This article was downloaded by: [129.93.16.

3] On: 27 December 2014, At: 10:45

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

INFORMS Journal on Computing

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

A Unified Modeling and Solution Framework for Vehicle

Routing and Local Search-Based Metaheuristics
S. Irnich,

To cite this article:

S. Irnich, (2008) A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics.
INFORMS Journal on Computing 20(2):270-287. http://dx.doi.org/10.1287/ijoc.1070.0239

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

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

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

Copyright © 2008, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 INFORMS Journal on Computing informs ®
Vol. 20, No. 2, Spring 2008, pp. 270–287
issn 1091-9856
eissn 1526-5528
08
2002
0270 doi 10.1287/ijoc.1070.0239
© 2008 INFORMS

A Unified Modeling and Solution Framework for

Vehicle Routing and Local Search-Based Metaheuristics
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

S. Irnich
Deutsche Post Endowed Chair of Optimization of Distribution Networks, RWTH, Aachen University,
Aachen, Germany, [email protected]

T his paper presents a new unified modeling and heuristic solution framework for vehicle-routing problems
(VRPs) with complex side constraints. The work is focused on strong modeling capabilities as well as efficient
solution procedures to be used in all kinds of metaheuristics. From the modeling point of view, the framework
covers a variety of standard VRP types with classical constraints such as capacity, distance, route length, time
window, pairing, and precedence constraints, but also nonstandard “rich” VRPs. From the methodological point
of view, local search (LS) is the key solver engine to be used in heuristic solution procedures. First and foremost,
the framework introduces two generic techniques for the efficient exploration of edge- and node-exchange
neighborhoods. New preprocessing methods allow

nk  neighborhoods to be searched in time complexity

nk ,
i.e., without an additional effort for feasibility testing in the worst case. Moreover, for accelerating LS in the
average case, Irnich et al. [Irnich, S., B. Funke, T. Grünert. 2006. Sequential search and its application to vehicle-
routing problems. Comput. Oper. Res. 33 2405–2429] have introduced sequential search that is here adapted to
cope with rich VRPs (complex side constraints). Computational tests on different types of VRPs indicate that the
proposed techniques are highly efficient. Sequential search procedures outperform the currently most efficient
search methods, which are based on lexicographic search, on large-scale instances and for nearly all types of
neighborhoods by factors of between 10 and 1,000.
Key words: local search; vehicle routing; rich vehicle-routing problems; resource-constrained paths
History: Accepted by Michel Gendreau, Area Editor for Heuristic Search and Learning; received August 2006;
revised March 2007; accepted August 2007. Published online in Articles in Advance January 4, 2008.

1. Introduction better quality. In this context, much progress has been

The diversity of models and solution approaches in made w.r.t. the design and analysis of metaheuristics,
vehicle routing is enormous (see, e.g., Laporte 1992, i.e., problem-independent top-level general strategies
1997; Toth and Vigo 2002). This can be estimated by that guide other heuristics to search for high-quality
the fact that, in 2006 alone, a few hundred scien- feasible solutions (Ribeiro and Hansen 2002). In addi-
tific papers were published. Many of these publica- tion, metaheuristic implementations become reusable
tions meet the challenge of extending known models using software libraries (Voß and Woodruff 2002).
and methods to cope with new or extended types of However, what is missing are powerful lower-level
vehicle-routing problems (VRPs). Under the name rich VRP algorithms that are efficient and, at the same
models, researchers summarize “non-idealized models time, general.
that represent the application at hand in an adequate Research on unifying approaches for VRPs has been
way by including all important optimization crite- undertaken in different directions: Formal schemes
ria, constraints, and preferences” (Hasle et al. 2006, like those of Desrochers et al. (1990) are helpful to
p. 1753). Plenty of contributions to “rich VRPs” exist classify different types of VRPs. Integrated models, as
in the form of specialized algorithms that incorporate presented by Desrosiers et al. (1995) and Desaulniers
different types of extensions into existing problems et al. (1998), provide comprehensive mixed-integer
(Janssens et al. 2006). However, it is not clear whether programming formulations. They can be used to
results presented there are transferable to other cases. devise powerful decomposition approaches, such
What is missing are unifying modeling and solution as column generation and Lagrangean relaxation
approaches that are general (=generic) and flexible (Desaulniers et al. 2005). These methods are primar-
enough to be used in a broad range of applications. ily intended to be used as exact solution procedures,
In many other publications, the focus is on enhanc- even if they can be redesigned into approximative
ing the efficiency of existing methods or on devis- algorithms (Desaulniers et al. 2002).
ing alternative approaches that solve larger instances, In practice, VRPs can almost never be solved with
compute solutions faster, or provide solutions of exact methods because instances are too large and
270
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 271

response times of decision-support systems have to a flexible and generic but well-defined representation
be short. Thus, metaheuristics have to be applied. of feasible and infeasible route plans.
Because the majority of metaheuristics in vehicle rout- Second, the framework is intended to support effi-
ing use local-search (LS) components, the efficiency of cient solution procedures that are based on LS. In
LS is crucial. fact, LS is the key component for finding improving
One way to cope with nonstandard side constraints solutions within nearly all metaheuristics for VRPs.
and options is to use LS in combination with constraint Any LS-based procedure iteratively builds a neighbor
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

programming (CP), as suggested by Shaw (1998) and solution first and checks its feasibility and gain after-
Kilby et al. (2000). CP-based methods appear attrac- wards. If implemented in a straightforward way, this
tive because new side constraints can easily be added feasibility check causes an extra effort bounded by
to existing solvers by stating additional rules (typi- the length of a longest tour. This length is in general
cally formulated in a high-level constraint program- only bounded by

n for instances of size n, where
ming language). The problem of identifying feasi- n is the number of nodes in the problem. Techniques
ble improving solutions is solved through a general- that avoid the additional factor for cost computa-
purpose search engine. A known drawback of CP- tions and feasibility checks are already known, but
based (VRP) solution methods is, however, that the they are intrinsically tied to the lexicographic search
additional flexibility in modeling is bought at the paradigm (see Kindervater and Savelsbergh 1997 and
expense of loosing efficiency, in particular, compared §3.3.1). Here, we present new techniques for search-
to traditional LS methods. The large neighborhood search ing neighborhoods of size

nk  in

nk  time. We give
(LNS) principle, which has been used in the context sufficient conditions on the update of resources that
of CP, is very successful in finding least-cost solutions. guarantee

1 feasibility tests. The new techniques
However, LNS neighborhoods can also be searched are more generic, compatible with classical node- and
directly (Røpke and Pisinger 2006). edge-exchange neighborhoods (Bräysy and Gendreau
Research on efficient LS methods for VRPs and 2005a) and any kind of neighborhood exploration
traditional k-edge exchange neighborhoods has been strategy (Funke et al. 2005a) and, thus, enable accel-
undertaken by Kindervater and Savelsbergh (1997). erated search methods.
It seems that these techniques are not widely used, Third, the goal of all efficient LS procedures is to
probably because they seem to be intricate. In addi- find a best or first improving neighbor solution as
tion, they were not explicitly presented in a way fast as possible, not only in the worst case but in
that allows a direct adaptation to different LS oper- the average case. Classical exchange procedures for
ators and to new types of side constraints (cf. Shaw VRPs generate the neighbor solutions of a (giant) tour
1998, p. 6). by removing  edges and replacing them by  others
This paper presents a new unified modeling and (even if it is a node-exchange procedure). The choice
heuristic solution framework for VRPs with complex of these edges is made by taking k ≤  independent
side constraints. The work is focused on strong mod- decisions. Hence, the associated LS procedure can be
eling capabilities and, first and foremost, on efficient considered a tree-search method where the search tree
solution procedures. The contribution is threefold: has depth k. The two main criteria for a reduction
First, the aim of the framework is to help model of the search space (for “terminating the search” or
different real-world VRPs in a generic way, so that “pruning the search tree”) are cost and feasibility con-
a broad class of standard problem types and also siderations. As discussed in Funke et al. (2005a) and
rich VRPs can be handled. The modeling capabil- Irnich et al. (2006), one can distinguish between two
ities cover all standard types of VRPs, such as efficient approaches. Sequential search is based on the
the capacitated and distance-constrained VRP (CVRP, idea of cost-based reductions; i.e., one tries to prove at
DCVRP), the VRP with multiple depots (MDVRP), an early stage i < k that no improvement can be found
time windows (VRPTW), simultaneous delivery and that includes the nodes or edges of the stages 1  i.
pickup (VRPSDP), backhauling (VRPB), pickup-and- Lexicographic search is driven by feasibility reductions;
delivery problems (PDP), the periodic VRP (PVRP), i.e., one tries to prove at an early stage i < k that
fleet mix problems (FMP), VRPs with site dependen- no feasible exchange exists that includes the nodes
cies, vehicle and request (in)compatibilities, multiple- or edges of the stages 1  i. This paper presents
start option, limited waiting times and times on duty, concepts for applying sequential-search procedures
as well as mixtures and extensions of these (see §4). to the generic modeling framework to further reduce
The framework is mainly based on the giant-tour the effort of evaluating a neighborhood of size

nk .
representation (Christofides and Eilon 1969) and the The goal is to perform less than

nk  operations
concept of resource-constrained paths (Desaulniers in the average case (for best improvement as well as
et al. 1998, Irnich and Desaulniers 2005). It provides first improvement pivoting strategies). Computational
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
272 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

results indicate the superiority of sequential search- has to cover typical node- and edge-exchange neigh-
based approaches for a variety of VRPs with side con- borhoods, but must still allow efficient algorithmic
straints over straightforward and also lexicographic procedures to explore neighborhoods. The basis for
implementations (Kindervater and Savelsbergh 1997). such a representation is a directed routing graph G =
Finally, this paper does not present a specific meta-
V  A. Any solution of the rich VRP is represented
heuristic. The presented research is a contribution to by a single cycle in G, the so-called giant tour.
the foundations of efficient search techniques. These Solutions with more than a single vehicle consist of a
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

efficient search techniques can be seen as basic build- collection of routes. Hence, the node set V = R ∪ O ∪ D
ing blocks that can easily be integrated into different of G consists of request nodes R and route-start O and
metaheuristics (see §6). route-end nodes D. A route is a path
v0  v1   vp ,
This paper is structured as follows. Section 2 pre- starting with a route-start node v0 = o ∈ O, continuing
sents the unified framework from a modeling point of with request nodes v1   vp−1 ∈ R, and ending with
view. Section 3 points out the major tasks that have a route-end node vp = d ∈ D. The interpretation of the
to be performed in an efficient LS procedure. These request nodes depends on the problem at hand. In
tasks include efficient cost computations and feasibil- the case of the VRP, request nodes correspond to cus-
ity testings as well as setting up well-suited search tomers that have to be visited. For the PDP, a request
strategies that match with these computational tasks. node is either a pickup or a delivery. In more com-
Section 4 presents real-world constraints fitting into plex routing applications, a request may even consist
the framework and also discusses limitations of the of more than a pair of nodes.
approach. The computational tests of §5 show the For those VRPs for which transportation tasks are
effectiveness of the new solution framework. Final uniquely represented by nodes, solutions coincide
conclusions are given in §6. with Hamiltonian cycles of the routing graph. VRPs
with other types of tasks (e.g., when alternative service
or delivery options exist) and arc-routing problems
2. Modeling Framework are discussed in the Online Supplement (available at
The proposed unified modeling and solution frame- http://joc.pubs.informs/org/ecompanion.html).
work for vehicle routing and LS-based metaheuristics
can be seen as a counterpart to the framework 2.2. Compatibility Relation Between Route-Start
of Desaulniers et al. (1998). Both frameworks and Route-End Nodes
follow the idea that resource-constrained paths The aim of route-start and route-end nodes is to intro-
capture which routes or schedules are feasible. While duce vehicle and depot characteristics into the prob-
lem. First and foremost, these nodes represent spatial
the unified framework of Desaulniers et al. (1998) is
points where vehicles start and end their trips. To
intended to be used with an exact column-generation
ensure that route-start and route-end nodes are com-
or Lagrangean-relaxation method, the framework
patible, we define a relation ∼ on O × D. Again,
presented here focuses on heuristic procedures
the compatibility of pairs
o d of route-start and
based on enumerative LS algorithms. Moreover,
route-end nodes depends on the problem at hand: For
in Desaulniers et al. (1998), only the feasibility of
single-depot problems with a homogeneous fleet, all
individual routes and schedules is encoded in the
o ∈ O and d ∈ D are compatible because all nodes rep-
definition of resource-feasible paths. Constraints that
resent the same physical location independent of the
couple together different routes form the constraints
vehicle. In multidepot problems, the sets O and D
of the master program (see Lübbecke and Desrosiers are partitioned according to the nD depots or garages,
2005). Here, the feasibility of individual routes as well e.g., O = O 1 ∪· · ·∪O nD , D = D1 ∪· · ·∪DnD . Pairs o ∈ O k ,
as several types of intertour constraints is defined by d ∈ Dl are compatible if and only if k = l. Sets O k × Dk ,
resource-constrained paths. The building blocks of consisting of a single pair, can be used to model VRPs
the representation are the routing graph, the giant-tour with individual vehicles departing from and going
representation, a compatibility relation between route- to different locations. In general, we assume that O
start and route-end nodes, and the consideration of and D have the same cardinality,
O
=
D
. The easi-
the entire giant route as a single resource-feasible path. est way to implicitly encode the compatibility relation
The following sections explain these building blocks into the routing graph is to define an arc
o d ∈ A if
in detail. and only if o ∼ d holds.

2.1. Routing Graph 2.3. Giant Route and Giant Tour

To describe neighborhoods and solution procedures A solution to a VRP is called a route plan. A route
formally, a concise representation of VRP solutions, plan can be written as x =
p1  p2   pH  with an
i.e., route plans, is needed. This representation has to H -tuple of disjoint routes in G. Note that this defini-
be flexible to model a wide range of rich VRPs and tion implies that every route-start and route-end node
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 273

o1 d1 o2 d2 o3 d3 o4 d4
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

Figure 1 Giant-Tour Representation

occurs in exactly one route. We will denote the (max- !ai  bi $ at nodes i ∈ V for the time resource. Let R be
imum) number of nodes in a route plan by n =
V
. the number of resources (such as time, load, cost, etc.).
The giant route is the path
p1 , p2 ,  pH  in which A vector T =
T 1   T R  ∈ R is called a resource
i
each route-end node d is connected to the next route- vector and its components resource variables. For two
start node oi+1 (for i = 1 2  H − 1). Similarly, the resource vectors a and b, the interval !a b$ is defined
giant tour is the cycle in which, additionally, d H is as the set T ∈ R & a ≤ T ≤ b (componentwise).
connected to o1 . In the following, P
p1  p2   pH  Resource intervals, also called resource windows, are
denotes the giant route and C
p1  p2   pH  the giant associated with nodes i ∈ V and are denoted by !ai  bi $
tour. The giant-tour representation of a route plan is a with ai , bi ∈ R , ai ≤ bi . (In the following, ari refers
generalization of the MTSP representation of the VRP to a resource vector of node i and its component
(Christofides and Eilon 1969) to more general VRPs. for the resource r.) The changes in the resource con-
It has the advantage of allowing single and multiple sumptions associated with an arc
i j ∈ A are given
route problems to be handled in a very similar way. by a vector fij =
fijr Rr=1 of so-called resource exten-
Figure 1 depicts such a representation for the case of sion functions (REFs). An REF for resource r, i.e.,
four routes, departing from two depots. fijr & R → , depends on a resource vector Ti ∈ R .
The vector Ti corresponds to the resource consump-
2.4. Resource-Constrained Paths tion accumulated along a path from a given start
Resource-constrained paths (RCPs) and associated node s to a node i, i.e., up to the tail node i of
shortest-path problems have been very successfully arc
i j. Hence, the result fij
Ti  ∈ R can be inter-
used in the context of column-generation methods, preted as a resource consumption accumulated along
not only applicable to VRP but also to vehicle- and the path
s  i j. Classical REFs are of the form
crew-scheduling problems (Desaulniers et al. 1998, fijr
Ti  = maxarj  Tir + tijr , where tijr are constants asso-
Irnich and Desaulniers 2005). The success of RCPs is ciated with the arc
i j and arj the lower bound of
based on the fact that the resource concept constitutes the resources r at node j. Classical REFs are sepa-
a very flexible tool for modeling complex cost struc- rable by resources; i.e., no interdependencies exists
tures for routes and schedules as well as a wide vari- between different resources. More general definitions
ety of rules that define their feasibility. In the context of REFs provide powerful instruments for model-
of VRPs, column generation and branch-and-price- ing practically relevant constraints over resources that
and-cut give rise to exact solution procedures that are are interdependent (see Irnich and Desaulniers 2005,
restricted to small and medium-sized instances of up Irnich 2008, and §4). A path P =
v0  v1   vp  is
to about 100 nodes (see, e.g., Fukasawa et al. 2006, resource-feasible if resource vectors Ti ∈ !avi  bvi $ exist
Desaulniers et al. 2006, and Jepsen et al. 2006). Here, for all positions i = 0 1  p such that fvi  vi+1
Ti  ≤
we propose to transfer the concept of RCPs from exact Ti+1 holds for all i = 0  p − 1. We denote by  the
to heuristic solution methods. The goal is to provide set of all resource-feasible paths.
LS components for metaheuristics, which are flexible Concluding, a route plan
p1  p2   pH  is feasi-
and at the same time powerful, so that they can be ble if and only if all of the following four condi-
applied to large-scale rich VRP instances to produce tions hold: (1) p1  p2   pH are node-disjoint routes,
high-quality solutions. (2) C
p1  p2   pH  is a Hamiltonian cycle in the
RCPs are defined over the routing graph G = routing graph G (or task-feasible cycle; see the Online

V  A. For the sake of convenience, we assume Supplement), (3) all route-start and route-end nodes
that G is simple, so that a path can be written of routes pi =
oi   d i  are compatible, i.e., oi ∼ d i
as P =
v0  v1   vp  with the understanding that for all i ∈ 1  H , and (4) P
p1  p2   pH  is a

v−1  v  ∈ A holds for all  ∈ 1  p. Resource resource-feasible path. The novelty in this definition
constraints can be formulated by means of
minimal is that the entire giant route P
p1  p2   pH  is con-
resource consumptions and resource intervals, e.g., the sidered as one RCP. This implies that particular REFs
travel times tij along arcs
i j ∈ A and time windows are needed to connect consecutive routes in the giant
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
274 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

tour. Whenever a route-end node d k ∈ D is connected x ∈ 
xt  and their costs c
x  are evaluated one by
to a route-start ok+1 ∈ O, all intratour resources r one), taking the first improving solution or taking a best
have to be reset. This fits in nicely with the defi- improving solution are two extreme strategies known
nition of classical REFs because a reset function is as first improvement and best improvement. Another
given by the REF fdrk  ok+1
T  = maxarok+1  T r − M (with well-known strategy, d-best improvement, terminates
an appropriate large number M). Note that intertour when d improving neighbors have been found and
resources r (such as cost) should not be reset but kept, returns a best of them. From the worst-case point
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

i.e., fdrk  ok+1
T  = max−M T r  = T r . of view, all search strategies are equivalent because
showing that xt is a local optimal solution requires
3. Efficient Local Search the entire neighborhood 
xt  to be scanned. How-
Local search is the most frequently used heuristic tech- ever, from an average case point of view, these strate-
nique for solving combinatorial optimization prob- gies might significantly differ in their efficiency (we
lems. It provides the basis for modern metaheuristics, expect best improvement to perform less iterations
such as tabu search (TS), GRASP, and variable neigh- with larger gains that take longer compared to first
borhood search (VNS) (see Hoos and Stützle 2005). improvement).
Most of the effort spent within an enumerative LS Typically, neighborhoods and neighbor solutions
algorithm is used for scanning the neighborhood (for are neither constructed by the function  & X → 2X nor
a classification of LS algorithms, see Funke et al. given by subsets 
x ⊂ X. Instead, they are defined
2005a). It is, therefore, desirable to use efficient algo- implicitly by a set of moves M. A move m ∈ M trans-
rithms within LS to speed up the procedure that forms a solution into a neighbor solution. Some of the
performs this scan. In this section, we first clarify moves m ∈ M might transform a feasible solution x
the relationship between neighborhoods, moves, the into an object m
x, which has a structure similar to
order in which the search tree is explored, and—in a feasible solution but does not necessarily satisfy all
detail—algorithms that compute costs and test the constraints that define feasible solutions. In the fol-
feasibility of neighbor solutions. lowing, we will refer to such an object as a solution.
Examples in the case of VRPs are the removal of a
3.1. Local Search, Neighborhoods, and Moves customer node and its insertion into another position
An instance
X c of a combinatorial optimization prob- or the swapping of two customers between two tours.
lem can be stated as minx∈X c
x, where X is the These moves might violate a constraint. Let Z ⊇ X be
set of feasible solutions and c is the cost function. the set of all solutions. In general, we denote by M the
The heart of an LS procedure is the definition of set of moves, where a move m ∈ M maps from Z to Z,
a neighborhood  , which is a mapping  & X → 2X . i.e., m& Z → Z. For a given x ∈ Z, the extended neighbor-
Each element x ∈ 
x is called neighbor of x. Neigh- hood  contains all neighbors of x, either feasible or
bors x with cost c
x  < c
x are improving neighbors. infeasible, i.e., 
x ⊇ 
x. Every move m ∈ M with
LS starts with an initial feasible solution x0 ∈ X. In m
x ∈ X is called a feasible move w.r.t. x.
each iteration t, it replaces the current solution xt by
an improving neighbor xt+1 ∈ 
xt . The LS proce-
3.2. Major Tasks in a Local-Search Procedure for
dure terminates with a local optimum, i.e., a solution xt
Vehicle Routing
without improving neighbor in 
xt . For further
The focus of this paper is on the efficient implemen-
details of LS, see the books by Aarts and Lenstra
tation of Step 4 of Algorithm 1. The major tasks that
(1997) and Hoos and Stützle (2005).
have to be performed are the implicit or explicit con-
Algorithm 1. Generic Local Search struction of neighbor solutions x ∈ 
x, for each of
1. Input: A feasible solution x0 ∈ X. them the computation of the cost c
x  or gain g
x  =
2. LET t = 0. c
x − c
x  compared to the current solution x, and
3. REPEAT the test of whether the newly constructed neighbor
4. SEARCH for an improving neighbor x in the is feasible or not (separating candidates x ∈ X from
neighborhood 
xt  of the current solution xt . those in Z\X).
5. IF there exists an improving neighbor The problem of checking the feasibility of a neigh-
solution x ∈ 
xt  THEN bor solution is best explained by an example. A swap
6. LET xt+1 = x and t = t + 1. move chooses two nodes wi and wj of the giant route
7. UNTIL no more improvements can be found. and exchanges them. Hence, the four arcs
wi−1  wi ,
8. Output: A local optimum xt .
wi  wi+1 ,
wj−1 , wj ,
wj  wj+1  are deleted and the
Note that there are several options for choosing four arcs
wi−1  wj ,
wj  wi+1 ,
wj−1  wi ,
wi  wj+1  are
improving neighbor solutions in Step 4. If the added to the current solution x; see Figure 2. The
search method is enumerative (i.e., neighbor solutions extended swap neighborhood 
x of giant route x
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 275

wi – 1 wi wi +1
allow constant time-cost computations provided that
d 1 o2 d 2 o3
REFs are generalizable to segments. Sequential search
o1 = w1
techniques, however, are not directly applicable then
d 5 = wn
(cf. §3.4).
o5 d 4 o4 d 3
wj + 1 wj wj – 1 Feasibility-based arguments try to identify branches
of the search tree that do not contain any feasible
Figure 2 Principle of a Swap Move, Giant Tour with Five Routes solution at all. Both types of arguments need tailored
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

search strategies in the sense that the sequence in

consists of all other giant routes that can be gener- which decisions are taken guarantees that all remain-
ated by choosing different nodes wi and wj , so that ing solutions of the branch under consideration are
the swap neighborhood is of size

n2 . A newly either more costly or “less” feasible. It is, therefore,
constructed neighbor solution x ∈ 
x can be hardly possible to directly mix both approaches. Con-
rejected if it is nonimproving or infeasible. Moreover, sidering feasibility of a swap, note first that one or
we see that a neighbor solution x is uniquely deter- two routes are affected, depending on whether wi and
mined after making k = 2 independent decisions (the wj are in the same route or different routes. Testing
decisions about the two nodes/positions to swap). a constructed route in a straightforward way means
In general, all enumerative search procedures for looping over the nodes of the route to compute mini-


nk  neighborhoods work on a search tree with mum resource consumptions which are then checked
(at least) k levels. They differ in two aspects: against upper bounds. This is at least possible if all
(1) In the order, in which objects, i.e., nodes or arcs REFs fij are nondecreasing (see Irnich 2008). The loop
defining the move, are determined (nodes wi and over the nodes of a single route causes an effort of
wj for the swap move). Enumeration rules for nodes

n if the length of a tour is not limited by a fixed
can consider nodes by increasing (decreasing) index, number, independent of n. Even if there is a maxi-
position in the giant tour, or ordered by an auxil- mum length of a tour, the presence of intertour con-
iary attribute (e.g., lower or upper limit of an associ- straints can require that resource consumptions have
ated resource interval). Alternative enumeration rules to be propagated along the entire giant route.
choose nodes one after another—sequentially—such
that distance, cost, or any other resource consumption 3.3. Feasibility Checks and Cost Computations in
of an associated arc is increasing. Lexicographic and Constant Time
sequential search approaches described in Irnich et al. If cost is one of the resources (this is no restriction,
(2006) differ exactly with respect to these orderings. but the standard case in Irnich and Desaulniers 2005),
(2) Different orderings allow tailored criteria for prun- feasibility checking and cost computation can be seen
ing the search tree. If some branch of the search tree as identical algorithmic procedures. Computing the
does not contain any feasible or improving neighbor cost of a giant tour C
p1   pH  is equivalent with
solution, it can be pruned. One does not have to build finding a least-cost resource vector at the destination
and evaluate the corresponding solutions x , but can node of P
p1   pH . Improving solutions w.r.t. x are
take a shortcut. This is the key idea for accelerating exactly those giant routes that respect an upper bound
enumerative search approaches for

nk -sized neigh- c
x − / for the cost resource (with / > 0 small). In
borhoods to be searched in less than nk operations. the following (if not stated otherwise), we speak of
In contrast to heuristic techniques like those used in “constant time feasibility tests” for both cost com-
Toth and Vigo (2003), we can be sure to find a best putations and for checking the remaining resource
neighbor solution. variables.
Note that gain-based criteria try to show that there Before we introduce our new approach, an al-
is no improving (or less strictly, no acceptable) solu- ternative method proposed by Kindervater and
tion relative to the quality of the current solution x Savelsbergh (1997) is explained along with its capabil-
and, possibly, relative to another improving neigh- ities and limitations.
bor solution x ∈ 
x already computed. Comput- 3.3.1. Global Variables Approach of Kindervater
ing the gain of a move is trivial as long as it can and Savelsbergh. According to Kindervater and
be expressed as the difference of the costs of all arcs Savelsbergh (1997, p. 348), “the basic idea is to use a
changed replacing x by x , e.g., g
x x  = cwi−1  wi + specific search strategy in combination with a set of
cwi  wi+1 + cwj−1 wj + cwj wj+1 − cwi−1 wj − cwj wi+1 − cwj−1  wi − global variables such that testing the feasibility of a
cwi wj+1 for the swapping of wi and wj . However, this is single exchange and maintaining the set [of] global
not the case if cost depends on other resources, such variables requires no more than constant time.” The
as load-dependent transport tariffs, wages for drivers specific search strategy they use is lexicographic search.
depending on the time on duty, etc. Then, the prepro- Note that the traditional node- and edge-exchange
cessing and search techniques presented in §3.3 still procedures split a given tour (or two or several
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
276 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

affected tours) into paths (from now on called seg- 3.3.2. Segment REFs. This section explains how
ments). These segments are permuted, some may be REFs can be inverted and generalized to segments,
inverted, and finally concatenated together again to so that extensions of the ideas of Kindervater and
form a new tour. Lexicographic search is character- Savelsbergh (1997) can be used (1) for more gen-
ized by the fact that, in the innermost loop of the search eral VRPs defined by nonstandard REFs, (2) in the
algorithm, from one iteration to the next, an inner seg- context of giant tours, i.e., when segments can also
ment grows by exactly one node. In this way, global contain nodes from more than just a single tour, and
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

variables for a segment
wi  wi+1   wj−1  wj  are (3) within different search strategies allowing more
computed by either concatenating
wi  wi+1   wj−1  flexibility than the lexicographic search approach.
with
wj−1  wj , or
wi  wi+1  with
wi+1   wj−1  wj . The key idea is to separate the search strategy from
Contrary, in an initialization phase and in outer loops the computation of global variables (or any similar
of the search algorithm, global variables for start- information, e.g., given by segment REFs). Note that
ing and ending segments, i.e.,
w1  w2   wi−1  and all the classical moves can be considered as k-edge

wj+1   wn , are computed and stored. Together, exchanges, even if their intention is to exchange
these global variables allow constant time feasibility nodes. The swap move, for example, is a special-
checks. For example, time window constraints require ized 4-opt move (except for the case where wi and
the computation of a total travel time, earliest depar- wj are adjacent yielding a 2-opt move; cf. legitimacy
ture time, and a latest arrival time. This is based on conditions, explained in Glover 1996 and Irnich et al.
certain forward and backward computations along 2006). Therefore, moves decompose the giant route
segments. Kindervater and Savelsbergh (1997) clar- into a small fixed number of segments. The swap
ify these procedures for 2-opt and Or-opt moves in move depicted in Figure 2 implies the segmenta-
connection with time windows and precedence con- tion P1 =
o1   wi−1 , P2 =
wi , P3 =
wi+1   wj−1 ,
straints as well as for problems with simultaneous P4 =
wj , and P5 =
wj+1   d 5 . The paths P1   P5
deliveries and pickups. depend on the giant tour
w1  w2   wn  w1  cur-
Their approach is intrinsically tied to the lexico- rently under consideration (the incumbent giant tour)
graphic order in which moves are considered because and the choice of the nodes wi and wj (or equiva-
a constant time update of the global variables from lently, their positions i and j). These five segments
one iteration to the next requires that only a fixed are permuted and constitute the new giant route P =
number of nodes (typically one node) is added to a P
P1  P4  P3  P2  P5  (cf. notation introduced in §2.3).
segment. In the case of a swap move, an outer loop The move is feasible if and only if P is resource fea-
considers nodes wi (at position i in the giant tour) sible and C
P  is a task-feasible cycle in which route-
in any order, e.g., in the order in which they appear start and route-end nodes are compatible. Testing the
in the tour. Contrary, the inner loop must choose the last two conditions is straightforward and possible
second customer nodes wj , one by one, at positions in

1. The following analysis, therefore, focuses on
i +2 i +3  n−1. The constant time computation of resource feasibility.
global variables is possible for the segments P2 =
wi , Our goal is now to determine attributes for each
P3 =
wi+1   wj−1  wj−1 , and P4 =
wj  because these of the possible segments such that one can decide
global variables are either computed from scratch (for in

R time whether the concatenation of two seg-
segments of length one) or from global variables of ments also forms a feasible or infeasible segment. Fur-
the previous segment P3 =
wi+1   wj−2 . The ini- thermore, we want to compute the attributes of the
tial phase has to provide global variables for all seg- concatenated segment in

R, so that, in summary,
ments P1 =
w1   wi−1  for i = 1 2  n − 3 and testing the feasibility of P can also be performed in
P5 =
wj+1   wn  for j = 3  n − 1. constant time

R. Irnich (2008) provides the theo-
Kindervater and Savelsbergh (1997, p. 350) point retical background for accomplishing this task. The
out that their global variables approach, combined attributes which have to be computed are the defining
with lexicographic search, can be used for multi- coefficients of the segment REFs as well as the inverse
ple constraints and all k-edge exchange neighbor- segment REFs for some of the segments underlying the
hoods. However, a unifying theory explaining which incumbent giant tour. For the sake of clarity, we start
types of constraints can and which cannot be dealt by pointing out the basic assumptions to hold for the
with is missing. For example, resource constraints rest of the paper:
with resources that depend on each other (such as Assumption (1) All REFs have a finite representa-
load-dependent travel times, etc.) are not considered. tion and allow function evaluations in

R time. This
On the other hand, resource extension functions, as is true for several types of nondecreasing REFs pre-
introduced by Desaulniers et al. (1998), provide a sented in §4.
well-defined, flexible, and generic formalism for the Assumption (2) All inverse REFs exist. The inverse
description of side constraints relevant for rich VRPs. of a nondecreasing REF fij & R → !aj   is a function
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 277

fijinv & R →
− bi $. It has to be nondecreasing and its of Pi+1 for all i ∈ 1  q − 1. Their concatenation P1 +
defining property is fij
T  ≤ T ⇐⇒ T ≤ fijinv
T  for P2 +· · ·+Pq is resource feasible if and only if all inequalities
all T ∈
− bi $ and all T ∈ !ai  .
fP1
aw0  ≤ fPinv
bw2 
Assumption (3) All inverse REFs have a finite 2

representation and allow function evaluations in fP2  fP1
aw0  ≤ fPinv
bw3 


R time.
3
(1)
Assumption (4) All REFs and inverse REFs can be
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

generalized to segments. Segment REFs also allow fPq−1  · · ·  fP2  fP1
aw0  ≤ fPinv
q

bwq 
function evaluations in

R time. hold. (Note that f  g
x is defined as f
g
x.)
Assumption (5) The concatenation of any two seg-
ments has a REF that can be computed in

R time A direct consequence of Proposition 1 is that the
problem of efficiently testing the feasibility and com-
from the REFs of the two segments.
puting gains is—at least partially—solved. A pre-
Obviously, if the number R of resources is fixed,
requisite is, however, that segment REFs must be
i.e., independent of the size n of the giant tour, all
available.
the above-mentioned operations can be performed in
constant time

1. Theorem 1. Let x be a feasible giant tour and let all
We refrain from giving a formal presentation of all segment REFs as well as inverse segment REFs w.r.t. x
be already computed for all possible segments. Then, any
the details concerning REFs and required properties,
neighbor solution x = m
x of an -edge exchange move
derivations, and proofs concerning finite representa-
m can be tested for feasibility in

R time.
tion, inversion, generalization to segments, function
evaluation, and concatenation. These details can be Because for all node- and edge-exchange neighbor-
found in Irnich (2008). However, some remarks for hoods that are explored with tree-search methods (cf.
explaining and interpreting the newly introduced Funke et al. 2005a) the number  of segments is con-
stant (and small), Theorem 1 implies

R time feasi-
segment REFs and inverse REFs seem appropri-
bility checks.
ate: We consider an arbitrary path P . The segment
REF fP & R → R gives for each initial minimum 3.3.3. Preprocessing. What remains to be done
resource consumption T at the start node the min- is to find efficient procedures to provide REFs and
imum resource consumption at the final node of P . inverse REFs for all or (at least) a suitable subset
of segments. Computing segment REFs and upper
Note first, that this is exactly the idea of arc REFs; i.e.,
bounds for a given giant tour can be undertaken
for P =
i j, the value fij
T  is the minimum resource
with a straightforward procedure requiring

Rn2 
consumption at j given the resource consumption T
time and space. The reason is that there are 2n2 seg-
at node i. Note further, that the term “the mini- ments and inverted segments spanned between the
mum resource consumption” is only well defined if n2 pairs of nodes (note that moves might invert some
the REFs are nondecreasing. While ordinary REFs for of the segments, so that inverted segments also have
arcs and segments propagate minimum resource con- to be considered). Segment REFs fP for a segment
sumptions forwards, inverse REFs propagate upper P =
vi   vj−1  vj  are generated from the segment
bounds for resource consumptions backwards. The REF fP of the segment P =
vi   vj−1  and the REF
inverse REF fijinv & R → R takes any upper bound T fvj−1 vj . Similarly, for the segment Q =
vi  vi+1   vj ,
for the resource consumption at node j and computes j > i, the inverse segment REF fQinv is computed
the value fijinv
T , which is an upper bound for the from the inverse REF fvinv i vi+1
and the inverse seg-
resource consumption on node i. Similarly, for the ment REF fQinv of Q =
vi+1   vj . With the gen-
inverse segment REF fPinv & R → R , the resource vec- eral Assumptions (1)–(5) on REF operations, each step
tor fPinv
T  is the upper bound for the resource con- requires

R time leading to the desired result.
sumption at the start node of P under the condition Proposition 2. Segment REFs and inverse segment
that one propagates resources along P and that T is REFs for all 2n2 segments and inverted segments of a giant
an upper bound for the resource consumption at the tour of length n can be computed by a straightforward pro-
final node. The importance of segment REFs and their cedure in

Rn2  time and space.
inverses is due to the following result: From a worst-case point of view, a quadratic pre-
processing is satisfactory if neighborhoods  of size
Proposition 1 (Irnich 2008, Theorem 3). Given

nk  with k ≥ 2 are inspected. However, we would
resource-feasible paths P1  P2   Pq ∈  , where the ith like to accelerate the average case and correspond-
path Pi starts with a node wi−1 and ends with a node wi , ing search strategies that scan less than

n2  neigh-
such that the end node of Pi coincides with the start node bors. Moreover, it has been shown by Funke (2003)
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
278 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

3.3.4. Generic Search Procedure. The following

pseudocode formalizes a generic search procedure for
searching an

nk  neighborhood 
x of a current
feasible solution x to determine a best neighbor solu-
tion x with a gain g
x x  > Gmin . The parameter
Section of nβ nodes Seed point
Gmin is chosen as Gmin = 0 for classical LS, Gmin = 
if any nonimproving neighbor could be accepted,
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

Figure 3 1-Level Hierarchy and Gmin > 0 for more selective procedures that con-
sider only substantial improvements. Independent of
that restricting the length of some segments can lead Gmin , the procedure guarantees a worst-case running
to interesting neighborhoods that can be searched time of

nmaxk h  and needs

nh  space, with h ∈
quickly. For example, restricting the length of inverted  43  78  depending on the type of hierarchy (1-level,
segments to a fixed value K for 2-opt moves yields 2-level) used.
a

Kn-sized neighborhood. Using first-improvement Algorithm 2. Generic Local Search (=Step 4 of
pivoting strategies in LS also requires accelerated Algorithm 1)
methods for the preprocessing phase. Our aim is, 1. Input: A feasible solution (giant route)
therefore, to reduce the number of segments that have x =
w0   wn  ∈ X; Gmin ∈  minimum gain
to be considered in feasibility testing procedures. (Phase 1—Preprocessing)
A solution to this problem is the definition of seed
2. LET  be the -level hierarchy of segment
points dividing the nodes of the giant tour uniformly
REFs fP , fPinv describing the current solution x
into sections. A 1-level hierarchy with parameter 4 ≤ 1
3. STORE positions iw , and positions ni , li of
uses equidistant sections of length n4 , so that n/n4 sec-
last/first route-start and route-end nodes
tions result; see Figure 3. The idea of a hierarchy of
(see remarks below)
REFs is that, instead of computing all 2n2 REFs for all
(Phase 2—Tree Search)
segments, only segments within a section (i.e., between
4. LET G∗ &= Gmin
two consecutive seed points) and between all pairs of
5. LOOP decision d1
seed points need to be considered. To compute a REF
6. LOOP decision d2
ranging from position i to position j, one has to con-
sider three cases: (1) If positions i and j fall into the 7.
same section, the REFs are already available. (2) If i 8. LOOP decision dk
and j are in two different but consecutive sections sur- 9. (Implicit construction of move and
rounding the unique seed point s, the REF between neighbor solution)
i and j can be computed as the concatenation of the 10. LET m &= md be the move implied by
REF from i to s and the REF from s to j. (3) Oth- decisions
d1  d2   dk 
erwise, there are at least two seed points between i 11. LET x =
P1  P2   Pq  the permutation
and j with s1 the first seed point following i, and s2 of the segments of
w0   wn  implied
the last seed point preceding j. The REF from i to j by m
can be computed as the concatenation of three REFs, (Feasible and Improving?)
i.e., from i to s1 , from s1 to s2 , and from s2 to j. In all 12.

LET  &=
P11   P1 1  P21   P2 2 ,


three cases, the segment REF from i to j is available q

 Pq1   Pq 
in, at the utmost,

3R =

R steps. The number of
segment REFs to compute is be the segmentation
13. implied by
P1  P2   Pq  and 


2 
n n 14. LET feasible &= Formula (1) is

2 4 n24 + 2 4 =

nmax1+4 2−24  fulfilled for 
n n
15. AND C
x  task-feasible
where factor 2 is for forward or inverted segments, 16. AND P
x  feasible w.r.t. route-start
the first term is the computation of all REFs inside and route-end nodes
sections, and the second term is for the REFs between 17. LET G &= g
x x  &= c
x
seed points. The effort is minimal for 4∗ = 1/3 result- −
fP q  · · ·  fPq1  · · ·  fP 2  · · ·  fP21
ing in

n4/3  computations. q 2
fP 1  · · ·  fP11
aw0 cost
Proposition 3. Segment REFs and inverse segment 18.
1
IF (feasible and G > G∗ ) THEN
REFs for a 1-level hierarchy of seed points for a giant tour 19. (Update of best neighbor solution
of length n can be computed in

Rn4/3  time and space.
found)
Generalizations to hierarchies with two and more LET G∗ &= G
levels can be found in the Online Supplement. 20. LET d ∗ &=
d1   dk 
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 279

21. Output: Gain G∗ and for G∗ > Gmin optimal last position j is smaller than its next route-start posi-
decisions d ∗ and best neighbor x = md∗
x tion ni for the start position i), replace P by its suc-
cessor segment in . Repeat, until P contains some
Remarks. (1) The preprocessing phase has to build
route-start node and let dP be the first route-start
REFs fP and inverse REFs fPinv for some segments P
node in P . Now that one knows route-start node oP
and some inverted segments P (if the neighborhood
and dP are linked (by request nodes or directly), one
also inverts one or several segments). Based on the
can check their compatibility. To iterate, replace P by
results of §3.3.3, the worst-case effort for the feasibil-
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

P and choose P as the successor segment.

ity test in Step 14 is

Rn2 ,

Rn4/3 , and

Rn8/7 
(for no hierarchy, a 1-level, or a 2-level hierarchy, 3.4. Sequential Search
respectively). Note that this first preprocessing phase Sequential search is a technique that allows neighbor-
is identical for any type of neighborhood, but the sec- hoods within local-search algorithms to be scanned
ond phase, the actual (tree) search, must be tailored in a highly efficient way. It was discovered indepen-
to the neighborhood. dently in the 1970s by Christofides and Eilon (1972)
(2) With an

n preprocessing (Step 3), we store and Lin and Kernighan (1973) in algorithms for the
for each position i ∈ 1  n of the giant tour the traveling salesman problem (TSP) and the graph-
node wi and, conversely, for each node w of the giant partitioning problem (Kernighan and Lin 1970).
tour its position iw . Apparently, the idea has since been forgotten and has
(3) The loops in Steps 5–8 exactly determine the not been tested for any type of constrained prob-
order in which moves and neighbor solutions are con- lem. Irnich et al. (2006) have introduced sequential
structed (cf. §3.2). search as a general method for accelerating LS proce-
(4) Some parts of the construction of the neighbor dures. It is based on the idea of decomposing moves
solution of Steps 10 and 11 might already be per- into so-called partial moves, so that partial moves are
formed in some of the outer loops in Steps 5–7. This cost independent and imply partial gains whose sum
can be useful for seeing that the resulting moves pro- is the overall gain of the move. Lin and Kernighan
duce infeasible or nonimproving neighbor solutions. (1973) proved that if the sum of a sequence of num-
(5) The tasks exchanged are determined by the bers (gains) is positive, then there exists a cyclic per-
nodes and edges that are removed and added. Hence, mutation of these numbers such that every partial
Step 14 can be performed in

1 if appropriate data sum is positive. This can be generalized to restrict
structures are used. a k-decision search procedure to consider only those
(6) In Step 12, the segmentation  results from x branches where the sum of the gains of the first p ≤
and the hierarchy  . For example, let n = 1000 and k partial moves has to be greater than pG∗ /k, where
 be the 1-level hierarchy introduced in §3.3.3. Then, G∗ is a lower bound for the overall gain. Details and
the n2/3 + 1 = 101 seed points are located at positions pseudocode of the application to several node- and
0 10 20 30  1000. Let m be the swap move that edge-exchange neighborhoods for CVRP can be found
exchanges the nodes at the positions 17 and 322. Then, in Irnich et al. (2006).
k = 5 and x =
P1  P2  P3  P4  P5  with P1 the segment Note that sequential search is directly applicable
from position 0 to 16, P2 the segment consisting of the only to those VRPs where the REFs are separable w.r.t.
node located at position 322, P3 the segment from 18 the cost resource, i.e., where the cost is given by the
to 321, P4 the single-node segment at position 17, and sum of the costs of all arcs in the giant tour. For more
P5 the segment from position 323 to 1000. Now, the complicated cost functions that are not separable (see
hierarchy  implies a split of P1 into P11 from position §3.2), the gain criterion might remain applicable if
0 to 10, and P12 from 10 to 16. P3 is split into three upper bounds for the resulting gain can be deduced
segments P31  P32  P33 from positions 18 to 20, 20 to 320, from removed arcs and lower bounds of the resulting
and 320 to 321, respectively. Finally, P5 is split into P51 loss can be determined for added arcs. As far as we
from position 323 to 330 and P52 from 330 to 1,000, know, these ideas have not been tested thus far.
while P2 and P4 are not split. Hence,  consists of We present the main idea of sequential search for
2 + 1 + 3 + 1 + 2 = 9 ≤ 3 · 5 =

k segments. the case of a swap move, depicted in Figure 2. We
(7) To check feasibility w.r.t. route-start and decompose the swap move into two parts: The first
route-end nodes in Step 16, one has to a priori record, part is the removal of the arcs
wi−1  wi ,
wi  wi+1 
for each position i of the giant tour, the next posi- and the addition of
wj−1  wi ,
wi  wj+1 . The second
tion ni of a route-start node and the last position li of part consists of removing
wj−1  wj ,
wj  wj+1  and
a route-end node. Along , consider pairs
P  P  of adding
wi−1  wj ,
wj  wi+1 . For the entire move to
(consecutive) segments in . Let the first segment P be improving, either the first or the second part has
contain the route-start node oP as the last route-start to be improving (i.e., the cost of the added arcs has
node. If P does not contain a route-end node (i.e., its to be smaller than the cost of the deleted arcs). In the
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
280 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

first case, starting the search at node wi , the cost of 21. (Inner Loop, Case 2: Arc
wj−1  wi  ∈ A
the removed arcs is given by B &= cwi−1  wi + cwi  wi+1 . must be short)
It follows that either cwj−1  wi < B/2 or cwj−1  wi < B/2 22. LOOP w ∈ N −
wi  AS LONG AS cwwi < B
must hold. By scanning the in-arcs
w wi  ∈ A and 23. LET j &= i
w + 1
out-arcs
wi  w ∈ A of node wi by increasing length,
the search can be terminated whenever an arc longer 24.
25. /∗ Steps 9–20 ∗ /
than B/2 is found. Because of the symmetry, identi-
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

cal arguments cover the second case for starting the 26.
search with node wj . 27. Output: Gain G∗ and for G∗ > Gmin optimal
swap
A prerequisite of this bounding procedure is that all decisions
i∗  j ∗  and best neighbor x = mi∗  j ∗
x
in-arcs and out-arcs of a given node wi are explored
The bound B computed in Step 5 limits the length
in an order where they are sorted by increasing cost.
of the out-arc
wi  w ∈ A, w ∈ N +
wi  in Step 7 or
Because in-arcs and out-arcs of wi are fully deter-
the in-arc
w wi  ∈ A, w ∈ N −
wi  in Step 22 for
mined by the other endpoint w of the arc, one can
any improving move. Because the neighbor lists are
retrieve the required information from so-called neigh-
sorted, the inner loop can be terminated whenever an
bor lists N +
wi  and N −
wi . N +
wi  is the list of head
arc not smaller than B comes up. Complete neighbor
nodes of out-arcs
wi  w of wi sorted by increasing
lists require

n2  space (for dense routing graphs)
cost. Analogous to this, N −
wi  is the sorted list of tail
which can be computationally prohibitive when VRP
nodes of in-arcs
w wi . instances with several thousands of nodes are con-
The sequential search algorithm for the swap neigh- sidered. Note that the neighbor list computation has
borhood can be formulated as follows. to be performed only once in an initial preprocess-
Algorithm 3. Sequential Search for Swap (Phase 2, ing. Its time complexity is

n2 log n, but anyway, this
Tree Search) time complexity is always dominated by the total run-
1. Input: A feasible solution (giant route) ning time of LS in practice. To reduce the required
x =
w1   wn  ∈ X; space, one can replace full neighbor lists by reduced
Gmin ∈  minimum gain neighbor lists, also called candidate lists (Glover 1996),
It is assumed that Phase 1 (=preprocessing) that contain only a subset of arcs (one hopes the rele-
is already performed vant ones!). A standard approach is to build candidate
2. LET G∗ &= Gmin lists NK+ , NK− that contain a fixed number K of request
3. (Outer Loop) nodes, but all route-start and route-end nodes (depot
4. LOOP i ∈ 1  n nodes) by default. Clearly, when using proper candi-
5. LET B &=
cwi−1  wi + cwi wi+1 /2 − G∗ /2 date lists, there is a trade-off between the accuracy
6. (Inner Loop, Case 1: Arc
wi  wj+1  ∈ A of the search and the computational burden. Irnich
must be short) et al. (2006) have compared this trade-off for the stan-
7. LOOP w ∈ N +
wi  AS LONG AS cwi w < B dard CVRP.
8. LET j &= i
w − 1 All infeasible arcs, i.e., arcs that cannot be part of
9. IF i > j THEN LET t &= i, i &= j, j &= t any feasible giant tour, can be omitted from the neigh-
bor lists. Using specialized probing techniques, as
10. (Implicit construction of move and
in Desrochers et al. (1992) and Ascheuer (1995), one
neighbor solution)
might substantially reduce the number of possible
11. LET P1 &=
w1   wi−1 , P2 &=
wj ,
arcs. The combination of both the static and a priori
P3 &=
wi+1   wj−1 ,
determination of relevant arcs and the dynamic prun-
P4 &=
wi , P5 &=
wj+1   wn  ing of the search tree based on partial gains, is—as
12. LET x &=
P1  P2  P3  P4  P5  far as we know—the first approach to effectively com-
13. LET G &= cwi−1  wi + cwi  wi+1 + cwj−1  wj bine feasibility-based and gain-based reductions. This
+ cwj  wj+1 − cwi−1  wj − cwj  wi+1 technique is not limited to the swap neighborhood
− cwj−1  wi − cwi  wj+1 but can be applied to all enumerate search procedures
14. LET feasible &= Formula (1) is fulfilled for for edge- and node-exchange VRP neighborhoods.

P1  P2  P3  P4  P5  A systematic explanation of move decomposition and,
15. AND C
x  task-feasible especially, of the gain criterion in sequential search
16. AND P
x  feasible w.r.t. procedures for different VRP neighborhoods can be
route-start and route-end nodes found in Funke et al. (2005a, b) and Irnich et al. (2006).
17. IF (G > G∗ and feasible and j = i + 1) THEN
18. (Update of best neighbor solution found) 4. Modeling Issues
19. LET G∗ &= G This section summarizes which types of VRPs can
20. LET
i∗  j ∗  &=
i j be handled with the unified framework. Before we
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 281

discuss particular types of constraints, we briefly 5.1. Preliminaries

repeat the basic assumptions: Before analyzing the proposed new techniques, we
(1) All feasible solutions of the given VRP can be have to explain and clarify the following aspects:
modeled as giant tours in the routing graph G = Which neighborhoods  are used? How are different

V  A. The length of a giant tour is bounded by n = neighborhoods combined to form a well-structured



V
. metaheuristic? How are sequential search and lexi-
(2) It must be possible to formulate the VRP as cographic search procedures compared; in particular,
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

a discrete task-partitioning or task-covering problem, how is the speedup measured? Finally, at least two
where tasks are associated with nodes and arcs or the significantly different implementation concepts exist
routing graph; see §2.1 and the Online Supplement. that constitute two extreme points w.r.t. the trade-
(3) The compatibility relation between route-start off between fast run time and economical use of
and route-end nodes must be given. memory (these are explained in §5.1.4∗ of the Online
(4) All intratour and intertour constraints have Supplement).
to be modeled as resource constraints on paths. The unified framework was coded in C++, and
The resulting REFs must fulfill Assumptions (1)–(5) different resource concepts and types of REFs were
of §3.3.2. integrated as template parameters. The algorithms were
(5) All moves m ∈ M of the neighborhood  under compiled in release mode (using MS-Visual C++
consideration decomposes a giant route into  seg- .NET 2003 version 7.1), and all runs were performed
ments. Any neighbor solution results from the per- on a standard PC (Intel × 86 family 15 model 2,
mutation, (partial) inversion, and concatenation of the 2.4 GHz, 1 GB main memory, on MS-Win 2000). Times
segments. were recorded using the time.h library. To be more
The complexity of the segment REF representation, precise, especially for times smaller than 10 ms, we
evaluation, and concatenation determines the effort performed multiple identical runs of the same proce-
for the preprocessing and the feasibility check in the dure. We made sure that running times of multiple
tree search. If all these operations can be performed runs exceeded 100 ms such that the average run time
in

R time, then any

nk  neighborhood can be is a rather accurate estimate for a single run.
fully explored in

Rnk  time and

Rn4/3  space. 5.1.1. Neighborhoods, Moves, and Sequential
These worst-case results are fully independent from Search Procedures. We have implemented lexico-
the search tree exploration strategy. If REF manipula- graphic and sequential search procedures for the
tions require more than

R time, additional factors neighborhoods listed in Table 1 and shown in Fig-
result in the above worst-case complexities (e.g., for ure 4. For a detailed description of the neighbor-
multiple time windows, see the Online Supplement). hoods and for pointers to the (original) literature, we
For the sequential search strategy, the only addi- refer the reader to the surveys (Funke et al. 2005a,
tional assumption needed is that the gain of a move is Bräysy and Gendreau 2005a), while the correspond-
directly associated with the exchanged arcs. Thus, for ing sequential search procedures with pseudocode are
any move m transforming x into x (i.e., x ∈ 
x), the explained in Irnich et al. (2006). To be self-contained,
gain G = g
x x  is given by the cost difference of the the Online Supplement briefly recalls basic properties
deleted and added arcs. In the case of more complex of these neighborhoods.
cost functions, the gain criterion remains applicable if
5.1.2. General Setup for Local Search. Our com-
bounds on gains can be computed (see §3.4).
parisons of sequential search and lexicographic search
For the sake of brevity, we had to shift all model-
procedures are always performed using the following
ing details into the Online Supplement. Section 4.1∗
setup that combines VND (Hansen and Mladenović
gives details on modeling combined collection and
2001, 2002) with LNS (Shaw 1998, Røpke and Pisinger
distribution, §4.2∗ on precedence constraints, §4.3∗ on
2006) strategies to escape from local optima. An ini-
limiting the number of vehicles, §4.4∗ on compat-
tial solution is computed by a problem-specific start
ibility constraints (e.g., site dependencies and haz-
heuristic. Starting from this solution, a local optimum
ardous material), §4.5∗ on interdependent resources,
w.r.t. all neighborhoods is computed. To seldom apply
§4.6∗ on heterogeneous fleets, §4.7∗ on periodic VRPs,
and §4.8∗ on intertour resources and constraints.
Table 1 Neighborhoods, Sizes, and Priorities in VND

Neighborhood Size

x
Priority

5. Computational Results
This section is intended to present empirical results Swap, 2-opt, (special) 2-opt∗ , node relocation

n 2  1
that show the effectiveness of the preprocessing and String exchange, Or-opt with and w/o inversion

n 2  2
acb-opt, request relocation (see the Online Supplement)

n 3  3
the sequential search procedures in practice.
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
282 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

(a) (b) A fair comparison of the running times by means

of a relative speedup factor is rather delicate to compute
for the following reasons. First, the preprocessing for
(d) the sequential search procedure has to be executed
(c)
only if the giant tour has changed; i.e., a preceding
search (of the same or another neighborhood) has
found an improving solution that has now become
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

(e) (f )
the incumbent solution. Hence, there is no intrinsic
connection between the current search procedure and
the preprocessing. Second, the ratio between success-
ful and unsuccessful searches strongly depends on
(g) the general setup, i.e., the start solutions, the mix of
neighborhoods, and the priorities for mixing them.
Third, the most frequently called procedures in the
search are the test of whether or not an arc exists and
Figure 4 Moves and Their Decomposition
the computation of the arc costs. Section 5.1.4∗ (see
Note. (a) Swap, (b) (Special) 2-opt∗ , (c) 2-opt, (d) Node Relocation, (e) Or-
opt, (f) Or-opt with String Inversion, (g) String-Exchange.
the Online Supplement) distinguishes between two
implementation principles for these procedures that
computationally costly operators, we have associ- also have an impact on the speedup factors.
ated priorities (see Table 1) to all neighborhoods. The most optimistic acceleration factor does not
More precisely, we alternate between the neighbor- consider the additional effort of the necessary prepro-
seq
hoods with priority 1 on the first search level. Here, cessing for sequential search at all. Let tlex and t
sequential and lexicographic search procedures are be the running times of lexicographic and sequen-
both applied to the same current solution x. Because tial search procedure (without time for preprocess-
seq
we are using a best-improvement pivoting strat- ing) for a neighborhood  . Then, fmax = tlex /t is the
egy, sequential and lexicographic search procedures maximum speedup or maximum acceleration factor. Note
return improving solutions with identical gain (due that running times might significantly vary depend-
to degeneracy, we cannot assure that identical solu- ing on the current giant tour x and whether a good
tions are computed; the improving solution found by bound B = B
G∗  (see Step 5 of Algorithm 3) is
sequential search is taken for the next search step). If available early in the sequential search procedure.
seq
none of the search procedures finds a move with posi- Therefore, only average values for t and tlex are con-
tive gain, the search is continued with neighborhoods sidered here. A very pessimistic factor is based on the
of priority 2. When improving solutions are found assumption that every sequential search is preceded
by a neighborhood of priority p > 1, then neighbor- by a preprocessing. Defining t pre as the (average) time
hoods of priority level 1 are tested again. This strat- of the preprocessing procedure (Steps 2 and 3 of Algo-
seq
egy is a minor modification of Hansen and Mladen- rithm 1), the factor fmin = tlex /
t pre + t  denotes the
ović’s VND metaheuristic (2001, 2002), which makes minimum speedup or minimum acceleration factor. This
the search more balanced for equally sized neighbor- factor applies to pure LS procedures in which only a
hoods with (empirically) identical search effort. single neighborhood N is incorporated such that the
For small-sized instances, VND with prioritized number of preprocessing and search steps coincide.
neighborhoods can result in only a few calls of search From our point of view, the most fair definition of the
procedures of priority 3. Therefore, three iterations speedup factor takes into account that only a fraction
of LNS with a random removal of 20 nodes (imple- of search steps is preceded by a preprocessing. Let
mented as suggested in §4.3∗ of the Online Supple- r pre ∈
0 1$ be the (instance and setup specific) ratio of
ment) and a simple cheapest-insertion procedure are the number of improvement steps performed to the
used to perturb the current solution such that one can overall number of search procedure calls. We define
seq
iteratively apply the above VND procedure. f = tlex /
r pre t pre + t  as the speedup or acceleration
min
factor. Note that f ≤ f < fmax holds, but that all
5.1.3. Relative Speedup of Sequential Search vs. values still depend on the initial solution, the choice
Lexicographical Search. The main part of the com- of neighborhoods, the VND/LNS strategy, as well as
putational study compares the running times of lex- several implementation issues.
icographic search and sequential search procedures
for the neighborhoods given above. Recall that both 5.2. Vehicle-Routing Problems with
approaches guarantee constant time feasibility checks. Time Windows
We will not compare them with a trivial implemen- The VRPTW is certainly the most studied variant
tation using straightforward feasibility tests because of VRPs and can be considered the prototype of
these techniques are obviously inferior. “rich” VRPs because time window constraints already
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 283

18

16

14

12

10
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

0
Swap 2-opt 2-opt* String-exch. Node reloc. Or-opt Inv Or-opt acb-opt
n = 200, short routes n = 400, short routes n = 600, short routes n = 800, short routes n = 1,000, short routes

Figure 5 Speedup of Sequential Search vs. Lexicographic Search for Homberger and Gehring (1999) VRPTW Instances with Short Routes

require sophisticated techniques for constant time fea- (average) number of nodes in a route. Initial solu-
sibility tests. Early work on VRPTW dates back to tions were computed using Solomon’s I1-heuristic
the 1960s, and since then, hundreds of scientific arti- (Solomon 1987), and REFs were a priori computed
cles have addressed modeling as well as methodolog- according to the first implementation concept pre-
ical aspects of developing exact and heuristic solution sented in the Online Supplement.
algorithms. For an overview, we refer to the surveys The main results for the Homberger and Gehring
(Cordeau et al. 2002; Kallehauge et al. 2005; Bräysy (1999) instances are depicted in the Figures 5 and 6
and Gendreau 2005a, b). for instances with short (C1, R1, RC1) and long (C2,
The Solomon (1987) and Homberger and Gehring R2, RC2) routes, respectively. Each column shows the
(1999) VRPTW instances have been used as bench- speedup factors for 30 VRPTW instances, reflecting
mark problems in numerous studies. While Solomon’s different problem characteristics (10 clustered, 10 ran-
(1987) instances have 100 customers, the Homberger domly distributed, and 10 mixed). The speedup fac-
and Gehring (1999) instances range from 200 to 1000 tor f is depicted as a bar, whereas fmin and fmax
customers. We therefore use the latter because we are are shown as error indicators. Both diagrams indicate
mainly interested in analyzing the behavior of the that there is always a speedup when a lexicographic
search procedures w.r.t. the number of tasks and the search approach is replaced by a sequential search
-->106.8
-->144.6
-->144.8
-->60.5
-->90.1

40

35

30

25

20

15

10

0
Swap 2-opt 2-opt* String-exch. Node reloc. Or-opt Inv Or-opt acb-opt
n = 200, long routes n = 400, long routes n = 600, long routes n = 800, long routes n = 1,000, long routes

Figure 6 Speedup of Sequential Search vs. Lexicographical Search for Homberger and Gehring (1999) VRPTW Instances with Long Routes
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
284 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

procedure. In the first group, capacities and time win- Very large-scale VRPTW instances are—as far as we
dows are chosen such that the average number of cus- know—not available. Hence, we created a small test
tomers in a route is about 10. Here, the acceleration set of 10 instances ranging from n = 1000 to 10000
factors vary from 2.4 to 4.5 for swap, from 5.9 to 12.2 customer nodes. The instances allow an average num-
for 2-opt, from 3.4 to 5.8 for 2-opt∗ , from 8.4 to 9.5 ber of about 45 customers per route. Results for these
for string exchange, from 5.0 to 8.3 for node reloca- instances are visualized in Figure 7. Note that we have
tion, from 3.9 to 5.2 for Or-opt, from 3.7 to 5.0 for considered only quadratic neighborhoods because the
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

Or-opt with segment inversion, and from 8.6 to 13.9 running times of the lexicographic search procedure
for acb-opt. For all neighborhoods, except for acb-opt, for the (cubic) acb-opt neighborhood were unaccept-
there is a positive correlation between the size of the ably long (more than five minutes for the largest
instance and the speedup. This contrasts with the instance and a single neighborhood search). We have
results for the CVRP in Irnich et al. (2006), where a used on-the-fly computations of REFs and candidate
clear negative correlation was only observed for the lists NK+
v and NK−
v. K is chosen such that each can-
string-exchange neighborhood. It remains unclear to didate list contains the 1000 closest customer nodes
us which characteristics of instances or properties of and all possible route-start and route-end nodes, i.e.,
neighborhoods imply such negative correlations. K ≤ 1000 +
O
.
The second group, depicted in Figure 6, contains For these large-scale instances, the speedup grows
instances with long routes, i.e., between 20 and even further. This is partly caused by the on-the-
40 customers per route. Similar to the results reported fly computation of costs and REFs, and also because
for the CVRP in Irnich et al. (2006), the speedup average route lengths increase. We also computed
grows when problems are less constrained. Here, the acceleration factors with the a priori computed REFs
acceleration factors vary from 3.5 to 9.7 for swap, for the instances with n ≤ 2000 and compared them
from 5.9 to 17.8 for 2-opt, from 4.5 to 12.3 for 2-opt∗ , with the results obtained for the on-the-fly implemen-
from 15.4 to 32.1 for string exchange, from 7.0 to 17.0 tation. The on-the-fly computation gives a contribu-
for node relocation, from 7.3 to 17.9 for Or-opt, from tion to the speedups by an average factor of between
6.9 to 17.3 for Or-opt with segment inversion, and 1.5 and 1.7 (but varying more strongly for different
from 60.5 to 144.8 for acb-opt. These are substantial move types).
speedups! The superiority of sequential search over
lexicographic search for less constrained instances 5.3. Multidepot VRPs with Time Windows
can be explained as follows. In more constrained The Solomon (1987) instances can be easily extended
problems (especially with tight time windows), opti- to multidepot VRPTW (MDVRPTW) instances. Be-
mal feasible routes can differ significantly from cost- cause all locations are given as pairs
x y in the
minimal TSP tours (and geometric intuition). Hence, Euclidean plane, these locations can be copied and
a larger fraction of moves seems improving (when shifted in space. Given that the locations of a VRPTW
looking at costs/gains only) but is in fact infeasible. instance are in a rectangle :x × :y , we shift these loca-
Consequently, gain-based arguments to terminate the tions by multiples of 0 9 · :x horizontally and by mul-
search apply less often. tiples of 0 9 · :y vertically. For example, a 20 = 5 × 4
-->157
-->173
-->162
-->191
-->185
-->206
-->191

150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
Swap 2-opt 2-opt* String-exch. Node reloc. Or-opt Inv Or-opt

n = 1,000 n = 2,000 n = 3,000 n = 4,000 n = 5,000 n = 6,000 n = 7,000 n = 8,000 n = 9,000 n = 10,000

Figure 7 Speedup of Sequential Search vs. Lexicographic Search for Large-Scale VRPTW Instances with n Between 1000 and 10000
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 285

-->1,358
-->283
-->357
-->427
-->463

-->390
-->619
200

180

160

140
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

120

100

80

60

40

20

0
Swap 2-opt 2-opt* String-exch. Node reloc. Or-opt Inv Or-opt acb-opt
n = 100 200 300 450 600 800 1,000 1,250 1,500 1,800 2,100 2,450

Figure 8 Speedup of Sequential Search vs. Lexicographic Search for MDVRPTW Instances

depot instance is created out of 20 copies, shifted neighborhood with factors of between fstr-exch = 15 4
by
0 9mx :x  0 9my :y  for
mx  my  ∈ 0 1 2 3 4 × and 575. This is, again, similar to the results for CVRP
0 1 2 3. Initial solutions for the separate VRPTW and VRPTW.
instances belonging to one depot are created with The results depicted in Figure 8 also indicate that
the VND approach of §5.2. Copies of these separate the on-the-fly implementation concept (for n ≥ 1250)
VRPTW solutions are then taken as initial solutions benefits more from the sequential search approach
for the MDVRPTWs. Because of the overlap created than the full representation (for n ≤ 1000) does. The
by the test generator (factor 0.9), the subsequent VND factor caused by the on-the-fly computation is approx-
and LNS procedures have the potential to create imately factor 2.0.
improving solutions.
Figure 8 depicts results for MDVRPTW instances 5.4. Additional Results
with n between 100 and 2450. All instances were cre- Simple versions of VRPs, such as the CVRP or the
ated from 12 selected 50 customer VRPTW instances distance-constrained VRP, have pure additive REFs
(c103, c109, r103, r112, rc101, rc106, c205, c208, r204, along the routes and globally fixed upper bounds (a
r208, rc202, and rc207). These instances were selected maximum load or travelled distance). Therefore, they
to yield a mix of clustered and unclustered instances, do not need the

n4/3  preprocessing as presented in
instances with tight and wide time windows, and §3.3.3. Instead, a linear time and space preprocessing
with short and long routes. We used both implemen- already allows constant time feasibility tests (Irnich
tation concepts, on-the-fly computation of REFs for et al. 2006). Very similar methods can be used for the
instances with n ≥ 1250, and full representation for “pure” multidepot VRPs with capacity and distance
smaller instances (see the Online Supplement). Can- constraints. Thus, these types of VRPs are beyond the
didate lists NK+
i and NK−
i were restricted to contain scope of this article and we refer to Irnich et al. (2006)
a maximum of 1,000 request nodes. for results on the CVRP.
The smallest speedup was found—as was to be Results for pickup-and-delivery problems and peri-
expected—for the smallest instances with n = 100 and odic VRPs can be found in the Online Supplement in
max
for the swap move. Here the factors fswap = 2 3, fswap = §§5.5∗ and 5.6∗ .
min
2 7, and fswap = 1 3 mean that there is still an accel-
eration. For medium-sized instances with n = 800, all
speedup factors are above 10.0. The largest speedups 6. Conclusions
were again observed for the cubic acb neighbor- The paper presents a new modeling framework and
hood with values facb = 29 for n = 100 and facb ≈ corresponding efficient LS methods for VRPs with
1350 for n = 800. Again, because of the high running classical and also nonstandard side constraints. One
times of the lexicographic search implementation, of the most important advantages of the framework is
we skipped the comparison for the acb neighbor- that it is generic and, therefore, allows various types
hood for n > 800. The conducted experiments also of VRPs to be handled in a similar and concise way.
gave remarkable speedups for the string-exchange The giant-tour representation is intuitive and enables
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
286 INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS

a unified view on moves, which can either be intra- apply directly. However, extensions of these sampling
tour moves or moves between different tours of the methods, such as the large-step Markov chain meta-
same or different depots, periods, vehicle types, etc. heuristic of Martin et al. (1992), use local optimal solu-
The unified framework also has advantages from a tions and present them to the acceptance algorithm.
software development point of view; once the search Hence, efficient LS procedures can also speed up these
procedures of the framework are implemented, addi- metaheuristics. Finally, metaheuristics, such as genetic
tional constraints can easily be integrated because algorithms, evolutionary strategies, or ant systems
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

feasibility is generically encoded by the routing do not use neighborhood-based search procedures, at
graph, start-route node and end-route node compat- least not in their “pure” versions. However, hybrid
ibilities, and—most important—resource-constrained versions use LS postprocessing procedures, which is
paths. Consequently, the framework separates the often the decisive device for designing a highly effec-
modeling (with instance-specific data and constraint tive metaheuristic.
formulation) from the actual search methods. The For the future, one challenge will be to model
addition or change of standard constraints becomes new real-world constraints or options and to integrate
“simply” a question of gathering input data and them by means of REFs that possess all the proper-
declaring constraints; it has no implications for the ties required for the unified framework. A first for-
search procedures. mal analysis of conditions for REFs to be invertible
Besides the powerful modeling capabilities of this and extendable to segments has been given in Irnich
framework, its main contribution is the incorporation (2008). Nevertheless, numerous real-world applica-
of highly efficient LS techniques. They allow constant tions, not only those sketched in §4, need to be
time feasibility tests as well as exact search-tree prun- examined in-depth. Tailored neighborhoods for spe-
ing based on sequential search (Irnich et al. 2006). cial routing applications not considered here (e.g.,
The extensive computational tests clearly show that arc routing, routing with choice of requests) need
sequential search procedures outperform the lexico- to be analyzed, and suitable search procedures have
graphic search methods. On large-scale instances and to be implemented. A better understanding of the
for nearly all types of neighborhoods, the speedup interplay between different start heuristics, neigh-
factors are between 10 and 1,000. We observed that borhoods, improvement, and diversification phases
the potential of large speedups grows with the size of of metaheuristics, considered in various scientific
neighborhoods. Hence, sequential search procedures and real-world applications, will certainly offer an
might become the only efficient technique for implic- interesting field for more theoretical and empirical
itly scanning even larger neighborhoods than those research. Finally, we hope that the unified framework
traditionally applied to VRPs thus far. We expect that will help researchers and practitioners get a more uni-
neighborhoods of size

nk  for k ≥ 3 will be used fied view on modeling and efficient search methods
more often.
for VRPs.
One key property of sequential search algorithms
for rich VRPs is the separation of the LS procedure
into two phases, namely, a preprocessing phase to References
compute

n4/3  segment REFs and the actual enu- Aarts, E., J. K. Lenstra. 1997. Local Search in Combinatorial Optimiza-
merative search phase. This separation also allows tion. Wiley, Chichester, UK.
alternative heuristic and exact search-tree pruning Ascheuer, N. 1995. Hamiltonian path problems in the on-line opti-
techniques for the second phase including, e.g., gran- mization of flexible manufacturing systems. PhD thesis, Tech-
nische Universität Berlin, Berlin.
ular edge selection procedures, as proposed by Toth
Bräysy, O., M. Gendreau. 2005a. Vehicle routing with time win-
and Vigo (2003), and methods to terminate the search dows, Part I: Route construction and local search algorithms.
on the basis of feasibility arguments. Transportation Sci. 39 104–118.
Obviously, the proposed LS techniques can be eas- Bräysy, O., M. Gendreau. 2005b. Vehicle routing with time win-
ily integrated into different metaheuristics, which is dows, Part II: Metaheuristics. Transportation Sci. 39 119–139.
substantial for producing high-quality solutions: First, Christofides, N., S. Eilon. 1969. An algorithm for the vehicle-
multistart and iterated LS, VND/VNS, and GRASP dispatching problem. Oper. Res. Quart. 20 309–318.
directly apply LS procedures (Hoos and Stützle 2005). Christofides, N., S. Eilon. 1972. Algorithms for large-scale travelling
salesman problems. Oper. Res. Quart. 23 511–518.
Second, tabu search and related methods also scan
Cordeau, J.-F., G. Desaulniers, J. Desrosiers, M. M. Solomon,
neighborhoods but ask for best non-tabu neighbors F. Soumis. 2002. VRP with time windows. The Vehicle Routing
(Glover and Laguna 1997). It is straightforward to Problem. SIAM Monographs on Discrete Mathematics and Applica-
integrate tabu constraints into the framework. All tions, Chapter 7. SIAM, Philadelphia, 155–194.
these metaheuristics will therefore directly benefit Desaulniers, G., J. Desrosiers, M. M. Solomon. 2002. Accelerat-
ing strategies in column generation for vehicle routing and
from accelerations of LS. Third, simulated anneal- crew scheduling problems. Essays and Surveys in Metaheuristics.
ing, threshold accepting, and related strategies sam- Operations Research/Computer Science Interface Series, Chapter 14.
ple from neighborhoods so that our methods do not Kluwer, Boston, 309–324.
 Irnich: A Unified Modeling and Solution Framework for Vehicle Routing and Local Search-Based Metaheuristics
INFORMS Journal on Computing 20(2), pp. 270–287, © 2008 INFORMS 287

Desaulniers, G., J. Desrosiers, M. M. Solomon, eds. 2005. Column Irnich, S. 2008. Resource extension functions: Properties, inversion,
Generation. Springer, New York. and generalization to segments. OR Spectrum 30(1) 113–148.
Desaulniers, G., F. Lessard, A. Hadjar. 2006. Tabu search, gen- Irnich, S., G. Desaulniers. 2005. Shortest path problems with
eralized k-path inequalities, and partial elementarity for the resource constraints. Column Generation, Chapter 2. Springer,
vehicle routing problem with time windows. Les Cahiers du New York, 33–65.
GERAD G-2006-45, HEC Montréal, Montréal. Irnich, S., B. Funke, T. Grünert. 2006. Sequential search and its
Desaulniers, G., J. Desrosiers, I. Ioachim, M. M. Solomon, F. Soumis, application to vehicle-routing problems. Comput. Oper. Res. 33
D. Villeneuve. 1998. A unified framework for deterministic 2405–2429.
time constrained vehicle routing and crew scheduling prob-
Downloaded from informs.org by [129.93.16.3] on 27 December 2014, at 10:45 . For personal use only, all rights reserved.

Janssens, G. K., R. Hartl, G. Hasle. 2006. Special issue on rich vehi-

lems. T. G. Crainic, G. Laporte, eds. Fleet Management and Logis- cle routing problems. Central Eur. J. Oper. Res. 14 103–104.
tics, Chapter 3. Kluwer Academic Publishers, Boston, 57–93.
Jepsen, M., S. Spoorendonk, B. Petersen, D. Pisinger. 2006. A non-
Desrochers, M., J. Desrosiers, M. Solomon. 1992. A new optimiza-
robust branch-and-cut-and-price algorithm for the vehicle rout-
tion algorithm for the vehicle routing problem with time win-
ing problem with time windows. DIKU Technical Report
dows. Oper. Res. 40 342–354.
06/03, Department of Computer Science, University of Copen-
Desrochers, M., J. K. Lenstra, M. W. P. Savelsbergh. 1990. A classi- hagen, Copenhagen.
fication scheme for vehicle routing and scheduling problems.
Kallehauge, B., J. Larsen, O. B. G Madsen, M. M. Solomon. 2005.
Eur. J. Oper. Res. 46 322–332.
Vehicle routing problem with time windows. Column Genera-
Desrosiers, J., Y. Dumas, M. M. Solomon, F. Soumis. 1995. Time con- tion, Chapter 3. Springer, New York, 67–98.
strained routing and scheduling. M. O. Ball, T. L. Magnanti,
C. L. Monma, G. L. Nemhauser, eds. Network Routing. Hand- Kernighan, B. W., S. Lin. 1970. An efficient heuristic procedure for
books in Operations Research and Management Science, Vol. 8. Else- partitioning graphs. Bell Syst. Tech. J. 49 291–307.
vier, Amsterdam, 35–139. Kilby, P., P. Prosser, P. Shaw. 2000. A comparison of traditional and
Fukasawa, R., J. Lysgaard, M. Poggi de Aragão, E. Uchoa, M. Reis, constraint-based heuristic methods on vehicle routing prob-
R. F. Werneck. 2006. Robust branch-and-cut-and-price for the lems with side constraints. Constraints 5 389–414.
capacitated vehicle routing problem. Math. Programming Ser. A. Kindervater, G. A. P., M. W. P. Savelsbergh. 1997. Vehicle routing:
106 491–511. Handling edge exchanges. Local Search in Combinatorial Opti-
Funke, B. 2003. Effiziente lokale suche für vehicle routing und mization, Chapter 10. Wiley, Chichester, UK, 337–360.
scheduling probleme mit ressourcenbeschränkungen. PhD the- Laporte, G. 1992. The vehicle routing problem: An overview of
sis, Rheinisch-Westfälische Technische Hochschule Aachen exact and approximate algorithms. Eur. J. Oper. Res. 59 345–358.
University, Aachen, Germany. Laporte, G. 1997. Vehicle routing. M. Dell’Amico, F. Maffioli,
Funke, B., T. Grünert, S. Irnich. 2005a. Local search for vehicle rout- S. Martello, eds. Annotated Bibliographics in Combinatorial Opti-
ing and scheduling problems: Review and conceptual integra- mization. Wiley, Chichester, UK, 223–240.
tion. J. Heuristics 11 267–306. Lin, S., B. W. Kernighan. 1973. An effective heuristic algorithm for
Funke, B., T. Grünert, S. Irnich. 2005b. A note on single alternating the traveling-salesman problem. Oper. Res. 21 498–516.
cycle neighborhoods for the TSP. J. Heuristics 11 135–146. Lübbecke, M., J. Desrosiers. 2005. Selected topics in column gener-
Glover, F. 1996. Ejection chains, reference structures and alternating ation. Oper. Res. 53 1007–1023.
path structures for traveling salesman problems. Discrete Appl.
Martin, O., S. W. Otto, E. W. Felten. 1992. Large-step Markov chains
Math. 65 223–253.
for the TSP incorporating local search heuristics. Oper. Res. Lett.
Glover, F., M. Laguna. 1997. Tabu Search. Kluwer, Dordrecht, the 11 219–224.
Netherlands.
Ribeiro, C. C., P. Hansen, eds. 2002. Essays and Surveys in Meta-
Hansen, P., N. Mladenović. 2001. Variable neighborhood search: heuristics. Operations Research/Computer Science Interfaces Series.
Principles and applications. Eur. J. Oper. Res. 130 449–467. Kluwer, Boston.
Hansen, P., N. Mladenović. 2002. Developments of variable neigh- Røpke, S., D. Pisinger. 2006. A unified heuristic for a large class of
borhood search. Essays and Surveys in Metaheuristics. Operations
vehicle routing problems with backhauls. Eur. J. Oper. Res. 171
Research/Computer Science Interface Series, Chapter 19. Kluwer,
750–775.
Boston, 415–439.
Shaw, P. 1998. Using constraint programming and local search
Hasle, G., A. Løkketangen, S. Martello. 2006. Rich models in dis-
methods to solve vehicle routing problems. Technical report,
crete optimization: Formulation and resolution (ECCO XVI).
Department of Computer Science, University of Strathclyde,
Eur. J. Oper. Res. 175 1752–1753.
Glasgow, Scotland.
Hempsch, C., S. Irnich. 2007. Vehicle-routing problems with
inter-tour resource constraints. Technical report 2007-01, Solomon, M. M. 1987. Algorithms for the vehicle routing and
Deutsche Post Endowed Chair of Optimization of Distribu- scheduling problem with time window constraints. Oper. Res.
tion Networks, Rheinisch-Westfälische Technische Hochschule 35 254–265.
Aachen University, Aachen, Germany. http://www.dpor.rwth- Toth, P., D. Vigo, eds. 2002. The Vehicle Routing Problem. SIAM
aachen.de. Monographs on Discrete Mathematics and Applications. SIAM,
Homberger, J., H. Gehring. 1999. Two evolutionary metaheuristics Philadelphia.
for the vehicle routing problem with time windows. Inform. Toth, P., D. Vigo. 2003. The granular tabu search and its applica-
Systems Oper. Res. 37 297–318. tion to the vehicle-routing problem. INFORMS J. Comput. 15
Hoos, H. H., T. Stützle. 2005. Stochastic Local Search Founda- 333–346.
tions and Applications. Morgan Kaufmann Publishers, Elsevier, Voß, S., D. Woodruff. 2002. Optimization Software Class Libraries.
San Francisco. Kluwer Academic Publishers, Boston.