost assignment of items to a given set of bins of capacity Q, and
where the items are characterized by a weight and an assignment cost for each bin. Each
vehicle is assigned a representative customer, called a seed, and the assignment cost of a
customer to a vehicle is equal to its distance to the seed. The GAP is then solved, either

12
optimally or heuristically, and the final routes are determined by solving a TSP on each
cluster.
Another two-phase method working with a fixed number m of vehicles was described
by Bramel and Simchi-Levi (1995). This algorithm determines route seeds by solving a
Capacitated Location Problem, where m customers are selected by minimizing the total
distance between each customer and its closest seed, and by imposing that the total de-
mand associated with each seed be at most Q. Once seeds have been determined and
the single-customer routes are initialized, the remaining customers are inserted in the cur-
rent routes by minimizing insertion costs. Various ways of approximating the insertion
cost are proposed and analyzed. It is worth noting that all three cluster-first-route-second
approaches just described allow for a direct control of the number of routes in the final
solution, whereas the sweep algorithm does not. The performance of these algorithms is
generally comparable to that of route construction algorithms in terms of effectiveness.
The location based approach of Bramel and Simchi-Levi produces better solutions but
requires much larger computing times.
A different family of two-phase methods is the class of so-called petal algorithms. These
generate a large set of feasible routes, called petals, and select the final subset by solving
a set partitioning model. Foster and Ryan (1976) and Ryan, Hjorring and Glover (1993)
have proposed heuristic rules for determining the set of routes to be selected, while Renaud,
Boctor and Laporte (1996b) have described an extension that considers more involved
configurations, called 2-petals, consisting of two embedded or intersecting routes. The
overall performance of these algorithms is generally superior to that of the sweep algorithm.
Finally, in route-first-cluster-second methods, a giant TSP tour over all customers is
constructed in a first phase and later subdivided into feasible routes. Examples of such
algorithms are given by Beasley (1983), Haimovich and Rinnooy Kan (1985), and Bertsimas
and Simchi-Levi (1996), but the performance of this approach is generally poor.

Route improvement heuristics

Local search algorithms are often used to improve initial solutions generated by other
heuristics. Starting from a given solution, a local search method applies simple modifi-
cations, such as arc exchanges or customer movements, to obtain neighbour solutions of
possibly better cost. If an improving solution is found, it then becomes the current solution
and the process iterates; otherwise a local minimum has been identified.
A large variety of neighbourhoods are available. These may be subdivided into intra-
route neighbourhoods, if they operate on a single route at a time, or inter-route neighbour-
hoods if they consider more than one route simultaneously. The most common neighbour-
hood type is the λ-opt heuristic of Lin (1965) for the TSP, where λ edges are removed from
the current solution and replaced by λ others. The computing time required to examine all

13
neighbours of a solution is proportional to nλ . Thus, only λ = 2 or 3 are used in practice.
As an alternative, one can use restricted neighbourhoods characterized by subsets of moves
associated with larger λ values, such as Or-exchanges (Or, 1976) or the 4-opt∗ neighbour-
hood of Renaud, Boctor and Laporte (1996a) which considers only a subset of all potential
4-opt exchanges. Laporte and Semet (2002) have conducted a computational comparison
of some basic route improvement procedures. More complex inter-route neighbourhoods
are analyzed by Thompson and Psaraftis (1993), Van Breedam (1994), and Kindervater
and Savelsbergh (1997).

2.3.2 Metaheuristics

Several metaheuristics have been applied to the VRP. With respect to classical heuristics,
they perform a more thorough search of the solution space and are less likely to end with
a local optimum. These can be broadly divided into three classes: 1) local search, includ-
ing simulated annealing, deterministic annealing, and tabu search; 2) population search,
including genetic search and adaptive memory procedures; 3) learning mechanisms, includ-
ing neural networks and ant colony optimization. The best heuristics often combine ideas
borrowed from different metaheuristic principles. Recent surveys of VRP metaheuristics
can be found in Gendreau, Laporte and Potvin (2002), Cordeau and Laporte (2004), and
Cordeau et al. (2005).
Local search algorithms explore the solution space by iteratively moving from a solution
xt at iteration t to a solution xt+1 in the neighbourhood N (xt ) of xt until a stopping
criterion is satisfied. If f (x) denotes the cost of solution x, then f (xt+1 ) is not necessarily
smaller than f (xt ). As a result, mechanisms must be implemented to avoid cycling. In
simulated annealing, a solution x is drawn randomly from N (xt ). If f (x) ≤ f (xt ), then
xt+1 := x. Otherwise,
(
x with probability pt
xt+1 :=
xt with probability 1 − pt ,

where pt is a decreasing function of t and of f (x) − f (xt ). This probability is often equal
to
pt = exp(−[f (x) − f (xt )]/θt ),
where θt is the temperature at iteration t, usually defined as a non-increasing function of t.
Deterministic annealing (Dueck, 1990, 1993) is similar. There are two main versions of this
algorithm: in a threshold-accepting algorithm, xt+1 := x if f (x) < f (xt ) + θ1 , where θ1 is
a user controlled parameter; in record-to-record travel, a record is the best known solution
x∗ , and xt+1 := x if f (xt+1 ) < θ2 f (x∗ ), where θ2 is also user controlled. In tabu search, in
order to avoid cycling, any solution possessing some given attribute of xt+1 is declared tabu
for a number of iterations. At iteration t, the search moves to the best non-tabu solution

14
x in N (xt ). These local search algorithms are rarely implemented in their basic version,
and their success depends on the careful implementation of several mechanisms. The rule
employed to define neighbourhoods is critical to most local search heuristics. In simulated
annealing several rules have been proposed to define θt (see Osman, 1993). Tabu search
relies on various strategies to implement tabu tenures (also known as short term memory),
search diversification (also known as long term memory), and search intensification which
accentuates the search in a promising region.
Population search algorithms operate on several generations of solution populations.
In genetic search it is common to repeat the following operation k times: extract two
parent solutions from the populations to create two offspring using a crossover operation,
and apply a mutation operation to each offspring; then remove the 2k worst elements
from the population and replace them with the 2k offspring. Several crossover rules are
available for sequencing problems (Bean, 1994; Potvin, 1996; Drezner, 2003; Prins, 2004).
In adaptive memory procedures, an offspring is created by extracting and recombining
elements of several parents. In the initial version proposed by Rochat and Taillard (1995)
for the VRP, non-overlapping routes are extracted from several parents to create a partial
solution. This solution is then gradually completed and optimized by tabu search.
Neural networks are models composed of richly interconnected units through weighted
links, like neurons in the brain. They gradually construct a solution through a feedback
mechanism that modifies the link weights to better match an observed output to a described
output. In the field of vehicle routing neural network models called the elastic net and the
self-organizing map are deformable templates that adjust themselves to the contour of the
vertices to generate a feasible VRP solution. An example is provided by Ghaziri (1993).
Ant colony algorithms (see Dorigo, Di Caro and Gambardella, 1999) also use a learning
mechanism. They are derived from an analogy with ants which lay some pheromone
on their trail when foraging for food. With time more pheromone is deposited on the
more frequented trails. When constructing a VRP solution a move can be assigned a
higher probability of being selected if it has previously led to a better solution in previous
iterations.
In what follows we summarize the most effective metaheuristics for the CVRP. Initially
the best methods were almost exclusively based on tabu search but in recent years several
excellent methods inspired from different paradigms have been proposed.

Local search heuristics

A limited number of simulated annealing heuristics for the CVRP were proposed in the
early 1990s. Osman’s implementation (Osman, 1993) is the most involved and also the most
successful. It defines neighbourhoods by means of a 2-interchange scheme and applies a
different rule of temperature changes. Instead of using a non-increasing function, as do

15
most authors in the field, Osman decreases θt continuously as long as the solution improves,
but whenever xt+1 = xt , θt is either halved or replaced by the temperature at which the
incumbent was identified. This algorithm succeeded in producing good solutions but was
not competitive with the best tabu search implementations available at the same period.
A large number of tabu search algorithms have been produced over the past fifteen years
(a survey is available in Cordeau and Laporte, 2004). In the first known implementation,
due to Willard (1989), a CVRP solution is represented as a giant tour containing several
copies of the depot and inter-depot chains corresponding to feasible vehicle routes, and
neighbourhoods are defined by means of 3-opt exchanges. The method was soon to be
superseded by more powerful algorithms, including those of Osman (1993), Taillard (1993),
and Gendreau, Hertz and Laporte (1994).
Taillard’s algorithm remains to this day one of the most successful tabu search im-
plementations for the CVRP. It is based on the use of an 1-interchange mechanism to
define neighbour solutions, combined with periodic route reoptimizations by means of an
exact TSP algorithm (Volgenant and Jonker, 1983). The algorithm also uses random tabu
durations. A continuous diversification mechanism that penalizes frequently performed
moves is implemented in order to provide a more thorough exploration of the search space.
Finally, Taillard’s algorithm employs a decomposition scheme that allows for the use of
parallel computing. In planar problems the customer set is partitioned into sectors and
then concentric rings, while in random instances the regions are defined by means of short-
est spanning arborescences rooted at the depot. The region boundaries are periodically
updated to produce a diversification effect.
The Taburoute algorithm of Gendreau, Hertz and Laporte (1994) moves at each itera-
tion a vertex from its current route to another route containing one of its closest neighbours.
Insertions are performed simultaneously with a local reoptimization of the route, based on
the GENI procedure (Gendreau, Hertz and Laporte, 1992). Only a subset of vertices are
considered for reinsertion at any given iteration. No vertex can return to its former route
during the next θ iterations, where θ is randomly selected in a closed interval. Taburoute
also uses a continuous diversification mechanism. During the course of the search infeasible
solutions are penalized. This mechanism is implemented by replacing the solution value
f (x) associated at solution x with a penalized objective f ′ (x) = f (x) + αQ(x) + βL(x),
where Q(x) is the total capacity violation of solution x and L(x) is the total route length
violation. The two parameters α and β self-adjust during the search to produce a mix of
feasible and infeasible solutions: every µ iterations, α (resp. β) is divided by 2 if the past
µ solutions were feasible with respect to capacity (resp. route length), or multiplied by
2 if they were all infeasible with respect to capacity (resp. route length). Other features
of Taburoute include the use of random tabu durations, periodic route reoptimizations by
means of the US procedure of Gendreau, Hertz and Laporte (1992), false starts to initialize
the search, and a final intensification phase around the best known solution.

16
 The Rego and Roucairol (1996) Tabuchain algorithm is based on the use of ejection
chains involving ℓ routes to define neighbourhoods. This process bumps a vertex from
one route of the chain to another route. The last bumped vertex may be relocated in the
position of the first bumped vertex or elsewhere. The process ensures that no arc or edge
is considered more than once in the solution. As in Taburoute, intermediate infeasible
solutions are allowed. The authors have also implemented a sequential and a parallel
version of their method. Another ejection scheme, called Flower, was later developed by
Rego (1998). It is based on the idea of exploiting the representation of routes as blossoms
and of paths as stems, and of performing ejection moves by means of edge deletions and
creations. This method was not as successful as Tabuchain. Another method employing
ejection chains was developed by Xu and Kelly (1996). It oscillates between ejection
chains and vertex swaps between two routes. The ejection chains are obtained by solving
an auxiliary network flow problem. On the whole this method succeeded in obtaining
several good CVRP solutions on benchmark instances but it is rather involved and time
consuming.
More recently, Ergun et al. (2003) have developed a very large neighbourhood search
(VLNS) algorithm for the VRP. This algorithm operates on several routes simultaneously,
not unlike what is done in cyclic transfers (Thompson and Psaraftis, 1993) or in ejection
chains. Neighbourhoods are defined by a combination of 2-opt moves, vertex swaps between
routes, and vertex insertions in different routes. The best choice of moves and of routes
involved in the moves is determined through the solution of a network flow problem on
an auxiliary graph. One advantage of VLNS is that it allows a broad search by acting on
several routes at once. Its main disadvantage lies in the effort required at each iteration
to perform moves.
A very useful concept put forward by Toth and Vigo (2003) is that of granular tabu
search (GTS). This algorithm a priori removes from the graph long edges that are unlikely
to belong to an optimal solution. To determine these edges, the problem is first solved by
means of a fast heuristic, e.g., the Clarke and Wright (1964) algorithm, and the average edge
cost c̄ in this solution is determined. Then only two families of edges are retained: those
incident to the depot, and those whose cost does not exceed βc̄, where β is a user-defined
sparsification parameter. The authors show that on benchmark instances, choosing β in
[1.0, 2.0] yields the elimination of between 80% and 90% of all edges. Granular tabu search
was implemented in conjunction with some of the features of Taillard’s algorithm (Taillard,
1993) and Taburoute (Gendreau, Hertz and Laporte, 1994), and neighbour solutions were
obtained by performing intra-route and inter-route exchanges.
Deterministic annealing was first applied to the VRP by Golden et al. (1998) and more
recently by Li, Golden and Wasil (2005). The latter algorithm combines the record-to-
record principle of Dueck (1993) with GTS. It works on a sparsified graph containing only
a proportion α of the 40 shortest edges incident to each vertex, where α varies throughout

17
the algorithm. The algorithm is applied several times from three initial solutions generated
by the Clarke and Wright (1964) algorithm, with savings sij defined as ci0 + c0j − λcij ,
and λ = 0.6, 1.4 and 1.6. Neighbours are defined by means of intra- and inter-route 2-opt
moves, and non-improving solutions are accepted as long as their cost does not exceed that
of the incumbent by more than 1%. Whenever the solution has not improved for a number
of iterations, a perturbation is applied to the best known solution to restart the search.
This is achieved by temporarily moving some vertices to different positions.

Population search heuristics

The adaptive memory procedure (AMP) put forward by Rochat and Taillard (1995) con-
stitutes a major contribution to the field of metaheuristics. Initially developed in the
context of the VRP, it is of general applicability and has been used, for example, to solve
political districting problems (Bozkaya, Erkut and Laporte, 2003). An adaptive memory
is a pool of good solutions which is updated by replacing its worst elements with better
ones. In order to generate a new solution, several solutions are selected from the pool and
recombined. In the context of the VRP, vehicle routes are extracted from these solutions
and used as the basis of a new solution. The extraction process is applied as long as it
is possible to identify routes that do not overlap with previously selected routes. When
this is no longer possible, a search process (e.g., tabu search) is initiated from a partially
constructed solution made up of the selected routes and some unrouted customers. Any
solution constructed in this fashion replaces the worst solution of the pool if it has a better
cost. Tarantilis and Kiranoudis (2002) have proposed a variant to this scheme. In a first
phase a solution is obtained by means of the Paessens (1988) constructive procedure, which
is an application of the Clarke and Wright savings heuristic followed by 2-opt moves, vertex
swaps between routes, and vertex reinsertions. In order to generate new solutions from
the adaptive memory, Tarantilis and Kiranoudis extract route segments, called bones, as
opposed to full vehicle routes as did Rochat and Taillard.
Prins (2004) has developed an algorithm combining two main features of evolutionary
search, namely crossovers and mutations. Crossovers consist of creating offspring solutions
from parents, while mutations are obtained here by applying a local search algorithm to
an offspring. This combination of solution recombination and local search is sometimes
referred to as a memetic algorithm (Moscato and Cotta, 2003). In this algorithm, solutions
are represented as a giant tour without trip delimeters. To create an offspring from two
parents, a chain (i, . . . , j) is first selected from the first parent and the vertices of the
second parent are scanned from position j + 1 by skipping those of the chain (i, . . . , j). A
second offspring is generated in a similar way by reversing the roles of the two parents.
Offspring are improved by applying a combination of vertex and edge reinsertions, vertex
swaps, combined vertex and edge swaps.
Two other memetic algorithms have recently been proposed by Berger and Barkaoui

18
(2004) and by Mester and Bräysy (2005). The first works on two populations whose sizes
are kept constant through the replacement of parents by newly created offspring, and
migrations take place between the two populations. Offspring are obtained by combining
routes from two parents as long as this can be done without overlapping, and by inserting
the unrouted customers according to a proximity criterion. A VLNS heuristic (Shaw,
1998) combining three insertion mechanisms is then applied to the offspring, followed by
an improvement scheme consisting of removing vertices from the solution and reinserting
them by means of the I1 procedure of Solomon (1987).
The active guided evolution strategies (AGES) of Mester and Bräysy was initially de-
veloped to solve the VRP with time windows and was later applied to the classical VRP. It
combines local search (Voudouris, 1997) with an evolution strategy (Rechenberg, 1973) to
produce an iterative two-stage procedure. The evolutionary strategy uses a deterministic
rule to select a parent solution and create a single offspring from a single parent. The
offspring replaces the parent if it improves upon it. Offspring are improved by means of
an elaborate search procedure combining granular tabu search, continuous diversification,
vertex swaps and moves, 2-opt∗ moves (Potvin and Rousseau, 1995), VLNS (Shaw, 1998),
and restarts.

Learning mechanisms

A limited number of heuristics based on learning mechanisms have been proposed for the
VRP. None of the known neural networks based methods is satisfactory, and the early ant
colony based heuristics could not compete with the best available approaches. Recently,
however, Reimann, Doerner and Hartl (2004) have proposed a well-performing heuristics
called D-ants. The method repeatedly applies two phases until a stopping criterion is
reached. In the first phase, a first generation of good solutions is generated through the
applications of a savings based heuristic (Clarke and Wright, 1964) and a 2-opt improve-
ment procedure is applied to each solution. New generations of solutions are then created
by benefitting from the knowledge gained in producing past generations. Thus, instead
of using the standard savings sij = ci0 + c0j − cij , an attractiveness value χij = τijα sβij is
now employed, where τijα contains information on how good linking i and j turned out to
be in previous generations, and α and β are user-controlled

parameters. Vertices i and
P
j are linked with probability pij = χij / χhℓ , where Ωk is the set of the feasible
(h,ℓ)∈Ωk
(i, j) pairs yielding the k best savings. In the second phase the best solution identified
in the first phase is decomposed into subproblems which are then reoptimized using the
procedure used in the first phase.

19
Computational comparison of metaheuristics

Cordeau et al. (2005) provide a computational comparison of recent VRP heuristics on

the 14 Christofides et al. (1979) instances (50 ≤ n ≤ 199) and on the 20 larger Li, Golden
and Wasil (2005) instances (200 ≤ n ≤ 480). Most metaheuristics used in the comparison
consistently yield solutions whose value lies within 1% of the best known value.
On the Christofides, Mingozzi and Toth (1979) instances, the best solutions are obtained
by Taillard (1993), Rochat and Taillard (1995) and Mester and Bräysy (2005). If the two
instance sets are taken together, the best performers, in terms of accuracy and computing
time are probably Mester and Bräysy (2005), Tarantilis and Kiranoudis (2002), and Prins
(2004). It should be noted that these three methods all combine population search and
local search, thus allowing for a broad and deep exploration of the solution space.
As noted by Cordeau et al. (2002b) heuristics should not be judged solely on speed
and accuracy. Simplicity and flexibility are also important. In this respect the Li, Golden
and Wasil (2005) record-to-record algorithm is rather interesting: this algorithm possesses
a simple structure and is capable of generating very high quality solutions. As far as
flexibility is concerned, the granularity principle (Toth and Vigo, 2003) and the adaptive
memory concept (Rochat and Taillard, 1995) are general and useful ideas which can easily
be applied to other problems.

3 The Vehicle Routing Problem with Time Windows

The Vehicle Routing Problem with Time Windows (VRPTW) is an important generaliza-
tion of the classical VRP in which service at every customer i must start within a given
time window [ai , bi ]. A vehicle is allowed to arrive before ai and wait until the customer
becomes available, but arrivals after bi are prohibited. The VRPTW has numerous appli-
cations in distribution management. Common examples are beverage and food delivery,
newspaper delivery, and commercial and industrial waste collection (see, e.g., Golden, As-
sad and Wasil, 2002).
The VRPTW is NP-hard since it generalizes the CVRP which is obtained when ai = 0
and bi = ∞ for every customer i. In the case of a fixed fleet size, even finding a feasible
solution to the VRPTW is itself an NP-complete problem (Savelsbergh, 1985). As a result,
research on the VRPTW has concentrated on heuristics. Nevertheless, when the problem
is sufficiently constrained (i.e., when time windows are sufficiently narrow), realistic size
instances can be solved optimally through mathematical programming techniques. This
section presents a mathematical formulation of the VRPTW followed by a description of
some of the most important available exact and heuristic algorithms. It is worth point-
ing out that while exact methods usually minimize distance, most heuristics consider a
hierarchical objective which first minimizes the number of vehicles used and then distance.

20
3.1 Formulation of the VRPTW
The VRPTW can be defined on a directed graph G = (V, A) where |V | = n + 2, and the
depot is represented by the two vertices 0 and n+1. Feasible vehicle routes then correspond
to paths starting at vertex 0 and ending at vertex n + 1. The set of vehicles is denoted by
K, with |K| = m. Let si denote the service time at i (with s0 = sn+1 = 0) and let tij be the
travel time from i to j. In addition to the time window [ai , bi ] associated with each vertex
i ∈ N = V \ {0, n + 1}, time windows [a0 , b0 ] and [an+1 , bn+1 ] can also be associated with
the depot vertex. If no particular restrictions are imposed on vehicle availability, one may
simply set a0 = mini∈N {ai − t0i }, b0 = maxi∈N {bi − t0i }, an+1 = mini∈N {ai + si + ti,n+1 }
and bn+1 = maxi∈N {bi +si +ti,n+1 }. As in the CVRP, let qi denote the demand of customer
i, and let Q be the vehicle capacity.
While several models are available for the VRPTW, this problem is often formulated as
a multicommodity network flow model with time window and capacity constraints. This
model involves two types of variables: binary variables xkij , (i, j) ∈ A, k ∈ K, equal to 1
if and only if arc (i, j) is used by vehicle k, and continuous variables wik , i ∈ N, k ∈ K,
indicating the time at which vehicle k starts servicing vertex i. Let δ+ (i) = {j : (i, j) ∈ A}
and δ− (j) = {i : (i, j) ∈ A}. The problem can then be stated as follows (see, e.g.,
Desrochers et al. (1988)) :

X X
Minimize cij xkij (11)
k∈K (i,j)∈A

subject to
X X
xkij = 1 (i ∈ N ) (12)
k∈K j∈δ+ (i)
X
xk0j = 1 (k ∈ K) (13)
j∈δ+ (0)
X X
xkij − xkji = 0 (k ∈ K, j ∈ N ) (14)
i∈δ− (j) i∈δ+ (j)
X
xki,n+1 = 1 (k ∈ K) (15)
i∈δ− (n+1)

xkij (wik + si + tij − wjk ) ≤ 0 (k ∈ K, (i, j) ∈ A) (16)

ai ≤ wik ≤ bi (k ∈ K, i ∈ V ) (17)
X X
qi xkij ≤ Q (k ∈ K) (18)
i∈N j∈δ+ (i)

xkij ∈ {0, 1} (k ∈ K, (i, j) ∈ A). (19)

21
 The objective function (11) minimizes the total routing cost. Constraints (12) state that
each customer is visited exactly once, while constraints (13)–(15) ensure that each vehicle
is used exactly once and that flow conservation is satisfied at each customer vertex. The
consistency of the time variables wik is ensured through constraints (16) while time windows
are imposed by (17). These constraints also eliminate subtours. Finally, constraints (18)
enforce the vehicle capacity restriction.
Formulation (11)–(19) is non-linear because of constraints (16). These constraints can,
however, be linearized as follows:

wjk ≥ wik + si + tij − Mij (1 − xkij ) (k ∈ K, (i, j) ∈ A), (20)

where Mij = max{0, bi + si + tij − aj } is a constant. As suggested by Desrochers and

Laporte (1991), the bounds on the time variables bki can also be strengthened:
X
wik ≥ ai + max{0, aj − ai + sj + tji }xkji (k ∈ K, i ∈ V ) (21)
j∈δ− (i)
X
wik ≤ bi − max{0, bi − bj + si + tij }xkij (k ∈ K, i ∈ V ). (22)
j∈δ+ (i)

3.2 Exact algorithms for the VRPTW

As for most other vehicle routing problems, it is difficult to solve the VRPTW exactly
through classical simplex-based branch-and-bound methods, even for small instances. This
is in large part explained by the fact that the LP relaxation of the problem provides a weak
lower bound. The first optimization algorithm for the VRPTW can be attributed to Kolen,
Rinnooy Kan and Trienekens (1987) who used dynamic programming coupled with state
space relaxation (Christofides, Mingozzi and Toth, 1981b) to compute lower bounds within
a branch-and-bound algorithm. Instances with n ≤ 15 were solved using this approach.
Most subsequent algorithms rely either on the generation of valid inequalities to strengthen
the LP relaxation or on mathematical decomposition techniques. This section reviews the
three main available approaches: Lagrangean relaxation, column generation and branch-
and-cut. Additional references on the subject can also be found in the Cordeau et al.
(2002a) review.

3.2.1 Lagrangean relaxation based algorithms

Lagrangean relaxation can be applied to the VRPTW in several ways. It is well known that
when the subproblem obtained by relaxing some of the constraints possesses the integrality
property, the best lower bound obtained by Lagrangean relaxation (i.e., the value of the
Lagrangean dual) is equal to the value of the linear programming relaxation of the original

22
problem. But as mentioned above, the LP relaxation of formulation (11)–(19) provides a
weak lower bound which will usually prevent the problem from being solved by branch-and-
bound. As a result, successful implementations of Lagrangean relaxation for the VRPTW
should retain at least some of the complicating constraints in the subproblem.
Fisher (1994) and Fisher, Jörnsten and Madsen (1997) described a Lagrangean relax-
ation based on m-trees (see Section 2.2.1). This approach relaxes the flow conservation
constraints as well as the capacity and time window constraints. Violated capacity con-
straints are handled by identifying subsets of customers S ⊆ N that must be visited by at
least κ(S) vehicles and imposing the constraint
X X X
xkij ≥ κ(S). (23)
k∈K i∈V \S j∈S

These constraints are relaxed in a Lagrangean fashion so that the resulting problem re-
mains an m-tree problem with modified costs. Time windows are handled similarly by
identifying infeasible paths and imposing the constraint that at least one arc in the path
be left out of the solution. This approach has solved a few of the Solomon (1987) test
instances with n = 100. In addition to the m-tree relaxation method, Fisher, Jörnsten and
Madsen (1997) have also experimented with a variable splitting approach in which addi-
tional variables yik , equal to 1 if and only if customer i is visited by vehicle k, are introduced
in the formulation, and the constraints j∈V xkij = yik (i ∈ N, k ∈ K) are dualized. The
P

Lagrangean subproblem decomposes into a semi-assignment problem in the yik variables

which is solvable by inspection, and a set of m elementary shortest path problems with
time windows and capacity constraints.
Another possible Lagrangean relaxation consists of dualizing the demand constraints.
Let λ = (λi ) (i ∈ N ) be the vector of multipliers associated with constraints (12) requiring
that each customer be visited exactly once. For given values of the multipliers, the La-
grangean subproblem L(λ) obtained by relaxing these constraints in the objective function
is
X X X
min (cij − λi ) xkij + λi , (24)
k∈K (i,j)∈A i∈N

subject to constraints (13)–(19).

This subproblem does not possess the integrality property. It does, however, decom-
pose into m disjoint elementary shortest-path problems with capacity and time window
constraints. When all vehicles are identical, a single problem can be solved to compute the
lower bound. The Lagrangean dual, i.e., the problem of finding optimal multipliers that
maximize L(λ), is a concave non-differentiable maximization problem. Using subgradient
and bundle methods, Kohl and Madsen (1997) were able to solve some instances with

23
up to 100 customers. They reported optimal solutions to each of the 27 clustered and
short-horizon Solomon instances.
Kallehauge, Larsen and Madsen (2006) have developed a stabilized cutting-plane al-
gorithm to solve the Lagrangean dual. Cutting planes are generated by solving the La-
grangean subproblem and are introduced in a master problem which imposes bounds (i.e.,
a trust region) on the dual variables to ensure the stability of their values from one itera-
tion to the next. Optimizing the relaxed master problem (a maximization linear program)
provides a lower bound on the value of the original problem. To obtain feasible integer
solutions, the cutting-plane algorithm is embedded within a branch-and-bound algorithm
and valid inequalities are introduced in the master problem. Because the relaxed master
problem is stated on the dual variables, violated subtour elimination constraints and 2-
path inequalities (see Section 3.2.2) are added as columns to this problem. This approach
has yielded good results on the Solomon test instances and was able to solve two large
instances with 400 and 1000 customers, respectively.

3.2.2 Column generation algorithms

Column generation is intimately related to constraint generation and can be seen as a

special way of updating the multipliers associated with the relaxed constraints. Let Ωk
denote the set of feasible paths for vehicle k ∈ K. For each path ω ∈ Ωk , let ckω be the
cost of this path and let θωk be a binary variable equal to 1 if and only if vehicle k uses
path ω. Let also aiω be the number of times customer i ∈ N is visited by path ω. As first
suggested by Balinski and Quandt (1964), the VRPTW can be stated as follows:
X X
Minimize ckω θωk (25)
k∈K ω∈Ωk

subject to
X X
aiω θωk = 1 (i ∈ N ) (26)
k∈K ω∈Ωk
X
θωk = 1 (k ∈ K) (27)
ω∈Ωk
θωk ∈ {0, 1} (k ∈ K, ω ∈ Ωk ). (28)

Because the sets Ωk are likely to have a very large cardinality, this problem can be solved
by a branch-and-bound algorithm in which the linear relaxations are solved by column
generation. At each node of the enumeration tree, a restricted column generation master
problem is solved over the current set of columns. New columns of negative reduced cost are
generated by solving a resource constrained shortest path problem (13)–(19) with modified

24
arc costs reflecting the current values of the dual variables associated with the constraints
of the column generation master problem. This process stops when no negative reduced
cost column can be generated. Because the column generation subproblem is equivalent
to the Lagrangean subproblem L(λ), the lower bound provided by column generation is
equal to the value of the Lagrangean dual. The dual of the LP relaxation of formulation
(25)–(28) is in fact equivalent to the Lagrangean dual defined in the previous section. This
formulation can also be obtained by applying the Dantzig-Wolfe decomposition principle
(Dantzig and Wolfe, 1960) to the original formulation (11)–(19).
Branching must be performed at each node of the branch-and-bound tree where the
optimal solution to the linear relaxation includes fractional path variables. While it is in
principle possible to branch directly on fractional θω variables, this approach is difficult to
implement in practice. Indeed, it is easy to set such variables equal to 1 but it is much
more difficult to impose the opposite decision. In the latter case, care must be taken
to ensure that the same path will not be generated more than once by the subproblem.
To this purpose, one could use a modified dynamic programming algorithm to implicitly
handle forbidden paths, or a p-shortest path algorithm where p is equal to the number of
forbidden paths plus one. This would ensure the generation of at least one valid path of
negative reduced cost whenever one exists. A more convenient branching scheme consists
of making decisions on the original arc flow variables xkij or on sums of these variables. For
example, binary decisions can be made on the following sum of variables:
X X
xkij ,
j∈N ′ k∈K ′

where i ∈ N , N ′ ⊆ δ+ (i) and K ′ ⊆ K. Forcing this sum to be equal to 1 requires that

some vertex in subset N ′ be visited immediately after i by some vehicle. If |N ′ | = 1, then
the corresponding vertex must be visited after i by some vehicle. If |K ′ | = 1, then vertex
i is implicitly assigned to vehicle k. The special case |N ′ | = 1 and |K ′ | = 1 is equivalent
to forcing xkij = 1 for some given j and k. It is worth pointing out that all such decisions
can be handled directly at the subproblem level through the simple elimination of arcs in
the networks.
Column generation was successfully applied to the VRPTW by Desrochers, Desrosiers
and Solomon (1992) and by Kohl et al. (1999). The latter authors also used valid inequal-
ities to strengthen the bounds obtained by column generation. More specifically, let
X X X
x(S) = xkij
k∈K i∈V \S j∈S

denote the flow into set S ⊆ N and denote by κ(S) the minimum number of vehicles needed
to serve all customers in S. Then the constraint

x(S) ≥ κ(S) (29)

25
is a valid inequality for the VRPTW and is called a κ-path inequality. Computing κ(S) is
a difficult problem which is equivalent to solving the VRPTW on a subset of vertices with
the objective of minimizing the number of vehicles used. Kohl et al. (1999) have in fact
restricted their attention to the case κ = 2. Determining whether κ(S) = 1 for a particular
subset S can be achieved by checking that the capacity of a single vehicle is sufficient and
the corresponding TSPTW is feasible. The latter problem is NP-hard but can be solved
relatively quickly by dynamic programming for small instances. The algorithm of Kohl
et al. was capable of solving 70 of the 87 Solomon short-horizon instances to optimality.
Cook and Rich (1999) have extended this approach to the case κ ≤ 6 by using parallel
computing and replacing the TSPTW feasibility problem with a VRPTW. They were thus
able to solve 80 of the short-horizon instances. They also solved 30 of the 81 long-horizon
instances.
While the constrained elementary shortest path problem is NP-hard, the relaxation
obtained by allowing cycles can be solved by a pseudo-polynomial labeling algorithm (see,
e.g., Desrochers and Soumis, 1988). Because of time windows and capacity constraints,
these cycles will nevertheless be of finite length. This relaxation will of course weaken the
value of the lower bound, but cycle elimination procedures can be used to circumvent this
difficulty. A procedure for eliminating 2-cycles (i.e., cycles of the form (i, j, i)) was first
proposed by Houck et al. (1980). More recently, Irnich and Villeneuve (2003) developed
an efficient approach to forbid cycles of length greater than 2. Experiments performed by
the authors show that k-cycle elimination with k ≥ 3 can substantially improve the lower
bounds. Embedding this technique within column generation enabled the exact solution
of 15 previously unsolved instances of the Solomon benchmark set.
Recently, Chabrier (2006) proposed a modified labeling algorithm to handle the con-
strained elementary shortest path problem and thus obtain improved lower bounds. In
this algorithm, both exact and heuristic dominance rules are considered. Whenever the
heuristic approach cannot find a path of negative reduced cost, the exact but slower im-
plementation is used. This approach has allowed the author to find the optimal solution
to 17 previously unsolved long-horizon instances from the Solomon benchmark set.
Promising results were also reported by Danna and Le Pape (2003) who developed a
cooperation scheme between column generation and local search applied to the VRPTW.
During the branch-and-price process, local search is regularly applied from the best known
integer solution. This often results in an improved upper bound that can then be used
to prune nodes in the enumeration tree. Furthermore, columns associated with solutions
identified during local search can be fed into the restricted master problem. The branch-
and-price algorithm thus benefits from local search by being provided at an early stage with
high quality upper bounds, resulting in a smaller search tree. In turn, local search benefits
from branch-and-price by working with a variety of different initial solutions, resulting in
an effective form of diversification.

26
3.2.3 A branch-and-cut algorithm

A branch-and-cut algorithm for the VRPTW was developed by Bard, Kontoravdis and Yu
(2002). As in most such algorithms for the VRP, the problem is formulated using two-
index variables xij equal to 1 if and only if a vehicle travels directly from vertex i to vertex
j. The algorithm incorporates five types of inequalities: subtour elimination constraints,
capacity constraints, comb inequalities, incompatible pair inequalities, and incompatible
path inequalities. At each node of the search tree an upper bound is computed by means
of the greedy randomized adaptive search procedure (GRASP) described by Kontoravdis
and Bard (1995).
Incompatible pair inequalities rely on the existence of vertex pairs that cannot be-
long to the same vehicle route. If i and j denote two incompatible vertices and P =
(i, h1 , . . . , h|P|−2 , j) is a path, then the following inequality is valid:

xi1 ,h1 + xh1 ,i + · · · + xh|P|−2 ,j + xj,h|P|−2 ≤ |P| − 2. (30)

Incompatible path inequalities are similar to infeasible pair inequalities but take arc
orientations into account. If i and j are two vertices such that i cannot precede j in a
feasible vehicle route then the following inequality is valid for any path P between i and j:

xi,h1 + xh1 ,h2 + · · · + xh|P|−2 ,j ≤ |P| − 2. (31)

The authors present four separation heuristics to identify violated capacity constraints.
The first is based on the computation of minimum cuts in G. The second applies a graph
shrinking heuristic similar to that proposed by Araque et al. (1994) for the VRP. The third
consists of identifying connected components in G that do not contain the depot. Finally,
the fourth is a heuristic proposed by Kohl et al. (1999) to identify violated 2-path inequal-
ities. Heuristic separation algorithms are also described for the identification of violated
comb inequalities, incompatible path inequalities and incompatible pair inequalities. The
branch-and-cut algorithm of Bard, Kontoravdis and Yu has obtained good results on the
Solomon test instances: all 50-customer instances and several 100-customer instances were
solved optimally.

3.3 Heuristics for the VRPTW

Because of the difficulty of the VRPTW and its high practical relevance, there is a genuine
need to develop fast algorithms capable of producing good quality solutions in short com-
puting times. Heuristics can also be used to provide upper bounds for the exact algorithms
described in the previous section. This section describes the three main classes of heuristics
for the VRPTW: construction heuristics, improvement heuristics and metaheuristics.

27
3.3.1 Construction heuristics

Route construction algorithms work by inserting customers one at a time into partial routes
until a feasible solution is obtained (see Section 2.3.1). Routes can either be constructed
sequentially or in parallel. Construction algorithms are mainly distinguished by the order
in which customers are selected and by the method used to determine where a customer
should be inserted.
Several sequential insertion heuristics for the VRPTW were proposed by Solomon
(1987). Among these heuristics, the most efficient, called I1, consists of first selecting
the farthest customer from the depot as a seed customer. The remaining customers are
then inserted one at a time into the current route by selecting at each iteration the cus-
tomer that maximizes a savings measure taking into account the distance from the depot
and the cost of insertion in the current route. The customer is then inserted in the position
minimizing a weighted combination of extra distance and extra time required to visit the
customer. The process is repeated until all customers have been inserted or it is no longer
possible to insert additional customers without violating either the capacity or time win-
dow constraints, at which point a new route is initialized by selecting a new seed customer
and the process repeats itself until no customers remain.
A parallel version of this heuristic was later developed by Potvin and Rousseau (1993)
who proposed a generalized regret measure to select the next customer for insertion. This
measure reflects the cost increase likely to result if a customer is not assigned to the route
minimizing the insertion cost. Further improvements to the sequential heuristic of Solomon
(1987) were also described by Ioannou, Kritikos and Prastacos (2001) who proposed to
modify the criteria for customer selection and insertion to take into account the impact of
the insertion on all routed and unrouted customers.

3.3.2 Improvement heuristics

Improvement heuristics iteratively improve an initial feasible solution by performing ex-

changes while maintaining feasibility. The process normally stops when no further exchange
can be made without deteriorating the solution. Improvement heuristics are mainly char-
acterized by the type of exchanges considered at each iteration. These define the neigh-
bourhood of a solution, i.e., the set of solutions reachable from the current solution by
performing a single exchange.
The first improvement heuristics for the VRPTW (see, e.g., Russell, 1977; Baker and
Schaffer, 1986) were adaptations of the 2-opt (Croes, 1958), 3-opt (Lin, 1965) and Or-opt
(Or, 1976) edge exchange mechanisms originally introduced for the TSP. Because of time
windows, checking whether a given exchange maintains feasibility of the solution can be
rather time consuming. Starting with the work of Savelsbergh (1985), several attempts have
been made to develop efficient implementations of neighbourhood evaluation procedures for

28
λ-exchanges (see also Solomon, Baker and Schaffer (1988) and Savelsbergh (1990, 1992)). A
comparison of 2-opt, 3-opt and Or-opt exchange heuristics for the VRPTW was performed
by Potvin and Rousseau (1995) who also introduced a new exchange, called 2-opt*, a special
case of 2-opt that maintains the orientation of the subroutes involved in the exchange. This
is accomplished by removing the last n1 customers from a route k1 , inserting them after
the first n2 customers of a route k2 and reconnecting the initial part of route k1 with the
terminal part of route k2 . Another exchange mechanism was described by Thompson and
Psaraftis (1993) who proposed transferring sets of customers in a cyclic fashion between
routes.
Several attempts have also been made to integrate construction and improvement
heuristics. Russell (1995) developed a procedure that embeds route improvement within
the solution construction process. More precisely, customers can be switched between
routes, and routes can be eliminated during the construction of the solution which is per-
formed by a procedure similar to that of Potvin and Rousseau (1993). More recently,
Cordone and Wolfler Calvo (2001) have proposed a composite heuristic in which a set of
initial solutions is first constructed by means of Solomon’s I1 insertion heuristic and an
improvement procedure is then applied to each of them. This procedure applies 2-opt and
3-opt exchanges and attempts to reduce the number of routes by relocating customers. To
escape from local optima, the heuristic alternates between an objective minimizing total
distance and an objective minimizing total route duration (the primary objective being in
both cases to minimize the number of routes). Several deterministic local search heuristics
were also proposed by Bräysy (2002), based on a new three-phase approach. In a first
phase, an initial solution is created with one of two proposed route construction heuristics
(a cheapest insertion-based heuristic with periodic route improvements and a parallel sav-
ings heuristic). The second phase attempts to reduce the number of routes by applying a
local search operator based on ejection chains (see, e.g., Glover, 1992). Finally, the third
phase applies Or-opt exchanges to reduce the total length of the routes.

3.3.3 Metaheuristics

Most of the recent research on approximate algorithms for the VRPTW has concentrated on
the development of metaheuristics. Unlike classical improvement methods, metaheuristics
usually incorporate mechanisms to continue the exploration of the search space after a
local minimum is encountered.

Tabu search heuristics

Some of the first applications of tabu search to the VRPTW can be attributed to Semet and
Taillard (1993) and to Potvin et al. (1996) who combined Solomon’s insertion heuristics
with improvement schemes based on vertex and chain exchange procedures.

29
 A more sophisticated algorithm was later developed by Taillard et al. (1997) for the
VRP with soft time windows in which vehicles are allowed to arrive late at customer lo-
cations but time window violations are penalized in the objective function. This heuristic
relies on the concept of adaptive memory introduced by Rochat and Taillard (1995) and on
the decomposition and reconstruction procedure developed by Taillard (1993) for the clas-
sical VRP. An adaptive memory is a pool of routes extracted from the best solutions found
during the search. This memory is first initialized with routes produced by a randomized
insertion heuristic. At each iteration of the metaheuristic, a solution is constructed from
the routes belonging to the adaptive memory and is improved through tabu search. The
routes of the resulting solution are then stored in the adaptive memory if this solution
improves upon the worst solution already stored. The tabu search heuristic uses an ex-
change operator, called CROSS exchange, which swaps sequences of consecutive customers
between two routes. Individual routes are also optimized by removing two edges from a
route and moving the segment between these two edges to another location within the
route. A parallel computing implementation of this approach is described in Badeau et al.
(1997).
A metaheuristic embedding reactive tabu search (see, e.g., Battiti and Tecchiolli, 1994)
within the parallel construction heuristic of Russell (1995) was developed by Chiang and
Russell (1997). In this implementation, the tabu list length is increased if identical so-
lutions occur too frequently and is decreased if no feasible solution can be found. Using
a variety of customer ordering rules and criteria for measuring the best insertion points,
the metaheuristic first constructs six different initial solutions by gradually inserting cus-
tomers and repeatedly applying tabu search to the the partial solutions. The best solution
obtained after this step is further improved through tabu search. Exchanges are performed
by using some of the λ-interchanges of Osman (1993): switch a customer from one route
to another and swap two customers belonging to different routes.
More recently, a tabu search heuristic was developed by Cordeau, Laporte and Mercier
(2001) for the VRPTW and two of its generalizations: the periodic VRPTW and the multi-
depot VRPTW (see also Cordeau, Gendreau and Laporte (1997)). In this heuristic, an
initial solution is obtained by means of a modified sweep heuristic. Infeasible solutions are
allowed during the search and violations of capacity, duration or time window constraints
are penalized in the objective function through dynamically updated penalty factors. At
each iteration of the tabu search, a customer is removed from its current route and inserted
into a different route by using a least cost insertion criterion. A continuous diversification
mechanism that penalizes frequently made exchanges is used to drive the search process
away from local optima. Finally, a post-optimizer based on a specialized TSPTW heuristic
(Gendreau et al. (1998)) is applied to individual routes. An improvement to this heuristic
for the handling of route duration constraints was recently described by Cordeau, Laporte
and Mercier (2004). The heuristic was also extended by Cordeau and Laporte (2001)

30
to handle heterogeneous vehicles. Other tabu search algorithms for the VRPTW were
proposed by Brandão (1998), Schulze and Fahle (1999) and Lau, Sim and Teo (2003).

Genetic algorithms

Homberger and Gehring (1999) have described two evolution strategies for the VRPTW.
Both are based on the (µ, λ) strategy: starting from a population with µ individuals,
subsets of individuals are randomly selected and recombined to yield a total of λ > µ
offspring. Each offspring is then subjected to a mutation operator, and the µ fittest are
selected to form the new population. In the first method, new individuals are generated
directly through mutations and no recombination takes place. Mutations are obtained by
performing one or several moves from the 2-opt, Or-opt and 1-interchange families. In
the second method, offspring are generated through a two-step recombination procedure
in which three individuals are involved. In both methods, the fitness of an individual
depends first on the number of vehicles used, and second on the total distance traveled.
Gehring and Homberger (2002) later proposed a two-phase metaheuristic in which the first
phase minimizes the number of vehicles through an evolution strategy, while the second
one minimizes the total distance through tabu search. A parallelization strategy is also
used to run several concurrent searches of the solution space with differently configured
metaheuristics cooperating through the exchange of solutions.
Berger, Barkaoui and Bräysy (2003) have developed a genetic algorithm that concur-
rently evolves two distinct populations pursuing different objectives under partial con-
straint relaxation. The first population aims to minimize the total distance traveled while
the second one focuses on minimizing the violations of the time window constraints. The
maximum number of vehicles imposed in the first population is equal to kmin whereas the
second population is allowed only kmin − 1 vehicles, where kmin refers to the number of
routes in the best known feasible solution. Whenever a new feasible solution emerges from
the second population, the first population is replaced with the second and the value of
kmin is updated accordingly. Two recombination operators and five mutation operators
are used to evolve the populations. This approach has proved to be rather efficient in
minimizing the number of vehicles used.
More recently, Mester and Bräysy (2005) have developed an iterative metaheuristic that
combines guided local search and evolution strategies. An initial solution is first created
by an insertion heuristic. This solution is then improved by the application of a two-stage
procedure. The first stage consists of a guided local search procedure in which 2-opt* and
Or-opt exchanges are performed together with 1-interchanges. This local search is guided
by penalizing long arcs appearing often in local minima. The second stage iteratively
removes a selected set of customers from the current solution and reinserts the removed
customers at minimum cost. These two stages are themselves repeated iteratively until no
further improvement can be obtained. Very good results are reported by the authors on

31
large-scale instances. According to Bräysy and Gendreau (2005b), the three approaches
just described seem to produce the best results among genetic algorithms. Other such
algorithms have also been proposed by a number of researchers including Potvin and Bengio
(1996), Thangiah and Petrovic (1998) and Tan, Lee and Ou (2001).

Other metaheuristics

Kontoravdis and Bard (1995) have described a two-phase GRASP for the VRPTW. A num-
ber of routes are first initialized by selecting seed customers. The remaining customers are
then gradually inserted in the routes by using a randomized least insertion cost procedure.
During this process, periodic attempts are made to improve the routes by local search. In
this phase certain routes may be eliminated by means of a deterministic procedure that
attemps to relocate the customers to a different route. To estimate the required number
of routes, the authors have proposed three lower bounds for fleet size. Two are based on
bin packing structures generated by the capacity or time window constraints. The other
is derived from the associated graph created by pairs of customers having incompatible
demands or time windows.
A guided local search algorithm for the VRPTW was introduced by Kilby, Prosser
and Shaw (1998). In guided local search, the objective function is augmented with a
penalty term reflecting the proximity of the current solution value to that of previously
encountered local minima. The method is used to drive a local search heuristic that
modifies the current solution by performing one of four moves: 2-opt exchanges within a
route, switching a customer from one route to another, exchanging customers belonging
to two different routes, and swapping the ends of two routes. All customers are first
assigned to a virtual vehicle and the routes for the actual vehicles are left empty. Because
a penalty is associated with not visiting a customer, a feasible solution will be constructed
in the process of minimizing cost. The local search algorithm starts from this solution and
performs a series of exchanges until a local minimum is reached. The objective function is
then modified by adding a term penalizing the presence of the arcs used in this solution.
The search iterates by finding new local minima and accumulating penalties until a stopping
criterion is met. This approach was later coupled with tabu search and embedded within
a constraint programming framework by De Backer et al. (2000).
Gambardella, Taillard and Agazzi (1999) have developed an ant colony optimization al-
gorithm for the VRPTW which associates an attractiveness measure to the arcs. Artificial
ants represent parallel processes whole role is to construct feasible solutions. To deal with
the hierarchical objective of first minimizing the number of vehicles and then minimizing
distance, two ant colonies are used, each dedicated to the optimization of a different objec-
tive. These colonies cooperate by exchanging information through pheromone updating.
Whenever a feasible solution with a smaller number of vehicles is found, both colonies are
reactivated with the reduced number of vehicles.

32
 Bent and Van Hentenryck (2004) have described a two-stage hybrid algorithm that first
minimizes the number of routes by simulated annealing and then minimizes total distance
traveled by using a large neighbourhood search (Shaw, 1998) which may relocate a large
number of customers. The first stage uses a lexicographic evaluation function to minimize
the number of routes, maximize the sum of the squares of the route sizes, and minimize the
minimal delay (a measure of time window tightness) of the solution. The neighbourhood
used in this stage consists of 2-opt, Or-opt, relocating, exchange and crossover moves. In
the second stage, subsets of customers are removed from their current route and reinserted
in possibly different routes. Customers selected for removal are randomly chosen but
the algorithm favors customers that are geographically close to each other and belong to
different routes. A branch-and-bound algorithm is then used to reinsert these customers.
A four-phase metaheuristic based on a modification of the variable neighbourhood
search was described by Bräysy (2003). In the first phase, an initial solution is created by
using route construction heuristics. During this process, the partial routes are periodically
reoptimized through Or-opt exchanges. In the second phase, an attempt is made to reduce
the number of routes by applying a route eliminination operator based on ejection chains.
In the third phase, four local search procedures embedded in a variable neighbourhood
search (see, e.g., Mladenović and Hansen, 1997) are applied to reduce the total distance
traveled. These procedures are based on modifications to the CROSS exchanges of Taillard
et al. (1997) and cheapest insertion heuristics. In the fourth phase, a modified objective
function considering waiting time is used by the local search operators in the hope of
further improving the solution.
More recently, a local search algorithm with restarts was also proposed by Li and Lim
(2003). This algorithm first constructs an initial solution by using an insertion heuristic.
Local search is then performed from this solution using three exchange operators that
move segments of customers either between routes or within the same route. Whenever a
local minimum is reached, multiple restarts are performed starting from the best known
solution, and a tabu list is used to prevent cycling.
A large number of other metaheuristics based on various paradigms have been described
in recent years. For additional references on approximate algorithms for the VRPTW as
well as detailed computational experiments, the reader is referred to recent surveys by
Bräysy and Gendreau (2005a,b).

4 The Inventory Routing Problem

The Inventory Routing Problem (IRP) is an important extension of the VRP which inte-
grates routing decisions with inventory control. The problem arises in environments where
vendor managed inventory (VMI) resupply policies are employed. These policies allow a
vendor to choose the timing and size of deliveries. In exchange for this freedom, the ven-

33
dor agrees to ensure that its customers do not run out of product. In a more traditional
relationship, where customers call in their orders, large inefficiencies can occur due to the
timing of customers’ orders (resulting in high inventory and distribution costs). Realiz-
ing the cost savings opportunities of vendor managed inventory policies, however, is not
a simple task, particularly with a large number and variety of customers. The inventory
routing problem achieves this goal by determining a distribution strategy that minimizes
long term distribution costs. This description of the inventory routing problem focuses
primarily on distribution. Inventory control is restricted to ensuring that no stockouts
occur at the customers. Inventory control takes a more prominent role when inventory
holding costs are considered. In the inventory control literature, the resulting environment
is usually referred to as a one warehouse multi-retailer system.
Inventory routing problems are very different from VRPs. Vehicle routing problems
occur when customers place orders and the vendor, on any given day, assigns the orders
for that day to routes for vehicles. In IRPs, the delivery company, not the customer, decides
how much to deliver to which customers each day. There are no customer orders. Instead,
the delivery company operates under the restriction that its customers are not allowed
to run out of product. Another key difference is the planning horizon. Vehicle routing
problems typically deal with a single day, the only requirement being that all orders have
to be delivered by the end of the day. Inventory routing problems are defined on a longer
horizon. Each day the delivery company makes decisions about which customers to visit
and how much to deliver to each of them, while keeping in mind that decisions made today
have an impact on what has to be done in the future. The objective is to minimize the
total cost over the planning horizon while ensuring that no customer runs out of product.

4.1 Definition of the IRP

The deterministic IRP is concerned with the repeated distribution of a single product from
a single facility, to a set of n customers over a planning horizon of length T , possibly infinity.
Customer i consumes the product at a rate ui (say volume per day) and can maintain a
local inventory of product of up to a maximum of Ci . The inventory at customer i is Ii0 at
time 0. A fleet of m homogeneous vehicles, with capacity D, is available for the distribution
of the product. If a quantity di is delivered at customer i, the vendor earns a reward equal
to ri di . It takes a vehicle a time tij to traverse arc (i, j) of the distribution network and a
cost cij is incurred when doing so. The objective is to maximize the profit (revenue minus
cost) over the planning horizon, without causing stockouts at any of the customers. (Note
that because product usage is assumed to be deterministic and no stockouts are allowed,
long run revenues are fixed and the key is to reduce delivery costs.) A dispatcher has to
decide when to serve a customer, how much to deliver, and which delivery routes to use to
serve customers.
In the stochastic IRP customer demands are defined at discrete time instants t by

34
means of random variables. Let Ut = (U1t , . . . , Unt ) denote the vector of random customer
demands at time t. Customer demands on different days are independent random vectors
with a joint probability distribution F that does not change with time; that is, U0 , U1 , . . . is
an independent and identically distributed sequence, and F is the probability distribution
of each Ut . The probability distribution F is known to the decision maker. The vendor
can measure the inventory level Xit of each customer i at any time t. At each time t,
the vendor makes a decision that controls the routing of vehicles and the replenishment of
customer inventories. Because demand is uncertain, there is often a positive probability
that a customer will run out of stock, and thus shortages cannot always be prevented.
Shortages result in a penalty pi si if the unsatisfied demand on day t at customer i is si .
Unsatisfied demand is treated as lost demand. The objective is to construct a distribution
policy maximizing the expected discounted profit over an infinite time horizon.

4.2 Motivating example

To illustrate the difficulty of inventory routing problems, we reproduce a small deterministic
example introduced by Fisher et al. (1982) and Bell et al. (1983). The relevant optimal
tour costs can be derived from the network shown in Figure 1, e.g., the optimal tour costs
for visiting customers 1 and 2, denoted by C1,2 , is equal to $210. The vehicle capacity is
5000 gallons and customer tank capacity and usage data, in gallons, are as follows:
Customer i di ui
1 5000 1000
2 3000 3000
3 2000 2000
4 4000 1500

A simple schedule jointly replenishes customers 1 and 2 as well as customers 3 and 4 on

a daily basis. This schedule is natural because customers 1 and 2 (3 and 4, respectively)
are near each other. Each customer i receives a quantity equal to its daily consumption
ui . The long-run average cost of this schedule is 420 miles per day. An improved schedule
consists of a cycle that repeats itself every two days. On the first day, one trip replenishes
3000 gallons to customer 2 and 2000 gallons to customer 3, at a cost of 340 miles. On the
second day, two trips are made. The first trip replenishes 2000 gallons to customer 1 and
3000 gallons to customer 2. The second trip replenishes 2000 gallons to customer 3 and
3000 gallons to customer 4. Each trip costs 210 miles. The average cost of this schedule is
380 miles per day, which is nearly 10% lower than the first schedule.

4.3 Observations on the IRP

Before describing solution approaches, we present some general observations concerning
inventory routing problems and some common elements found in most solution approaches.

35
 10
1 2

140

100 100

100 3

10

100
4

Figure 1: A four-customer example with distances shown on edges.

The IRP is a long-term dynamic control problem which is extremely difficult to solve.
Therefore, most of the available algorithms solve only a short-term planning problem. In
early publications, it was often just a single day but later, short-term was expanded to a few
days. Two key issues need to be resolved with such approaches: how to model the long-term
effect of short-term decisions, and which customers to consider in the short-term planning
period. A short-term approach that only minimizes costs has the tendency to defer as many
deliveries as possible to future planning periods, which may lead to an undesirable situation
in the future. Therefore, a proper incorporation of the long-term objective into the short-
term planning problem is essential. The long-term effect of short-term decisions needs to
capture the costs and benefits of delivering to a customer earlier than necessary. This
usually means delivering less and may lead to higher future distribution costs, but reduces
the risk of a stockout and may thus reduce future shortage costs. Decisions regarding which
customers need to be considered in the short-term planning period are usually guided by
some measure of the urgency to make a delivery to a customer and the quantity that can
be delivered. Usually, it is assumed that customers considered in the short-term planning
period may actually be visited, but the decision whether or not to actually visit them still
has to be made.
When the short-term planning problem consists of a single day, the problem can be
viewed as an extension of the VRP and solution techniques for the VRP can be adapted.
For example, Campbell and Savelsbergh (2004c) have discussed efficient implementations
of insertion heuristics to handle situations where the delivery amount has to lie between
a lower and an upper bound, as opposed to being fixed. In related work, Campbell and
Savelsbergh (2004b) have studied the problem of determining an optimal delivery schedule
for a route, i.e., given a sequence of customer visits, determine the timing of the visits so
as to maximize the total amount of the product delivered on the route. Because single day

36
approaches usually base decisions on the latest inventory measurement and a predicted
usage for that day, they avoid the difficulty of forecasting long-term usage, which makes
the problem much simpler.

4.4 Single customer analysis

It is insightful to analyze the “simple” situation in which there is only a single customer.
The results of this type of analysis can be used effectively to guide decisions on which
customers to consider in a short term planning horizon. The material presented in this
subsection is primarily based on Jaillet et al. (2002), although much of it dates back to
the work of Dror and Ball (1987). We first consider the deterministic case. For ease of
notation, let the usage rate of the customer be u, the storage capacity of the customer
be C, the initial inventory level be I 0 , the delivery cost to the customer be c, and the
vehicle capacity be Q. It is easy to see that an optimal policy is to fill up the storage space
precisely at the time when it becomes empty. Therefore the cost vT for a planning period
of length T is
T u − I0
  
vT = max 0, c.
min{C, Q}
Now consider the stochastic case in which one decides daily whether to make a delivery
to the customer or not. The demand U between consecutive decision points, i.e., the
demand per day, is a random variable with known probability distribution and finite mean.
Assuming that the storage capacity at each customer is at least as large as the vehicle
capacity and the vendor can only monitor the inventory in the storage space at the time of
a delivery, it can be shown that for the infinite horizon case, there exists an optimal policy
that fills up the storage space at each delivery and, following any scheduled or stock-out
delivery, plans the next delivery d days after. The optimal replenishment interval d is a
constant chosen to minimize the expected daily cost.
A d-day policy makes a delivery to the customer every d days and delivers as much as
possible, unless a stockout occurs earlier. In such a case, the vehicle is sent right away,
which generates a cost S. It is assumed that deliveries are instantaneous, so that no
additional stockout penalties are incurred. Furthermore, assume that initially the storage
space is full. Let pj be the probability that a stockout first occurs on day j (1 ≤ j ≤ d − 1).
Then p = p1 +p2 +. . .+pd−1 is the probability that there is a stockout in period [1, ..., d−1].
Furthermore, let vT (d) be the expected total cost of this policy over a planning period of
length T . We now have for d > T
X
vT (d) = pj (vT −j (d) + S),
1≤j≤T

37
and for d ≤ T
X
vT (d) = pj (vT −j (d) + S) + (1 − p)(vT −d (d) + c).
1≤j≤d−1

As a consequence, the expected total cost of filling up a customer’s tank every d days
over a T -day period (T ≥ d) is given by

vT (d) = α(d) + β(d)T + f (T, d),

where α(d) is a constant depending only on d, f (T, d) a function that tends to zero expo-
nentially fast as T tends to infinity, and

pS + (1 − p)c
β(d) = P ,
1≤j≤d jpj

with pd = 1 − p. The value β(d) is the long-run average cost per day. To determine the
best policy in this class, we need to minimize vT (d) which for large T means finding a value
of d minimizing β(d).

4.5 The two-customer IRP

When more than one customer is served, the problem becomes significantly harder. Not
only is it necessary to decide which customers to visit next, but one must also determine
how to combine them into vehicle tours, and how much to deliver to each of them. Even
if there are only two customers, these decisions may not be easy. The material in the
remainder of this section is primarily based on Campbell et al. (1998).
If the two customers are visited together, it is intuitively clear that given the amount
delivered at the first customer, it is optimal to deliver as much as possible at the second
one (determined by the remaining amount in the vehicle, and the remaining capacity at
the second customer). Thus the problem of deciding how much to deliver to each customer
involves a single decision. However, making that decision may not be easy, as the following
two-customer stochastic IRP example shows.
Assume the product is delivered and consumed in discrete units and that each customer
has a storage capacity of 20 units. The daily demands of the customers are independent and
identically distributed (across customers as well as across time), with P (U = 0) = 0.4 and
P (U = 10) = 0.6. The shortage penalty is s1 = 1000 per unit at customer 1 and s2 = 1005
per unit at customer 2. The vehicle capacity is 10 units. At the beginning of each day
the inventory at the two customers is measured, and the decision maker determines how
much to deliver to each customer. There are three possible vehicle tours, namely tours
exclusively to customers 1 and 2, of cost 120 each, and a tour to both customers 1 and 2, of

38
cost 180. Only one vehicle tour can be completed per day. This situation can be modeled as
an infinite horizon Markov decision process, with the objective of minimizing the expected
total discounted cost. Because of the small size of the state space, it is possible to compute
the optimal expected value and an optimal policy.
Figure 2 shows the expected value (total discounted cost) as a function of the amount
delivered at customer 1 (and therefore also at customer 2), when the inventory at each
customer is 7, and both customers are to be visited in the next vehicle tour (which is the
optimal decision in the given state). The figure shows that the objective function is not
unimodal, with a local minimum at 3, and a global minimum at 7. Consequently, deciding
just how much to deliver to each customer may require solving a non-linear optimization
problem with a non-unimodal objective function. This is a hard problem for which available
search methods may not converge to an optimal solution.
39350

39300

39250
Expected value

39200

39150

39100

39050
0 1 2 3 4 5 6 7 8 9 10
Amount delivered at customer 1

Figure 2: Non-unimodal objective function for determining the optimal delivery quantity

4.6 Literature review on the IRP

Rather than providing a comprehensive review of the IRP literature, we discuss several
research streams representing a variety of solution approaches that have been proposed and
investigated. We encourage the reader to examine the referenced papers for more elaborate
and precise coverage.
A first stream of research uses time-discretized integer programming models to deter-

39
mine the set of customers to be visited in a short-term planning horizon as well as the
amount of product to deliver to them. In order to accurately reflect costs and time related
aspects, the integer linear programs work with a set of potential delivery routes. Fisher
et al. (1982) and Bell et al. (1983) pioneered this approach when they studied the IRP at
Air Products, a producer of industrial gases. Their formulation determines the delivery
volumes to customers, the assignment of customers to routes, the assignment of vehicle
to routes, and the start times of routes. The core structure of their model is presented
below, where the variable xirtv represents the amount of product delivered to customer
i ∈ N on route r ∈ R starting at time t ∈ T , the variable yrt is 1 if route r starts at time
t and 0 otherwise, Sr the set of customers visited on route r, pi the value of delivering a
unit of product to customer i, Fr the fixed cost of executing route r, q it a lower bound on
the cumulative amount delivered to customer i by time t, and q it an upper bound on the
cumulative amount that can be delivered to customer i by time t.
X X X 
Maximize pi xirt − Fr yrt
r∈R t∈T i∈Sr

subject to
XX
q it ≤ xirs ≤ q it (i ∈ N, t ∈ T )
r∈R s≤t
X
xirt ≤ Qyrt (r ∈ R, t ∈ T )
i∈Sr
xirt ≥ 0, yrt ∈ {0, 1}.

In the model, the per unit value of a delivery to a customer is used to represent the effect of
decisions on events occuring beyond the planning horizon of the model. In the short-term
planning period considered by the model, there is considerable discretion in the amount of
product to deliver. In the long run this amount is determined by customer usage. Hence,
each unit scheduled for delivery to a customer within the planning horizon reduces the
amount to be delivered in the future. This is accounted for by setting the unit value to
an estimate of the cost of delivering to a customer at a point in time outside the planning
horizon of the model. Furthermore, rather than explicitly incorporating customer usage
rates into the model, lower and upper bounds on the cumulative amount to be delivered
to each customer in each time period in the planning horizon are used. It is simple,
of course, to convert customer usage rates into bounds, i.e., qit = max{0, tui − Ii0 } and
q it = tui + Ci − Ii0 . Lagrangean relaxation was a central tool in developing an effective
heuristic for solving the integer program. The size of the integer programs to be solved
depends on the chosen time discretization as well as on the size of set of routes.
Campbell and Savelsbergh (2004a) use an integer linear program with a similar structure
to determine which customers to visit in the next few days (even though the integer program
covers several weeks) and to suggest quantities and delivery times to these customers.

40
However, then, in a second phase, they use modified insertion heuristics to determine the
actual delivery routes and quantities. The advantage of such a two-phase approach is
that a higher degree of accuracy (in terms of timing of events) can be provided in the
second phase and other practical details, such as drivers shifts, can be considered. The
delivery quantities and times specified by the solution to the integer program are good
from a long-term perspective; they may need to be modified somewhat to also be good
from a short-term perspective. When constructing the actual delivery routes, Campbell
and Savelsbergh consider delivering more to the customers than the quantity suggested
by the integer program (and slightly altering the delivery time if needed) since this may
result in higher vehicle utilization and thus higher revenues. As in Bard et al. (1998b)
their approach is embedded into a rolling horizon framework.
A second stream of research is based on the single customer analysis presented above.
This approach was pioneered by Dror, Ball and Golden (1985) and Dror and Ball (1987).
The optimal replenishment day t∗i minimizing the expected total cost for customer i is used
to determine the set of customers considered in a short-term planning problem for the next
t̄ days. If t∗i ≤ t̄, then the customer will be included and will definitely be visited. A value
ct is computed for each day of the planning period to reflect the expected increase in future
cost if the delivery is made on day t instead of t∗ . If t∗ ≥ T , i.e., the optimal replenishment
day falls outside the short-term planning period, then a future benefit gt can be computed
for making an early delivery to the customer on day t of the short-term planning period.
These computed values reflect the long term effects of short term decisions. An integer
linear program is subsequently solved to assign customers to a vehicle and a day, or just
a day, that minimizes the sum of these costs plus the transportation costs. (It was shown
by Adelman (2004) that this objective function is in fact equivalent to that used by Fisher
et al. (1982).) The delivery amount to a customer on a specific day is fixed and set to the
quantity needed to fill up the storage tank on that day. This leaves either TSPs or VRPs
to be solved in the second stage. These ideas are extended and improved in Trudeau and
Dror (1992). The most recent work along these lines is that of Bard et al. (1998a,b) who
work with a rolling horizon approach in which a short term planning problem is defined for
a two-week period, but only the decisions for the first week are implemented. In addition,
satellite facilities are considered, i.e., locations other than the depot where vehicles can be
refilled.
A third stream of research focuses on the asymptotic analysis of delivery policies. Anily
and Federgruen (1990, 1991, 1993) analyze fixed partition policies for the IRP with an
unlimited number of vehicles. Customers within the same partition are divided into regions
so as to make the demand of each region roughly equal to a vehicle load. A customer may
appear in more than one region, but then a certain percent of his demand is allocated
to each region. When one customer in a region is also visited, all customers in that
region are visited. The authors determine lower and upper bounds on the minimum long-

41
run average cost over all fixed partition policies, and propose a heuristic, called modified
circular regional partitioning, to choose a fixed partition. Using similar ideas, Gallego and
Simchi-Levi (1990) evaluate the long-run effectiveness of direct deliveries (one customer
on each route). Direct shipping is shown to be at least 94% effective over all inventory
routing strategies whenever the minimal economic lot size is at least 71% of vehicle capacity.
This indicates that direct shipping becomes an undesirable and costly policy when many
customers require significantly less than a vehicle load, making more complicated routing
policies the appropriate choice. Another adaptation of these ideas can be found in Bramel
and Simchi-Levi (1995) who consider a variant of the IRP in which customers can hold
an unlimited amount of inventory. To obtain a solution, they transform the problem into
a Capacitated Concentrator Location Problem (CCLP), solve it, transform the solution
back into a solution to the IRP, and heuristically improve it. The CCLP solution will
partition the customers into disjoint sets, which in the IRP will become the fixed partitions.
Chan, Federgruen and Simchi-Levi (1998) analyze zero-inventory ordering policies, in which
a customer’s inventory is replenished only when it has been depleted, in combination
with fixed partitioning routing policies and derive asymptotic worst-case bounds on their
performance. Gaur and Fisher (2004) consider an IRP with time varying demand. They
propose a randomized heuristic to find a fixed partition policy with periodic deliveries.
Their method was implemented for a supermarket chain.
The fourth stream of research is based on formulating the stochastic IRP as a Markov
decision process and thus explicitly incorporating demand uncertainty. This approach was
pioneered by Minkoff (1993) who proposed a decomposition heuristic to overcome the com-
putational difficulties caused by large state spaces. The heuristic solves a linear program
to allocate joint transportation costs to individual customers and then solves individual
customer subproblems. The value functions of the subproblems are added to approximate
the value function of the original problem. The main limitation of the proposed approach is
that it assumes the availability of a set of delivery routes with fixed delivery quantities for
the customers on a route and the dispatcher only has to decide which of the delivery routes
to use at each decision point. This limitation is removed in the work of Kleywegt, Nori
and Savelsbergh (2002, 2004) on approximate dynamic programming approaches and in
that of Adelman (2003a, 2004) on price-directed approaches. Let state x = (x1 , x2 , . . . , xn )
represent the current inventory at each customer, and let A(x) denote the set of all feasible
decisions when the process is in state x. A decision a ∈ A(x) specifies which customer in-
ventories to replenish, how much to deliver at each customer location, and how to combine
customers into vehicle routes. Let Q be the Markov transition function according to which
transitions occur. Let g(x, a) denote the expected single stage net reward if the process is
in state x at time t and decision a ∈ A(x) is implemented. The objective is to maximize
the expected total discounted value over an infinite horizon. Let V ∗ (x) denote the optimal

42
expected value given that the initial state is x. Then, for any state x,
 Z 
∗ ∗
V (x) = sup g(x, a) + α V (y)Q[dy | x, a] . (32)
a∈A(x)

∗
A policy π ∗ is called optimal if V π = V ∗ , where V π represents the value function of policy
π. Solving a Markov decision process involves computing the optimal value function V ∗
and an optimal policy π ∗ by solving the optimality equation (32). This requires performing
the following major computational tasks:

1) The computation of the optimal value function V ∗ . Most algorithms for computing
V ∗ involve the computation of successive approximations to V ∗ (x) for every state x.
These algorithms are practical only if the state space is small.

2) The estimation of the expected value (the integral in (32)). For the stochastic IRP,
this is a high dimensional integral. Conventional numerical integration methods are
not practical for the computation of such high dimensional integrals.

3) The maximization problem on the right hand side of (32) has to be solved to deter-
mine the optimal decision for each state. For the stochastic IRP, this means solving
a complex variant of the VRP.

Kleywegt, Nori, and Savelsbergh develop approximation methods to efficiently perform

these computational tasks. Furthermore, their approach has the ability to handle a finite
fleet of vehicles, whereas in other Markov decision process based approaches it is assumed
that there exists an infinite fleet of vehicles. The optimal value function V ∗ is approximated
by V̂ as follows. First, the stochastic IRP is decomposed into subproblems defined for
specific subsets of customers. Each subproblem is also a Markov decision process. The
subsets of customers do not necessarily partition the set of customers, but must cover it.
The idea is to define each subproblem so that it provides an accurate representation of
the overall process as experienced by the subset of customers. To do so, the parameters
of each subproblem are determined by simulating the overall stochastic IRP process, and
by constructing simulation estimates of subproblem parameters. Next, each subproblem is
solved optimally. Finally, for any given state x, the approximate value V̂ (x) is determined
by choosing a partition of the customers and by setting V̂ (x) equal to the sum of the optimal
value functions of the subproblems corresponding to the partition at states corresponding
to x. The partition is chosen to maximize V̂ (x). Randomized methods, incorporating
variance reduction techniques to limit the required sample size, are used to estimate the
expected value on the right hand side of (32). Action determination involves deciding which
customers to visit on a route and how much to deliver to them. This is achieved through a
heuristic. An initial solution consisting of only direct delivery routes is constructed. This
is followed by a local search procedure that examines the benefit of adding a customer to

43
an existing route and modifying the delivery quantities. Using their approach Kleywegt,
Nori, and Savelsbergh can solve problems involving up to 50 customers.
More recently, Adelman (2003a, 2004) proposed a price-directed operating policy based
on a simple economic mechanism to determine routing and delivery decisions for a given
inventory state. Suppose management specifies a value Vi for replenishing one unit of
product at customer i. A dispatcher can now evaluate a feasible delivery route as follows.
If a set S = {s1 , ..., sn } of customers is visited, quantities d1 , ..., dn are delivered, and a cost
P
cS is incurred. Then the net value of the route equals i∈S Vi di −cS . The dispatcher has to
choose delivery routes so as to maximize his total net value without stockouts at customers.
This mechanism motivates the dispatcher to replenish a customer i whose current inventory
level is low, because then di can be set large. When faced with the option of expanding the
set S of customers to visit on a route which does not yet use the full vehicle capacity, the
dispatcher will consider the incremental cost cS∪{k} − cS and determine if a quantity
dk
can be replenished that is large enough to justify it, i.e., whether dk Vk − cS∪{k} − cS > 0
or dk ≥ (cS∪{k} − CS )/Vk .
The key to success in solving management’s problem is to set the Vi s in such a way that
the dispatcher is motivated to (ideally) minimize the long-run time average replenishment
costs. If the dispatcher’s total net value is regularly positive, then his performance exceeds
management’s long range expectations. Management should decrease the Vi s to make them
consistent with actual performance. On the other hand, if the dispatcher’s total net value
is regularly negative, then the Vi s impose unrealistic expectations on the dispatcher and
management should increase them. Ideally, management should set the Vi s equal to the
lowest achievable marginal costs.
Starting from a dynamic control model of the inventory routing problem, Adelman
(2003b) derives the following non-linear programming relaxation, which computes a long
run “average” solution to the inventory routing problem. Let zR be a decision variable
representing the rate at which a subset R of customers is visited together. Furthermore,
let di,R for all i ∈ R be a decision variable representing the average quantity delivered to
customer i on a delivery route visiting subset R. This yields the following formulation:
X
(NLP) Minimize CR zR (33)
R⊆N

subject to
X
di,R zR = ui (i ∈ N ) (34)
R⊆N
X
di,R ≤ Q (R ⊆ N ) (35)
i∈R
di,R ≤ Ci (R ⊆ N, i ∈ R) (36)

44
 zR , di,R ≥ 0 (R ⊆ N, i ∈ R). (37)

The objective (33) minimizes the long run average replenishment cost. Constraints (34)
state that for each customer i the rate at which quantities are replenished must equal the
rate at which they are consumed. Constraints (35) state that on average vehicle capacity
is satisfied, and constraints (36) state that on average the quantity delivered at customer
i is less than the storage capacity. Consider the following linear program
X
(D) Maximize ui Vi (38)
i∈N

subject to
X
di,R Vi ≤ CR (R ⊆ N ), (39)
i∈R

with decision variables Vi . Adelman shows that this semi-infinite linear program is dual
to the non-linear program in that there is no duality gap between them and a version of
complementary slackness holds. In (NLP) di,R is a decision variable while in (D) it is part
of the input. The decision variables Vi at optimality are the marginal costs associated
with satisfying constraints (34) of (NLP). This means that at optimality ui Vi is the total
allocated cost rate for replenishing customer i in an optimal solution to (NLP). Each Vi
can be interpreted as the payment management transfers to the dispatcher for replenishing
one unit of product of customer i. Hence, the objective (38) maximizes the total transfer
rate, subject to the constraint (39) that the payments can be no larger than the cost of any
replenishment. NLP can be solved effectively by means of column generation techniques.
We have opted to focus on only a few research streams with an emphasis on more
recent efforts. However, many other researchers have contributed to the inventory routing
literature, including Federgruen and Zipkin (1984), Golden, Assad and Dahl (1984), Burns
et al. (1985), Larson (1988), Chien, Balakrishnan and Wong (1989), Webb and Larson
(1995), Barnes-Schuster and Bassok (1997), Herer and Roundy (1997), Viswanathan and
Mathur (1997), Christiansen and Nygreen (1998a), Christiansen and Nygreen (1998b),
Christiansen (1999), Reimann, Rubio and Wein (1999), Waller, Johnson and Davis (1999),
Çetinkaya and Lee (2000), Lau, Liu and Ono (2002), Bertazzi, Paletta, and Speranza
(2002), Savelsbergh and Song (2005), and Song and Savelsbergh (2005).

5 Stochastic Vehicle Routing Problems

Stochastic Vehicle Routing Problems (SVRPs) are extensions of the deterministic VRP
in which some components are random. The three most common cases are 1) stochastic
customers: customer i is present with probability pi and absent with probability 1 − pi ; 2)

45
stochastic demands (to be collected, say): the demand ξi of customer i is a random variable;
3) stochastic times: the service time si of customer i and the travel time tij of edge (i, j)
are random variables. Because some of the data are random it is no longer required to
satisfy the constraints for all realizations of the random variables, and new feasibility and
optimality concepts are required. With respect to their deterministic counterparts, SVRPs
are considerably more difficult to solve. Not only is the notion of a solution different, but
some of the properties that were valid in a deterministic context no longer hold in the
stochastic case (see, e.g., Dror, Laporte and Trudeau, 1989 and Gendreau, Laporte and
Séguin, 1996).
Applications of SVRP arise in a number of settings such as the delivery of meals on
wheels (Bartholdi et al., 1983) or of home heating oil (Dror, Ball and Golden, 1985), sludge
disposal (Larson, 1988), forklift routing in warehouses (Bertsimas, 1992), money collection
in bank branches (Lambert, Laporte and Louveaux, 1993), and general pickup and delivery
operations (Hvattum, Løkketangen and Laporte, 2003).
Stochastic VRPs can be formulated and solved in the context of stochastic program-
ming: a first stage or a priori solution is computed, the realizations of the random variables
are then disclosed and, in a second stage, a recourse or corrective action is applied to the
first stage solution. The recourse action usually generates a cost or a saving which may be
taken into account when designing the first stage solution. To illustrate, consider a planned
vehicle route in an SVRP with stochastic demands. Because demands are stochastic, the
vehicle capacity may be attained or exceeded at some customer j before the route is com-
pleted. In this case several possible recourse policies are possible. For example, the vehicle
could return to the depot to unload and resume collections at customer j (if the vehicle
capacity was exceeded at j) or at the successor of j on the route (if the vehicle capacity
was attained exactly at j). Another policy would be to plan preventive return trips to the
depot in the hope of avoiding higher costs at a later stage (see, e.g., Laporte and Louveaux,
1990; Dror, Laporte and Louveaux, 1993 and Yang, Mathur and Ballou, 2000). A more
radical policy would be to reoptimize the route segment following j upon arrival at the
depot (see, e.g., Bastian and Rinnooy Kan, 1992; Secomandi, 1998 and Haughton, 1998,
2000). The best choice of a recourse policy depends on the time at which information
becomes available. For example, information about a customer demand may only be avail-
able upon arriving at that customer or when visiting the previous customer, thus allowing
for a wider range of recourse actions, such as returning to the depot in anticipation of
failure or postponing the visit of a high demand customer. An extensive discussion of
recourse policies in the context of availability of information is provided in Dror, Laporte
and Trudeau (1989).
There exist two main solution concepts in stochastic programming. In chance con-
strained programming (CCP) the first stage problem is solved under the condition that the
constraints are satisfied with some probability. For example one could impose a failure

46
threshold α, i.e., planned vehicle routes should fail with probability at most equal to α.
The cost of failure is typically disregarded in this approach. Stewart and Golden (1983)
have proposed the first CCP formulation for the VRP with stochastic demands. Using a
three-index model they showed that probabilistic constraints could be transformed into a
deterministic equivalent form. Laporte, Louveaux and Mercure (1989) later proposed a
similar transformation for a two-index model. The interest of such transformations is that
the chance constrained SVRP can then be solved using any of the algorithms available for
the deterministic case. In stochastic programming with recourse (SPR) two sets of variables
are used: first-stage variables characterize the solution generated before the realization of
the random variables, while second-stage variables define the recourse action. The solution
cost is defined as the sum of the cost of the first-stage solution and that of the recourse
action. The aim of SPR is to design a first-stage solution of least expected total cost.
Stochastic VRPs are usually modeled and solved with the framework of a priori op-
timization (Bertsimas, Jaillet and Odoni, 1990) or as Markov decision processes (Dror,
Laporte and Trudeau, 1989). A priori optimization computes a first-stage solution of least
expected cost under a given recourse policy. The most favoured a priori optimization
methodology is the Integer L-Shaped method (Laporte and Louveaux, 1993, 1998) which
belongs to the same class as Benders decomposition (Benders, 1962) and the L-Shaped
method for continuous stochastic programming (Van Slyke and Wets, 1969). While route
reoptimization is preferable to a priori optimization from a solution cost point of view, it is
computationally more cumbersome. In contrast, a priori optimization entails solving only
one instance of an NP-hard problem and produces a more stable and predictable solution
(Bertsimas, Jaillet and Odoni, 1990). It is also superior to solving a deterministic VRP
instance with expected demands (Louveaux, 1998).
The Integer L-Shaped method is essentially a variant of branch-and-cut. It operates on
a current problem obtained by relaxing integrality requirements and subtour elimination
constraints, and by replacing the cost of recourse Q(x) of first-stage solution x by a lower
bound θ on its value. Integrality and subtour elimination constraints are gradually satisfied
as is commonly done in branch-and-cut algorithms for the deterministic VRP (see, e.g.,
Naddef and Rinaldi, 2002) while lower bounding functionals on θ, called optimality cuts,
are introduced into the problem at integer or fractional solutions. The method assumes
that a lower bound L on θ is available. In the following description xij is a binary variable
equal to 1 if and only if edge (i, j) is used in the first stage solution.

Step 0. Set the iteration count ν := 0 and introduce the bounding constraint θ ≥ L into
the current problem. Set the value z̄ of the best known solution equal to ∞. At this
stage, the only active node corresponds to the initial current problem.

Step 1. Select a pendent node from the list. If none exists stop.

Step 2. Set ν := ν + 1 and solve the current problem. Let (xν , θ ν ) be an optimal solution.

47
Step 3. Check for any subtour elimination constraint violation. If at least one violation
can be identified, introduce a suitable number of subtour elimination constraints into
the current problem, and return to Step 2. Otherwise, if cxν + θ ν ≥ z̄, fathom the
current node and return to Step 1.

Step 4. If the solution is not integer, branch on a fractional variable. Append the corre-
sponding subproblems to the list of pendent nodes and return to Step 1.

Step 5. Compute Q(xν ) and set z ν := cxν + Q(xν ). If z ν < z̄, set z̄ := z ν .

Step 6. If θ ν ≥ Q(xν ), then fathom the current node and return to Step 1. Otherwise,
impose the optimality cut
X X 
θ ≥ L + (Q(xν ) − L) xij − xνij + 1 (40)
1<i<j 1<i<j
xν =1
ij

into the current problem and return to Step 2.

The optimality cut (40) uses the fact that a feasible solution is fully characterized by the
xij variables associated with edges non-incident to the depot. They state that either the
current solution must be maintained, in which case the cut becomes θ ≥ Q(xν ), or a
new solution must be identified, in which case the cut becomes θ ≥ L or less and is thus
redundant.
Markov decision models are defined on a state space. The system is observed at various
transition times corresponding to moments at which a new customer is visited, and new
decisions are taken at these moments. The state of the system at a given transition time
is described by the set of customers already visited by the vehicle and by its current
load. Because the state space is typically very large, this approach can only be applied to
relatively small scale instances.
Heuristics for SVRPs are adaptations of methods originally designed for the determinis-
tic case, which can be rather intricate because of the probability computations involved. In
particular, computing the expected cost of a vehicle route is itself complicated and it may
be advisable to use approximations if such computations are to be performed repeatedly
within a search process (see, e.g., Gendreau, Laporte and Séguin, 1996). In what follows
we study some particular classes of SVRPs.

5.1 The vehicle routing problem with stochastic customers

In vehicle routing problems with stochastic customers each vertex i is present with prob-
ability pi . A first-stage solution consists of a set of vehicle routes visiting the depot and

48
each customer exactly once. The set of absent customers is then revealed and the second-
stage solution consists of following the first-stage routes while skipping the absent vertices.
Jaillet (1985) laid the foundations of this line of research in his study of the Traveling Sales-
man Problem with Stochastic Customers (TSPSC). He proposed mathematical models and
bounds, and he investigated a number of properties of the problem. For example, he showed
that the solution of a deterministic TSP can be arbitrarily bad for the TSPSC. Also, even
if the TSPSC is defined in a plane with Euclidean distances, an optimal cycle may cross
itself, contrary to what happens for the TSP (Flood, 1956). Jézéquel (1985) and Rossi
and Gavioli (1987) have proposed a number of heuristics for the TSPSC based on adapta-
tions of the Clarke and Wright (1964) savings principle. Bertsimas (1988) and Bertsimas
and Howell (1993) later investigated further properties of the TSPSC and proposed new
heuristics, namely methods based on space filling curves (Bartholdi and Platzman, 1982)
and on a 2-opt edge interchange mechanism. The first exact algorithm for the TSPSC is
an integer L-shaped algorithm developed by Laporte, Louveaux and Mercure (1994) and
capable of solving instances involving up to 50 customers. An extension of the TSPSC,
called the Pickup and Delivery Traveling Salesman Problem with Stochastic Customers
(PDTSPSC), was recently investigated by Beraldi et al. (2005). In this problem there are
n requests, each consisting of a pickup location and of a delivery location, but request i
only materializes with probability pi . The authors show how to efficiently implement a low
complexity interchange heuristic for this problem.
The Vehicle Routing Problem with Stochastic Customers (VRPSC) has been mostly
studied in the context of unit demand customers. As in the TSPSC, vehicles follow the
first-stage routes while skipping the absent customers and return to the depot to unload
when their capacity is reached. This problem was first studied by Jézéquel (1985), Jaillet
(1987) and Jaillet and Odoni (1988). The latter reference states two interesting properties
of the VRPSC: 1) even if travel costs are symmetric the overall solution cost is dependent
on the direction of travel; 2) larger vehicle capacities may yield larger solution costs.
Bertsimas’s Ph.D. thesis (1988) is an excellent source of information on this problem. It
describes several properties, bounds and heuristics. Waters (1989) has studied the case of
general integer demands and has compared three simple heuristics for this problem.

5.2 The vehicle routing problem with stochastic demands

The Vehicle Routing Problem with Stochastic Demands (VRPSD) has been the most stud-
ied stochastic VRP. In this problem customer demands are random and usually (but not
always) independent. Tillman (1969) was probably the first to study this problem in a
multi-depot context. He proposed a savings based heuristic for its solution. The first ma-
jor study of the VRPSD can be attributed to Golden and Stewart (1978) who presented
a chance constrained model and two recourse models. In the first of these a penalty pro-
portional to vehicle overcapacity is imposed; in the second the penalty is proportional to

49
the expected demand in excess of the vehicle capacity. Several basic heuristics were im-
plemented and tested. Dror and Trudeau (1986) developed further heuristics and showed
that for this problem expected travel cost depends on the direction of travel even in the
symmetric case. Again, Bertsimas’s thesis (1988) constitutes a major contribution to the
study of the VRPSD. It proposes several bounds, asymptotic results and properties for the
case where ξi is equal to 1 with probability pi , and equal to 0 otherwise. In their survey
paper, Dror, Laporte and Trudeau (1989) have shown that some properties established by
Jaillet (1985, 1988) and Jaillet and Odoni (1988) extend to the VRPSC, namely 1) in an
optimal solution a vehicle route may intersect itself; 2) in a Euclidean problem customers
are not necessarily visited in the order in which they appear on the convex hull of vertices;
3) segments of an optimal route are not necessarily optimal when considered separately.
The latter property can have a major impact on the design of a dynamic algorithm for the
VRPSD.
Laporte, Louveaux and Mercure (1989) proposed a two-index chance constrained model
for the VRPSC as well as an associated branch-and-cut algorithm capable of solving in-
stances with n ≤ 30. They also introduced a bounded penalty model in which the cost
of recourse associated with a given route cannot exceed a preset proportion of the first-
stage route cost. The best exact solution approach for the VRPSD is again the integer
L-shaped algorithm. Séguin (1994) and Gendreau, Laporte and Séguin (1995) proposed
the first implementation of this method for the solution of the VRPSD and were able to
solve instances of up to 70 vertices. The most difficult case arises when the expected filling
rate f of the vehicles is large. For example, when f = 0.3 instances with n = 70 can be
solved optimally, but when f = 1.0 instances with n = 10 can rarely be solved. Using
a similar approach, Hjorring and Holt (1999) solved one-vehicle instances (m = 1) with
0.95 ≤ f ≤ 1.05 and n = 90. Laporte, Louveaux and Van hamme (2002) imposed an
additional restriction, namely that the expected demand of a route does not exceed the
vehicle capacity, and they also exploited properties of the demand under known distribu-
tions (Poisson and normal) in the generation of lower bounding functionals on the cost of
recourse. This enabled them to solve larger instances: for Poisson demands they solved
instances with f = 0.9, m = 4 and n = 25, or with m = 2 and n = 100; for normal
demands they solved instances with f = 0.9, m = 3 and n = 50.
Dynamic programming was applied by Secomandi (1998) to the VRP with stochastic
demands. The largest instance solved to optimality with this method contained only 10
customers. The author also developed neuro-dynamic programming (NDP) algorithms
(Secomandi, 1998, 2000, 2003) for the same problem. Neuro-dynamic programming (see,
e.g., Bertsekas, 1995) is a heuristic approach used to solve large-scale dynamic programs.
It replaces the “cost-to-go” computations by proxys based on simulation and parametric
function approximations. Secomandi (2000) compared two NDP implementations for the
VRP with stochastic demands: an Optimistic Approximate Iteration Policy in which a

50
neural network methodology is used to compute the approximations, and a Rollout Policy
in which the cost-to-go is approximated by means of a heuristic. Computational results
show that the second of these two policies is consistently and substantially superior to the
first.

5.3 The vehicle routing problem with stochastic customers and demands
The VRP with stochastic customers and demands combines two difficult cases. This prob-
lem was first mentioned by Jézéquel (1985), Jaillet (1987), Jaillet and Odoni (1988), and
was later formally defined by Bertsimas (1992). A first-stage solution visiting all customers
is first constructed, the set of present customers is then revealed and their demand becomes
known upon the arrival of the vehicle at the customer’s location, routes are followed as
planned but absent customers are skipped and the vehicle returns to the depot to unload
whenever its capacity becomes attained. Benton and Rossetti (1992) proposed an algo-
rithm which performs route reoptimizations whenever demands are revealed. One major
difficulty in solving this problem lies in the computation of the objective function value.
Recursions, bounds, asymptotic results and a comparison of various reoptimization poli-
cies are provided by Bertsimas (1992). Séguin (1994) and Gendreau, Laporte and Séguin
(1995) developed the first exact algorithm for this problem, based again on the Integer L-
shaped approach. They solved instances involving up to 46 customers and concluded that
stochastic customers are a far more complicating factor than are stochastic demands. In a
different study, Gendreau, Laporte and Séguin (1996) developed a tabu search algorithm
which uses an approximation of the objective function cost in order to ease computations.
On a set of 825 instances with 6 ≤ n ≤ 46 for which the optimum was known, an optimal
solution was identified in 89.45% of the cases and the average optimality gap was 0.38%.

5.4 The vehicle routing problem with stochastic travel times

In the Vehicle Routing Problem with Stochastic Travel Times (VRPSTT) travel times on
the edges and service times at the vertices are random variables. Vehicles follow their
planned routes and may incur a penalty if the route duration exceeds a given deadline.
It is natural to make this penalty proportional to the elapsed route duration in excess of
the deadline (Laporte, Louveaux and Mercure, 1992). Another possibility is to define a
penalty proportional to the uncollected demand within the time limit, as is the case in a
money collection application studied by Lambert, Laporte and Louveaux (1993).
The simplest case of the VRPSTT is the Traveling Salesman Problem with Stochastic
Travel Time (TSPSTT) in which there is only one vehicle. It was first studied by Leipälä
(1978) who computed the expected length of tours with random travel times. A common
version of the TSPSTT is the case where the objective is to design a tour having the largest
probability of being completed within the deadline. Kao (1978) proposed two heuristics for

51
this problem: one based on dynamic programming, and the other on implicit enumeration.
Sniedovich (1981) has shown that dynamic programming applied to the same problem
can be suboptimal because the monotonicity property required for this method is not
satisfied in the TSPSTT. Carraway, Morin and Moskowitz (1989) later developed a so-
called generalized dynamic programming algorithm that overcomes this difficulty. Kenyon
and Morton (2003) have shown that an optimal TSPSTT can be identified by solving a
deterministic TSP in which the travel and service times are replaced by their mean values.
Verweij et al. (2003) have developed a heuristic for the case where a penalty proportional
to route duration in excess of the deadline is incurred. The method uses a sample average
approximation technique in which a sample of instance realizations is drawn and each is
solved optimally by means of a deterministic technique. By repeating the method with
different samples a statistical estimate of the optimality gap can be computed.
Laporte, Louveaux and Mercure (1992) were probably the first to provide exact algo-
rithms for the VRPSTT. They formulated the chance constrained version of the problem,
and they modeled a recourse version of the problem in which a penalty proportional to
route duration in excess of the deadline is incurred. The problem was solved optimally
by means of an Integer L-shaped algorithm for 10 ≤ n ≤ 20 and two to five travel time
scenarios (each scenario corresponds to a different travel speed for the entire network). In
a more recent study, Kenyon and Morton (2003) have investigated properties of VRPSTT
solutions and have developed bounds on the objective function value. They have devel-
oped a heuristic that combines branch-and-cut and Monte-Carlo simulation which, if run
to completion, terminates with a solution value within a preset percentage of the optimum.
Finally, vehicle routing with stochastic travel time is frequently encountered in pickup
and delivery problems such as those arising in truckload operations. Wang and Regan
(2001) have proposed models for this class of problems under the presence of time windows.

Acknowledgments
This work has been supported by the Canadian Natural Sciences and Engineering Research
Council under grants 227837-00 and OGP0039682, by the Ministero dell’Università e della
Ricerca (MIUR), and by the Consiglio Nazionale delle Ricerche (CNR), Italy. This support
is gratefully acknowledged.

References
Achuthan, N.R., L. Caccetta and S.P. Hill (1996). A new subtour elimination constraint
for the vehicle routing problem. European Journal of Operational Research 91, 573–586.

52
Achuthan, N.R., L. Caccetta and S.P. Hill (2003). An improved branch and cut algorithm
for the capacitated vehicle routing problem. Transportation Science 37, 153–169.
Adelman, D. (2003a). Price-directed replenishment of subsets: Methodology and its appli-
cation to inventory routing. Manufacturing & Service Operations Management 5, 348–
371.
Adelman, D. (2003b). Internal transfer pricing for a decentralized operation with a shared
supplier. Working paper, Graduate School of Business, The University of Chicago.
Adelman, D. (2004). A price-directed approach to stochastic inventory/routing. Operations
Research 52, 499–514.
Agarwal, Y., K. Mathur and H.M. Salkin (1989). A set-partitioning-based exact algorithm
for the vehicle routing problem. Networks 19, 731–749.
Altinkemer, K., and B. Gavish (1991). Parallel savings based heuristic for the delivery
problem. Operations Research 39, 456–469.
Anily, S., and A. Federgruen (1990). One warehouse multiple retailer systems with vehicle
routing costs. Management Science 36, 92–114.
Anily, S., and A. Federgruen (1991). Rejoinder to “one warehouse multiple retailer systems
with vehicle routing costs”. Management Science 37, 1497–1499.
Anily, S., and A. Federgruen (1993). Two-Echelon Distribution Systems with Vehicle Rout-
ing Costs and Central Inventories. Operations Research 41, 37–47.
Araque, J.R., L. Hall and T.L. Magnanti (1990). Capacitated trees, capacitated routing and
associated polyhedra. Discussion Paper 90-61, CORE, University of Louvain-la-Neuve,
Belgium.
Araque, J.R., G. Kudva, T. Morin and J.F. Pekny (1994). A branch-and-cut algorithm for
vehicle routing problems. Annals of Operations Research 50, 37–59.
Augerat, P., J.M. Belenguer, E. Benavent, A. Corberán and D. Naddef (1999). Separating
capacity inequalities in the CVRP using tabu search. European Journal of Operational
Research 106, 546–557.
Augerat, P., J.M. Belenguer, E. Benavent, A. Corberán, D. Naddef and G. Rinaldi (1995).
Computational results with a branch and cut code for the capacitated vehicle routing
problem. Technical Report RR 949-M, Université Joseph Fourier, Grenoble.
Badeau, P., M. Gendreau, F. Guertin, J.-Y. Potvin and É.D. Taillard (1997). A parallel
tabu search heuristic for the vehicle routing problem with time windows. Transportation
Research C 5, 109–122.

53
Baker, E., and J. Schaffer (1986). Computational experience with branch exchange heuris-
tics for vehicle routing problems with time window constraints. American Journal of
Mathematical and Management Sciences 6, 261–300.

Baldacci, R., L. Bodin and A. Mingozzi (2004). The multiple disposal facilities and multiple
inventory locations rollon-rolloff vehicle routing problem. Technical Report, Department
of Mathematics, University of Bologna, Italy.

Baldacci, R., E. Hadjiconstantinou and A. Mingozzi (2004). An exact algorithm for the ca-
pacitated vehicle routing problem based on a two-commodity network flow formulation.
Operations Research 52, 723–738.

Balinski, M., and R. Quandt (1964). On an integer program for a delivery problem. Oper-
ations Research 12, 300–304.

Bard, J.F., L. Huang, M. Dror and P. Jaillet (1998a). A branch and cut algorithm for the
VRP with satellite facilities. IIE Transactions 30, 831–834.

Bard, J.F., L. Huang, P. Jaillet and M. Dror (1998b). A decomposition approach to the
inventory routing problem with satellite facilities. Transportation Science 32, 189–203.

Bard, J.F., G. Kontoravdis and G. Yu (2002). A branch-and-cut procedure for the vehicle
routing problem with time windows. Transportation Science 36, 250–269.

Barnes-Schuster, D. and Bassok, Y. (1997). Direct Shipping and the Dynamic Single-
depot/Multi-retailer Inventory System. European Journal of Operational Research 101,
509–518.

Bartholdi, J.J., and L.K. Platzman (1982). An N log N planar traveling salesman heuristic
based on spacefilling curves. Operations Research Letters 1, 121–125.

Bartholdi, J.J., L.K. Platzman, R.L. Collins and W.H. Warden (1983). A minimal tech-
nology routing system for meals on wheels. Interfaces 13(3), 1–8.

Bastian, C., and A.H.G. Rinnooy Kan (1992). The stochastic vehicle routing problem
revisited. European Journal of Operational Research 56, 407–412.

Battiti, R., and G. Tecchiolli (1994). The reactive tabu search. ORSA Journal on Com-
puting 6, 126–140.

Bean, J.C. (1994). Genetic algorithms and random keys for the sequencing and optimiza-
tion. ORSA Journal on Computing 6, 154–160.

Beasley, J.E. (1983). Route-first cluster-second methods for vehicle routing. Omega 11,
403–408.

54
Bell, W., L. Dalberto, M.L. Fisher, A. Greenfield, R. Jaikumar, P. Kedia, R. Mack and
P. Prutzman (1983). Improving the distribution of industrial gases with an on-line com-
puterized routing and scheduling optimizer. Interfaces 13(6), 4–23.

Benders, J.F. (1962). Partitioning procedures for solving mixed variables programming
problems. Numerische Mathematik 4, 238–252.

Bent, R., and P. Van Hentenryck (2004). A two-stage hybrid local search for the vehicle
routing problem with time windows. Transportation Science 38, 515–530.

Benton, W.C., and M.D. Rossetti (1992). The vehicle scheduling problem with intermittent
customer demands. Computers & Operations Research 19, 521–531.

Beraldi, P., G. Ghiani, G. Laporte and G. Musmanno (2005). Efficient neighbourhood

search for the probabilistic pickup and delivery travelling salesman problem. Networks
46, 195–198.

Berger, J. and M. Barkaoui (2004). A new hybrid genetic algorithm for the capacitated
vehicle routing problem. Journal of the Operational Research Society 54, 1254–1262.

Berger, J., M. Barkaoui and O. Bräysy (2003). A route-directed hybrid genetic approach
for the vehicle routing problem with time windows. INFOR 41, 179–194.

Bertazzi, L. and G. Paletta and M. G. Speranza (2002). Deterministic Order-Up-To Level

Policies in an Inventory Routing Problem. Transportation Science 36, 119–132.

Bertsekas, D.P. (1995). Dynamic Programming and Optimal Control, Athena Scientific,
Belmont, MA.

Bertsimas, D.J. (1988). Probabilistic combinatorial optimization problems. Ph.D. thesis,

Operations Research Center, Massachusetts Institute of Technology, Cambridge, MA.

Bertsimas, D.J. (1992). A vehicle routing problem with stochastic demand. Operations
Research 40, 574–585.

Bertsimas, D.J., and L.H. Howell (1993). Further results on the probabilistic traveling
salesman problem. European Journal of Operational Research 65, 68–95.

Bertsimas, D.J., P. Jaillet and A.R. Odoni (1990). A priori optimisation. Operations Re-
search 38, 1019–1033.

Bertsimas, D.J., and D. Simchi-Levi (1996). A new generation of vehicle routing research:
Robust algorithms addressing uncertainty. Operations Research 44, 286–304.

55
Blasum, U., and W. Hochstättler (2000). Application of the branch and cut method to
the vehicle routing problem. Technical Report ZPR2000-386, ZPR, Universität zu Köln.
URL: http://www.zaik.uni-koeln.de/ paper.

Bozkaya, B., E. Erkut and G. Laporte (2003). A tabu search algorithm and adaptive
memory procedure for political districting. European Journal of Operational Research
144, 12–26.

Bramel, J., and D. Simchi-Levi (1995). A location based heuristic for general routing
problems. Operations Research 43, 649–660.

Bramel, J., and D. Simchi-Levi (1997). On the effectiveness of set covering formulations
for the vehicle routing problem with time windows. Operations Research 45, 295–301.

Bramel, J., and D. Simchi-Levi (2002). Set-covering-based algorithms for the capacitated
VRP, in: P. Toth and D Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs
on Discrete Mathematics and Applications, Philadelphia, pp. 85–108.

Brandão, J. (1998). Metaheuristic for the vehicle routing problem with time windows, in:
S. Voss, S. Martello, I.H. Osman and C. Roucairol (eds.), Meta-Heuristics: Advances
and Trends in Local Search Paradigms for Optimization, Kluwer, Boston, pp. 19–36.

Bräysy, O. (2002). Fast local searches for the vehicle routing problem with time windows.
INFOR 40, 319–330.

Bräysy, O. (2003). A reactive variable neighborhood search for the vehicle routing problem
with time windows. INFORMS Journal on Computing 15, 347–368.

Bräysy, O., and M. Gendreau (2005a). Vehicle routing problem with time windows, Part
I: Route construction and local search algorithms. Transportation Science 39, 104–118.

Bräysy, O., and M. Gendreau (2005b). Vehicle routing problem with time windows, Part
II: Metaheuristics. Transportation Science 39, 119–139.

Burns, L.D., R.W. Hall, D.E. Blumenfeld and C.F. Daganzo (1985). Distribution strategies
that minimize transportation and inventory costs. Operations Research 33, 469–490.

Campbell, A.M., L. Clarke, A.J. Kleywegt and M.W.P. Savelsbergh (1998). The inven-
tory routing problem, in: T.G. Crainic and G. Laporte (eds.), Fleet Management and
Logistics, Kluwer, Boston, pp. 95–112.

Campbell, A.M., and M.W.P. Savelsbergh (2004a). Delivery volume optimization. Trans-
portation Science 38, 210–223.

56
Campbell, A.M., and M.W.P. Savelsbergh (2004b). A decomposition approach for the
inventory-routing problem. Transportation Science, 38, 488–502.

Campbell, A.M., and M.W.P. Savelsbergh (2004c). Efficient insertion heuristics for vehicle
routing and scheduling problems. Transportation Science. 38, 269–378.

Campos, V., A. Corberán and E. Mota (1991). Polyhedral results for a vehicle routing
problem. European Journal of Operational Research 52, 75–85.

Carraway, R.L., T.L. Morin and H. Moskowitz (1989). Generalized dynamic programming
for stochastic combinatorial optimization. Operations Research 37, 819–829.

Çetinkaya, S. and Lee, C. Y. (2000). Stock Replenishment and Shipment Scheduling for
Vendor Managed Inventory Systems. Management Science 46, 217–232.

Chabrier, A. (2006). Vehicle routing problem with elementary shortest path based column
generation. Computers & Operations Research, forthcoming.

Chan, L.M., A. Federgruen and D. Simchi-Levi (1998). Probabilistic analyses and practical
algorithms for inventory-routing models. Operations Research 46, 96–106.

Chiang, W.-C., and R.A. Russell (1997). A reactive tabu search metaheuristic for the
vehicle routing problem with time windows. INFORMS Journal on Computing 9, 417–
430.

Chien, T., A. Balakrishnan and R. Wong (1989). An integrated inventory allocation and
vehicle routing problem. Transportation Science 23, 67–76.

Christiansen, M. (1999). Decomposition of a Combined Inventory and Time Constrained

Ship Routing Problem. Transportation Science 33, 3–16.

Christiansen, M. and Nygreen, B. (1998a). A Method for Solving Ship Routing Problems
with Inventory Constraints. Annals of Operations Research 81, 357–378.

Christiansen, M. and Nygreen, B. (1998b). Modelling Path Flows for a Combined Ship
Routing and Inventory Management Problem. Annals of Operations Research 82, 391–
412.

Christofides, N., and A. Mingozzi (1989). Vehicle routing: Practical and algorithmic as-
pects, in: C.F.H. Van Rijn (ed.), Logistics: Where Ends Have to Meet, Pergamon Press,
Oxford, pp. 30–48.

Christofides, N., A. Mingozzi and P. Toth (1979). The vehicle routing problem, in: N.
Christofides, A. Mingozzi, P. Toth and C. Sandi (eds.), Combinatorial Optimization,
Wiley, Chichester, pp. 315–338.

57
Christofides, N., A. Mingozzi and P. Toth (1981a). Exact algorithms for the vehicle rout-
ing problem based on the spanning tree and shortest path relaxations. Mathematical
Programming 20, 255–282.

Christofides, N., A. Mingozzi and P. Toth (1981b). State-space relaxation procedures for
the computation of bounds to routing problems. Networks 11, 145–164.

Clarke, G., and J.W. Wright (1964). Scheduling of vehicles from a central depot to a
number of delivery points. Operations Research 12, 568–581.

Cook, W., and J.L. Rich (1999). A parallel cutting-plane algorithm for the vehicle routing
problem with time windows. Technical Report TR99-04, Computational and Applied
Mathematics Department, Rice University, TX.

Cordeau, J.-F., G. Desaulniers, J. Desrosiers, M.M. Solomon and F. Soumis (2002a). VRP
with Time Windows, in: P. Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM
Monographs on Discrete Mathematics and Applications, Philadelphia, pp. 157–193.

Cordeau, J.-F., M. Gendreau, A. Hertz, G. Laporte and J.-S. Sormany (2005). New heuris-
tics for the vehicle routing problem, in: A. Langevin and D. Riopel (eds.), Logistics
Systems: Design and Optimization, Springer, New York, 279-297, 2005.

Cordeau, J.-F., M. Gendreau and G. Laporte (1997). A tabu search heuristic for periodic
and multi-depot vehicle routing problems. Networks 30, 105–119.

Cordeau, J.-F., M. Gendreau, G. Laporte, J.-Y. Potvin and F. Semet (2002b). A guide to
vehicle routing heuristics. Journal of the Operational Research Society 53, 512–522.

Cordeau, J.-F., and G. Laporte (2001). A tabu search algorithm for the site dependent
vehicle routing problem with time windows. INFOR 39, 292–298.

Cordeau, J.-F., and G. Laporte (2004). Tabu search heuristics for the vehicle routing
problem, in: C. Rego and B. Alidaee (eds.), Metaheuristic Optimization via Memory
and Evolution: Tabu Search and Scatter Search, Kluwer, Boston, pp. 145–163.

Cordeau, J.-F., G. Laporte and A. Mercier (2001). A unified tabu search heuristic for vehicle
routing problems with time windows. Journal of the Operational Research Society 52,
928–936.

Cordeau, J.-F., G. Laporte and A. Mercier (2004). An improved tabu search algorithm
for the handling of route duration constraints in vehicle routing problems with time
windows. Journal of the Operational Research Society 55, 542–546.

Cordone, R., and R. Wolfler Calvo (2001). A heuristic for the vehicle routing problem with
time windows. Journal of Heuristics 7, 107–129.

58
Cornuéjols, G., and F. Harche (1993). Polyhedral study of the capacitated vehicle routing
problem. Mathematical Programming 60, 21–52.

Croes, A. (1958). A method for solving traveling salesman problems. Operations Research
6, 791–812.

Danna, E., and C. Le Pape (2003). Accelerating branch-and-price with local search: A
case study on the vehicle routing problem with time windows. Technical Report 03-006,
ILOG.

Dantzig, G.B. and J.M. Ramser (1959). The truck dispatching problem. Management Sci-
ence 6, 81–91.

Dantzig, G.B., and P. Wolfe (1960). Decomposition principle for linear programming. Op-
erations Research 8, 101–111.

De Backer, B., V. Furnon, P. Kilby, P. Prosser and P. Shaw (2000). Solving vehicle routing
problems using constraint programming and metaheuristics. Journal of Heuristics 6,
501–523.

Dell’Amico, M., and P. Toth (2000). Algorithms and codes for dense assignment problems:
The state of the art. Discrete Applied Mathematics 100, 17–48.

Desrochers, M., J. Desrosiers and M.M. Solomon (1992). A new optimization algorithm
for the vehicle routing problem with time windows. Operations Research 40, 342–354.

Desrochers, M., and G. Laporte (1991). Improvements and extensions to the Miller-Tucker-
Zemlin subtour elimination constraints. Operations Research Letters 10, 27–36.

Desrochers, M., and F. Soumis (1988). A generalized permanent labeling algorithm for the
shortest path problem with time windows. INFOR 26, 191–212.

Desrochers, M., J.K. Lenstra, M.W.P. Savelsbergh and F. Soumis (1988). Vehicle Routing
with Time Windows: Optimization and Approximation, in: B.L. Golden and A.A. Assad
(eds.), Vehicle Routing: Methods and Studies, North-Holland, Amsterdam, pp. 65–84.

Desrochers, M., and T.W. Verhoog (1989). A matching based savings algorithm for the
vehicle routing problem. Les Cahiers du GERAD G–89–04, HEC Montréal.

Dorigo, M., G. Di Caro and L.M. Gambardella (1999). Ant algorithms for discrete opti-
mization. Artificial Life 5, 137–172.

Drezner, Z. (2003). A new genetic algorithm for the quadratic assignment problem. IN-
FORMS Journal on Computing 15, 320–330.

59
Dror, M., and M.O. Ball (1987). Inventory/routing: Reduction from an annual to a short
period problem. Naval Research Logistics Quarterly 34, 891–905.

Dror, M., M.O. Ball and B.L. Golden (1985). A computational comparison of algorithms
for the inventory routing problem. Annals of Operations Research 4, 3–23.

Dror, M., G. Laporte and F.V. Louveaux (1993). Vehicle routing with stochastic demands
and restricted failures. Zeitschrift für Operations Research 37, 273–283.

Dror, M., G. Laporte and P. Trudeau (1989). Vehicle routing with stochastic demands:
Properties and solution frameworks. Transportation Science 23, 166–176.

Dror, M., and P. Trudeau (1986). Stochastic vehicle routing with modified savings algo-
rithm. European Journal of Operational Reserach 23, 228–235.

Dueck, G. (1990). New optimization heuristics, the great deluge algorithm and the record-
to-record travel. Technical Report, IBM Germany, Heidelberg Scientific Center.

Dueck, G. (1993). New optimization heuristics: the great deluge algorithm and the record-
to-record travel. Journal of Computational Physics 104, 86–92.

Ergun, Ö., J.B. Orlin and A. Steele-Feldman (2003). Creating very large scale neigh-
borhoods out of smaller ones by compounding moves: A study on the vehicle routing
problem. MIT Sloan Working Paper no. 4393-02, Massachusetts Institute of Technology,
Cambridge, MA.

Federgruen, A., and P. Zipkin (1984). A combined vehicle routing and inventory allocation
problem. Operations Research 32, 1019–1036.

Fischetti, M., and P. Toth (1989). An additive bounding procedure for combinatorial op-
timization problems. Operations Research 37, 319–328.

Fischetti, M., P. Toth and D. Vigo (1994). A branch-and-bound algorithm for the capaci-
tated vehicle routing problem on directed graphs. Operations Research 42, 846–859.

Fisher, M.L. (1994). Optimal solution of vehicle routing problems using minimum k-trees.
Operations Research 42, 626–642.

Fisher, M.L., A. Greenfield, R. Jaikumar and P. Kedia (1982). Real-time scheduling of a

bulk delivery fleet: Practical application of Lagrangean relaxation. Technical Report,
The Wharton School, University of Pennsylvania.

Fisher, M.L., and R. Jaikumar (1981). A generalized assignment heuristic for the vehicle
routing problem. Networks 11, 109–124.

60
Fisher, M.L., K.O. Jörnsten and O.B.G. Madsen (1997). Vehicle routing with time windows
- Two optimization algorithms. Operations Research 45, 488–492.

Flood, M.M. (1956). The travelling salesman problem. Operations Research 4, 61–75.

Foster, B.A., and D.M. Ryan (1976). An integer programming approach to the vehicle
scheduling problem. Operations Research 27, 367–384.

Fukusawa, R., M. Poggi de Aragão, M. Reis and E. Uchoa (2003). Robust branch-and-cut-
and-price for the capacitated vehicle routing problem. Technical Report RPEP Vol.3
no.8, Universidade Federal Fluminense, Niterói, Brasil.

Gallego, G., and D. Simchi-Levi (1990). On the effectiveness of direct shipping strategy for
the one-warehouse multi-retailer r-systems. Management Science 36, 240–243.

Gambardella, L.M., É.D. Taillard and G. Agazzi (1999). MACS-VRPTW: A multiple ant
colony system for vehicle routing problems with time windows, in: D. Corne, M. Dorigo
and F. Glover (eds.), New Ideas in Optimization, McGraw-Hill, London, pp. 63–76.

Gaur, V. and M.L. Fisher (2004). A Periodic Inventory Routing Problem at a Supermarket
Chain. Operations Research 52, 813–822.

Gehring, H., and J. Homberger (2002). Parallelization of a two-phase metaheuristic for

routing problems with time windows. Journal of Heuristics 8, 251–276.

Gendreau, M., A. Hertz and G. Laporte (1992). New insertion and post-optimization pro-
cedures for the traveling selesman problem. Operations Research 40, 1083–1094.

Gendreau, M., A. Hertz and G. Laporte (1994). A tabu search heuristic for the vehicle
routing problem. Management Science 40, 1276–1290.

Gendreau, M., A. Hertz, G. Laporte and M. Stan (1998). A generalized insertion heuristic
for the traveling salesman problem with time windows. Operations Research 43, 330–335.

Gendreau, M., G. Laporte and J.-Y. Potvin (2002). Metaheuristics for the capacitated
VRP, in: P. Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs
on Discrete Mathematics and Applications, Philadelphia, pp. 129–154.

Gendreau, M., G. Laporte and R. Séguin (1995). An exact algorithm for the vehicle routing
problem with stochastic customers and demands. Transportation Science 29, 143–155.

Gendreau, M., G. Laporte and R. Séguin (1996). A tabu search algorithm for the vehicle
routing problem with stochastic demands and customers. Operations Research 44, 469–
477.

61
Ghaziri, H. (1993). Algorithmes connexionistes pour l’optimisation combinatoire. Thèse de
doctorat, École Polytechnique Fédérale de Lausanne, Switzerland.

Ghiani, G., G. Laporte and F. Semet (2004). The black and white traveling salesman
problem. Operations Research, forthcoming.

Gillett, B.E., and L.R. Miller (1974). A heuristic algorithm for the vehicle-dispatch prob-
lem. Operations Research 21, 340–349.

Glover, F. (1992). New ejection chain and alternating path methods for traveling salesman
problems, in: O. Balci, R. Sharda and S. Zenios (eds.), Computer Science and Operations
Research: New Developments in Their Interfaces, Pergamon Press, Oxford, pp. 449–509.

Golden, B.L., A.A. Assad and R. Dahl (1984). Analysis of a large scale vehicle routing
problem with an inventory component. Large Scale Systems 7, 181–190.

Golden, B.L., A.A. Assad and E.A. Wasil (2002). Routing vehicles in the real world:
Applications in the solid waste, beverage, food, dairy, and newspaper industries, in: P.
Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs on Discrete
Mathematics and Applications, Philadelphia, pp. 245–286.

Golden, B.L., T.L. Magnanti and H.Q. Nguyen (1977). Implementing vehicle routing algo-
rithms. Networks 7, 113–148.

Golden, B.L., and W.R. Stewart (1978). Vehicle routing with probabilistic demands, in: D.
Hogben and D. Fife (eds.), Computer Science and Statistics: Tenth Annual Symposium
on the Interface, NBS Special Publication 503, pp. 252–259.

Golden, B.L., E.A. Wasil, J.P. Kelly and I-M. Chao (1998). Metaheuristics in vehicle
routing, in: T.G Crainic and G. Laporte (eds.), Fleet Management and Logistics, Kluwer,
Boston, pp. 33–56.

Gouveia, L. (1995). A result on projection for the vehicle routing problem. Journal of
Operational Research Society 85, 610–624.

Hadjiconstantinou, E., N. Christofides and A. Mingozzi (1995). A new exact algorithm for
the vehicle routing problem based on q-paths and k-shortest paths relaxations. Annals
of Operations Research 61, 21–43.

Haimovich, M., and A.H.G. Rinnooy Kan (1985). Bounds and heuristics for capacitated
routing problems. Mathematics of Operations Research 10, 527–542.

Haughton, M.A. (1998). The performance of route modification and demand stabilization
strategies in stochastic vehicle routing. Transportation Research 32, 551–566.

62
Haughton, M.A. (2000). Quantifying the benefits of route reoptimisation under stochastic
customer demands. Journal of the Operational Research Society 51, 320–332.
Held, M., and R.M. Karp (1971). The traveling salesman problem and minimum spanning
trees: Part II. Mathematical Programming 1, 6–25.
Held, M., P. Wolfe and M.P. Crowder (1974). Validation of the subgradient optimization.
Mathematical Programming 6, 62–88.
Herer, Y. and Roundy, R. (1997). Heuristics for a One-Warehouse Multiretailer Distribu-
tion Problem with Performance Bounds. Operations Research 45, 102–115.
Hjorring, C., and J. Holt (1999). New optimality cuts for a single-vehicle stochastic routing
problem. Annals of Operations Research 86, 569–585.
Homberger, J., and H. Gehring (1999). Two evolutionary metaheuristics for the vehicle
routing problem with time windows. INFOR 37, 297–318.
Houck, D.J., J.-C. Picard, M. Queyranne and R.R. Vemuganti (1980). The traveling sales-
man problem as a constrained shortest path problem: Theory and computational expe-
rience. Opsearch 17, 93–109.
Hvattum, L.M., A. Løkketangen and G. Laporte (2003). Solving a dynamic and stochastic
vehicle routing problem with a sample scenario hedging heuristic, submitted.
Ioannou, G., M. Kritikos and G. Prastacos (2001). A greedy look-ahead heuristic for the
vehicle routing problem with time windows. Journal of the Operational Research Society
52, 523–537.
Irnich, S., and D. Villeneuve (2003). The shortest path problem with resource constraints
and k-cycle elimination for k ≥ 3. Technical Report, Rheinisch-Westfälische Technische
Hochschule, Aachen, Germany.
Jaillet, P. (1985). Probabilistic traveling salesman problem. Ph.D. thesis, Operations Re-
search Center, Massachusetts Institute of Technology, Cambridge, MA.
Jaillet, P. (1987). Stochastic routing problems, in: G. Andreatta, F. Mason and P. Serafini
(eds.), Stochastics in Combinatorial Optimization, Word Scientific, Singapore, pp. 197–
213.
Jaillet, P. (1988). A priori solution of a traveling salesman problem in which a random
subset of the customers are visited. Operations Research 36, 929–936.
Jaillet, P., J.F. Bard, L. Huang and M. Dror (2002). Delivery cost approximations for
inventory routing problems in a rolling horizon framework. Transportation Science 3,
292–300.

63
Jaillet, P., and A.R. Odoni (1988). The probabilistic vehicle routing problem, in: B.L.
Golden and A.A. Assad (eds.), Vehicle Routing: Methods and Studies, North-Holland,
Amsterdam, pp. 293–318.

Jézéquel, A. (1985). Probabilistic vehicle routing problems. M.Sc. Dissertation, Department

of Civil Engineering, Massachusetts Institute of Technology, Cambridge, MA.

Kallehauge, B., J. Larsen and O.B.G. Madsen (2006). Lagrangean duality applied to the
vehicle routing with time windows. Computers & Operations Research, forthcoming.

Kao, E.P.C. (1978). A preference order dynamic program for a stochastic traveling salesman
problem. Operations Research 26, 1033–1045.

Kenyon, A.S., and D.P. Morton (2003). Stochastic vehicle routing with random travel
times. Transportation Science 37, 69–82.

Kilby, P.J., P. Prosser and P. Shaw (1998). Guided local search for the vehicle routing
problem with time windows, in: S. Voss, S. Martello, I.H. Osman and C. Roucairol (eds.),
Meta Heuristics: Advances and Trends in Local Search Paradigms for Optimisation,
Kluwer, Boston, pp. 473–486.

Kindervater, G.A.P., and M.W.P. Savelsbergh (1997). Vehicle routing: Handling edge
exchanges, in: E.H.L. Aarts and J.K. Lenstra (eds.), Local Search in Combinatorial
Optimization, Wiley, Chichester, pp. 337–360.

Kleywegt, A.J., V. Nori and M.W.L. Savelsbergh (2002). The stochastic inventory routing
problem with direct deliveries. Transportation Science 36, 94–118.

Kleywegt, A.J., V. Nori and M.W.L. Savelsbergh (2004). Dynamic programming approxi-
mations for a stochastic inventory routing problem. Transportation Science 38, 42–70.

Kohl, N., J. Desrosiers, O.B.G. Madsen, M.M. Solomon and F. Soumis (1999). 2-path cuts
for the vehicle routing problem with time windows. Transportation Science 33, 101–116.

Kohl, N., and O.B.G. Madsen (1997). An optimization algorithm for the vehicle routing
problem with time windows based on Lagrangean relaxation. Operations Research 45,
395–406.

Kolen, A.W.J., A.H.G. Rinnooy Kan and H.W.J.M. Trienekens (1987). Vehicle routing
with time windows. Operations Research 35, 256–273.

Kontoravdis, G., and J.F. Bard (1995). A GRASP for the vehicle routing problem with
time windows. ORSA Journal on Computing 7, 10–23.

64
Lambert, V., G. Laporte and F.V. Louveaux (1993). Designing collection routes through
bank branches. Computers & Operations Research 20, 783–791.

Laporte, G., and F.V. Louveaux (1993). The integer L-shaped method for stochastic integer
programs with complete recourse. Operations Research Letters 13, 133–142.

Laporte, G., and F.V. Louveaux (1998). Solving stochastic routing problems with the
Integer L-Shaped method, in: T.G. Crainic and G. Laporte (eds.), Fleet Management
and Logistics, Kluwer, Boston, pp. 159–167.

Laporte, G., and F.V. Louveaux (1990). Formulations and bounds for the stochastic capac-
itated vehicle routing problem with uncertain supplies, in: J. Gabzewicz, J.-F. Richard
and L. Wolsey (eds.), Economic Decision Making: Games, Econometrics and Optimisa-
tion. North-Holland, Amsterdam, pp. 443–455.

Laporte, G., F.V. Louveaux and H. Mercure (1989). Models and exact solutions for a class
of stochastic location-routing problems. European Journal of Operational Research 39,
71–78.

Laporte, G., F.V. Louveaux and H. Mercure (1992). The vehicle routing problem with
stochastic travel times. Transportation Science 26, 161–170.

Laporte, G., F.V. Louveaux and H. Mercure (1994). A priori optimization of the proba-
bilistic traveling salesman problem. Operations Research 42, 543–549.

Laporte, G., F.V. Louveaux and L. Van hamme (2002). An integer L-shaped algorithm for
the capacitated vehicle routing problem with stochastic demands. Operations Research
50, 415–423.

Laporte, G., H. Mercure and Y. Nobert (1986). An exact algorithm for the asymmetrical
capacitated vehicle routing problem. Networks 16, 33–46.

Laporte, G., and Y. Nobert (1987). Exact algorithms for the vehicle routing problem.
Annals of Discrete Mathematics 31, 147–184.

Laporte, G., Y. Nobert and M. Desrochers (1985). Optimal routing under capacity and
distance restrictions. Operations Research 33, 1050–1073.

Laporte, G., and F. Semet (2002). Classical heuristics for the capacitated VRP, in: P.
Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs on Discrete
Mathematics and Applications, Philadelphia, pp. 109–128.

Larson, R.C. (1988). Transportation of sludge to the 106-mile site: An inventory routing
problem for fleet sizing and logistic system design. Transportation Science 22, 186–198.

65
Lau, H.C., Q. Liu and H. Ono (2002). Integrating local search and network flow to solve
the inventory routing problem. American Association for Artificial Intelligence 2, 9–14.
Lau, H.C., M. Sim and K.M. Teo (2003). Vehicle routing problem with time windows and
a limited number of vehicles. European Journal of Operational Research 148, 559–569.
Leipälä, T. (1978). On the solutions of stochastic traveling salesman problems. European
Journal of Operational Research 2, 291–297.
Letchford, A., R.W. Eglese and J. Lysgaard (2002). Multistars, partial multistars and the
capacitated vehicle routing problem. Mathematical Programming 94, 21–40.
Li, F., B.L. Golden and E.A. Wasil (2005). Very large-scale vehicle routing: New test
problems, algorithms and results. Computers & Operations Research 32, 1165–1179.
Li, H., and A. Lim (2003). Local search with annealing-like restarts to solve the VRPTW.
European Journal of Operational Research 150, 115–127.
Lin, S. (1965). Computer solutions of the travelling salesman problem. Bell System Tech-
nical Journal 44, 2245–2269.
Louveaux, F.V. (1998). An introduction to stochastic transportation models, in: M. Labbé,
G. Laporte, K. Tanczos, P. Toint (eds.), Operations Research and Decision Aid Method-
ologies in Traffic and Transportation Management. NATO ASI Series, Series F: Com-
puter and Systems Sciences, Vol. 166, Springer-Verlag, Berlin and Heidelberg, pp. 244–
263.
Lysgaard, J., A. Letchford and R.W. Eglese (2004). A new branch-and-cut algorithm for
the capacitated vehicle routing problem. Mathematical Programming 100, 423–445.
Martinhon, C., A. Lucena and N. Maculan (2000). A relax and cut algorithm for the vehicle
routing problem. Technical Report RT-05/00, Universidade Federal Fluminense, Niterói,
Brasil.
Mester, D., and O. Bräysy (2005). Active guided evolution strategies for large scale vehicle
routing problem with time windows. Computers & Operations Research 32, 1593–1614.
Miller, D.L. (1995). A matching based exact algorithm for capacitated vehicle routing
problems. ORSA Journal on Computing 7, 1–9.
Miller, D.L., and J.F. Pekny (1995). A staged primal-dual algorithm for perfect b-matching
with edge capacities. ORSA Journal on Computing 7, 298–320.
Mingozzi, A., N. Christofides and E. Hadjiconstantinou (1994). An exact algorithm for
the vehicle routing problem based on the set partitioning formulation. Technical Report,
Department of Mathematics, University of Bologna, Italy.

66
Minkoff, A. (1993). A markov decision model and decomposition heuristic for dynamic
vehicle dispatching. Operations Research 41, 77–90.

Mladenović, N., and P. Hansen (1997). Variable neighborhood search. Computers & Oper-
ations Research 24, 1097–1100.

Mole, R.H., and S.R. Jameson (1976). A sequential route-building algorithm employing a
generalized savings criterion. Operational Research Quarterly 27, 503–511.

Moscato, P. and C. Cotta (2003). A gentle introduction to memetic algorithms, in: F.

Glover and G.A. Kochenberger (eds.), Handbook of Metaheuristics, Kluwer, Boston,
pp. 105–144.

Naddef, D., and G. Rinaldi (2002). Branch-and-cut algorithms for the capacitated VRP,
in: P. Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs on
Discrete Mathematics and Applications, Philadelphia, pp. 53–84.

Nelson, M.D., K.E. Nygard, J.H. Griffin and W.E. Shreve (1985). Implementation tech-
niques for the vehicle routing problem. Computers & Operations Research 12, 273–283.

Or, I. (1976). Traveling salesman-type combinatorial problems and their relation to the
logistics of blood banking. Ph.D. thesis, Department of Industrial Engineering and Man-
agement Science, Northwestern University, Evanston, IL.

Osman, I.H. (1993). Metastrategy simulated annealing and tabu search algorithms for the
vehicle routing problem. Annals of Operations Research 41, 421–451.

Paessens, H. (1988). The savings algorithm for the vehicle routing problem. European
Journal of Operational Research 34, 336–344.

Potvin, J.-Y. (1996). Genetic algorithms for the traveling salesman problem. Annals of
Operations Research 63, 339–370.

Potvin, J.-Y., and S. Bengio (1996). The vehicle routing problem with time windows - Part
II: Genetic search. INFORMS Journal on Computing 8, 165–172.

Potvin, J.-Y., T. Kervahut, B.L. Garcia and J.-M. Rousseau (1996). The vehicle routing
problem with time windows - Part I: Tabu search. INFORMS Journal on Computing 8,
158–164.

Potvin, J.-Y., and J.-M. Rousseau (1993). A parallel route building algorithm for the vehicle
routing and scheduling problem with time windows. European Journal of Operational
Research 66, 331–340.

67
Potvin, J.-Y., and J.-M. Rousseau (1995). An exchange heuristic for routing problems with
time windows. Journal of the Operational Research Society 46, 1433–1446.

Prins, C. (2004). A simple and effective evolutionary algorithm for the vehicle routing
problem. Computers & Operations Research 31, 1985–2002.

Ralphs, T.K., L. Kopman, W.R. Pulleyblank and L.E. Trotter Jr. (2003). On the capaci-
tated vehicle routing problem. Mathematical Programming B 94, 343–359.

Rechenberg, I. 1973. Evolutionsstrategie, Fromman-Holzboog, Stuttgart, Germany.

Rego, C. (1998). A subpath ejection method for the vehicle routing problem. Management
Science 44, 1447–1459.

Rego, C., and C. Roucairol (1996). A parallel tabu search algorithm using ejection chains
for the vehicle routing problem, in: I.H Osman and J.P. Kelly (eds.), Meta-Heuristics:
Theory and Applications, Kluwer, Boston, pp. 661–675.

Reimann, M., K. Doerner and R.F. Hartl (2004). D-ants: Savings based ants divide and
conquer for the vehicle routing problem. Computers & Operations Research 31, 563–591.

Reimann, M., R. Rubio and L.M. Wein (1999). Heavy traffic analysis of the dynamic
stochastic inventory-routing problem. Transportation Science 33, 361–380.

Renaud, J., F.F. Boctor and G. Laporte (1996a). A fast composite heuristic for the sym-
metric traveling salesman problem. INFORMS Journal on Computing 8, 134–143.

Renaud, J., F.F. Boctor and G. Laporte (1996b). An improved petal heuristic for the
vehicle routing problem. Journal of the Operational Research Society 47, 329–336.

Rochat, Y., and É.D. Taillard (1995). Probabilistic diversification and intensification in
local search for vehicle routing. Journal of Heuristics 1, 147–167.

Rossi, F.A., and I. Gavioli (1987). Aspects of heuristic methods in the “Probabilistic Trav-
eling Salesman Problem”, in: G. Andreatta, F. Mason and P. Serafini (eds.), Stochastics
in Combinatorial Optimization, World Scientific, Singapore, pp. 214–227.

Russell, R.A. (1977). An effective heuristic for the M -tour traveling salesman problem with
some side conditions. Operations Research 25, 517–524.

Russell, R.A. (1995). Hybrid heuristics for the vehicle routing problem with time windows.
Transportation Science 29, 156–166.

Ryan, D.M., C. Hjorring and F. Glover (1993). Extensions of the petal method for vehicle
routing. Journal of Operational Research Society 44, 289–296.

68
Savelsbergh, M.W.P. (1985). Local search in routing problems with time windows. Annals
of Operations Research 4, 285–305.

Savelsbergh, M.W.P. (1990). En efficient implementation of local search algorithms for

constrained routing problems. European Journal of Operational Research 47, 75–85.

Savelsbergh, M.W.P. (1992). The vehicle routing problem with time windows: Minimizing
route duration. ORSA Journal on Computing 4, 146–154.

Savelsbergh, M.W.P. and J.-H. Song (2005). Inventory Routing with Continuous Moves.
Computers and Operations Research, forthcoming.

Schulze, J., and T. Fahle (1999). A parallel algorithm for the vehicle routing problem with
time window constraints. Annals of Operations Research 86, 585–607.

Secomandi, N. (1998). Exact and heuristic dynamic programming algorithms for the vehicle
routing problem with stochastic demands. Ph.D. Dissertation, Faculty of the College of
Business Administration, University of Houston, TX.

Secomandi, N. (2000). Comparing neuro-dynamic programming algorithms for the vehicle

routing problem with stochastic demands. Computers & Operations Research 27, 1201–
1225.

Secomandi, N. (2003). Analysis of a rollout approach to sequencing problems with stochas-

tic routing applications. Journal of Heuristics 9, 321–352.

Séguin, R. (1994). Problèmes stochastiques de tournées de véhicules. Ph.D. Thesis,

Département d’informatique et de recherche opérationnelle, Université de Montréal,
Canada.

Semet, F. and É.D. Taillard (1993). Solving real-life vehicle routing problems efficiently
using tabu search. Annals of Operations Research, 41, 469–488.

Shaw, P. (1998). Using constraint programming and local search methods to solve vehi-
cle routing problems, in: M. Maher and J.-F. Puget (eds.), Principles and Practice of
Constraint Programming, Springer-Verlag, New York, pp. 417–431.

Sniedovich, M. (1981). Analysis of a preference order traveling salesman problem. Opera-

tions Research 29, 1234–1237.

Solomon, M.M. (1987). Algorithms for the vehicle routing and scheduling problems with
time window constraints. Operations Research 35, 254–265.

69
Solomon, M.M., E.K. Baker and J.R. Schaffer (1988). Vehicle routing and scheduling prob-
lems with time window constraints: Efficient implementations of solution improvement
procedures, in: B.L. Golden and A.A. Assad (eds.), Vehicle Routing: Methods and Stud-
ies, North-Holland, Amsterdam, pp. 85–106.

Song, J.-H. and M. W. P. Savelsbergh (2005). Performance Measurement for Inventory

Routing. Transportation Science, forthcoming.

Stewart, W.R., and B.L. Golden (1983). Stochastic vehicle routing: A comprehensive
approach. European Journal of Operational Research 14, 371–385.

Taillard, É.D. (1993). Parallel iterative search methods for vehicle routing problems. Net-
works 23, 661–673.

Taillard, É.D., P. Badeau, M. Gendreau, F. Guertin and J.-Y. Potvin (1997). A tabu search
heuristic for the vehicle routing problem with soft time windows. Transportation Science
31, 170–186.

Tan, K.C., L.H. Lee and K. Ou (2001). Hybrid genetic algorithms in solving vehicle routing
problems with time window constraints. Asia-Pacific Journal of Operational Research
18, 170–186.

Tarantilis, C.-D. and C.T. Kiranoudis (2002). Bone Route: Adaptive memory method for
effective fleet management. Annals of Operations Research 115, 227–241.

Thangiah, S.R., and P. Petrovic (1998). Introduction to genetic heuristics and vehi-
cle routing problems with complex constraints, in: Advances in Computational and
Stochastic Optimization, Logic Programming, and Heuristic Search, Operations Re-
search/Computer Science Interfaces, Series 9, Kluwer, Boston, pp. 253–286.

Thompson, P.M., and H.N. Psaraftis (1993). Cyclic transfer algorithms for multi-vehicle
routing and scheduling problems. Operations Research 41, 935–946.

Tillman, F. (1969). The multiple terminal delivery problem with probabilistic demands.
Transportation Science 3, 192–204.

Toth, P., and D. Vigo (1995). An exact algorithm for the capacitated shortest spanning
arborescence. Annals of Operations Research 61, 121–142.

Toth, P., and D. Vigo (1997). An exact algorithm for the vehicle routing problem with
backhauls. Transportation Science 31, 372–385.

Toth, P., and D. Vigo (1998). Exact algorithms for vehicle routing, in: T. Crainic and G.
Laporte (eds.), Fleet Management and Logistics, Kluwer, Boston, pp. 1–31.

70
Toth, P., and D. Vigo, eds. (2002a). The Vehicle Routing Problem, SIAM Monographs on
Discrete Mathematics and Applications, Philadelphia.

Toth, P., and D. Vigo (2002b). An overview of vehicle routing problems, in: P. Toth and D.
Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics
and Applications, Philadelphia, pp. 1–26.

Toth, P., and D. Vigo (2002c). Branch-and-bound algorithms for the capacitated VRP,
in: P. Toth and D. Vigo (eds.), The Vehicle Routing Problem, SIAM Monographs on
Discrete Mathematics and Applications, Philadelphia, pp. 29–51.

Toth, P., and D. Vigo (2002d). Models, relaxations and exact approaches for the capaci-
tated vehicle routing problem. Discrete Applied Mathematics 123, 487–512.

Toth, P., and D. Vigo (2003). The granular tabu search and its application to the vehicle
routing problem. INFORMS Journal on Computing 15, 333–346.

Trudeau, P., and M. Dror (1992). Stochastic inventory routing: Route design with stockouts
and route failures. Transportation Science 26, 171–184.

Van Breedam, A. (1994). An analysis of the behavior of heuristics for the vehicle routing
problem for a selection of problems with vehicle-related, customer-related, and time-
related constraints. Ph.D. Dissertation, University of Antwerp, Belgium.

Van Slyke, R. and R. J-B.Wets (1969). L-Shaped programs with applications to optimal
control and stochastic programming. SIAM Journal of Applied Mathematics 17, 638–653.

Verweij, B., S. Ahmed, A.J. Kleywegt, G.L. Nemhauser and A. Shapiro (2003). The sample
average approximation method applied to stochastic routing problems: A computational
study. Computational Optimization and Applications 24, 289–333.

Vigo, D. (1996). A heuristic algorithm for the asymmetric capacitated vehicle routing
problem. European Journal of Operational Research 89, 108–126.

Viswanathan, S. and Mathur, K. (1997). Integrating Routing and Inventory Decisions in

One-Warehouse Multiretailer Multiproduct Distribution Systems. Management Science
43, 294–312.

Volgenant, A. and R. Jonker (1983). The symmetric traveling salesman problem and edge
exchange in minimal 1-trees. European Journal of Operationsl Research 12, 395–403.

Voudouris, C. (1997). Guided local search for combinatorial problems. Dissertation, Uni-
versity of Essex, United Kingdom.

71
 Waller, M., M.E. Johnson and T. Davis (1999). Vendor-managed inventory in the retail
supply chain. Journal of Business Logistics 20, 183–203.

Wang, X. and A.C. Regan (2001). Assignment models for local truckload trucking prob-
lems with stochastic service times and time window constraints. Transportation Research
Record 1171, 61–68.

Wark, P., and J. Holt (1994). A repeated matching heuristic for the vehicle routing problem.
Journal of Operational Research Society 45, 1156–1167.

Waters, C.D.J. (1989). Vehicle-scheduling problems with uncertainty and omitted cus-
tomers. Journal of the Operational Research Society 40, 1099–1108.

Webb, R. and R. Larson (1995). Period and Phase of Customer Replenishment: A New
Approach to the Strategic Inventory/Routing Problem. European Journal of Operational
Research 85, 132–148.

Willard, J.A.G. (1989). Vehicle routing using r-optimal tabu search. M.Sc. dissertation,
The Management School, Imperial College, London.

Wren, A. (1971). Computers in Transport Planning and Operation, Ian Allan, London.

Wren, A., and A. Holliday (1972). Computer scheduling of vehicles from one or more depots
to a number of delivery points. Operational Research Quarterly 23, 333–344.

Xu, J., and J.P. Kelly (1996). A network flow-based tabu search heuristic for the vehicle
routing problem. Transportation Science 30, 379–393.

Yang, W.H., K. Mathur and R.H. Ballou (2000). Stochastic vehicle routing with restocking.
Transportation Science 34, 99–112.

72

View publication stats