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

net/publication/226255592

Vehicle Routing Problem with Time Windows

Chapter · March 2006

DOI: 10.1007/0-387-25486-2_3

CITATIONS READS
152 5,183

4 authors, including:

Jesper Larsen Oli B. G. Madsen

Technical University of Denmark Technical University of Denmark
68 PUBLICATIONS   2,543 CITATIONS    43 PUBLICATIONS   2,644 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tramp Ship Routing and Scheduling - Incorporating Additional Complexities View project

Integrated Rolling Stock Planning for Suburban Passenger Railways View project

All content following this page was uploaded by Jesper Larsen on 22 May 2014.

The user has requested enhancement of the downloaded file.

On the vehicle routing problem with
time windows

Brian Kallehauge

Submitted in partial fulfillment of the requirements

of the degree of Doctor of Philosophy

Centre for Traffic and Transport

Technical University of Denmark

January 2006
 Abstract
The vehicle routing problem with time windows is concerned with the optimal routing of a fleet of ve-
hicles between a depot and a number of customers that must be visited within a specified time interval,
called a time window. The purpose of this thesis is to develop new and efficient solution techniques for
solving the vehicle routing problem with time windows (VRPTW). The thesis consists of a section of
introductory remarks and four independent papers.
The first paper ‘Formulations and exact approaches for the vehicle routing problem with time win-
dows’ (Kallehauge, 2005, unpublished) is a review of the exact algorithms proposed in the last three
decades for the solution of the vehicle routing problem with time windows. A detailed analysis of the
formulations of the VRPTW is presented together with a review of the literature related to the different
formulations. We present the two main lines of development in relation to the exact approaches for the
VRPTW. One is concerned with the general decomposition approach and the solution to certain dual
problems associated with the VRPTW. Another more recent direction is concerned with the analysis of
the polyhedral structure of the VRPTW. We conclude by examining possible future lines of research in
the area of the VRPTW.
In the second paper ‘Lagrangian duality applied to the vehicle routing problem with time windows’
(Kallehauge, Larsen, and Madsen, Computers & Operations Research, 33:1464-1487, 2006) we consider
the Lagrangian relaxation of the constraint set requiring that each customer must be served by exactly
one vehicle yielding a constrained shortest path subproblem. We present a stabilized cutting-plane algo-
rithm within the framework of linear programming for solving the associated Lagrangian dual problem.
This algorithm creates easier constrained shortest path subproblems because less negative cycles are in-
troduced and it leads to faster multiplier convergence due to a stabilization of the dual variables. We have
embedded the stabilized cutting-plane algorithm in a branch-and-bound search and introduce strong valid
inequalities at the master problem level by Lagrangian relaxation. The result is a Lagrangian branch-and-
cut-and-price (LBCP) algorithm for the VRPTW. Making use of this acceleration strategy at the master
problem level gives a significant speed-up compared to algorithms in the literature based on traditional
column generation. We have solved two test problems introduced in 2001 by Gehring and Homberger
with 400 and 1000 customers respectively, which to date are the largest problems ever solved to optimal-
ity. We have implemented the LBCP algorithm using the ABACUS open-source framework for solving
mixed-integer linear-programs by branch, cut, and price.
In the third paper ‘Path inequalities for the vehicle routing problem with time windows’ (Kallehauge,
Boland, and Madsen, 2005, submitted) we introduce a new formulation of the VRPTW involving only
binary variables associated with the arcs in the underlying digraph. The new formulation is based on
a formulation of the asymmetric traveling salesman problem with time windows and has the advantage
of avoiding additional variables and linking constraints. In the new formulation of the VRPTW time
windows are modeled using path inequalities. The path inequalities eliminate time and capacity infeasible
paths. We present a new class of strengthened path inequalities based on polyhedral results obtained
in the context of the asymmetric traveling salesman problem with replenishment arcs. We study the
VRPTW polytope and determine the polytope dimension. We show that the lifted path inequalities are
facet defining under certain assumptions. We also introduce precedence constraints in the context of the
VRPTW. Computational experiments are performed with a branch-and-cut algorithm on the Solomon
test problems with wide time windows. Based on results on 25-node problems the outcome is that the
algorithm shows promising results compared to leading algorithms in the literature. In particular we report
a solution to a previously unsolved 50-node Solomon test problem R208. The conclusion is therefore
that the path formulation of the VRPTW is no longer the unchallenged winning strategy for solving the
VRPTW.
The fourth and final paper ‘Vehicle routing problem with time windows’ (Kallehauge, Larsen, Mad-
sen, and Solomon. In Desaulniers, Desrosiers, and Solomon, editors, Column generation, pages 67-98,
Springer, New York, 2005) is a contribution to a book on column generation edited by G. Desaulniers,

i
J. Desrosiers, and M. M. Solomon. The focus of the paper is on the VRPTW as one of the important
applications of column generation in integer programming. We discuss the VRPTW in terms of its math-
ematical modeling, its structure and decomposition alternatives. We then present the master problem and
the subproblem for the column generation approach, respectively. Next, we illustrate a branch-and-bound
framework and address acceleration strategies used to increase the efficiency of branch-and-price meth-
ods. Then, we describe generalizations of the problem and report computational results for the classic
Solomon test sets. Finally, we present our conclusions and discuss some open problems.

ii
 Resumé
Den danske titel på denne afhandling er ‘Ruteplanlægningsproblemet med tidsvinduer’. Dette prob-
lem omhandler den optimale styring af en flåde af lastbiler mellem et lager og et antal kunder, der
skal besøges inden for et bestemt tidsinterval, et såkaldt tidsvindue. Formålet med denne afhandling
er udvikling af nye og effektive metoder til løsning af ruteplanlægningsproblemet med tidsvinduer (ve-
hicle routing problem with time windows - VRPTW). Afhandlingen består af et afsnit af introducerende
bemærkninger og fire separate artikler. Det introducerende afsnit beskriver artiklernes videnskabelige
bidrag. I det følgende vil artiklernes engelske titel ikke blive oversat til dansk.
Den første artikel ‘Formulations and exact approaches for the vehicle routing problem with time
windows’ (Kallehauge 2005, ikke publiceret) er en gennemgang af eksakte metoder udviklet gennem
de seneste tre årtier til løsning af ruteplanlægningsproblemet med tidsvinduer. Der præsenteres en de-
taljeret analyse af formuleringerne af VRPTW, samt en gennemgang af den relevante litteratur for de
forskellige formuleringer. De to hovedretninger i forskningen inden for eksakte metoder til VRPTW
beskrives. En af retningerne omhandler generel dekomposition og løsning af visse duale problemer for-
bundet med VRPTW. En anden og nyere retning omhandler analyse af den polyhedrale struktur forbundet
med VRPTW. Artiklen afsluttes med en diskussion af mulige fremtidige forskningsområder inden for
VRPTW.
I den anden artikel ‘Lagrangian duality applied to the vehicle routing problem with time windows’
(Kallehauge, Larsen, and Madsen, Computers & Operations Research, 33:1464-1487, 2006) betragter vi
en Lagrange relaksering af VRPTW. De restriktioner der kræver at hver kunde besøges af netop en lastbil
relakseres, hvilket resulterer i et korteste vej subproblem med bibetingelser. Der præsenteres en stabilis-
eret snitplansalgoritme inden for rammerne af lineær programmering. Algoritmen anvendes til løsning
af det Lagrange duale problem. Denne algoritme resulterer i nemmere korteste vej subproblemer, fordi
færre negative kredse introduceres, og den resulterer i bedre konvergens af løsningen, fordi de duale vari-
able stabiliseres. Vi har inkluderet den stabiliserede snitplansalgoritme i en branch-and-bound søgning
og introducerer stærke gyldige uligheder i master problemet ved hjælp af Lagrange relaksering. Resul-
tatet er en Lagrangian branch-and-cut-and-price (LBCP) algoritme til løsning af VRPTW. Benyttelsen af
denne algoritme giver en signifikant forbedring af løsningstiden sammenlignet med algoritmer i littera-
turen baseret på sædvanlig søjlegenerering. Vi har løst to test problemer introduceret i 2001 af Gehring
og Homberger med henholdsvist 400 og 1000 kunder. Disse problemer er de til dato største problemer
løst til optimalitet. Algoritmen er implementeret ved hjælp af open-source rammesystemet ABACUS, der
er beregnet til løsning af heltalsproblemer ved hjælp af branch-and-cut-and-price.
I den tredje artikel ‘Path inequalities for the vehicle routing problem with time windows’ (Kalle-
hauge, Boland, and Madsen, 2005, indsendt til publicering) præsenterer vi en ny formulering af VRPTW
der kun involverer binære variable tilknyttet kanterne i en underliggende orienteret graf. Denne nye for-
mulering er baseret på en formulering af traveling salesman problemet (TSP) med tidsvinduer og har
den fordel at man undgår ekstra variable og koblende begrænsninger. I den nye formulering af VRPTW
modelleres tidsvinduer ved hjælp af uligheder, der eliminerer ugyldige veje i netværket. En ugyldig vej
kan skyldes overskridelse af kapacitet eller tidsbegrænsninger. Vi præsenterer en ny klasse af uligheder
baseret på polyhedrale resultater opnået inden for TSP med replenishment begrænsninger. Vi bestemmer
dimensionen af VRPTW polytopen. Vi beviser under visse antagelser at de nye uligheder er facetter
for VRPTW polytopen. Vi introducerer også precedence begrænsinger for VRPTW. Beregningsmæs-
sige eksperimenter udføres med en branch-and-cut algoritme. Vi betragter Solomons test problemer med
brede tidsvinduer. Baseret på resultater for problemer med 25 kunder er konklusionen at algoritmen er
lovende sammenlignet med førende algoritmer i litteraturen. Vi præsenterer også en løsning til et hidtil
uløst problem med 50 kunder, nemlig R208. Konklusionen er derfor at korteste vej dekompositionen af
VRPTW ikke længere er den absolutte vinderstrategi til løsning af VRPTW.
Den fjerde og sidste artikel ‘Vehicle routing problem with time windows’ (Kallehauge, Larsen, Mad-
sen, and Solomon. In Desaulniers, Desrosiers, and Solomon, editors, Column generation, pages 67-98,

iii
Springer, New York, 2005) er et bidrag til en bog om søjlegenerering redigeret af G. Desaulniers, J.
Desrosiers og M. M. Solomon. Fokus for denne artikel er VRPTW som et vigtigt eksempel på anvendelse
af søjlegenerering i heltalsprogrammering. Vi diskuterer VRPTW i henhold til modelleringsmæssige as-
pekter, problemstrukturen og dekompositionsalternativer. Derefter præsenterer vi master problemet og
subproblemet ved søjlegenereringsalgoritmen. Efterfølgende beskriver vi branch-and-bound strukturen
og forskellige strategier til accelerering af branch-and-price metoder. Vi beskriver generaliseringer af
problemet og rapporterer beregningsmæssige resultater for de klassiske Solomon test problemer. Afslut-
ningsvist præsenterer vi vore konklusioner og diskuterer åbne problemstillinger.

iv
 Introductory remarks
This thesis is not really meant to be read from cover to cover. Instead, the material is organized
in four independent papers. This type of organization should not get in the way of a reader with some
acquaintance with the subject, however, in this section it is described just what the contributions of each
paper are and how the papers are related. Paper 1 and paper 4 are survey papers. In paper 1 an attempt
has been made to provide a complete survey of the exact approaches for the VRPTW. The focus of paper
4 is on the path formulation of the VRPTW and column generation. Paper 2 and paper 3 represent the
main part of the work during my graduate studies and describe the development of respectively a dual
algorithm and a branch and cut algorithm for the solution of the VRPTW.

Paper 1: Formulations and exact approaches for the vehicle routing problem with time windows

Although this is the first paper of the thesis, it is actually the last paper I wrote during my graduate
studies and it represents my view on the research of this field in the fall of 2005. The purpose of this
paper is to provide a complete survey of the exact approaches proposed in the last three decades for
the solution of the VRPTW. The material in the survey is strictly organized according to four different
formulations of the VRPTW. This is somewhat different from other surveys and in my opinion one of
the contributions of the survey. The relationship between the research on the VRPTW and the TSP
is discussed because of the fundamental role of the TSP in combinatorial optimization in general and
routing and scheduling in particular. I have included some historical references in this survey that I
believe improve the understanding of this field. The survey includes new material related to the polyhedral
approach that is not described in other surveys. In particular it presents my recent work with Boland and
Madsen, which is the subject of paper 3 of this thesis. I also believe that the treatment of the spanning
tree formulation of the VRPTW is more thorough than elsewhere and it illustrates that further work is
required in this area. Finally, the survey presents the research on the decomposition approach of the path
formulation including my work on acceleration strategies for the Lagrangian dual problem with Larsen
and Madsen, which is the subject of paper 2 of this thesis. A question that the survey also discusses is
this: What are the gaps between the bounds of the elementary path formulation and the non-elementary
relaxation with 2-cycle elimination? I made some measurements in relation to this question that are
included in the survey and I believe they are interesting because they made a clear case for eliminating
higher order of cycles. Following this the 2-cycle elimination scheme was generalized by other authors.
The survey concludes by discussing future directions of research that in my opinion are not as clearly
expressed elsewhere.

Paper 2: Lagrangian duality applied to the vehicle routing problem with time windows

This is the first paper I wrote during my graduate studies and it is an extension of the work for my
Masters’ thesis. The topic of my Masters’ thesis was also Lagrangian relaxation and its application to
the VRPTW; particularly how the Lagrangian multipliers could be determined in an efficient way. I
was motivated by the work on acceleration strategies at the master problem level by Kohl and Madsen.
The authors had implemented a method for solving the Lagrangian dual problem that combined the use
of the subgradient algorithm and a version of the bundle algorithm but had only applied the method to
a set of test problems with clustered customer locations. During my Masters’ thesis I had access to a
software package for the solution of convex nondifferentiable optimization problems. I therefore only
had to provide a subroutine for evaluating a single subgradient at each trial point. In the Lagrangian
relaxation of the VRPTW I considered this corresponds to solving a resource-constrained shortest path
problem. Larsen had just completed the development of a state-of-the-art column generation algorithm
for the VRPTW and provided me with the subroutine for solving the shortest path problem. I applied this
dual algorithm to all of the Solomon problem classes including the problems with wide time windows and

v
the results were promising. The method created easier shortest path subproblems because one could start
with small values of the Lagrangian multipliers and increase them to the optimal level. This is different
from standard column generation were one typically finds an initial feasible basis with high objective
function coefficients which in turn gives high dual values. Another drawback of standard column gen-
eration is instability of the solution. Instability refers to a situation where the current solution is closer
(with respect to some norm) to the optimal solution than the next solution. Identifying the difficulty of
the subproblem relative to the behavior of the Lagrangian multipliers and addressing the drawbacks of
standard column generation was the main contributions of the Masters’ thesis. In order to find integer
solutions we combined the dual algorithm with Larsen’s column generation algorithm creating a hybrid
algorithm. However, in 2000 it turned out that we could not use the package for nondifferentiable op-
timization and we therefore developed our own trust-region method for maximizing concave piece-wise
linear functions, which are encountered in Lagrangian relaxation of integer linear programming prob-
lems. The development of this algorithm is the subject of paper 2. Kaj Madsen was also involved in
the development of the trust-region method, which was based on some of his previous work in nonlinear
optimization. We embedded the trust-region method in the branch and bound framework ABACUS. The
dual algorithm is a row generation algorithm similar to polyhedral approaches, however, it is concerned
with the characterization of the objective function of the combinatorial optimization problem instead of
its polytope. Trust-region methods and polyhedral methods are cutting-plane algorithms and we denoted
our method a Lagrangian branch and cut algorithm where the cutting-planes corresponds to the subgradi-
ents. We also introduced valid inequalities for the VRPTW polytope in the master problem but because
the master problem is stated on the dual variables these inequalities are added as columns to the problem
through a pricing step. One of the benefits of the trust-region method was that we avoided solving the
quadratic problems of the bundle algorithms previously applied. It is also interesting to note that Thienel,
the creator of ABACUS, thought that it would require a generalization of ABACUS to use the system
for Lagrangian relaxation but we showed that by remaining within the context of linear programming
when solving the Lagrangian dual problem it was already possible to embed Lagrangian relaxation in the
system. The main contribution of paper 2 is the development of a trust-region method for solving the
Lagrangian dual problem and embedding the trust-region method in a branch and bound algorithm. The
method of paper 2 gives a significant speed-up compared to algorithms in the literature based on stan-
dard column generation and we also succeeded in solving two large-scale problems with 400 and 1000
customers, which to date are the largest problems ever solved to optimality.

Paper 3: Path inequalities for the vehicle routing problem with time windows

Through September 2001 to March 2002 I visited Dr. N. Boland of the University of Melbourne and
paper 3 grew out of the work I carried out during that visit. The work was inspired by the work of Boland
and Mak on polyhedral approaches for a variant of the traveling salesman problem. Because of the links
to the TSP the material of paper 3 very much reflects the research on exact approaches for the TSP. We
proposed a new formulation of the VRPTW that is based on a formulation of the TSP with time windows.
We introduced new and stronger path inequalities for modeling the time windows of the VRPTW based
on the polyhedral results of Mak for the TSP with replenishment constraints. We presented the first
polyhedral results for the VRPTW by determining the dimension of the polytope and proving that the
new path inequalities under certain conditions are facet defining. Furthermore, we take advantage of
the precedence structure that the time windows induce and transfer precedence inequalities from the TSP
context to the VRPTW. Finally, we make use of classes of inequalities for the ATSP in our implementation
of a branch and cut algorithm. The contributions of paper 3 are therefore both theoretical and algorithmic,
i.e. it is the first branch and cut algorithm for this variant of the VRPTW. The computational results of
our algorithm shows that the polyhedral approach is a promising direction of research and in my opinion
the conclusion is that the path formulation is no longer the unchallenged winning strategy for solving the
VRPTW. However, it is clear that the amount of research effort spent to solve VRPTW by the polyhedral

vi
approach is not comparable with what has been dedicated to the decomposition approach of the path
formulation. Paper 3 is only the beginning of the development of the polyhedral approach for the VRPTW
and a better understanding of its polytope and further work on developing efficient separation routines
should yield much better computational results than those reported here.

Paper 4: Vehicle routing problem with time windows

In 2005 GERAD celebrated its 25th anniversary. The Gencol team is one of the best known research
teams of the center and has made significant advances in the integer programming column generation
area. The group originally focused on the vehicle routing problem with time windows and has made very
substantial achievements in this area so it was a great privilege to contribute to a chapter on the VRPTW
for the 25th GERAD anniversary volume on column generation. Paper 4 represents this contribution to
the book on column generation edited by Desaulniers, Desrosiers, and Solomon, all part of the Gencol
team. Paper 4 focuses on the methodological evolution, including cutting planes, parallelism, acceleration
strategies for the master problem, novel subproblem approaches, and generalizations of the VRPTW. It
also reports computational results for the classic Solomon test problems. It is clear that there is some
overlap between paper 1 and paper 4 but in my opinion it is still valuable to include both surveys in this
thesis because they are structured differently and with a different focus thereby highlighting different
aspects of the research on the VRPTW.

vii
 Acknowledgements
I would like to express my deep thanks to my principal supervisor Professor Oli B. G. Madsen of the
Centre for Traffic and Transport, Technical University of Denmark, whose support has been a mainstay
since I started my graduate studies. I would also like to express my gratitude to my supervisor Associate
Professor Jesper Larsen of the department of Informatics and Mathematical Modelling, Technical Uni-
versity of Denmark, whose advise and help over the course of my project has been extremely valuable. I
would also like to thank my supervisor Professor Kaj Madsen of the department of Informatics and Math-
ematical Modelling, Technical University of Denmark, for all his suggestions regarding the development
of a trust-region algorithm for the Lagrangian dual problem.
I am very grateful to Dr. Natashia Boland of University of Melbourne for giving me the opportunity
to visit University of Melbourne through September 2001 to March 2002 and benefit from her broad
research experience. I am also very grateful to Dr. Vicky Mak of Deakin University for all the fruitful
discussions with her during my stay in Melbourne and subsequently. I feel very indebted to Boland and
Mak for giving me the opportunity to benefit from their knowledge about polyhedral combinatorics.
The scholarship for my graduate studies was provided by the Technical University of Denmark. The
research presented in this thesis was carried out in the Centre for Traffic and Transport of the Technical
University of Denmark. I appreciate all the support these institutions offered.

January 10, 2006

Brian Kallehauge

viii
Contents

1 Formulations and exact approaches for the vehicle routing problem with time windows 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem definition and notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Polytope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Subtour and path inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Resource inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Lagrangian duality applied to the vehicle routing problem with time windows 31
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 An ILP formulation of the VRPTW . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3 A Lagrangian relaxation of the VRPTW . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4 Solving the Lagrangian problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 Solving the Lagrangian dual problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6 The Lagrangian branch-and-cut-and-price algorithm . . . . . . . . . . . . . . . . . . . . 40
2.7 Computational results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.7.1 Comparison of column generation and stabilized cutting-planes . . . . . . . . . 41
2.7.2 Solutions for the Solomon problems . . . . . . . . . . . . . . . . . . . . . . . . 42
2.7.3 Solutions for the Homberger problems . . . . . . . . . . . . . . . . . . . . . . . 45
2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3 Path inequalities for the vehicle routing problem with time windows 57
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 A BIP formulation of the VRPTW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 Lifted path inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3.1 Facet proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 Precedence constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 ATSP inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6.1 Tightening of the time windows . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6.2 Precedence relationships and elimination of arcs . . . . . . . . . . . . . . . . . 68
3.7 Test problems and computational platform . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.8 The branch-and-cut algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.8.1 Formulation of the initial binary integer program . . . . . . . . . . . . . . . . . 71

ix
 3.8.2 Separation routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.8.3 Branching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.8.4 Enumeration strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.9 Computational results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.9.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.9.2 Solomon’s test problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

4 Vehicle routing problem with time windows 93

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.3 Structure and decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 The master problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.5 The subproblem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.6 Branch-and-bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.6.1 Branching on the number of vehicles . . . . . . . . . . . . . . . . . . . . . . . 103
4.6.2 Branching on flow variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.6.3 Branching on resource windows . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.7 Acceleration strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.7.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.7.2 Subproblem strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.7.3 Master problem strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.7.4 Cutting planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.8 Generalizations of the VRPTW model . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.8.1 Non-identical vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.8.2 Multiple depots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.8.3 Multiple or soft time windows . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.9 Computational experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.9.1 The Solomon instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.9.2 Computational results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

x
Chapter 1

Formulations and exact approaches for

the vehicle routing problem with time
windows

Brian Kallehauge
Centre for Traffic and Transport, Technical University of Denmark

Abstract
In this paper we review the exact algorithms proposed in the last three decades for the solution of the
vehicle routing problem with time windows (VRPTW). The exact approaches for the VRPTW are in
many aspects inherited from work on the traveling salesman problem (TSP). In recognition of this fact
this paper is structured relative to four seminal papers concerning the formulation and exact solution of
the TSP, i.e. the arc formulation, the arc-node formulation, the spanning tree formulation, and the path
formulation. We give a detailed analysis of the formulations of the VRPTW and a review of the literature
related to the different formulations. There are two main lines of development in relation to the exact
approaches for the VRPTW. One is concerned with the general decomposition approach and the solution
to certain dual problems associated with the VRPTW. Another more recent direction is concerned with
the analysis of the polyhedral structure of the VRPTW. We conclude by examining possible future lines
of research in the area of the VRPTW.

1.1 Introduction
In 1959, a paper by Dantzig and Ramser [18] appeared in the journal Management Science concerning
the routing of a fleet of gasoline delivery trucks between a bulk terminal and a number of service stations
supplied by the terminal. The distance between any two locations is given and a demand for a certain
product is specified for the service stations. The problem is to assign service stations to trucks such
that all station demands are satisfied and total mileage covered by the fleet is minimized. The authors
imposed the additional conditions that each service station is visited by exactly one truck and that the
total demand of the stations supplied by a certain truck does not exceed the capacity of the truck. The
problem formulated in the paper by Dantzig and Ramser [18] was given the name ‘truck dispatching
problem’. I do not know who coined the name ‘vehicle routing problem’ (VRP) for Dantzig and Ramser’s

1
problem but it caught on in the literature and is the title of the most recent book on the problem, and some
of its main variants, edited by Toth and Vigo [85]. In this book, Toth and Vigo [86] considered branch
and bound algorithms for the VRP, Naddef and Rinaldi [72] branch and cut algorithms for the VRP and
polyhedral studies, Bramel and Simchi-Levi [8] set covering based approaches for the VRP, Cordeau,
Desaulniers, Desrosiers, Solomon, and Soumis [15] the VRP with time windows, Toth and Vigo [87] the
VRP with backhauls, and Desaulniers, Desrosiers, Erdmann, Solomon, and Soumis [20] the VRP with
pickup and delivery. Furthermore, the book reviews heuristic approaches and issues arising in real-world
applications. Now the basic version of the VRP is often given the name ‘capacitated vehicle routing
problem’ (CVRP) to distinguish it from its variants. In this paper we consider the variant of the VRP with
time windows (VRPTW), where each customer must be visited within a specified time interval, called a
time window. We consider the case of hard time windows where a vehicle must wait if it arrives before
the customer is ready for service and it is not allowed to arrive late. In the case of soft time windows a
violation of the time window constraints is accepted but then a price must be paid.
Dantzig and Ramser [18] described how the VRP may be considered as a generalization of the trav-
eling salesman problem (TSP). They described the generalization of the TSP with multiple salesmen and
called this problem the ‘clover leaf problem’, a name that is the very picture of the problem. If there are
m salesmen we will refer to the clover leaf problem as the m-TSP, a less lucid name. If in the m-TSP we
impose the condition that specified deliveries be made at every location, excepting the start location, we
get Dantzig and Ramser’s problem. Obviously the VRP is identical with the m-TSP if the total demand
of all locations is less than the capacity of a single vehicle. The standard reference book on the TSP
was edited by Lawler, Lenstra, Rinnooy Kan, and Shmoys [65]. In this book Hoffman and Wolfe [42]
describe how the importance of the TSP comes from the fact that it is typical of other problems of its
genre: combinatorial optimization.
Dantzig had previously collaborated with Fulkerson and Johnson in developing an exact approach to
the TSP. The appearance of their paper ‘Solution of a large-scale traveling-salesman problem’ (Dantzig,
Fulkerson, and Johnson, 1954) in the journal Operations Research was according to Hoffman and Wolfe
[42] "one of the principal events in the history of combinatorial optimization". In this paper the authors
first associated with every tour a vector whose entries are indexed by the roads between the cities. An
entry of this vector is 1 whenever the road between a pair of cities is traveled, otherwise it is 0. They also
defined the linear equations that ensure all cities are visited exactly once in all representations of tours.
These equations are called the degree constraints. Second, they defined a linear objective function that
expressed the cost of a tour as the sum of road distances of successive pairs of cities in the tour. The
problem is then to minimize the linear objective function such that the degree constraints are satisfied
and the solution forms a tour. Third, the authors made a linear programming problem out of this integer
programming problem by identifying just enough additional linear constraints on the vectors to assure
that the minimum is assumed by some tour. This lead to the introduction of the subtour elimination
constraints, which excludes solutions where cities are visited exactly once, but in a set of disconnected
subtours. However, the authors pointed out that there are other types of constraints which sometimes must
be added in addition to subtour elimination constraints in order to exclude solutions vectors involving
fractional entries.
By now the approach of Dantzig, Fulkerson, and Johnson is basic in combinatorial optimization. The
approach is concerned with identifying linear inequalities or cutting planes describing the polytope de-
fined by the convex full of the points in the Euclidean space that represents the set of feasible solutions of
the combinatorial optimization problem. No full description of the TSP polytope is known and because
the TSP belongs to the class of NP-complete combinatorial optimization problems there is no hope for a
polynomial-time cutting plane method for the TSP, unless NP = P. However, as Dantzig, Fulkerson, and
Johnson showed the cutting plane approach can still be applied to the TSP by including the TSP polytope
in a larger polytope (a relaxation) over which we minimize a linear objective function. In this way the
TSP is formulated as a linear program that gives a lower bound for the TSP which can be useful in a

2
branch and bound algorithm. Padberg and Rinaldi [75] refined the integration of the enumeration ap-
proach of classical branch and bound algorithms with the polyhedral approach of cutting planes to create
the solution technique called branch and cut. This method has been very successful in solving large-scale
instances of the TSP and different authors have therefore applied the polyhedral approach to other hard
combinatorial optimization problems. Laporte, Nobert, and Desrochers [62] were the first to apply the
polyhedral approach to the VRP. Finally, we note that the field of discrete mathematics where combina-
torial optimization problems are formulated as linear programs is called polyhedral combinatorics and
we refer to the recent work of Schrijver [80] for a detailed treatment of this subject. For a treatment of
polyhedral theory we refer to Nemhauser and Wolsey [73].
Now we consider another basic method in combinatorial optimization which is concerned with the
characterization of the objective function of the combinatorial optimization problem instead of its poly-
tope. Using relaxation and duality we can determine the optimal objective function value, or at least a
good lower bound on it (assuming minimization), without explicitly solving the integer problem. In par-
ticular, we are concerned with Lagrangian relaxation and duality. A related technique is Dantzig-Wolfe
decomposition, which provides an equivalent bound to the Lagrangian dual bound. In Lagrangian relax-
ation a set of complicating constraints are dualized into the objective function by associating Lagrangian
multipliers with them. This gives us an infinite family of relaxations with respect to the Lagrangian multi-
pliers. For a given set of values of the Lagrangian multipliers the relaxed problem is called the Lagrangian
subproblem. The problem of determining the largest lower bound for this family is called the Lagrangian
dual problem. A fundamental result in mathematical programming is that the Dantzig-Wolfe (gener-
alized) linear programming problem of finding a convex combination of solutions to the (Lagrangian)
subproblem that also satisfy the complicating constraints is dual to the Lagrangian dual problem. The
book by Shapiro [81] marked the first appearance of the term Lagrangian relaxation in a textbook. In this
book the treatment of duality takes a constructive rather than existential approach to Lagrangian multipli-
ers. Everett [27] was the first to take this constructive point of view of Lagrangian multipliers, which is
different from the Karush-Kuhn-Tucker point of view of optimality involving dual variables. For a treat-
ment of Lagrangian duality we refer to Hiriart-Urruty and Lemaréchal [41, Chapter XII]. There exist two
classical algorithms for solving the Lagrangian dual problem. The simplest algorithm for the Lagrangian
dual problem is the subgradient algorithm. The other classical algorithm is the cutting-plane algorithm
(a row-generation algorithm), which in the primal version is the column-generation algorithm. These
algorithms are convex minimization algorithms and belong to the field of nonsmooth or nondifferentiable
optimization. For a treatment of nonsmooth optimization we also refer to Hiriart-Urruty and Lemaréchal
[41]. The combination of branch and bound and column generation was by analogy to branch and cut
called branch and price by Savelsbergh [83]. Finally, when both variables and constraints are generated
in the nodes of the search tree the procedure is called branch, cut, and price. In the last decade a number
of frameworks for implementing branch, cut, and price has appeared, e.g. ABACUS [49], SYMPHONY
[78], and BCP [13].
The use of Lagrangian relaxation in combinatorial optimization was in fact also inspired by the suc-
cessful application of it to the TSP by Held and Karp [38, 39]. Lagrangian relaxation translates the
problem of minimizing our objective function over a set of linear inequalities to finding the maximum
of a concave piecewise affine function. There is a relationship between polyhedral combinatorics and
Lagrangian relaxation. It is defined by the set of inequalities describing the convex hull of the incidence
vectors of solutions to the (Lagrangian) subproblem. Held and Karp [38] proved using general linear
programming theory that the maximization of the bound provided by the 1-tree Lagrangian relaxation
of the TSP gives the same bound as the linear programming relaxation of the TSP proposed by Dantzig
and Ramser [18] using the subtour inequalities. In this way the relationship between these two seminal
contributions [18, 38, 39] is established, and thereby also an example of the relationship between polyhe-
dral combinatorics and Lagrangian relaxation. If the complete set of inequalities of the subproblem was
known it could be included in the integer programming problem and minimizing our objective function

3
over the set of complicating constraints and the inequalities of the subproblem would give the Lagrangian
dual value.
The methods for the vehicle routing problem with time windows are in many aspects inherited from
the work done for the traveling salesman problem. In recognition of this fact this paper is structured
relative to the four seminal papers on the TSP formulations, i.e. the arc formulation, the arc-node formu-
lation, the spanning tree formulation, and the path formulation. We only give a detailed analysis of the
formulations in this paper but we do give a full review of the literature related to the different formula-
tions. There are two main lines of development in relation to the exact approaches. One is concerned with
the general decomposition approach and the solution to a certain dual problem associated with the primal
VRPTW. Another direction is concerned with the analysis of the polyhedral structure of the VRPTW. The
idea of convexity is central to both directions.
This paper is structured according to the four different formulations that have formed the starting
point of the exact approaches to the VRPTW. The four formulations have also been considered in the
context of the TSP. In what follows we give the complete list of references for the VRPTW relative to
the four seminal papers on the TSP. We give the name of the authors of the papers concerning the TSP
followed by a list of the papers on the VRPTW that consider a generalization of the approach for the TSP.
• Arc formulation, Dantzig, Fulkerson, and Johnson [17], [52, 69]
• Arc-node formulation, Miller, Tucker, and Zemlin [71], [4]
• Spanning tree formulation, Held and Karp [38, 39], [31]
• Path formulation, Houck, Picard, Queyranne, and Vemuganti [43], [60, 23, 37, 31, 58, 59, 63, 64,
14, 29, 9, 16, 28, 46, 54, 48]
Cordeau, Desaulniers, Desrosiers, Solomon, and Soumis [15] and Kallehauge, Larsen, Madsen, and
Solomon [53] also give recent surveys in relation to the VRPTW. The survey of Kallehauge et al. [53] is
given in the context of column generation in general and the focus is therefore the path formulation of the
VRPTW which has been studied by several authors. These surveys also give a status on the computational
success of the state of the art algorithms proposed in the literature.
This paper is organized as follows. In Section 1.2 we define the VRPTW as a graph theoretic problem
and introduce notation used throughout the paper. We also describe the complexity of the VRPTW
and define its polytope. In Section 1.3 we consider the arc formulation of the VRPTW involving only
binary variables associated with arcs of an underlying directed graph. In Section 1.4 we review the arc-
node formulation of the VRPTW where we also associate variables with nodes of the directed graph.
Section 1.5 considers a method to find lower bounds for the VRPTW, with the help of time and capacity
constrained shortest spanning trees and Lagrangian relaxation. In Section 1.6 we consider a method to
find lower bounds for the VRPTW, with the help of time and capacity constrained shortest paths and
Lagrangian relaxation. Finally, in Section 1.7 we present some conclusions and discuss future directions
of research.

1.2 Problem definition and notation

In this section we define the VRPTW as a graph theoretic problem and introduce notation used throughout
the paper.

Definition 1.2.1 A time and capacity constrained digraph D = (V, A, c,t, a, b, d, q) is defined by a node
set V = V∗ ∪ {0, n + 1} for V∗ = {1, . . . , n} the set of customer nodes and 0 and n + 1 respectively the start
and destination depot node, arc set A = A∗ ∪ δ + (0) ∪ δ − (n + 1) for A∗ = A(V∗ ) the set of arcs spanned

4
by the customer nodes and δ + (0) = {(0, i) | i ∈ V∗ } the set of arcs leaving the start depot node and
δ − (n + 1) = {(i, n + 1) | i ∈ V∗ } the set of arcs entering the destination depot node, costs on arcs c ∈ ZA
where ci j ≤ cik + ck j for i, j, k ∈ V , durations on arcs t ∈ NA where ti j ≤ tik + tk j for i, j, k ∈ V , release and
due times on nodes a, b ∈ {Z+ ∪ {+∞}}V where a0 = an+1 = 0, b0 = bn+1 = +∞, ai ≥ t0i and bi ≥ ai
for i ∈ V∗ and b j ≥ ai + ti j for (i, j) ∈ A∗ , demands on nodes d ∈ ZV+ where d0 = dn+1 = 0, and capacity
q ∈ Z+ where q ≥ di for i ∈ V∗ and q ≥ di + d j for (i, j) ∈ A∗ .

Definition 1.2.2 For any path P = (v1 , . . . , vk ) in D, the arrival times of the set of nodes V (P) of the path
V (P)
is the vector s ∈ Z+ defined by:

sv1 = av1 ,
svi = max{svi−1 + tvi−1 vi , avi } for i = 2, . . . , k,

the demand of the path is d(V (P)), and the cost of the path is c(A(P)).

Definition 1.2.3 We say that a path P = (v1 , . . . , vk ) in D is feasible if

(1.1) svi ≤ bvi for i ∈ V (P)

and

(1.2) d(V (P)) ≤ q.

Definition 1.2.4 We say that a path P = (v1 , . . . , vk ) in D is infeasible if

(1.3) svi > bvi for any i ∈ V (P)

or

(1.4) d(V (P)) > q.

Definition 1.2.5 An infeasible path P = (v1 , . . . , vk ) in D is said to be minimal infeasible if the subpaths
of P induced by depriving V (P) of respectively the starting node

(1.5) V (P) \ {v1}

and the end node

(1.6) V (P) \ {vk }

are feasible. We denote by PD the set of all minimal infeasible paths in D.

Definition 1.2.6 A route in D is defined as a feasible path from 0 to n + 1

R = (0, v2 , . . . , vk−1 , n + 1).

We denote by R the set of all routes in D.

Definition 1.2.7 A k-route in D is the union of k routes

κ k = R1 ∪ R2 ∪ · · · ∪ Rk ,

such that each node v ∈ V∗ belongs to exactly one set V (Ri ), 1 ≤ i ≤ k. The cost of a k-route is c(A(κk )) =
c(∪A(Ri ) | i = 1, . . . , k).

5
For any W ⊆ V∗ , computing the number

(1.7) k(W ) = min{k | a κk exists in D(W ∪ {0, 1})},

represents the problem of finding the minimum number of routes required to visit the subset of customer
nodes W ; we stress the fact that the notation k(W ) represents at the same time a number and a problem to
solve.

Definition 1.2.8 A partition of the set of customer nodes V∗ induced by a k-route κk is called a feasible
k-partition. We denote by K = {k(V∗ ), . . . , n} the set of feasible partition sizes.

Definition 1.2.9 For each k ∈ K we denote by Rk the set of k-routes with corresponding partition size k.

Definition 1.2.10 The set of all k-routes in D for k ∈ K is denoted by RK = {∪Rk | k = k(V∗ ), . . . , n}.
The vehicle routing problem with time windows is defined as follows. Given a time and capacity
constrained digraph D, find a k-route of minimum cost, i.e.

(VRPTW) min{c(A(κk )) | κk ∈ RK }.

1.2.1 Complexity
If we place the restrictions on the instances of the VRPTW that ai = 0 and bi = +∞ for every i ∈ V∗ the
resulting restricted problem will be identical to the CVRP. If we furthermore place the restriction on the
instances of the CVRP that q ≥ d(V∗ ) the resulting restricted problem will be identical to the TSP with
multiple salesmen [71], sometimes denoted by m-TSP where m is the number of salesmen, and that can
be transformed to the standard TSP [7]. On the other hand, the question of whether there exists a feasible
solution for a given instance of the CVRP is an instance of the bin packing problem (BPP). Finally, if we
place the restrictions on the instances of the VRPTW that K = {1}, q ≥ d(V∗ ), and ci j = 0 for (i, j) ∈ A
the resulting restricted problem is the nonpreemptive single machine scheduling problem with release
dates and deadlines (SS1). Garey and Johnson [32] proved the NP-completeness of TSP, BPP, and SS1.
This implies by proof of restriction the NP-completeness of the VRPTW.

1.2.2 Polytope
Definition 1.2.11 With every k-route κk ∈ RK in D, we associate an incidence vector xκk ∈ RA defined
by:
(
κk 1 if (i, j) ∈ A(κk ),
xi j =
/ A(κk ).
0 if (i, j) ∈

Definition 1.2.12 The VRPTW polytope of a time and capacity constrained digraph D = (V, A, c,t, a, b, d, q)
is the convex hull of the incidence vectors of the k-routes in RK :

PVRPTW = conv{xκk ∈ RA | κk ∈ RK }.

The vehicle routing problem with time windows is equivalent to minimizing the function cT x over the
VRPTW polytope. The NP-completeness of the VRPTW also implies that no description in terms of
inequalities of the VRPTW polytope may be expected. However, polynomial-time computable lower
bounds for the VRPTW can be obtained by including the VRPTW polytope in a larger polytope (a relax-
ation) over which cT x can be minimized in polynomial time.

6
1.3 Subtour and path inequalities
In this section we consider a formulation of the VRPTW involving only binary variables associated with
the arcs in D.
The VRPTW polytope is the set of those x ∈ BA satisfying the degree equations

(1.8) x(δ + (i)) = 1 for i ∈ V∗ ,

(1.9) x(δ (i)) = 1
−
for i ∈ V∗ ,

the subtour inequalities

(1.10) x(A(W )) ≤ |W | − 1 for W ⊆ V∗ with |W | ≥ 2,

and the path inequalities

(1.11) x(A(P)) ≤ |A(P)| − 1 for P ∈ PD .

The formulation (1.8)-(1.11) of the VRPTW was proposed by Kallehauge, Boland, and Madsen [52].
The subtour inequalities were proposed by Dantzig, Fulkerson, and Johnson [17] in their seminal paper
on the TSP. The idea of using path inequalities to model time window restrictions was presented by
Ascheuer, Fischetti, and Grötschel [1] in their paper on the ATSPTW.
Laporte, Nobert, and Desrochers [62] generalized the subtour inequalities for the CVRP
 
d(W )
(1.12) x(A(W )) ≤ |W | − for W ⊆ V∗ with W 6= 0.
/
q
Naddef and Rinaldi [72] reviewed capacity inequalities of the CVRP polytope including the rounded
capacity inequalities (1.12). Kohl, Desrosiers, Madsen, Solomon, and Soumis [59] further generalized
the subtour inequalities for the VRPTW

(1.13) x(A(W )) ≤ |W | − k(W ) for W ⊆ V∗ with W 6= 0,

/

and denoted them k-path inequalities. If we in the formulation of the VRPTW replace the subtour inequal-
ities (1.10) with the capacity inequalities (1.12) then it is sufficient to only include condition (1.1) and
(1.3) in Definition 1.2.3 and Definition 1.2.4, respectively. Then we denote by PDTW the set of minimal
time infeasible paths of Definition 1.2.5 and redefine (1.11) to

(1.14) x(A(P)) ≤ |A(P)| − 1 for P ∈ PDTW .

However, further replacing the capacity inequalities (1.12) with the k-path inequalities (1.13) is not suffi-
cient to drop (1.14) in the formulation (1.8), (1.9), (1.12), and (1.14).
Kallehauge et al. [52] presented a class of strengthened path inequalities S1 for the VRPTW based on
the polyhedral results obtained by Mak [70] in the context of the asymmetric traveling salesman problem
with replenishment arcs. Furthermore, Kallehauge et al. [52] determined the dimension of the VRPTW
polytope and proved that the S1 inequalities are facet defining under certain assumptions. These were the
first polyhedral results for the VRPTW. Kallehauge et al. [52] also transferred the precedence constraints
of Balas, Fischetti, and Pulleyblank [3] to the VRPTW context. Finally, the authors implemented a branch
and cut algorithm that showed promising results and reported a solution to a previously unsolved 50-node
test problem of Solomon [82].
Mak and Ernst [69] have also studied a formulation of the VRPTW similar to (1.8)-(1.11) and pre-
sented five new classes of valid inequalities for the VRPTW. The first four classes are based on the
well-known Dk cycle inequalities [35] and the last is a class of path inequalities related to the S1 class.
The authors also proved that the new classes of inequalities are facet defining under certain assumptions.

7
1.4 Resource inequalities
Next we introduce a formulation of the VRPTW where we also associate variables with the nodes in D.
The integer solutions of (1.8)-(1.11) are exactly the incidence vectors of k-routes, so it gives an inte-
ger programming formulation of the VRPTW. The class of subtour inequalities (1.10) have a cardinality
growing exponentially with n. An equivalent class of inequalities with polynomial cardinality was pro-
posed by A. W. Tucker in 1960 [71]. He introduced node variables u ∈ ZV∗ and proposed the inequalities
(1.15) ui − u j + pxi j ≤ p − 1 for (i, j) ∈ A∗ .
where p ≤ n. The node variables ui play the role of node potentials in an electrical network and the
inequalities involving them serve to eliminate routes that do not begin at the start depot node 0 and end
at the destination depot node n + 1. This is already achieved by the subtour inequalities of Dantzig et al.
[17]. The ui can be adjusted so that ui = j if customer i is the jth customer visited in the route which
includes customer i. The node variables ui therefore represent the accumulated number of visits along a
route. The inequalities (1.15) ensure that no more than p customers are visited in one route. For p ≥ n
we have the standard VRPTW. In this way p is a resource limit on the number of visits in a route and we
generally denote (1.15) resource inequalities.
Kulkarni and Bhave [61] generalized A. W. Tucker’s inequalities for the CVRP, i.e. introduced a class
equivalent to (1.12) but with polynomial cardinality. If every demand di for i ∈ V∗ represents a pick-up at
customer i and y ∈ ZV∗ then
(1.16) yi − y j + qxi j ≤ q − d j for (i, j) ∈ A∗ ,
where di ≤ yi ≤ q for i ∈ V∗ , are denoted pick-up inequalities. For any route R = (0, v2 , . . . , vk−1 , n + 1)
where k ≥ 3, the node variables of the route yvi ∈ Z, i = 2, . . . , k − 1, can be adjusted so that:
i
(1.17) yvi = ∑ dv j ,
j=2

where yvk−1 ≤ q. In case every demand di for i ∈ V∗ represents a delivery to customer i and y′ ∈ ZV∗ then
(1.18) y′i − y′j − qxi j ≥ d j − q for (i, j) ∈ A∗ ,
where yi ≤ q − di for i ∈ V∗ , are denoted delivery inequalities. For any route R = (0, v2 , . . . , vk−1 , n + 1)
where k ≥ 3, the node variables of the route y′vi ∈ Z, i = 2, . . . , k − 1, can be adjusted so that:
i
(1.19) y′vi = d(V (R)) − ∑ dv j ,
j=2

where yv2 ≤ q − dv2 . In the standard VRPTW it is required that all demands represent a pick-up or
alternatively that all demands represent a delivery.
Desrochers and Laporte [22] further generalized A. W. Tucker’s inequalities for the VRPTW, but in
the case of time windows the resource inequalities are not only equivalent to the generalized subtour
inequalities (1.13) of Dantzig et al. [17], but also equivalent to the path inequalities (1.14). If s ∈ ZV∗ then
(1.20) si − s j + (bi + ti j − a j )xi j ≤ bi − a j for (i, j) ∈ A∗ ,
where ai ≤ si ≤ bi for i ∈ V∗ , are denoted the time inequalities.
For any route R = (0, v2 , . . . , vk−1 , n + 1) where k ≥ 3, the node variables of the route svi ∈ Z, i = 2, . . . , k −
1, can be adjusted so that:
(1.21) sv1 = av1 ,
(1.22) svi = max{svi−1 + tvi−1 vi , avi } for i = 2, . . . , k,

8
where svi ≤ bvi for i = 2, . . . , k − 1.
The VRPTW polytope is the set of those x ∈ BA , y ∈ ZV∗ , and s ∈ ZV∗ satisfying the degree equations

(1.23) x(δ + (i)) = 1 for i ∈ V∗ ,

(1.24) x(δ (i)) = 1
−
for i ∈ V∗ ,

the pick-up inequalities

(1.25) yi − y j + qxi j ≤ q − d j for (i, j) ∈ A∗ ,

the time inequalities

(1.26) si − s j + (bi + ti j − a j )xi j ≤ bi − a j for (i, j) ∈ A∗ ,

and the bounds

(1.27) si ≤ bi for i ∈ V∗ ,
(1.28) si ≥ ai for i ∈ V∗ ,
(1.29) yi ≤ q for i ∈ V∗ ,
(1.30) yi ≥ d i for i ∈ V∗ .

The formulation (1.23)-(1.30) of the VRPTW was proposed by Bard, Kontoravdis, and Yu [4]. How-
ever, the authors considered the problem (1.7) of finding the minimum number of routes required to visit
the set of customers V∗ . This problem is equivalent to minimizing the function x(δ + (0)) over the VRPTW
polytope. Bard et al. [4] proposed the first branch and cut algorithm for the VRPTW based on this for-
mulation. They considered a number of well-known inequalities from the TSP and VRP and proposed
two new types of path inequalities taking into account the time windows of the problem. However, from
a computational point of view the generalized subtour inequalities were the most effective. A reason for
this is that the authors developed efficient heuristics for the separation of subtour inequalities. Further-
more, these path inequalities are quite weak compared to e.g. the S1 inequalities proposed in the context
of the VRPTW by Kallehauge et al. [52]. Bard et al. [4] also used a so-called greedy randomized adaptive
search procedure (GRASP) for finding feasible solutions or upper bounds in the search tree. The branch
and cut method of Bard et al. [4] showed promising computational results.
Formulating the integer programming model is only the first step when hard optimization problems
are solved by branch and cut. The crucial part is the subset one considers of the finite family of defining
inequalities of the associated polytope. It is well-known that A. W. Tucker’s inequalities generally provide
worse linear programming bounds than families of inequalities with exponential cardinality. In the context
of the ATSPTW Ascheuer, Fischetti, and Grötschel [2] noted that their model involving only binary
variables e.g. cannot handle as general objective functions as a model involving node variables. One
example could be a makespan type of objective where the total time spent is minimized, i.e. including
waiting time. Depending on the application this should therefore be the criterion for considering node
variables or not because strengthened path inequalities dominate Tucker’s inequalities.

1.5 Trees
In this section we consider a method to find lower bounds for the VRPTW, with the help of time and
capacity constrained shortest spanning trees and Lagrangian relaxation.

Definition 1.5.1 A 0-arborescence in D is a subset B of A such that B forms a shortest spanning tree on
V \ {n + 1} rooted at node 0 and such that for each node v ∈ V∗ there is a feasible path from 0 to v.

9
Definition 1.5.2 A routed arborescence, or just arborescence, in D is a subset T of A such that T \ δ − (n +
1) is a 0-arborescence and such that T contains a subset of arcs entering n + 1, say F = T ∩ δ − (n + 1),
where |F| = |T ∩ δ + (0)|. We denote by T the collection of all arborescences in D. For any arborescence
T in D, the cost is defined by c(T ).
The shortest (= minimum cost) arborescence problem with time windows and capacity constraints is
defined as follows. Given a time and capacity constrained digraph D, find an arborescence T of minimum
cost, i.e.

(SAPTWCC) min{c(T ) | T ∈ T }.

Papadimitriou [76] proved the NP-completeness of the capacitated tree problem and therefore by
proof of restriction the SAPTWCC is also NP-complete.

Definition 1.5.3 With every arborescence T ∈ T in D, we associate an incidence vector xT ∈ RA defined

by:
(
T 1 if (i, j) ∈ T ,
xi j =
0 if (i, j) ∈
/ T.

Definition 1.5.4 The SAPTWCC polytope of a time and capacity constrained digraph D is the convex
hull of the incidence vectors of the arborescences in T :

PSAPTWCC = conv{xT ∈ RA | T ∈ T }.

The shortest arborescence problem with time windows and capacity constraints is equivalent to minimiz-
ing the function cT x over the SAPTWCC polytope.
The SAPTWCC polytope is the set of those x ∈ BA satisfying the indegree equations

(1.31) x(δ − (i)) = 1 for i ∈ V∗ ,

the cut-set inequalities

(1.32) x(δ − (W )) ≥ 1 for W ⊆ V∗ with W 6= 0/

the path inequalities

(1.33) x(A(P)) ≤ |A(P)| − 1 for P ∈ PD

and the flow balance equation

(1.34) x(δ + (0)) − x(δ − (n + 1)) = 0.

The SAPTWCC can be solved by considering two separate problems:

• the determination of a shortest 0-arborescence B∗ in D, defined by those x ∈ BA satisfying (1.31),
(1.32), and (1.33), and
• the determination of a subset F ∗ of minimum cost arcs entering the destination depot n + 1, defined
by those x ∈ BA satisfying (1.34).

10
 To see the relationship between the VRPTW and the SAPTWCC we consider a slightly different
formulation of the VRPTW equivalent to (1.8)-(1.11). The outdegree (1.8) and indegree (1.9) equations
give us a mean to alter the "inside" form of the subtour inequalities (1.10). By subtracting (1.8) from
(1.10) we obtain the "outside" form defined by the set of arcs in D leaving W

(1.35) x(δ + (W )) ≥ 1 for W ⊆ V∗ with W 6= 0,

/

and by subtracting (1.9) from (1.10) we obtain the outside form defined by the set of arcs in D entering
W

(1.36) x(δ − (W )) ≥ 1 for W ⊆ V∗ with W 6= 0.

/

Furthermore, the degree equations (1.8) and (1.9) together with the subtour inequalities (1.10) imply that

(1.37) x(δ + (0)) − x(δ − (n + 1)) = 0.

The formulation (1.8)-(1.11) of the VRPTW is therefore equivalent to

(1.38) x(δ + (i)) = 1 for i ∈ V∗ ,

(1.39) x(δ (i)) = 1
−
for i ∈ V∗ ,
(1.40) x(δ (W )) ≥ 1 for W ⊆ V∗ with W 6= 0,
−
/
(1.41) x(A(P)) ≤ |A(P)| − 1 for P ∈ PD ,
(1.42) x(δ + (0)) − x(δ − (n + 1)) = 0.

In the formulation (1.38)-(1.42) of the VRPTW, the outdegree equations (1.38) appear as the compli-
cating constraints. If these constraints were not present the VRPTW would reduce to the SAPTWCC. To
take advantage of this problem structure we therefore consider the Lagrangian relaxation with respect to
the outdegree equations (1.38). For any λ ∈ RV∗ we consider the Lagrangian function defined by:
!
(1.43) L(λ , x) = ∑ ∑ ci j xi j − ∑ λ i ∑ xi j − 1
i∈V j∈V i∈V∗ j∈V

(1.44) = c̃ x + λ (V∗),
T

where
(
ci j − λ i for i ∈ V∗ , j ∈ V ,
(1.45) c̃i j =
ci j for i = 0, j ∈ V∗ .

The Lagrangian problem associated with λ is defined by:


(1.46) z(λ ) = min c̃T x | x ∈ BA satisfies conditions (1.39) − (1.42) + λ (V∗)

where λ is fixed in RV∗ . Problem (1.46) is a SAPTWCC with the cost function given by c̃ : A → R.
The Lagrangian dual problem is defined by:

(1.47) max{z(λ ) | λ ∈ RV∗ }.

Definition 1.5.5 Let the set of arborescences T be indexed with k = 1, . . . , |T | so Tk is the kth arbores-
cence and define the cost of the kth arborescence

∑
T
ck = ci j xi jk
(i, j)∈Tk

11
and the outdegree of node i ∈ V∗ in the kth arborescence

∑ xi jk
T
aik = for i = 1, . . . , n.
j∈V

The Lagrangian problem (1.46) is then defined by:


(1.48) z(λ ) = min ck − aT k λ + λ (V∗ )
1≤k≤|T |

and the Lagrangian dual problem is defined by:


(1.49) max { min ck − aT k λ + λ (V∗ )}.
λ ∈RV∗ 1≤k≤|T |

Since T is finite it allows us to express (1.49) as the following linear program with many constraints or
rows:

(1.50) max θ + ∑ λi
i∈V∗
subject to
θ 6 ck − ∑ aik λi for k = 1, . . . , |T |,
i∈V∗
λi ∈ R for i ∈ V∗ ,
θ ∈ R.

The LP dual of (1.50) is a linear program with many variables or columns:

|T |
(1.51) min ∑ ck yk
k=1
subject to
|T |
∑ aik yk = 1 for i ∈ V∗ ,
k=1
|T |
∑ yk = 1,
k=1
yk > 0 for k = 1, . . . , |T |.

Problem (1.51) with yk required to be integral is equivalent to the VRPTW. In case of integrality
constraints an optimal solution to (1.51) must satisfy yk∗ = 1 for some Tk∗ ∈ T and yk = 0 for all Tk ∈
T \ {Tk∗ }. Problem (1.51) is the LP relaxation of the Dantzig-Wolfe decomposition obtained when
any solution to the VRPTW is expressed as a non-negative convex combination of resource-constrained
directed spanning trees. The relaxation of the VRPTW presented in this section has never been used
directly in a branch and bound algorithm. The idea of using shortest spanning trees has been considered
in the VRPTW context but only one paper [31] in the literature on the VRPTW has considered this
classical approach in vehicle routing.
Held and Karp [38] explored the relationship between the symmetric and asymmetric traveling sales-
man problem and shortest spanning trees in undirected and directed graphs, respectively. Consider the
symmetric TSP and the complete undirected graph G = (V, E) on n nodes. A 1-tree is a subgraph T of
G with nodes 1, 2, . . . , n consisting of a tree on the nodes 2, 3, . . . , n together with two edges incident with

12
node 1. In fact, a 1-tree T in G consists of exactly n arcs, whereas a tree in G consists of n − 1 arcs.
So a 1-tree is a tree with an additional arc added explaining the term 1-tree. A solution to the STSP
is precisely a 1-tree in which all nodes have degree 2. The Lagrangian relaxation of the STSP with re-
spect to the degree constraints x(δ (v)) = 2, v ∈ V \ {1}, gives a shortest 1-tree Lagrangian problem and
the Lagrangian dual problem can be expressed as a pair of linear programs similar to (1.50) and (1.51).
Held and Karp [38] gave a column generation method and an ascent method for finding the Lagrangian
dual value. The column generation method was able to solve the program (1.51) for most problems with
n = 12 and some problems with 13 ≤ n ≤ 20. On larger problems the convergence was always too slow
and the authors noted that this was consistent with the behavior of other column generation techniques at
that time referring to the work of Gilmore and Gomory [34]. Held and Karp [38] also described how the
approach for the symmetric TSP carry over to the asymmetric case. In this case the authors introduced a
type of directed subgraph they called a 1-arborescence, defined as an arborescence (directed tree) rooted
at node 1 plus an arc (v, 1) joining some node v ∈ V \ {1} to node 1. We remark that the authors notion of
a 1-arborescence is slightly different from our r-arborescence, which is defined as an arborescence rooted
at node r. Held and Karp [38] also described the extension of the 1-tree approach to the m-STSP in which
the degree at node 1 is 2m. This would later be further generalized to the symmetric CVRP. Held and
Karp [38] already added in the proof of their paper that a new method for computing the Lagrangian dual
value would be presented in a sequel to the paper. The following paper [39] was a milestone in the sub-
ject of Lagrangian relaxation in integer programming. Held and Karp [39] successfully introduced what
became known as the subgradient algorithm (a term introduced by Held, Wolfe, and Crowder [40]) and
influenced future research dramatically [5]. In 1974 Geoffrion [33] coined the term Lagrangian relaxation
to describe the method of Held and Karp [38, 39]. Because of the initial use of the subgradient algorithm,
Lagrangian relaxation to some extent became synonymous with the subgradient algorithm, which is un-
fortunate because this algorithm is the simplest algorithm for concave maximization and suffers from
several drawbacks [41]. Indeed the method is mainly attractive because it is so simple to implement.
The earliest generalization of the approach by Held and Karp [38] was proposed by Christofides, Min-
gozzi, and Toth [11] for the SCVRP based on the k-degree center tree (k-DCT) relaxation of the SCVRP.
The approach allow for the possibility of single customer routes. Fisher [30] presented a different relax-
ation of the SCVRP using shortest k-trees. Consider the symmetric CVRP and the complete undirected
graph G = (V, E) on n nodes. A k-tree is a subgraph of G consisting of n − 1 + k edges that span the n
nodes. The degree of the depot node 1 is 2k. As the name suggests this is a generalization of the 1-tree
approach of Held and Karp [38]. The author dualized the capacity constraints (1.12) of Laporte, Nobert,
and Desrochers [62] and solved the Lagrangian dual problem using the subgradient algorithm and gener-
ation of violated capacity constraints. The Lagrangian dual problem expressed as an LP similar to (1.50)
is exponential in size since it has exponentially many variables as well as constraints corresponding to
the number of capacity constraints and k-trees, respectively. Instead of using the subgradient algorithm
a cut and column generation algorithm similar to the one proposed by Kallehauge, Larsen, and Madsen
[54] could be used for solving the dual problem.
Fisher [30] also described an extension of his method to the VRPTW. He introduced path inequalities
(1.11) in a formulation of the VRPTW and relaxed these to obtain the same k-tree Lagrangian problem
considered for the CVRP. He did not report any computational results. The extension to time windows
was developed with K. Jörnsten and O.B.G. Madsen and together Fisher, Jörnsten, and Madsen [31] later
presented computational results using the k-tree method. However, the shortest tree relaxation of the
VRPTW has not been the subject of the same amount of research as the shortest path relaxation and in
our view it is at this point an open question whether the formulation described in this section is effective.
We do not consider a formulation of the SAPTWCC with a fixed degree at the root node because we
consider the variant of the VRPTW with a free number of vehicles. This is different from the TSP (VRP)
where the authors consider rooted trees with a degree constraint at the root node because they consider
problems with a fixed number of tours (routes). Toth and Vigo [84] proposed an algorithm for the shortest

13
arborescence problem with capacity constraints (SAPCC), however the extension to time windows has
not been considered in the literature. Finally, we refer to Schrijver [80, Section 50.6a] for a complexity
survey for shortest spanning trees.

1.6 Paths
We next consider a method to find lower bounds for the VRPTW, with the help of time and capacity
constrained shortest paths and Lagrangian relaxation.

Definition 1.6.1 We denote by D0,n+1 the time and capacity constrained digraph obtained from D by
adding the arc (0, n + 1) to the set of arcs A of D. For notational convenience we denote the extended set
of arcs by A. The cost c0,n+1 and duration t0,n+1 on arc (0, n + 1) is zero.
The elementary shortest path problem with time windows and capacity constraints is defined as fol-
lows. Given a time and capacity constrained digraph D0,n+1 , find a path from 0 to n + 1 of minimum cost,
i.e.

(ESPPTWCC) min{c(A(R)) | R ∈ R}.

Dror [26] proved the NP-completeness of the ESPPTWCC. Next we define a relaxation of the ESPPTWCC.
A walk in D0,n+1 from 0 to n + 1 is a sequence of m nodes

(1.52) R(k) = (0, v2 , . . . , vm−1 , n + 1),

where v2 , . . . , vm−1 are not necessarily distinct. If vi+p 6= vi for 2 ≤ p ≤ k the walk is called a non-
elementary route with no k-cycles. The (non-elementary) shortest path problem with time windows and
capacity constraints and no k-cycles is defined as follows. Given a time and capacity constrained digraph
D0,n+1 , find a walk from 0 to n + 1 of minimum cost containing no k-cycles, i.e.

(k-SPPTWCC) min{c(A(R(k) )) | R(k) ∈ R(k) }.

Pseudo-polynomial time algorithms exist for the k-SPPTWCC [21, 43, 23, 46]. If no cycle elimination is
performed the problem is called the SPPTWCC. In D0,n+1 we have that R(2) ⊇ R(3) ⊇ · · · ⊇ R(n−1) ⊇ R.

Definition 1.6.2 With every route R ∈ R in D0,n+1 , we associate an incidence vector xR ∈ RA defined by:
(
1 if (i, j) ∈ A(R),
xRij =
0 if (i, j) ∈
/ A(R).

Definition 1.6.3 The ESPPTWCC polytope of a time and capacity constrained digraph D0,n+1 is the
convex hull of the incidence vectors of the routes in R:

PESPPTWCC = conv{xR ∈ RA | R ∈ R}.

The elementary shortest path problem with time windows and capacity constraints is equivalent to mini-
mizing the function cT x over the ESPPTWCC polytope.
The ESPPTWCC polytope is the set of those x ∈ BA satisfying the degree equations

(1.53) x(δ + (0)) = 1,

(1.54) x(δ − (n + 1)) = 1,

14
the balance equations

(1.55) x(δ − (i)) − x(δ + (i)) = 0, for i ∈ V∗ ,

the subtour inequalities

(1.56) x(A(W )) ≤ |W | − 1 for W ⊆ V∗ with |W | ≥ 2,

and the path inequalities

(1.57) x(A(P)) ≤ |A(P)| − 1 for P ∈ PD .

To see the relationship between the VRPTW and the ESPPTWCC we define the VRPTW in a slightly
different way. Given a time and capacity constrained digraph D0,n+1 , find a collection of routes {Rk | k =
1, . . . , n} of minimum cost such that each node v ∈ V∗ is visited exactly once, i.e.

(1.58) min ∑ c(A(Rk ))

1≤k≤n
subject to
|∪1≤k≤n A(Rk ) ∩ δ + (v)| = 1 for v ∈ V∗ ,
Rk ∈ R for k = 1, . . . , n.

Remark 1.6.1 Since we have introduced the arc (0, n + 1) in D0,n+1 the solution to (1.58) may contain a
number of zero-cost routes not visiting any customer nodes and problem (1.58) is therefore equivalent to
the VRPTW.

Definition 1.6.4 With every collection of routes κn = {Rk | k = 1, . . . , n} in D0,n+1 , we associate an

incidence vector xκn ∈ RA defined by:
n

(
1 if (i, j) ∈ A(Rk ),
xiκjkn =
0 if (i, j) ∈
/ A(Rk ).
n
The polytope of (1.58) is the set of those x ∈ BA where xk ∈ BA satisfy (1.53)-(1.57) for k = 1, . . . , n
and:

(1.59) ∑ ∑ xi jk = 1 for i ∈ V∗ .
1≤k≤n j∈V

The vehicle routing problem with time windows is therefore equivalent to minimizing the function
∑1≤k≤n cT xk over the polytope of (1.58).
In the formulation (1.53)-(1.59) of the VRPTW, the constraints (1.59) appear as coupling constraints,
which link the individual variables xk . If these constraints were not present the VRPTW would reduce to
n ESPPTWCC problems, each with the simpler formulation (1.53)-(1.57), and thus become considerable
simpler. To take advantage of this problem structure we therefore consider the Lagrangian relaxation with
respect to the constraints (1.59). For any λ ∈ RV∗ we consider the Lagrangian function defined by:
!
(1.60) L(λ , x) = ∑ ∑ ∑ ci j xi jk − ∑ λi ∑ ∑ xi j − 1
1≤k≤n i∈V j∈V i∈V∗ 1≤k≤n j∈V

(1.61) = ∑ c̃T xk + λ (V∗),

1≤k≤n

15
where
(
ci j − λ i for i ∈ V∗ , j ∈ V ,
(1.62) c̃i j =
ci j for i = 0, j ∈ V∗ .

The Lagrangian problem associated with λ is defined by:

(1.63) z(λ ) = min{ ∑ c̃T xk | xk ∈ BA satisfies conditions (1.53) − (1.57) for k = 1, . . . , n} + λ (V∗ )
1≤k≤n

where λ is fixed in RV∗ . Problem (1.63) can be solved by considering n ESPPTWCC problems with the
cost function c̃ : A → R. Since the n ESPPTWCC subproblems are identical, one only needs to consider
one subproblem and the Lagrangian problem takes the form:

(1.64) z(λ ) = n min{c̃T x | x ∈ BA satisfies conditions (1.53) − (1.57)} + λ (V∗)

The Lagrangian dual problem is defined by:

(1.65) max{z(λ ) | λ ∈ RV∗ }.

Definition 1.6.5 Let the set of routes R be indexed with k = 1, . . . , |R| so Rk is the kth route and define
the cost of the kth route

∑
R
ck = ci j xi jk
(i, j)∈A(Rk )

and the number of times node i ∈ V∗ is visited by the kth route

∑ xi jk
R
aik = for i = 1, . . . , n.
j∈V

The Lagrangian problem (1.64) is then defined by:

(1.66) z(λ ) = n min {ck − aT

k λ } + λ (V∗ )
1≤k≤|R|

and the Lagrangian dual problem is defined by:

(1.67) k λ } + λ (V∗)}.
zLD (R) = max {n min {ck − aT
λ ∈RV∗ 1≤k≤|R|

In (1.67) we have that zLD (R(2) ) ≤ zLD (R(3) ) ≤ · · · ≤ zLD (R(n−1) ) ≤ zLD (R).
Since R is finite it allows us to express (1.67) as the following linear program with many constraints
or rows:

(1.68) max nθ + ∑ λi
i∈V∗
subject to
θ 6 ck − ∑ aik λi for k = 1, . . . , |R|,
i∈V∗
λi ∈ R for i ∈ V∗ ,
θ ∈ R.

16
The LP dual of (1.68) is a linear program with many variables or columns:
|R|
(1.69) min ∑ ck yk
k=1
subject to
|R|
∑ aik yk = 1 for i ∈ V∗ ,
k=1
|R|
∑ yk = n,
k=1
yk > 0 for k = 1, . . . , |R|.

Problem (1.69) with yk required to be integral is equivalent to the VRPTW. This also holds true if
we formulate the linear program with respect to R(k) for some 2 ≤ k ≤ n − 1. Problem (1.69) is the
LP relaxation of the Dantzig-Wolfe decomposition obtained when any solution to the VRPTW is ex-
pressed as a non-negative convex combination of routes, however, because the subproblems are identical
the convexity constraints have been aggregated [53]. The aggregated formulation is equivalent to the
standard set-partitioning formulation of vehicle routing problems [8]. There is a benefit in working with
the decomposed formulation; it does not suffer from the drawback of symmetry present in the original
formulation using the three-index variables xi jk where a given solution can be represented in several ways
by permuting the k indexing. The formulation of the VRPTW presented in this section using three-index
variables have never been used in a branch and bound algorithm. Problem (1.69) have been an important
construct in the formulation of algorithms for the VRPTW but more recently the dual point of view of
(1.68) has also been considered.
Houck, Picard, Queyranne, and Vemuganti [43] presented a relaxation of the symmetric and asymmet-
ric traveling salesman problem based on resource-constrained paths. The resource in their path definition
is the number of arcs contained in the path and the limit on the consumption of this resource is the number
of nodes in the graph n. They called a path containing n arcs an n-path. They called a path elementary if
its nodes are all distinct except possibly the first and last node. If we fix a node 1 of the (directed) graph
an n-path from node 1 to node 1 is a Hamilton tour if and only if it is elementary. The problem of finding
an elementary n-path of minimum cost from node 1 to node 1 is equivalent to the traveling salesman
problem and hence NP-complete. The authors therefore relaxed the condition that the path should be
elementary. If P = (1, i1 , . . . , in−1 , 1) denotes an n-path and there exists a k such that ik = ik+2 for some k
then the path P is said to contain a 2-cycle. It was observed that many n-paths contained 2-cycles. They
proposed a tighter relaxation by forbidding paths containing 2-cycles and presented a dynamic program-
ming algorithm for finding an n-path of minimum cost which does not contain 2-cycles. Houck et al. [43]
also showed that the problem of maximizing the bound derived by this relaxation could be expressed as
linear programs similar to (1.68) and (1.69). They proposed a column generation scheme for solving the
master problem but noted that the column generation was very slow in converging to the optimal solu-
tion. This is a typical observation in early research involving column generation. An American pioneer in
linear programming computing techniques, Orchard-Hays [74, 34, p. 240] said: "Nevertheless, the D-W
(Dantzig-Wolfe) (generalized programming) algorithm presents difficulties, and overall experience with
its use has been mixed. This has led to some disappointment with decomposition algorithms". Houck
et al. [43] then proposed to use the subgradient algorithm following the earlier approach of Held and
Karp [39] and embedded this method in a branch and bound algorithm. The authors made an impor-
tant concluding remark that the n-path relaxation can easily accommodate extra conditions. In fact the
computational work required in the dynamic programming algorithm is just reduced when additional con-
straints are handled. In relation to this observation Picard and Queyranne [77] had previously considered

17
a time-constrained variant of the TSP.
Christofides, Mingozzi, and Toth [11, 12] generalized the n-path relaxation of Houck et al. [43]. They
formulated the capacitated vehicle routing problem using resource-constrained paths where the resource
is the accumulated demand q along the path and called these paths q-paths. They also considered the
set-partitioning formulation of the CVRP similar to (1.69) but instead of solving the linear programming
relaxation of this master problem they proposed a relaxation which could be solved by dynamic pro-
gramming. In this way their method is a two-level dynamic programming approach. In 1987 Kolen,
Rinnooy Kan, and Trienekens [60] introduced the first method for the exact solution of the VRPTW.
Kolen et al. [60] extended the two-level dynamic programming approach of Christofides et al. for the
VRPTW by introducing the accumulated time along the paths as an additional resource. The importance
of this research comes from the introduction of the shortest path problem with time windows and capacity
constraints SPPTWCC, which has played a prominent role in the research on the VRPTW. The method
of Kolen et al. [60] was only capable of solving problems with up to 15 customers. The reason for this
is the relaxation of the master level problem and the use of dynamic programming for solving the master
problem.
The appearance of ‘A new optimization algorithm for the vehicle routing problem with time windows’
(Desrochers, Desrosiers, and Solomon, 1992) in the journal Operations Research was a breakthrough in
the history of the VRPTW and furthermore an important paper in relation to the successful application of
Dantzig-Wolfe decomposition and column generation in general. The method of Desrochers, Desrosiers,
and Solomon is also based on the resource-constrained path formulation of the VRPTW but they used
column generation to solve the linear programming relaxation of the set-partitioning master problem
(1.69). The idea of embedding column generation in a branch and bound algorithm was previously
introduced by Desrosiers, Soumis, and Desrochers [24] for the m-TSP with time windows. Another
important contribution of Desrochers, Desrosiers, and Solomon was the introduction of the set of test
problems developed by Solomon [82]. The introduction of a standard set of test problems is important
because it enables relative evaluation of competing approaches. Of course different authors also need
to consider the same problem variant and adhere to certain conventions with respect to the precision
of problem data. Finally, the Desrochers, Desrosiers, and Solomon paper introduced the label setting
algorithm of Desrochers [21] in the context of the VRPTW for solving the shortest path problem with time
windows and capacity constraints. The authors extended the algorithm to include the 2-cycle elimination
scheme of Houck, Picard, Queyranne, and Vemuganti [43]. Desrochers’ algorithm has been an important
component in the solution of a large class of resource constrained routing and scheduling problems [25]
but as far as we are aware the manuscript ‘An algorithm for the shortest path problem with resource
constraints’ (Desrochers, 1988) has not yet been published in an international journal.
Lagrangian decomposition is an approach that attempts to strengthen the bounds of Lagrangian re-
laxation [36]. This approach splits the original problem into two or more different types of subproblems.
Halse [37] describes three different Lagrangian decompositions of the VRPTW: VS1, VS2, and VS3.
In the VS1 decomposition the VRPTW is formulated using a shortest path problem with time windows
(SPPTW) and a generalized assignment problem (GAP). The VS2 decomposition considers a shortest
path problem with time windows and capacity constraints (SPPTWCC) and a semi-assignment problem
(SAP). Finally, the VS3 decomposition considers an SPPTWCC and a GAP by including vehicle capacity
constraints in both subproblems. The Lagrangian dual problems of the three Lagrangian decomposition
approaches involves more multipliers than in the Lagrangian relaxation of the VRPTW (1.61). In fact,
if n is the number of vehicles then the Lagrangian decompositions requires n times as many multipliers
as the Lagrangian relaxation (1.61). Kohl [57] made an analytical comparison of the bounds provided
by the Lagrangian decompositions and the Lagrangian relaxation and proved that the VS1 and VS2 de-
compositions give the same bound as the Lagrangian relaxation. Furthermore, he also proved that the
VS3 decomposition gives the same bound as the Lagrangian relaxation under the assumptions that the
vehicles are identical and a feasible solution exists for the VRPTW. This proof is non-trivial since the

18
two subproblems SPPTWCC and GAP do not have the integrality property. Since the Lagrangian de-
compositions increase the dimension of the Lagrangian dual problem, but do not strengthen the bound of
the Lagrangian relaxation, the conclusion of the analysis by Kohl [57] is that the Lagrangian decompo-
sitions are not attractive for the identical-vehicle VRPTW. Fisher, Jörnsten, and Madsen [31] presented
computational results based on the VS2 decomposition solving the Lagrangian dual problem using the
subgradient algorithm. Although the Lagrangian decomposition considered by Fisher et al. [31] is not
attractive compared to the Lagrangian relaxation the authors were the first to consider a path formulation
of the VRPTW from the dual point of view. It would later become clear that it was the choice of the
subgradient algorithm that impeded the dual approach.
Kohl and Madsen [58] proposed a method for the VRPTW based on the Lagrangian relaxation (1.61).
They implemented a bundle algorithm of Lemaréchal, Strodiot, and Bihain [67] with an Euclidean steep-
est descent direction finding problem and the line-search of Lemaréchal [66] for determining the step-
size. In relation to line-searches Hiriart-Urruty and Lemaréchal [41, p. 403] mentioned that "the modern
tendency goes towards the so-called trust-region technique". A number of important issues were ad-
dressed by Kohl and Madsen [58]. First, the dimension of the Lagrangian dual problem is smaller in the
Lagrangian relaxation than in the Lagrangian decompositions. Second, the convergence of the bundle
algorithm is better compared to the subgradient algorithm. Third, using a bundle algorithm it is possi-
ble to obtain a primal solution equivalent to the variables of the Dantzig-Wolfe master problem (1.69).
Finally, in the bundle algorithm one can choose a starting point with relatively small multiplier values
and gradually increase the multipliers to the optimal level. The dual approach therefore creates easier
shortest path subproblems because the modified arc costs are less negative, hence less negative cycles are
introduced. Kohl and Madsen [58] only considered problems that required very little branching so a full
computational study of the path formulation from a dual point of view was not performed.
Kohl [57] and Kohl, Desrosiers, Madsen, Solomon, and Soumis [59] addressed the need for improving
the bounds provided by the non-elementary path formulation with 2-cycle elimination. They introduced
the k-path inequalities (1.13) for k = 2 in the Dantzig-Wolfe master problem (1.69). The path formulation
then has exponentially many variables as well as constraints and a column and cut generation approach
was used for solving the master problem. They embedded the column and cut generation in a branch and
bound algorithm and their work is one of the early examples of what became known as branch, price, and
cut algorithms.
Larsen [63] parallelized the branch and bound search in the algorithm of Kohl et al. [59]. Furthermore,
Larsen [63, 64] proposed column deletion and forced early stop for improving solution times. The column
deletion procedure deletes columns in the master problem and is similar to the concept of the bundle
reduction technique used by Kohl and Madsen [58], i.e. to limit the size of the coordinating master
problem. The forced early stop terminates the SPPTWCC pricing algorithm as soon as one path with
negative reduced cost is generated. The forced early stop is motivated from a dual point of view. In
the column generation algorithm the initial master problem must be initialized with a feasible solution.
If this primal solution corresponds to a dual solution with relatively high multiplier values compared
to the optimal level then the difficulty of the SPPTWCC subproblem is relatively higher in the initial
phase of the column generation algorithm. In the first iterations forced early stop therefore gives rapid
improvements of the dual solution cutting down on solution times in the subproblem. Forced early stop
is also called partial pricing whereas solving the subproblem to optimality is called full pricing.
Rich [79] and Cook and Rich [14] extended the work of Kohl et al. [59]. They proposed a new
separation algorithm for the k-path inequalities for k ≥ 2. The authors used the randomized algorithm
given by Karger and Stein [55] that finds α -minimal cuts in undirected graphs in O(n2α log2 n) time
where n is the number of nodes. By setting α = k, sets W for which x(W ) < k are found. If k(W ) ≥ k in
(1.7) then W induces a valid k-path inequality (1.13). Moreover, they parallelized the separation of k-path
inequalities and the branch and bound search.

19
 The bounds provided by the path formulation is also improved if we require the solution to the
resource-constrained shortest path problem to be elementary. Beasley and Christofides [6] proposed
to add a visitation resource for each node v ∈ V∗ with a lower and upper limit of zero and one, respec-
tively. The visitation resource usage when passing through node v is one. This new resource definition
ensures we generate only elementary paths. However, the size of the state space increases dramatically
and Beasley and Christofides [6] expected that this formulation would only be suitable for solving prob-
lems with a small number of resources and made no further studies. The approach of Desrochers et al.
[23] is attractive from a computational point of view because the SPPTWCC can be solved in pseudo-
polynomial time. However, the method has the disadvantage of weakening the lower bound provided
by the path formulation. In an effort to address this issue Feillet, Dejax, Gendreau, and Gueguen [29],
Chabrier [9], and Danna and Le Pape [16] adapted the idea of Beasley and Christofides [6] to Desrochers’
algorithm. The authors also generalized the consumption of the visitation resource. A node v is called
unreachable with respect to a path P if the path already has visited v or there is no way to extend the path
to v due to other resource limitations, i.e. time or capacity. The visitation resources of a path are therefore
consumed either because the nodes have already been visited or because of other resource constraints.
The concept of unreachable nodes is attractive because it sharpens the dominance relation. Chabrier [9],
Danna and Le Pape [16], and Feillet, Gendreau, and Rousseau [28] proposed other heuristic and exact
improvements incorporated in their algorithms for the solution of the elementary shortest path problem.
The developments in the solution of the subproblem are interesting because the VRPTW polytope in the
elementary approach is embedded in a larger polytope but not a polytope over which cT x can be mini-
mized in polynomial time. In fact, the authors propose to solve another NP-complete problem instead of
the VRPTW. In our view this raises the following research question: Can one find an algorithm for the
VRPTW that compute polynomial-time lower bounds that are not dominated by the bounds obtained by
the elementary path formulation? We believe that further investigations of the VRPTW polytope PVRPTW
of Definition 1.2.12 will prove valuable in answering this question.
Kallehauge [50] presented some measurements of the effects of relaxing the condition that paths must
be elementary. The measurements followed a suggestion by Natashia Boland, who raised the following
research question: What are the gaps between the solutions to (1.69) with elementary routes and solutions
to (1.69) with non-elementary routes and 2-cycle elimination for Solomon’s data sets? She suggested that
a good start would be to look at the non-elementary LP solutions that are produced, and check how many
non-elementary paths are assigned non-zero LP values, and measuring the sum of the non-elementary
LP variables over the sum of the LP-variables, i.e. the fraction of the non-elementary paths. Kallehauge
[51] made these measurements for all Solomon’s short-horizon 100 customer problems (excluding C1
problems) and it showed that the problems that remained unsolved have a high fraction of flow on non-
elementary paths. In Table 1.1 we present these measurements, which are made in the root node of the
branch-and-bound tree before inserting any cuts. NE CUSTOMERS is the number of customers that are visited
more than once. NE PATHS/ ALL PATHS is the number of corresponding paths in the LP-solution that contain
cycles compared to the total number of paths, i.e. the number of variables greater than zero. NE FLOW/
TOTAL is the sum of the LP-variables that correspond to paths with cycles and the total sum. %NE FLOW
shows the percentage of the non-elementary compared to the total flow.
Following a suggestion by Jacques Desrosiers, Kallehauge [51] presented the same measurements as
in Table 1.1 after the total flow is integer, i.e. when the number of vehicles is integer. It is also possible
to include e.g. subtour inequalities and 2-path inequalities in the LP-model before branching on vehicles.
Table 1.2 shows these measurements after subtour inequalities and 2-path inequalities are generated for
the LP problem of the root node and we have branched on vehicles. We keep generating these inequalities
as long as we only branch on vehicles. The number of times we branch on vehicles is relatively small
because the number of vehicles quickly becomes integer. If we compare the column %NE FLOW in Table
1.1 with the equivalent column in Table 1.2 we see that the flow on the non-elementary paths is decreased
after inserting cuts and branching on vehicles but it is still significant, i.e. above 20% for all unsolved

20
 NE CUSTOMERS NE PATHS/ ALL PATHS NE FLOW/ TOTAL FLOW %NE FLOW
R101 0 0 / 26 0 / 19.5 0
R102 0 0 / 18 0 / 18.0 0
R103 8 7 / 39 1.22 / 14.06 8.7
R104* 49 41 / 78 4.18 / 10.16 41.1
R105 0 0 / 59 0 / 14.88 0
R106 8 9 / 74 1.37 / 13.00 10.5
R107 23 29 / 78 2.42 / 10.95 22.1
R108* 34 32 / 73 3.63 / 9.82 37.0
R109 14 14 / 78 1.70 / 12.23 14.0
R110 24 25 / 89 2.86 / 10.96 26.1
R111 17 20 / 91 1.87 / 11.43 16.4
R112* 41 36 / 86 2.79 / 9.49 29.4
RC101 0 0 / 60 0 / 14.58 0
RC102 11 10 / 61 1.88 / 13.51 14.0
RC103 20 24 / 72 3.11 / 10.69 29.0
RC104* 53 48 / 85 4.22 / 9.88 42.7
RC105 2 2 / 35 0.40 / 13.53 3.0
RC106* 29 32 / 89 3.42 / 12.34 27.7
RC107* 31 30 / 74 2.87 / 11.48 25.0
RC108* 40 38 / 65 3.58 / 10.73 33.4
* : unsolved instance by September 2000

Table 1.1: Measuring the effects of the non-elementary relaxation in the root node before any inequalities
are generated.

instances.
Table 1.3 shows for each instance the number of k-cycles on the paths with non-zero LP-values in the
optimal solution to the LP problem of the root node. k-CYCLES shows the number of cycles by type: 3-
cycle/4- cycle/5-cycle/etc. It is characteristic for the instances that remained unsolved that a high fraction
of the total number of cycles in the non-elementary LP solution is 3-cycles. At that time we therefore
were interested in trying 3-cycle elimination on these instances [51]. Oli B. G. Madsen [68] presented our
results at a GERAD seminar in Montreal where Stefan Irnich [44] was visiting at that time. Following
this seminar Stefan Irnich [45] generalized the 2-cycle elimination of Houck et al. [43] and Irnich and
Villeneuve [46] further extended this work with detailed computational results available in [47]. Indeed
the k-cycle elimination gave improvements in the lower bounds and was still computationally attractive
as long as k is not too large. Irnich and Villeneuve [46] presented computational results for values of
k = 2, . . . , 5. Irnich and Villeneuve [46] gave an upper bound on the increase in the number of labels.
This bound is (k − 1)!2 for k > 3 so there are limits to how large values of k should be used. For k = 2
the increased number of labels is a factor 2, for k = 3 it is a factor 6.
Kallehauge, Larsen, and Madsen [54] presented a stabilized cutting-plane algorithm for the La-
grangian dual problem. The idea is to force the next dual solution of the cutting-plane algorithm to
be a priori in a ball or trust-region associated with the given norming. The authors use the max-norm so
the master problem is an LP problem with bounds on the dual variables. This is an acceleration of the
cutting-plane algorithm of Kelley [56] and Cheney and Goldstein [10]; the original reference for the col-
umn generation variant is Dantzig and Wolfe [19]. The trust-region ensures stability of the dual solution
from one iteration to the next. Instability refers to the situation where the current iterate is closer (with
respect to some norm) to the optimal solution than the next iterate. Kallehauge et al. [54] was motivated

21
 NE CUSTOMERS NE PATHS/ ALL PATHS NE FLOW/ TOTAL FLOW %NE FLOW
R101 0 0 / 29 0 / 20.00 0
R102 0 0 / 18 0 / 18.00 0
R103 4 4 / 42 1.00 / 15.00 6.7
R104* 45 29 / 81 2.95 / 11.00 26.8
R105 0 0 / 74 0 / 15.00 0
R106 6 7 / 74 1.11 / 13.00 8.5
R107 16 20 / 87 1.69 / 12.00 14.1
R108* 35 35 / 85 2.90 / 10.00 29.0
R109 6 6 / 65 0.85 / 13.00 6.5
R110 23 23 / 92 2.51 / 11.00 22.8
R111 18 19 / 90 1.97 / 12.00 16.4
R112* 40 31 / 91 2.70 / 10.00 27.0
RC101 0 0 / 77 0 / 16.00 0
RC102 12 19 / 97 1.91 / 14.00 13.6
RC103 22 25 / 85 2.97 / 11.00 27.0
RC104* 42 37 / 93 2.93 / 10.00 29.3
RC105 0 0 / 25 0 / 15.00 0
RC106* 28 30 / 99 3.03 / 13.00 23.3
RC107* 37 35 /101 2.85 / 12.00 23.8
RC108* 40 48 / 97 2.95 / 11.00 26.8
* : unsolved instance by September 2000

Table 1.2: Measuring the effects of the non-elementary relaxation after inequalities are generated and
branching on vehicles.

by the work on acceleration strategies at the master problem level by Kohl and Madsen [58]. However,
using the simple max-norm trust-region method the authors avoided solving the quadratic problems of the
version of the bundle algorithm Kohl and Madsen [58] considered and they also avoided the line-searches
associated with the bundle algorithm. To obtain feasible integer solutions the cutting-plane algorithm is
embedded in the branch, cut, and price framework ABACUS [49]. ABACUS is a C++ class library for
solving mixed-integer linear-programs (MILP) by branch, cut, and price. It is interesting to note that the
authors’ formulation of the MILP presented to ABACUS only involves continuous variables (sic!). Obvi-
ously, that is because it is the dual problem (1.68) that is presented to ABACUS. Branching decisions are
then based on the dual variables of the master problem, i.e. the path variables of (1.69). The authors also
introduce inequalities in the master problem. Because the master problem is stated on the dual variables,
subtour and 2-path inequalities are added as columns to this problem. Thienel [83] noted that although
ABACUS was designed for linear programming relaxations there were no reasons that the branch and
bound algorithm of ABACUS was restricted to this type of relaxation. However, Thienel [83] thought
that it would require a generalization of ABACUS to use the system for Lagrangian relaxation. In fact
Kallehauge et al. [54] showed that by remaining within the context of linear programming when solving
the Lagrangian dual problem it is already possible to embed Lagrangian relaxation in this system. The
trust-region method of Kallehauge et al. [54] can be used to solve any Lagrangian dual problem associ-
ated with a hard optimization problem and its implementation in ABACUS would allow the developer to
concentrate on the problem specific parts, i.e. the cutting plane and the column generation, the branching
rules, and the primal heuristics.
Recently Jepsen, Spoorendonk, and Petersen [48] introduced 3-customer clique inequalities that are
valid for the set-partitioning polytope of (1.69). These inequalities change the structure of the shortest

22
 k-CYCLES
R101 0
R102 0
R103 4/1/1/1/1/0/0/1
R104* 29/7/11/10/6/6/8/4/7
R105 0
R106 1/0/1/1/3/3
R107 7/5/1/9/6/5/3/2
R108* 26/2/4/5/2/8/6/8/5
R109 19/0/1
R110 24/4/2/1/2/0/1
R111 10/4/3/0/2/1/2/1
R112* 54/7/14/8/0/0/0/1
RC101 0
RC102 5/2/2/1/2/2/3/1
RC103 18/5/6/3/2/1/2
RC104* 55/9/8/9/5/9/6/7/4
RC105 2
RC106* 36/8
RC107* 38/3/6/3/3/2
RC108* 55/7/3/1/5/1/3/3
* : unsolved instance by September 2000

Table 1.3: Number of k-cycles (3-cycles/4-cycles/5-cycles/etc.) on paths with non-zero LP-values in root
node.

path subproblem and the authors describe how the dominance relation of the subproblem is modified to
incorporate these clique inequalities. This is the first example of strengthening the path formulation by
introducing inequalities defined directly with respect to the path variables of the master problem.

1.7 Conclusions
In this paper we have reviewed four different formulations of the VRPTW and the exact approaches as-
sociated with them. We have identified and organized a total of 24 references on the VRPTW relative
to four seminal papers on formulations of the TSP: arc formulation, arc-node formulation, spanning tree
formulation, and path formulation. Out of these 24 references are 20 references related to the path for-
mulation of the VRPTW. The polyhedral approach of the arc formulation is in our opinion promising and
relatively little research has been conducted along these lines compared to the decomposition approach of
the path formulation. Furthermore, the spanning tree formulation of the VRPTW has not been the subject
of the same amount of research as the path formulation and the extension of the shortest spanning tree
subproblem to time windows has not been considered in the literature. In our view it is at this point there-
fore an open question whether the spanning tree formulation described in this paper is effective compared
to the path formulation.
The exact approaches based on the path formulation has been very successful and the most important
contributions of the research on the VRPTW lies in this area. The developments in the solution of the
subproblem are interesting because the VRPTW polytope in the elementary path formulation approach
is embedded in a larger polytope but not a polytope over which the objective function can be minimized
in polynomial time. In fact, the authors propose to solve another NP-complete problem instead of the

23
VRPTW, i.e. a resource-constrained elementary shortest path problem. In our view this raises the ques-
tion of one can find an algorithm for the VRPTW that compute polynomial-time lower bounds that are
not dominated by the bounds obtained by the elementary path formulation. We believe that further inves-
tigations of the VRPTW polytope will prove valuable in answering this question. The developments in
the solution of the (dual) master problem associated with the path formulation are also very interesting.
There are at least three important developments. First, introduction of strong valid inequalities for the
VRPTW polytope in the master problem, e.g. generalized subtour inequalities. Second, development
of acceleration techniques that addresses the instability issues with the cutting-plane algorithm for con-
vex minimization or equivalently the column generation algorithm of the Dantzig-Wolfe decomposition.
Third, and more recently, strong valid inequalities have been introduced for the set-partitioning polytope
and thereby strengthening the lower bounds provided by this relaxation. The inequalities can also be used
from a dual point of view in the Lagrangian dual problem.
It is clear that ‘A new optimization algorithm for the vehicle routing problem with time windows’
(Desrochers, Desrosiers, and Solomon, 1992) was a very substantial achievement. It is important both for
introducing the path formulation and the column generation algorithm to the VRPTW and for the future
developments it inspired. It is remarkable how much of the scope and methodology of combinatorial op-
timization has been applied in the attack on the VRPTW. The importance of the research described in this
paper comes not from the number of applications where the mathematical model of the VRPTW precisely
fits, but from the fact that the vehicle routing problem with time windows is typical of other resource-
constrained problems in combinatorial optimization.

References
[1] N. Ascheuer, M. Fischetti, and M. Grötschel. A polyhedral study of the asymmetric traveling sales-
man problem with time windows. Networks, 36:69–79, 2000. [7]
[2] N. Ascheuer, M. Fischetti, and M. Grötschel. Solving the asymmetric travelling salesman problem
with time windows by branch-and-cut. Mathematical Programming Series A, 90:475–506, 2001.
[9]
[3] E. Balas, M. Fischetti, and W. R. Pulleyblank. The precedence-constrained asymmetric traveling
salesman polytope. Mathematical Programming, 68:241–265, 1995. [7]
[4] J. F. Bard, G. Kontoravdis, and G. Yu. A branch-and-cut procedure for the vehicle routing problem
with time windows. Transportation Science, 36:250–269, 2002. [4, 9]
[5] J. E. Beasley. Lagrangean relaxation. In C. R. Reeves, editor, Modern Heuristic Techniques for Com-
binatorial Problems, Advanced topics in computer science, pages 243–303. McGraw-Hill, London,
1995. [13]
[6] J. E. Beasley and N. Christofides. An algorithm for the resource constrained shortest path problem.
Networks, 19:379–394, 1989. [20]
[7] M. Bellmore and S. Hong. Transformation of multisalesmen problem to the standard traveling
salesman problem. Journal of the Association for Computing Machinery, 21:500–504, 1974. [6]
[8] J. Bramel and D. Simchi-Levi. Set-covering-based algorithms for the capacitated VRP. In P. Toth
and D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics
and Applications, pages 85–108. SIAM, Philadelphia, 2002. [2, 17]
[9] A. Chabrier. Vehicle routing problem with elementary shortest path based column generation. To
appear in Computers & Operations Research. [4, 20]

24
[10] E. W. Cheney and A. A. Goldstein. Newton’s method for convex programming and tchebycheff
approximation. Numerische Mathematik, 1:253–268, 1959. [21]
[11] N. Christofides, A. Mingozzi, and P. Toth. Exact algorithms for the vehicle routing problem, based
on spanning tree and shortest path relaxations. Mathematical Programming, 20:255–282, 1981. [13,
18]
[12] N. Christofides, A. Mingozzi, and P. Toth. State-space relaxation procedures for the computation of
bounds to routing problems. Networks, 11:145–164, 1981. [18]
[13] COIN/BCP User’s Manual. 2001. [3]
[14] W. Cook and J. L. Rich. A parallel cutting-plane algorithm for the vehicle routing problem with
time windows. Technical report, Rice University, 1999. [4, 19]
[15] J.-F. Cordeau, G. Desaulniers, J. Desrosiers, M. M. Solomon, and F. Soumis. VRP with time win-
dows. In P. Toth and D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete
Mathematics and Applications, pages 157–193. SIAM, Philadelphia, 2002. [2, 4]
[16] E. Danna and C. Le Pape. Branch-and-price heuristics: A case study on the vehicle routing prob-
lem with time windows. In G. Desaulniers, J. Desrosiers, and M. M. Solomon, editors, Column
generation, GERAD 25th Anniversary Series, pages 99–129. Springer, New York, 2005. [4, 20]
[17] G. Dantzig, R. Fulkerson, and S. Johnson. Solution of a large-scale traveling-salesman problems.
Operations Research, 2:393–410, 1954. [2, 4, 7, 8]
[18] G. B. Dantzig and J. H. Ramser. The truck dispatching problem. Management Science, 6:80–91,
1959. [1, 2, 3]
[19] G. B. Dantzig and P. Wolfe. A decomposition principle for linear programs. Operations Research,
8:101–111, 1960. [21]
[20] G. Desaulniers, J. Desrosiers, A. Erdmann, M. M. Solomon, and F. Soumis. VRP with pickup and
delivery. In P. Toth and D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on
Discrete Mathematics and Applications, pages 225–242. SIAM, Philadelphia, 2002. [2]
[21] M. Desrochers. An algorithm for the shortest path problem with resource constraints. Technical
report, GERAD, 1988. [14, 18]
[22] M. Desrochers and G. Laporte. Improvements and extensions to the Miller-Tucker-Zemlin subtour
elimination constraints. Operations Research Letters, 10:27–36, 1991. [8]
[23] M. Desrochers, J. Desrosiers, and M. Solomon. A new optimization algorithm for the vehicle routing
problem with time windows. Operations Research, 40:342–354, 1992. [4, 14, 18, 20, 24]
[24] J. Desrosiers, F. Soumis, and M. Desrochers. Routing with time windows by column generation.
Networks, 14:545–565, 1984. [18]
[25] J. Desrosiers, Y. Dumas, M. M. Solomon, and F. Soumis. Time constrained routing and scheduling.
In M. Ball, T. Magnanti, C. Monma, and G. L. Nemhauser, editors, Network Routing, volume 8
of Handbooks in Operations Research and Management Science, pages 35–139. Elsevier Science
Publishers B.V. (North-Holland), Amsterdam, The Netherlands, 1995. [18]
[26] M. Dror. Note on the complexity of the shortest path models for column generation in VRPTW.
Operations Research, 42:977–978, 1994. [14]

25
[27] H. Everett, III. Generalized lagrange multiplier method for solving problems of optimum allocation
of resources. Operations Research, 11:399–417, 1963. [3]
[28] D. Feillet, M. Gendreau, and L.-M. Rousseau. New refinements for the solution of vehicle routing
problems with branch and price. Submitted. [4, 20]
[29] D. Feillet, P. Dejax, M. Gendreau, and C. Gueguen. An exact algorithm for the elementary shortest
path problem with resource constraints: Application to some vehicle routing problems. Networks,
44:216–229, 2004. [4, 20]
[30] M. L. Fisher. Optimal solution of vehicle routing problems using minimum K-trees. Operations
Research, 42:626–642, 1994. [13]
[31] M. L. Fisher, K. O. Jörnsten, and O. B. G. Madsen. Vehicle routing with time windows: Two
optimization algorithms. Operations Research, 45:488–492, 1997. [4, 12, 13, 19]
[32] M. R. Garey and D. S. Johnson. Computers and Intractability, A Guide to the Theory of NP-
Completeness. Freeman, New York, 1979. [6]
[33] A. M. Geoffrion. Lagrangian relaxation for integer programming. Mathematical Programming
Study, 2:82–114, 1974. [13]
[34] P. C. Gilmore and R. E. Gomory. A linear programming approach to the cutting-stock problem –
part II. Operations Research, 11:863–888, 1963. [13]
[35] M. Grötschel and M. W. Padberg. Polyhedral theory. In E. L. Lawler, J. K. Lenstra, A. H. G.
Rinnooy Kan, and D. B. Shmoys, editors, The Traveling Salesman Problem - A Guided Tour of
Combinatorial Optimization, pages 251–305. Wiley, New York, 1985. [7]
[36] M. Guignard and S. Kim. Lagrangean decomposition: a model yielding stronger Lagrangean
bounds. Mathematical Programming, 39:215–228, 1987. [18]
[37] K. Halse. Modeling and solving complex vehicle routing problems. PhD thesis, Department of
Mathematical Statistics and Operations Research, Technical University of Denmark, 1992. [4, 18]
[38] M. Held and R. M. Karp. The traveling-salesman problem and minimum spanning trees. Operations
Research, 18:1138–1162, 1970. [3, 4, 12, 13]
[39] M. Held and R. M. Karp. The traveling-salesman problem and minimum spanning trees: Part II.
Mathematical Programming, 1:6–25, 1971. [3, 4, 13, 17]
[40] M. Held, P. Wolfe, and P. Crowder. Validation of subgradient optimization. Mathematical Program-
ming, 6:62–88, 1974. [13]
[41] J.-B. Hiriart-Urruty and C. Lemaréchal. Convex Analysis and Minimization Algorithms I-II.
Grundlehren der matematischen Wissenschaften 304-305. Springer-Verlag, Berlin Heidelberg,
1993. [3, 13, 19]
[42] A. J. Hoffman and P. Wolfe. History. In E. L. Lawler, J. K. Lenstra, A. H. G. Rinnooy Kan, and D. B.
Shmoys, editors, The Traveling Salesman Problem - A Guided Tour of Combinatorial Optimization,
pages 1–15. Wiley, New York, 1985. [2]
[43] D. J. Houck, Jr., J. C. Picard, M. Queyranne, and R. R. Vemuganti. The travelling salesman problem
as a constrained shortest path problem: Theory and computational experience. OPSEARCH, 17:
93–109, 1980. [4, 14, 17, 18, 21]

26
[44] S. Irnich. The mixed direct and hub flight problem, a branch-and-price approach for simultaneous
network design and scheduling, September, 2000. GERAD seminar. [21]
[45] S. Irnich. The shortest path problem with k-cycle elimination (k > 3): Improving a branch and price
algorithm for the VRPTW. Technical report, Lehr- und Forschungsgebiet Unternehmensforschung,
Rheinisch-Westfälische Technische Hochschule, November 22, 2000. [21]
[46] S. Irnich and D. Villeneuve. The shortest-path problem with resource constraints and k-cycle elim-
ination for k ≥ 3. To appear in INFORMS Journal on Computing, . [4, 14, 21]
[47] S. Irnich and D. Villeneuve. Online supplement for the shortest-path problem with resource con-
straints and k-cycle elimination for k ≥ 3, .
http://joc.pubs.informs.org/OnlineSupplements.html. [21]
[48] M. Jepsen, S. Spoorendonk, and B. Petersen. A branch-and-cut-and-price framework for VRP
applied on CVRP and VRPTW. Master’s thesis, Department of Computer Science, University of
Copenhagen, 2005. [4, 22]
[49] M. Jünger and S. Thienel. The ABACUS system for branch-and-cut-and-price algorithms in integer
programming and combinatorial optimization. Software: Practice and Experience, 30:1325–1352,
2000. [3, 22]
[50] B. Kallehauge. A hybrid optimal method to the vehicle routing problem with time windows.
In ROUTE 2000 - International workshop on vehicle routing, Skodsborg, Denmark, August 16-
19, 2000. Presentation available at http://www1.ctt.dtu.dk/ROUTE2003/route2000/
presentations/presentation_brian.pdf. [20]
[51] B. Kallehauge. Ideas for solving VRPTW, September 21, 2000. Private communication to Jacques
Desrosiers, Oli B.G. Madsen and Jesper Larsen. [20, 21]
[52] B. Kallehauge, N. Boland, and O. B. G. Madsen. Path inequalities for the vehicle routing problem
with time windows. Submitted. [4, 7, 9]
[53] B. Kallehauge, J. Larsen, O. B. G. Madsen, and M. M. Solomon. Vehicle routing problem with
time windows. In G. Desaulniers, J. Desrosiers, and M. M. Solomon, editors, Column generation,
GERAD 25th Anniversary Series, pages 67–98. Springer, New York, 2005. [4, 17]
[54] B. Kallehauge, J. Larsen, and O. B. G. Madsen. Lagrangian duality applied to the vehicle routing
problem with time windows. Computers & Operations Research, 33:1464–1487, 2006. [4, 13, 21,
22]
[55] D. R. Karger and C. Stein. A new approach to the minimum cut problem. Journal of the Association
for Computing Machinery, 43:601–640, 1996. [19]
[56] J. E. Kelley. The cutting-plane method for solving convex programs. Journal of SIAM, 8:703–712,
1960. [21]
[57] N. Kohl. Exact methods for time constrained routing and related scheduling problems. PhD thesis,
Department of Mathematical Modelling, Technical University of Denmark, 1995. [18, 19]
[58] N. Kohl and O. B. G. Madsen. An optimization algorithm for the vehicle routing problem with time
windows based on Lagrangean relaxation. Operations Research, 45:395–406, 1997. [4, 19, 22]
[59] N. Kohl, J. Desrosiers, O. B. G. Madsen, M. M. Solomon, and F. Soumis. 2-path cuts for the vehicle
routing problem with time windows. Transportation Science, 33:101–116, 1999. [4, 7, 19]

27
[60] A. W. J. Kolen, A. H. G. Rinnooy Kan, and H. W. J. M. Trienekens. Vehicle routing with time
windows. Operations Research, 35:266–273, 1987. [4, 18]
[61] R. V. Kulkarni and P. R. Bhave. Integer programming formulations of vehicle routing problems.
European Journal of Operational Research, 20:58–67, 1985. [8]
[62] G. Laporte, Y. Nobert, and M. Desrochers. Optimal routing under capacity and distance restrictions.
Operations Research, 33:1050–1073, 1985. [3, 7, 13]
[63] J. Larsen. Parallelization of the Vehicle Routing Problem with Time Windows. PhD thesis, Depart-
ment of Mathematical Modelling, Technical University of Denmark, 1999. [4, 19]
[64] J. Larsen. Refinements of the column generation process for the vehicle routing problem with time
windows. Journal of Systems Science and Systems Engineering, 13:326–341, 2004. [4, 19]
[65] E. L. Lawler, J. K. Lenstra, A. H. G. Rinnooy Kan, and D. B. Shmoys, editors. Wiley, New York,
1985. [2]
[66] C. Lemaréchal. A view of line-searches. In A. Auslender, W. Oettli, and J. Stoer, editors, Optimiza-
tion and Optimal Control, volume 30 of Lecture Notes in Control and information Sciences, pages
59–78. Springer, Berlin Heidelberg, 1981. [19]
[67] C. Lemaréchal, J. J. Strodiot, and A. Bihain. On a bundle algorithm for nonsmooth optimization.
In O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, editors, Nonlinear Programming 4, pages
245–282. Academic Press, New York, 1981. [19]
[68] O. B. G. Madsen. Vehicle routing problems with time windows - some recent developments,
September, 2000. GERAD seminar. [21]
[69] V. Mak and A. Ernst. New cutting-planes for the time and/or precedence constrained ATSP and
directed VRP. Submitted. [4, 7]
[70] V. H. Mak. On the Asymmetric Travelling Salesman Problem with Replenishment Arcs. PhD thesis,
Department of Mathematics and Statistics, The University of Melbourne, 2001. [7]
[71] C. E. Miller, A. W. Tucker, and R. A. Zemlin. Integer programming formulation of traveling sales-
man problems. Journal of the Association for Computing Machinery, 7:326–329, 1960. [4, 6, 8]
[72] D. Naddef and G. Rinaldi. Branch-and-cut algorithms for the capacitated VRP. In P. Toth and
D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and
Applications, pages 49–78. SIAM, Philadelphia, 2002. [2, 7]
[73] G. L. Nemhauser and L. A. Wolsey. Integer and Combinatorial Optimization. Wiley, New York,
1988. [3]
[74] W. Orchard-Hays. Advanced Linear-Programming Computing Techniques. McGraw-Hill, New
York, 1968. [17]
[75] M. Padberg and G. Rinaldi. A branch-and-cut algorithm for the resolution of large-scale symmetric
traveling salesman problems. SIAM Review, 33:60–100, 1991. [3]
[76] C. H. Papadimitriou. The complexity of the capacitated tree problem. Networks, 8:217–230, 1978.
[10]
[77] J. C. Picard and M. Queyranne. The time-dependent traveling salesman problem and its application
to the tardiness problem in one-machine scheduling. Operations Research, 26:86–110, 1978. [17]

28
[78] T. Ralphs and M. Guzelsoy. The SYMPHONY callable library for mixed integer programming.
Technical report, Department of Industrial and Systems Engineering, Leheigh University, 2004. [3]
[79] J. L. Rich. A Computational Study of Vehicle Routing Applications. PhD thesis, Rice University,
1999. [19]
[80] A. Schrijver. Combinatorial Optimization, Polyhedra and Efficiency. Algorithms and Combinatorics
24. Springer, Berlin Heidelberg, 2003. [3, 14]
[81] J. F. Shapiro. Mathematical Programming: Structures and Algorithms. Wiley, New York, 1979. [3]
[82] M. M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35:254–265, 1987. [7, 18]
[83] S. Thienel. ABACUS A Branch-And-CUt System. PhD thesis, Mathematisch-Naturwissenschaft-
lichen Fakultät, Universität zu Köln, 1995. [3, 22]
[84] P. Toth and D. Vigo. An exact algorithm for the capacitated shortest spanning arborescence. Annals
of Operations Research, 61:121–141, 1995. [13]
[85] P. Toth and D. Vigo, editors. The Vehicle Routing Problem. SIAM Monographs on Discrete Mathe-
matics and Applications. SIAM, Philadelphia, 2002. [2]
[86] P. Toth and D. Vigo. Branch-and-bound algorithms for the capacitated VRP. In P. Toth and D. Vigo,
editors, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and Applica-
tions, pages 29–51. SIAM, Philadelphia, 2002. [2]
[87] P. Toth and D. Vigo. VRP with backhauls. In P. Toth and D. Vigo, editors, The Vehicle Routing
Problem, SIAM Monographs on Discrete Mathematics and Applications, pages 195–224. SIAM,
Philadelphia, 2002. [2]

29
30
Chapter 2

Lagrangian duality applied to the

vehicle routing problem with time
windows

Brian Kallehauge
Centre for Traffic and Transport, Technical University of Denmark

Jesper Larsen
Informatics and Mathematical Modelling, Technical University of Denmark

Oli B. G. Madsen
Centre for Traffic and Transport, Technical University of Denmark

Abstract
This paper considers the vehicle routing problem with time windows, where the service of each cus-
tomer must start within a specified time interval. We consider the Lagrangian relaxation of the constraint
set requiring that each customer must be served by exactly one vehicle yielding a constrained shortest
path subproblem. We present a stabilized cutting-plane algorithm within the framework of linear pro-
gramming for solving the associated Lagrangian dual problem. This algorithm creates easier constrained
shortest path subproblems because less negative cycles are introduced and it leads to faster multiplier
convergence due to a stabilization of the dual variables. We have embedded the stabilized cutting-plane
algorithm in a branch-and-bound search and introduce strong valid inequalities at the master problem
level by Lagrangian relaxation. The result is a Lagrangian branch-and-cut-and-price (LBCP) algorithm
for the VRPTW. Making use of this acceleration strategy at the master problem level gives a signifi-
cant speed-up compared to algorithms in the literature based on traditional column generation. We have
solved two test problems introduced in 2001 by Gehring and Homberger with 400 and 1000 customers
respectively, which to date are the largest problems ever solved to optimality. We have implemented the
LBCP algorithm using the ABACUS open-source framework for solving mixed-integer linear-programs
by branch, cut, and price.

31
2.1 Introduction
This paper considers the following variant of the vehicle routing problem with time windows (VRPTW):
Consider a directed graph D = (V, A) on n + 2 nodes. The node 0 corresponds to a depot and the node
n + 1 is a copy of node 0. Let the set K correspond to a fleet of identical vehicles available at the depot
source node 0 and each with a capacity q. Nodes i ∈ N = {1, . . . , n} correspond to customers to be served
by a vehicle k ∈ K. With each customer node i ∈ N we associate a demand 0 6 di 6 q, a service time
sti > 0, a release time ai > 0 and a due time bi > ai . The release time ai and the due time bi is the
earliest possible and the latest possible starting time respectively for serving customer i ∈ N. We refer
to the time interval [ai , bi ] as the time window of customer i ∈ N and bi − ai as the width of the time
window. We assume that d0 = st0 = 0 and w.l.o.g that [a0 , b0 ] = [0, ∞). Let A = {(0, j) | j ∈ N} ∪ {(i, j) |
i ∈ N, j ∈ N, ai + ti j 6 b j and di + d j 6 q} ∪ {(i, n + 1) | i ∈ N}. With each arc (i, j) ∈ A, we associate
an arc cost ci j > 0 and an arc travel time ti j > 0, which include any service time sti at node i. We
assume that demands, service times, release times, due times, costs, travel times, and vehicle capacity are
integer values. It is also assumed that the triangle inequality on the costs and travel times is satisfied, i.e.
ci j 6 cih + ch j and ti j 6 tih + th j , for all (i, j) ∈ A. Given a vehicle k ∈ K a path pk consists of the arc set
{(0, v1 )} ∪ {(vi , vi+1 ) | i = 1, . . . , h − 1} ∪ {(vh, n + 1)}, where vi 6= v j , for i 6= j. Such a path represents
the trip of one vehicle k ∈ K leaving the depot source node 0, collecting the demand di of the nodes vi ,
i = 1, . . . , h, and going back to the depot sink node n + 1. We consider the case where a vehicle k ∈ K
waits if it arrives at a node i ∈ N before the release time ai , i.e. given a path pk let svi denote the earliest
start of service of node vi , i = 1, . . . , h, defined as follows.

sv1 = av1 ,
svi = max{svi−1 + tvi−1 ,vi , avi } for i = 2, . . . , h.

This gives a waiting time wvi = max{0, avi − (svi−1 + tvi−1 vi )}, which is positive when a vehicle arrives at
a node before its release time. We allow vehicles to remain at the depot. This is modeled by introducing
the arc (0, n + 1) ∈ A, with c0,n+1 = t0,n+1 = 0 and thereby the "empty path" pk = {(0, n + 1)}, which
represents an unused vehicle k ∈ K. We assume that the number of vehicles used is free. This is modeled
by setting |K| = n or another valid upper bound. A path pk is called feasible with respect to a vehicle
k ∈ K if the total demand ∑hi=1 di of the nodes in pk does not exceed the vehicle capacity q and each node
vi is visited within its time window, i.e. avi 6 svi 6 bvi , for i = 1, . . . , h. Given a vehicle k ∈ K we denote
the set of all feasible paths Pk .
For a subset F of A, D(F) denotes the subdigraph (V (F), F) induced by F, where V (F) is the set of
nodes incident to at least one edge of F.
The cost of a path is the sum of the costs of arcs used in the path, i.e. ∑(i, j)∈pk ci j . An m-path is the
union of m paths p1 , . . . , pm , such that each node i ∈ N belongs to exactly one set N(pk ), k = 1, . . . , m,
and |K| − m paths representing unused vehicles, i.e. m 6 |K| denotes the number of vehicles used to serve
the customers i ∈ N. An m-path is feasible if it consists only of feasible paths. The cost of an m-path is
∑mk=1 ∑(i, j)∈pk ci j . The problem is to find a minimum cost feasible m-path. We would like to note that our
presentation of the VRPTW problem is inspired by Ascheuer et al. [1] and Naddef and Rinaldi [30].
The VRPTW reduces to the VRP if sti = 0, ai = 0, and bi = ∞ for every i ∈ N. Therefore the VRPTW
is NP-hard. Indeed, it is strongly NP-complete to find a feasible solution for the VRPTW with a fixed
number of vehicles [33].
The VRPTW variant we are considering has been attacked by various methods of integer program-
ming. In Cordeau et al. [6] a review of the research until 2000 on exact methods for this variant is given.
The most successful methods in this period are based on the path formulation of Desrochers et al. [10],
involving binary variables associated with feasible paths in the underlying time and capacity constrained
digraph. Kohl and Madsen [23] proposed an equivalent method based on Lagrangian relaxation. Both

32
decomposition methods split the constraints into the same two sets, yielding the same constrained short-
est path subproblem and hence the same lower bound on the VRPTW. The subproblem is NP-hard in the
strong sense [11] and a non-elementary state-space relaxation allowing cycles of length greater than two
is solved. In the Dantzig-Wolfe decomposition method the master problem is a set-partitioning problem.
The linear programming relaxation of the master problem is solved using the column generation algo-
rithm. In the Lagrangian relaxation method the master problem can be formulated as the maximization
of a concave nonsmooth function, piecewise affine in the Lagrangian multipliers [14]. This problem is
denoted the Lagrangian dual problem. Kohl and Madsen [23] solved the Lagrangian dual problem using
a combination of a subgradient algorithm and a bundle algorithm. A bundle algorithm can be viewed as
an acceleration of the column generation algorithm [17, Chapter XII and XV]. Kohl et al. [24] proposed
a strengthening to the path formulation involving constraints for general subtour elimination and Cook
and Rich [5] extended this approach. This strengthened path formulation is exponential in size since it
has exponentially many variables as well as constraints and a column and cut generation approach was
used for solving the problem. Recently Irnich and Villeneuve [19] proposed a strengthening to the bound
provided by the path formulation involving elimination of cycles of length greater than two in the sub-
problem. Algorithms for the elementary shortest path subproblem have also recently been proposed in
Feillet et al. [13] and Chabrier [2].
The research in this paper was motivated by the work on acceleration strategies at the master problem
level by Kohl and Madsen [23]. However we have chosen to remain within the context of linear pro-
gramming and the focus is therefore on accelerating the classical cutting-plane algorithm of Kelley [21]
and Cheney and Goldstein [3]; the original reference for the column generation variant is Dantzig and
Wolfe [7]. In a different application du Merle et al. [12] proposed ways to accelerate the cutting-plane
and column generation algorithm. The algorithm we present creates easier constrained shortest path sub-
problems because less negative cycles are introduced and it leads to faster multiplier convergence due to
a stabilization of the dual variables.
The main contributions of this paper are:
• Developing a stabilized cutting-plane algorithm within the framework of linear programming for
solving the Lagrangian dual problem associated with the Lagrangian relaxation of the assignment
constraints of the VRPTW.
• Embedding the stabilized cutting-plane algorithm in a branch-and-bound algorithm and introducing
strong valid inequalities at the master problem level by Lagrangian relaxation. The result is a
Lagrangian branch-and-cut-and-price (LBCP) algorithm for the VRPTW.
• Decreasing the solution times significantly for a large number of Solomon VRPTW problems com-
pared with a traditional column generation based algorithm.
The paper is organized as follows. In Section 2, we give an integer linear programming (ILP) for-
mulation of the VRPTW. Section 3 presents a Lagrangian relaxation of the VRPTW and shows that the
Lagrangian dual problem can be formulated as a linear programming problem. Section 4 briefly describes
the algorithm used for solving the Lagrangian problem, which can be split into a subproblem for each
vehicle. In Section 5, we present a stabilized cutting-plane algorithm for solving the Lagrangian dual
problem and in Section 6 we present the LBCP algorithm used for finding integer solutions. Section 7
presents computational results. We compare the numerical performance of the stabilized cutting-plane
algorithm to a traditional column generation algorithm and we test the LBCP algorithm on the Solomon
[34] problems and on a number of larger problems created by Gehring and Homberger [15]. Finally,
Section 8 presents our conclusions.

33
2.2 An ILP formulation of the VRPTW
The model involves binary arc variables xi jk as well as integer node variables sik for each vehicle. For a
given vehicle k ∈ K the binary arc variables xi jk for each arc (i, j) ∈ A are defined as follows.

1 (i, j) ∈ A is used by vehicle k in the m-path,
xi jk =
0 otherwise.

The decision variables sik are defined for each node i ∈ N and each vehicle k ∈ K and represent the time
vehicle k starts to service customer i. In case the given vehicle k does not service customer i sik does not
mean anything. The VRPTW is formulated as the following integer linear program:

(2.1) zVRPTW = minimize ∑ ∑ ci j xi jk

k∈K (i, j)∈A

subject to
(2.2) ∑ ∑ xi jk = 1 ∀i ∈ N,
k∈K j∈V

(2.3) ∑ x0 jk = 1 ∀k ∈ K,
j∈V

(2.4) ∑ xihk − ∑ xh jk = 0 ∀h ∈ N, ∀k ∈ K,
i∈V j∈V

(2.5) ∑ xi,n+1,k = 1 ∀k ∈ K,
i∈V
(2.6) ∑ di ∑ xi jk 6 q ∀k ∈ K,
i∈N j∈V

(2.7) sik + ti j − Li j (1 − xi jk ) 6 s jk ∀i, j ∈ N, ∀k ∈ K,

(2.8) ai 6 sik 6 bi ∀i ∈ N, ∀k ∈ K,
(2.9) xi jk ∈ {0, 1} ∀(i, j) ∈ A, ∀k ∈ K,
(2.10) sik ∈ Z+ ∀i ∈ N, ∀k ∈ K,

where Li j = bi − a j and Z+ denotes the set of nonnegative integers.

The objective function (2.1) expresses the total cost. Equalities (2.2) are the assignment constraints
requiring each customer to be served by exactly one vehicle. Equalities (2.3)-(2.5) are the out-degree,
flow balance, and in-degree constraints forcing the solution to consist of a set of paths. Inequalities (2.6)
are the capacity constraints and inequalities (2.7) and (2.8) are used to model the time window restrictions.
Finally, (2.9) and (2.10) are the binary and integer constraints.

2.3 A Lagrangian relaxation of the VRPTW

We consider the Lagrangian relaxation of the VRPTW with respect to the constraints (2.2), by introducing
a vector of Lagrangian multipliers λ = (λ1 , . . . , λn ) ∈ Rn , where λi is associated with the ith constraint in
(2.2):
!
(2.11) zD (λ ) = minimize ∑ ∑ ci j xi jk − ∑ λi ∑ ∑ xi jk − 1 .
subject to k∈K i∈N k∈K j∈V
(2.3)-(2.10) (i, j)∈A

We call (2.11) the Lagrangian problem. The minimal value in the Lagrangian problem (2.11) is called
the dual function and is denoted zD . The set of feasible solutions to the Lagrangian problem (2.11)

34
defined by (2.3)-(2.10) is denoted Pλ . Pλ splits into |K| disjoint subsets, i.e. Pλ = P1 × P2 × · · · × P|K| =
{(p1 , . . . , p|K| ) | p1 ∈ P1 , . . . , p|K| ∈ P|K| }, where each Pk is defined by (2.3)-(2.10) for a given k ∈ K. The
Lagrangian problem (2.11) therefore splits into |K| “simpler” subproblems, one for each vehicle, and the
kth subproblem is then the integer linear program:
(2.12) zk (λ ) = minimize ∑ c̃i j xi j
(i, j)∈A

subject to
(2.13) ∑ x0 j = 1,
j∈V

(2.14) ∑ xih − ∑ xh j = 0 ∀h ∈ N,
i∈V j∈V

(2.15) ∑ xi,n+1 = 1,
i∈V
(2.16) ∑ di ∑ xi j 6 q,
i∈N j∈V

(2.17) si + ti j − Li j (1 − xi j ) 6 s j ∀i, j ∈ N,
(2.18) ai 6 si 6 bi ∀i ∈ N,
(2.19) xi j ∈ {0, 1} ∀i, j ∈ V,
(2.20) si ∈ Z+ ∀i ∈ N,
where c̃i j = ci j − λi for i ∈ N, j ∈ V , c̃i j = ci j otherwise, and admitting that xi j stands for xi jk , but with k
fixed.
Constraints (2.13)-(2.15) force the solution pk ∈ Pk to represent a path starting in the depot source node
0 and ending in the depot sink node n + 1, whereas the capacity constraints (2.16) and the time constraints
(2.17) and (2.18) forbid infeasible paths to be part of the solution. This subproblem (2.12)-(2.20) is an
elementary shortest path problem with time windows and capacity constraints (ESPPTWCC), where each
node can participate at most once in the path pk ∈ Pk . Note that while ci j , (i, j) ∈ A, is a non-negative
integer, c̃i j , i ∈ N, j ∈ V , may be any real number.
The subproblems (2.12)-(2.20) for k ∈ K are identical, which means the Lagrangian problem (2.11)
is expressed as:
 
(2.21) zD (λ ) = |K| minimize ∑ c̃i j xi j + ∑ λi
subject to i∈N
(2.13)-(2.20) (i, j)∈A

Since the subsets Pk for k ∈ K are identical Pλ = P|K| , where P denotes the set of feasible solutions to
the ESPPTWCC subproblem (2.12)-(2.20) for any vehicle k ∈ K. Let z(λ ) denote the minimal solution
value for λ ∈ Rn . We describe each path p ∈ P with the integer variables xi j p , (i, j) ∈ A. Given a path
p ∈ P let c p be the cost of path p ∈ P and let aip be the number of times customer i is served on path p:
cp = ∑ ci j xi j p , for p = 1, . . . , |P|,
(i, j)∈A

aip = ∑ xi j p, for i = 1, . . . , |N|, p = 1, . . . , |P|.

j∈V

Since P is finite, we can consider z(λ ) to be determined by minimization over the set P of constrained
shortest paths. Then (2.11) is expressed as:
 
(2.22) zD (λ ) = |K| minimize c p − ∑ aip λi + ∑ λi
p∈P i∈N i∈N

35
 The fact that zD (λ ) 6 zVRPTW , for λ ∈ Rn , provides us with lower bounds in a branch-and-bound
algorithm for the VRPTW. Clearly, we wish to find the best lower bound by solving the Lagrangian dual
problem:
(2.23) zLD = maximize zD (λ )
λ ∈Rn
 
= maximize |K| minimize c p − ∑ aip λi + ∑ λi .
λ ∈Rn p∈P i∈N i∈N

Since P is finite it allows us to express (2.23) as the following linear program with many constraints
or rows:
(2.24) zLD = maximize |K|θ + ∑ λi
λ ∈R
n
i∈N
θ ∈R
subject to
θ 6 c p − ∑ aip λi for all p ∈ P.
i∈N

The LP dual of (2.24) is a linear program with many variables or columns:

(2.25) zLD = minimize ∑ c p y p
p∈P

subject to
∑ aipy p = 1 for all i ∈ N,
p∈P

∑ y p = |K|,
p∈P

y p > 0 for all p ∈ P.

Problem (2.25) with y p required to be integral is equivalent to the original VRPTW formulation (2.1)-
(2.10). Problem (2.25) is the LP relaxation of the Dantzig-Wolfe decomposition obtained when any
solution to the VRPTW is expressed as a non-negative convex combination of constrained paths [6].
The method of Desrochers et al. [10] can be characterized as column generation on the problem (2.25)
and their formulation of the VRPTW can therefore be viewed as a Dantzig-Wolfe decomposition of the
formulation (2.1)-(2.10). Note that {(0, n + 1)} ∈ P, c0,n+1 = 0, and requiring a fixed number of vehicles
|K| is equivalent with having a free number of vehicles in (2.25).
Let RA be the set of real vectors whose components are indexed by A. Let x = (x1 , . . . , x|K| ), where
xk ∈ RA . Now consider the Lagrangian function or Lagrangian:
 
(2.26) L(λ , x) = ∑ ∑ ci j xi jk + ∑ λi 1 − ∑ ∑ xi jk
k∈K (i, j)∈A i∈N k∈K j∈V
 
= |K| ∑ ci j xi j + ∑ λi 1 − |K| ∑ xi j ,
(i, j)∈A i∈N j∈V

which shows that the Lagrangian is a function of say x1 ∈ RA only. For a fixed solution x p = x1p ∈ P, the
Lagrangian is an affine function in λ :
(2.27) L(λ , x p ) = hs p , λ i + |K|c p ,
where h·, ·i denotes the ordinary dot-product and s p = (s1p , . . . , snp ) ∈ Rn , where:
(2.28) sip = 1 − |K| ∑ xi j p .
j∈V

36
The dual function becomes:

(2.29) zD (λ ) = minimizehs p , λ i + |K|c p,

16p6|P|

which shows that the dual function zD (2.11) is the minimum of a finite number of affine functions and is
therefore piecewise affine and concave [32].
The dual function is non-differentiable or non-smooth at any point λ ∈ Rn where the solution of the
Lagrangian problem (2.11) is not unique. This corresponds to the subproblem (2.12) - (2.20) having
several shortest path solutions.
The subdifferential of the dual function at λ is given by the convex hull of the gradients of the La-
grangian functions that give the minimal value:

(2.30) ∂ zD (λ ) = conv{∇L(λ , x p ) | L(λ , x p ) = zD (λ )}

= conv{s p | hs p , λ i + |K|c p = zD (λ )}

and the elements of the subdifferential are called subgradients.

However, any suboptimal solution of the Lagrangian problem may also be used. Neame et al. [31]
presents an outer approximation of the subdifferential in connection with Lagrangian duality. First define
an index set:

(2.31) PE (λ ) = {p | hs p , λ i + |K|c p 6 zD (λ ) + E, E > 0},

which is the set of solutions to the Lagrangian problem where the solution values are less or equal to the
dual function value plus a positive constant. Then the outer approximation to the subdifferential is:

(2.32) ∂E zD (λ ) = conv{s p | p ∈ PE (λ )}.

2.4 Solving the Lagrangian problem

We compute a solution to the Lagrangian problem (2.11) by solving an ESPPTWCC. However, Dror [11]
proves that the ESPPTWCC is NP-hard in the strong sense, which justifies the approach of considering
the non-elementary state-space relaxation SPPTWCC in the papers by Desrochers et al. [10], Kohl and
Madsen [23] and Kohl et al. [24] on decompositions methods for the VRPTW.
In the SPPTWCC non-simple or non-elementary paths are allowed, i.e. paths including cycles, and
for this problem pseudo-polynomial dynamic programming algorithms are known [9, 8].
In order to tighten the state-space relaxation and increase the lower bound provided by the SPPTWCC
subproblem a standard procedure is to use the 2-cycle elimination scheme proposed by Houck et al. [18],
in which cycles of the form i − j − i are eliminated.
In this paper we also consider the non-elementary relaxation SPPTWCC of the subproblem (2.12)-
(2.20). The SPPTWCC is solved using a label setting algorithm with 2-cycle elimination developed by
Larsen [25]. The algorithm not only computes the shortest path solution, but may also return a number
of non-dominated paths with negative reduced cost in the sink depot.

2.5 Solving the Lagrangian dual problem

We state our Lagrangian dual problem (2.23) as the following maximin problem:

(2.33) zLD = maximize minimize L(λ , x p )

λ ∈Rn 16p6|P|

37
which is equivalent to solving the linear program:

(2.34) zLD = maximize θ

λ ∈Rn
subject to
θ 6 L(λ , x p ) p = 1 . . . , |P|,

Now, since the Lagrangian function (2.27) is an affine function in λ for a fixed solution x p , we have:

(2.35) L(λ , x p ) = zD (λ p ) + hs p, λ − λ pi

so the dual problem (2.34) can be written using only dual objects, i.e. dual function values and subgradi-
ents:

(2.36) zLD = maximize θ

subject to
θ 6 zD (λ p ) + hs p , λ − λ pi for p = 1, . . . , |P|.

The idea of the cutting-plane algorithm is to accumulate the constraints one after the other in (2.36) [17].
Now we will describe the rules that our algorithm follows in order to generate the constraints of (2.36).
Ideally we want to only generate the constraints that intersect at the optimal point. Suppose at iteration u
we have generated the iterates µ1 , . . . , µu . At each iterate µq , q = 1, . . . , u we generate the information:

(2.37) zDp = zDp (µq ), s p = s p ( µq ) p = 1, . . . , |PE (µq )|,

where PE (µq ) is an index set that contains at least one “optimal subgradient” corresponding to an optimal
solution to (2.12)-(2.20) and then any “suboptimal subgradient” corresponding to any suboptimal solution
to (2.12)-(2.20). This approach is equivalent to “multiple pricing” in column generation in which one may
choose not only the non-basic variable with the most negative reduced cost but a set of non-basic variables
with negative reduced costs [4]. Note that it is possible to control the number of subgradients generated
|PE (µq )| by adjusting E but we simply fix the maximum number of returned gradients to the dimension
of µ .
To generate the next sampling point µu+1 we maximize the cutting-plane model of our original prob-
lem (2.34):

(2.38) ẑuLD (µ ) = maximize θ

subject to
θ 6 zDp (µq ) + hs p , µ − µq i for q = 1, . . . , u p = 1, . . . , |PE (µq )|,

but we add bounds on the variables, which is similar to the bounds used by Griffith and Stewart [16]:

(2.39) µ 6 λ u + ∆u ,
µ > λ u − ∆u ,

where λu is the current solution at iteration u, which we refer to as the stability center. Instability refers
to if the current iterate is closer (with respect to some norm) to the optimal solution than the next iterate.
We refer to the parameter ∆u as the trust-region since we assume the cutting-plane approximation of the
dual function is good within this region. The idea of imposing the bounds of Griffith and Stewart [16] is
in a nonlinear programming context used by Madsen [27] in an algorithm for minimizing a function that
is expressed as the maximum of a finite number of explicit nonlinear functions. Madsen [27] also gives
rules for updating ∆u that we will fit to the case of linear functions we are considering. Independently,

38
Marsten et al. [29] also imposes the bounds of Griffith and Stewart [16] in the context of maximizing a
concave function but do not assume that the function considered is explicitly available. However the box
is kept constant, i.e. ∆u is not updated. The authors conclude that the method may find fruitful application
in Lagrangian relaxation and indeed our method can be viewed as an extension of the BOXSTEP method.
When we have computed µu+1 we compute the decrease δu of our model (2.38) and (2.39):

δu = ẑuLD (µu+1 ) − z1D (λu ),

where z1D (λu ) is the current best dual solution value and ẑuLD (µu+1 ) the LP-solution value of (2.38) and
(2.39). Now, if δu is smaller than some user provided tolerance δ we have found an δ -optimal solution.
Marsten et al. [29] shows that in the case of linear functions δ may be zero.
Next we solve the Lagrangian problem with respect to the sampling point µu+1 and get the informa-
tion:

zDp (µu+1 ), s p = s p (µu+1 ), for all p ∈ PE (µu+1 ).

The trust-region parameter is adjusted automatically, depending on the ratio between the actual de-
crease of the dual function and the decrease predicted by the cutting-plane model (2.38) and (2.39):

z1D (µu+1 ) − z1D (λu )

ρ= .
δu
If ρ is equal to 1 then we have just taken a step along one of the pieces of the dual function and no new
information has been included. In this case we increase the size of the trust-region in order to quickly
discover new constraints. If ρ is less than zero we have taken a step into a region where our current
model is not a good approximation of the dual function and we decrease the trust-region in order to
collect additional information in this region.
The algorithm is an ascent method since we use the condition ρ > 0.01 in order to move to the next
sampling point µu+1 , i.e. to take a serious step. If the ascent condition is not satisfied we stay at the
current point, i.e. we take a null-step. The complete algorithm is as follows.

Algorithm 1 (Stabilized cutting-plane Algorithm) Choose an initial point λ1 , a stopping tolerance δ >
0 and a trust-region size ∆1 > 0. Initialize the iteration-counter u = 1 and µ1 = λ1 ; compute zDp (µ1 ) and
s p , for all p ∈ PE (µ1 ).
STEP 1 (Master problem). Solve the following relaxation of the dual problem:

(2.40) ẑuLD (µ ) = maximize θ

subject to
θ 6 zDp (µq ) + hs p, µ − µq i for q = 1, . . . , u p = 1, . . . , |PE (µq )|,
µ 6 λ u + ∆u ,
µ > λ u − ∆u ,

to get a solution µu+1 and compute:

δu = ẑuLD (µu+1 ) − z1D (λu ).

STEP 2 (Stopping criterion). If δu 6 δ then stop.

STEP 3 (Local problem). Find a solution to the Lagrangian problem to get zDp (µu+1 ) and s p , for all
p ∈ PE (µu+1 ).

39
STEP 4 (Update). Compute the gain ratio ρ = (z1D (µu+1 ) − z1D (λu ))/δu .
If ρ = 1 then set ∆u+1 = ∆u ∗ 1.5.
If ρ < 0 then set ∆u+1 = ∆u /1.1.
If 0 < ρ < 1 then set ∆u+1 = ∆u .
If ρ > 0.01 then set λu+1 = µu+1 (ascent-step). Otherwise set λu+1 = λu (null-step).
Set u = u + 1 and go to step 1. 

This type of algorithm is called a trust-region algorithm, introduced by Levenberg [26] and Marquardt
[28] in connection with least squares calculations. For a recent presentation on ways to stabilize the
cutting-plane algorithm we refer to Hiriart-Urruty and Lemaréchal [17, Chapter XV].

2.6 The Lagrangian branch-and-cut-and-price algorithm

In order to find integer solutions we embedded the stabilized cutting-plane algorithm in the open-source
framework ABACUS [20]. ABACUS is a C++ class library for solving mixed-integer linear-programs
by branch, cut, and price. Note that because we are using the dual Lagrangian based decomposition
principle we perform cutting instead of pricing and vice versa compared to the primal Dantzig-Wolfe
decomposition principle. ABACUS provides a general interface to linear programming solvers and we
used the ILOG CPLEX solver (www.ilog.com). The overall solution procedure can be described as
follows. First we solve the root which corresponds to solving the dual problem. If the dual optimum
corresponds to an integer solution then we have found the optimal solution to the VRPTW. Else we start
to generate strong valid inequalities. We generate subtour elimination constraints (SECS) and 2-path cuts
using separation algorithms developed by Kohl [22]. However, the strong valid inequalities are added
as columns in our linear program and the dual variables corresponding to the valid inequalities are also
stabilized. We generate 2-path cuts only in the root node and SECS in all nodes of the branch-and-bound
tree. If the generation of valid inequalities in the root node does not terminate with an integer solution
branching is performed on the number of vehicles and on the xi jk variables of the original formulation
(2.1)-(2.10) [25]. We set the dual multipliers of child nodes to the optimal values from their parent node.
In ABACUS each node in the enumeration tree also inherits the final constraint and variable system of the
father node. This avoids tedious recomputations and is the reason why a cutting-plane method is superior
to a simple subgradient based method. However we modify this system according to the branching rule
on the arc variables since paths can be viewed as locally valid constraints in our method.

2.7 Computational results

The LBCP algorithm presented in this paper is implemented using standard C++ except the SPPTWCC
algorithm developed by Larsen [25] and the separation algorithms for the SECS and 2-path cuts developed
by Kohl [22], which are implemented in standard C. The computational experiments were performed on
two different machines. The hardware and software configuration of the machines is given in Table 2.1.
The LBCP algorithm was tested on the 56 Solomon [34] problems with 100 customers. This set of test
problems was enlarged by only considering the first 25 and 50 customers of each original problem. This
brings the total number of problems up to 168. In addition we have also tested the algorithm on a number
of the problems created by Gehring and Homberger [15], who extended the Solomon problems to sizes of
up to 1000 customers. In the Solomon and Homberger test problems the nodes i = 0, . . . , n are specified by
integer coordinates (x′i , y′i ) in the plane and the vehicle capacity by an integer q. For each node i = 0, . . . , n
the following integer values are given: di , a′i , b′i , and sti′ . In order to fulfill the assumptions stated in this
paper regarding the model parameters in (2.1)-(2.10) we perform the following transformations of the
test problem data [22]. Step 1. Create a copy of node 0 and call it n + 1. Step 2. For i = 0, . . . , n + 1 set

40
 Machine Sun Fire 15K Dell Inspiron 7500
CPU UltraSPARC III Cu 900MHz Intel Pentium III 600MHz
RAM 384Gb 256Mb
Operating system Solaris Microsoft Windows XP
Compiler g++ 2.95 Microsoft Visual C++ 6.0
Compiler options -O NDEBUG, Maximize Speed
ABACUS version 2.4 2.3
ILOG CPLEX version 8.1 7.5

Table 2.1: Hardware and software configuration for the computational experiments.

p
xi = 10 x′i , yi = 10 y′i , ai = 10 a′i , bi = 10 b′i , and sti = 10 sti′ . Step 3. ci j = ⌊ (xi − x j )2 + (yi − y j )2 ⌋ and
ti j = sti + ci j for i, j = 0, . . . , n + 1, i 6= j. Step 4. Add 1 to all ci j for i 6= 0 in order to fulfill the triangle
inequality. In steps 3 and 4 we have for notational convenience assumed that we have a fully connected
digraph on the n + 2 nodes. Clearly in the algorithms we only consider arcs (i, j) ∈ A. The solution value
for the original problem is calculated as (zVRPTW − n)/10, where zVRPTW denotes the solution value for
the transformed problem. Note that we also apply the preprocessing of the time windows described in
Desrochers et al. [10].

2.7.1 Comparison of column generation and stabilized cutting-planes

First we wish to compare the numerical efficiency of the stabilized cutting-plane algorithm to an unsta-
bilized column generation algorithm by solving respectively the Lagrangian dual problem (2.23) and the
relaxed set-partitioning problem (2.25). This corresponds to solving the root node in the branch-and-
bound tree. All computational experiments described in this section were performed on the Sun Fire 15K
machine.
Let B denote the basis of the constraint matrix of (2.25) and cBS the cost coefficients of the basis
variables yBS .
The column generation algorithm is as follows.

Algorithm 2 (Column generation Algorithm) The Dantzig-Wolfe master problem (2.25) is initialized
with a feasible basis. Initialize the column counter r = n, the iteration counter u = 1 and compute the
initial simplex multipliers π = cBS B−1 .

STEP 1 (Subproblem). Return a solution of SPPTWCC (and any negative cost non-dominated path in
the sink depot) with respect to the modified costs c̃i j = ci j − πi to obtain candidate columns y p , with
reduced costs z̄ p , for p ∈ Pu .
STEP 2 (Stopping criterion). If z̄ p > −δ , for p ∈ Pu , then stop (all variables price out correctly). Oth-
erwise adjoin the |Pu | columns with negative reduced cost to the restricted master problem and set
r = r + |Pu |.
STEP 3 (Master problem). Compute a solution to (2.25), i.e. determine a new basis and calculate the
new simplex multipliers π = cBS B−1 , set u = u + 1 and go to step 1. 

Assume we have initialized (2.25) with the paths {(0, i, n + 1)}, ∀i ∈ C, and suppose c p = 10000 for
p = 0, . . . , n − 1, then πi = 10000 for i ∈ C. Figure 2.1 illustrates the effect of the size of the multipliers
on the computational difficulty of the SPPTWCC subproblems. In the Dantzig-Wolfe column generation
algorithm the multipliers are large, compared to the optimal level, in the beginning of the solution process,
while the multipliers in the cutting-plane algorithm are small. As would be expected the number of labels

41
 Labels Stabilized Cutting-Plane Algorithm
100,000
Column Generation Algorithm
80,000

60,000

40,000

20,000

0 u
0 20 40 60 80 100 120

Figure 2.1: Development of the number of labels generated in the SPPTWCC subproblem against the
iteration number when solving the dual problem (2.23) for a 100 customer Solomon problem (R104.100).

generated for solving the subproblems in the two algorithms are almost the same when the optimal level
is reached.
Figure 2.2 illustrates the instability of the column generation algorithm compared to the stabilized
cutting-plane algorithm. We have observed the same behavior as illustrated in figure 2.1 and 2.2 when
solving several other problems. In fact, the total computational time for solving the dual problems for the
R1, C1 and RC1 problems with 100 customers was decreased respectively by a factor 15, 46, and 2 by
using the stabilized cutting-plane algorithm compared to the column generation algorithm. We present
the results in Table 2.2-2.4, which show the following: Name of Solomon problem with suffix indicating
the number of customers (Problem), value of lower bound (LB1), number of times the subproblem is
solved (Iterations), and total CPU-time in seconds to reach the specified stopping tolerance (Seconds).
The stabilized cutting-plane algorithm uses the initial information λ1 = 0, δ = 10−7 and ∆1 = 1. The
maximum number of constrained paths with negative reduced cost returned in each call of the SPPTWCC
algorithm is fixed to the number of customers. For the column generation algorithm we used the stopping
tolerance δ = 10−7 and the maximum number of constrained paths returned after each execution of the
SPPTWCC algorithm was fixed to 100.

2.7.2 Solutions for the Solomon problems

Next we perform computational experiments with the LBCP algorithm on the Solomon problems. Our
experiments were performed on the Sun Fire 15K machine. The reason for this is that we initially experi-
enced extensive memory use in ABACUS 2.3 when the number of subproblems becomes large. ABACUS
2.4 provides a new interface to ILOG CPLEX 8 which solves the memory problem. However currently
this version is available on Unix platforms only. Given the amount of memory available on the Sun Fire
15K machine memory is not a restriction in our experiments. However we restricted the maximal CPU
time to 1 h. Making use of the stabilized cutting-plane algorithm in the branch-and-bound search we were
able with the one hour restriction to solve 119 out of the 168 Solomon test problems. Table 2.5 gives an
overview of the results compared to leading algorithms in the literature. We present a detailed overview
of the results in Table 2.7-2.12 and describe the associated columns in Table 2.6.
Compared to our method the additional R1 and RC1 solutions in the literature are found using exten-
sive parallel branching [5] or 3-cycle elimination [19]. The additional R2 and RC2 solutions are found
using 3-, 4-, or 5-cycle elimination [19] or elementary shortest paths [2]. The C204.100 problem is solved
by Irnich and Villeneuve [19] using 2-cycle elimination. However the computational time is more than

42
 Distance
125,000
100,000
75,000
50,000
25,000
0 Seconds
0 200 400 600 800
(a) Column Generation Algorithm

Distance

1,000

750

500

250

0 Seconds
0 10 20 30 40 50 60 70
(b) Stabilized Cutting-Plane Algorithm

Figure 2.2: The Euclidian distance between the current dual variables and the optimum dual variables.
Observe the different scales.

Column Generation Cutting-Plane

Problem LB1 Iterations Seconds Iterations Seconds
R101.100 1631.150 30 1.36 85 1.08
R102.100 1466.600 48 42.08 102 4.70
R103.100 1206.312 64 183.82 143 20.05
R104.100 949.134 57 782.75 122 67.22
R105.100 1346.142 33 4.95 46 1.47
R106.100 1226.440 50 78.21 81 9.76
R107.100 1051.844 58 1538.29 108 31.38
R108.100 907.162 51 628.50 83 67.35
R109.100 1130.587 42 40.46 87 7.40
R110.100 1048.482 40 150.35 79 22.87
R111.100 1032.028 51 178.84 71 17.45
R112.100 919.192 41 872.47 59 45.56
Total 565 4502.08 1066 296.29

Table 2.2: Comparing the stabilized cutting-plane algorithm to a column generation algorithm on R1
problems with 100 customers.

43
 Column Generation Cutting-Plane
Problem LB1 Iterations Seconds Iterations Seconds
C101.100 827.300 83 8.53 53 0.84
C102.100 827.300 123 45.59 30 1.15
C103.100 826.300 115 172.05 55 5.58
C104.100 822.900 82 1195.66 73 18.50
C105.100 827.300 92 9.61 47 0.93
C106.100 827.300 85 7.77 34 0.81
C107.100 827.300 113 18.59 32 0.71
C108.100 827.300 92 16.09 47 1.44
C109.100 825.640 55 17.50 42 2.20
Total 840 1491.39 413 32.16

Table 2.3: Comparing the stabilized cutting-plane algorithm to a column generation algorithm on C1
problems with 100 customers.

Column Generation Cutting-Plane

Problem LB1 Iterations Seconds Iterations Seconds
RC101.100 1584.094 31 1.92 50 1.13
RC102.100 1403.646 37 13.82 98 7.23
RC103.100 1218.495 44 50.46 72 19.76
RC104.100 1094.333 47 321.49 112 148.09
RC105.100 1471.160 38 6.94 106 6.82
RC106.100 1308.781 31 7.77 81 6.96
RC107.100 1170.689 35 42.67 58 21.23
RC108.100 1063.011 44 133.77 78 77.44
Total 307 578.84 655 288.66

Table 2.4: Comparing the stabilized cutting-plane algorithm to a column generation algorithm on RC1
problems with 100 customers.

Author R1 C1 RC1 Type 1 R2 C2 RC2 Type 2 Total

Cook and Rich [5] 33 27 20 80 8 20 2 30 110
Irnich and Villeneuve [19] 29 27 20 76 21 24 14 59 135
Chabrier [2] 19 24 17 60
This paper 28 27 18 73 17 23 6 46 119
Total solved 34 27 22 83 21 24 17 62 145
Total number of problems 36 27 24 87 33 24 24 81 168

Table 2.5: The number of Solomon problems solved.

44
 Problem Name of Solomon problem with suffix indicating number of customers.
LB1 Value of root node before any strong valid inequalities are generated. For some problems we were not
able to solve the root node to optimality. This is indicated by a blank entry.
LB2 Value of root node after strong valid inequalities are generated. Only printed if LB2>LB1.
LB3 Value of global lower bound in the search tree when optimization was terminated. Only printed if
LB3>max{LB1, LB2}. If the problem is solved max{LB1, LB2, LB3 } is equal to the optimal integer
solution value.
Vehicles Number of vehicles corresponding to UB.
B&B nodes Number of selected nodes in the branch-and-bound tree.
Cuts Number of generated strong valid inequalities in the root node and the search tree.
Iterations Number of times the SPPTWCC algorithm is executed.
Seconds Total CPU-time in seconds to solve problem to optimality (-*: time limit of 3600 CPU seconds exceeded).

Table 2.6: Description of columns in Table 2.7-2.12 and 2.17.

10 h. Compared to Irnich and Villeneuve [19] and Chabrier [2] the conclusion is that our algorithm is
not competitive with respect to the R2 and RC2 problems, since the bound provided by the SPPTWCC
with 2-cycle elimination is not tight enough. However compared to Cook and Rich [5] the acceleration
strategy made it possible to solve more and larger problems in the R2 and RC2 sets.
Irnich and Villeneuve [19] performed all their computational experiments on a Pentium III 600MHz
machine with 512Mb RAM. The Dell Inspiron 7500 machine is similar except that it only has 256Mb
RAM. We should therefore be able to compare computational times for experiments performed on the
Dell Inspiron 7500 with the results reported in Irnich and Villeneuve [19]. We tried to solve the 119
problems on the Dell Inspiron 7500 machine. However because of lack of memory we solved only 117
out of the 119 problems (R112.50 and R205.50 were not solved); 91 out of the 117 solved problems
were solved faster than the minimum time reported in Irnich and Villeneuve [19] corresponding to a total
decrease of 20437 seconds; 26 were solved with a longer computational time corresponding to a total
increase of 940 seconds. Note that Irnich and Villeneuve [19] report solution times using 2-, 3-, and
4-cycle elimination. We compare our method with the k-cycle elimination method with the minimum
solution time. In Table 2.13 we compare the total solution times per problem set for the LBCP algorithm
with Irnich and Villeneuve [19]. In Table 2.15 and 2.16 we show the 10 problems for which we observed
respectively the largest increase and decrease in the solution time by our method compared to the minimal
solution time reported in Irnich and Villeneuve [19]. We describe the associated columns in Table 2.14.

2.7.3 Solutions for the Homberger problems

We have solved 9 problems from the Homberger test sets introduced in Gehring and Homberger [15],
among them problems with 400 and 1000 customers. In 8 of the problems the customers are clustered
(C-problems), while we succeeded in solving a 200 customer problem where the customers are randomly
located. The results are presented in Table 2.17.

2.8 Conclusions
The algorithm has been tested on the Solomon VRPTW test problems and a range of extended Solomon
problems created by Homberger. Using a stabilized cutting-plane algorithm in a branch-and-bound
scheme gives a significant speed-up compared to an algorithm based on traditional column generation.
We have solved two Homberger problem with 400 and 1000 customers respectively, which to date is the
largest problems ever solved to optimality. The conclusion is therefore that it is an efficient acceleration
strategy that performs significantly better than a traditional column generation based algorithm on a large
number of test problems.

45
Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
R101.25 617.100 8 1 0 30 0.1
R101.50 1043.367 1044.000 12 1 2 43 0.2
R101.100 1631.150 1634.000 1637.700 20 15 9 385 8.4
R102.25 546.333 547.100 7 1 3 32 0.1
R102.50 909.000 11 1 0 53 0.5
R102.100 1466.600 18 1 0 99 4.8
R103.25 454.600 5 1 0 39 0.2
R103.50 765.950 767.300 772.900 9 57 7 516 10.0
R103.100 1206.312 1206.376 1208.700 14 69 3 821 123.0
R104.25 416.900 4 1 0 32 0.2
R104.50 616.500 620.758 625.400 6 103 13 1129 303.9
R104.100 949.134 950.987 956.496 100 158 10 4057 -*
R105.25 530.500 6 1 0 24 0.1
R105.50 892.120 893.650 899.300 9 23 15 181 1.9
R105.100 1346.142 1348.632 1355.300 15 151 23 1573 102.9
R106.25 457.300 465.400 5 1 15 41 0.3
R106.50 791.367 793.000 8 1 7 57 0.9
R106.100 1226.440 1227.404 1234.600 13 1457 11 15173 2187.5
R107.25 422.925 423.800 424.300 4 3 3 41 0.3
R107.50 704.438 704.814 711.100 7 83 3 776 21.5
R107.100 1051.844 1052.714 1061.883 100 421 12 9568 -*
R108.25 396.139 396.720 397.300 4 3 2 61 0.5
R108.50 588.926 595.624 611.785 50 209 33 4854 -*
R108.100 907.162 910.603 915.265 100 234 17 4927 -*
R109.25 441.300 5 1 0 21 0.1
R109.50 775.096 775.890 786.800 8 247 7 1904 25.8
R109.100 1130.587 1133.164 1141.979 100 1011 34 26904 -*
R110.25 437.300 437.938 444.100 5 25 5 263 1.4
R110.50 692.577 694.150 697.000 7 5 4 83 2.2
R110.100 1048.482 1049.939 1058.607 100 491 8 9892 -*
R111.25 423.788 424.583 428.800 4 5 3 83 0.5
R111.50 691.812 692.635 707.200 7 461 14 4348 114.5
R111.100 1032.028 1041.850 100 528 6 10646 -*
R112.25 384.200 385.391 393.000 4 13 25 208 6.6
R112.50 607.219 612.374 630.200 6 5263 36 65983 3166.4
R112.100 919.192 922.398 930.128 100 171 41 4063 -*

Table 2.7: Solution overview for the R1 problems.

46
Problem LB1 LB2 Vehicles B&B nodes Cuts Iterations Seconds
C101.25 191.300 3 1 0 24 0.1
C101.50 362.400 5 1 0 19 0.2
C101.100 827.300 10 1 0 50 0.8
C102.25 190.300 3 1 0 20 0.2
C102.50 361.400 5 1 0 38 0.5
C102.100 827.300 10 1 0 27 1.2
C103.25 190.300 3 1 0 26 0.2
C103.50 361.400 5 1 0 39 1.1
C103.100 826.300 10 1 0 52 5.7
C104.25 186.900 3 1 0 27 0.4
C104.50 357.250 358.000 5 1 3 67 7.3
C104.100 822.900 10 1 0 70 18.9
C105.25 191.300 3 1 0 26 0.1
C105.50 362.400 5 1 0 18 0.2
C105.100 827.300 10 1 0 44 0.9
C106.25 191.300 3 1 0 25 0.1
C106.50 362.400 5 1 0 28 0.2
C106.100 827.300 10 1 0 31 0.8
C107.25 191.300 3 1 0 24 0.1
C107.50 362.400 5 1 0 22 0.2
C107.100 827.300 10 1 0 29 0.7
C108.25 191.300 3 1 0 26 0.2
C108.50 362.400 5 1 0 34 0.3
C108.100 827.300 10 1 0 44 1.6
C109.25 191.300 3 1 0 25 0.2
C109.50 362.400 5 1 0 27 0.4
C109.100 825.640 827.300 10 1 3 104 6.1

Table 2.8: Solution overview for the C1 problems.

47
Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
RC101.25 406.625 461.100 4 1 9 53 0.2
RC101.50 850.021 942.625 944.000 8 3 38 99 1.9
RC101.100 1584.094 1617.276 1619.800 15 27 68 340 32.9
RC102.25 351.800 3 1 0 22 0.2
RC102.50 719.902 813.037 822.500 7 683 12 6804 115.0
RC102.100 1403.646 1437.000 1450.327 100 1152 40 30698 -*
RC103.25 332.050 332.800 3 3 0 42 0.4
RC103.50 643.133 710.667 710.900 6 5 10 118 4.3
RC103.100 1218.495 1241.705 1250.706 100 423 44 12506 -*
RC104.25 305.825 306.600 3 7 0 75 0.7
RC104.50 543.750 545.800 5 17 5 247 18.4
RC104.100 1094.333 1112.354 100 3 32 360 -*
RC105.25 410.950 411.300 4 3 0 46 0.3
RC105.50 754.443 852.858 855.300 8 33 23 441 7.3
RC105.100 1471.160 1509.800 1513.700 15 41 34 686 65.6
RC106.25 342.829 343.200 345.500 3 13 1 135 0.6
RC106.50 664.433 714.788 723.200 6 37 12 532 11.9
RC106.100 1308.781 1332.510 1347.342 100 927 46 22511 -*
RC107.25 298.300 3 1 0 40 0.4
RC107.50 591.476 632.336 642.700 6 123 7 2095 193.6
RC107.100 1170.689 1178.484 1193.337 100 392 39 8850 -*
RC108.25 293.791 294.500 3 1 5 48 0.9
RC108.50 538.957 596.867 598.100 6 9 10 144 29.8
RC108.100 1063.011 1091.555 1096.177 100 95 52 2319 -*

Table 2.9: Solution overview for the RC1 problems.

48
Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
R201.25 460.100 463.300 4 3 0 45 0.2
R201.50 788.425 791.900 6 1 1 67 0.8
R201.100 1136.248 1138.650 1143.200 8 183 2 3000 253.4
R202.25 406.350 408.350 410.500 4 5 1 87 0.8
R202.50 692.737 696.525 698.500 5 11 2 291 11.4
R202.100 1009.828 1009.859 1012.776 100 46 12 2143 -*
R203.25 379.882 381.625 391.400 3 37 3 426 5.4
R203.50 590.930 593.430 605.300 5 491 20 7906 923.4
R203.100 846.489 847.097 847.379 100 6 6 507 -*
R204.25 333.075 335.350 355.000 2 777 42 11052 190.6
R204.50 474.562 482.324 487.949 50 64 38 3874 -*
R204.100 100 1 0 378 -*
R205.25 381.283 388.425 393.000 3 15 4 205 1.1
R205.50 666.604 672.350 690.100 4 5255 34 98061 2558.7
R205.100 916.976 923.025 931.290 100 179 34 9884 -*
R206.25 363.132 365.908 374.400 3 85 12 946 14.2
R206.50 609.590 611.363 624.260 50 617 54 25518 -*
R206.100 835.326 840.751 844.619 100 16 12 1170 -*
R207.25 347.592 349.741 361.600 3 125 20 1568 30.2
R207.50 539.067 544.324 554.200 50 194 44 8821 -*
R207.100 100 1 13 588 -*
R208.25 318.105 318.911 328.200 1 75 24 1531 58.2
R208.50 462.412 471.563 472.779 50 14 30 1184 -*
R208.100 100 1 0 222 -*
R209.25 353.875 358.321 370.700 2 65 8 781 6.5
R209.50 582.926 588.413 600.600 4 525 49 9077 530.4
R209.100 819.847 823.734 829.175 100 45 28 2582 -*
R210.25 395.844 397.906 404.600 3 71 8 760 7.5
R210.50 624.421 638.524 50 1086 67 35729 -*
R210.100 849.477 855.852 860.328 100 21 17 1231 -*
R211.25 330.140 330.472 350.900 2 1513 84 23178 515.3
R211.50 507.950 512.567 523.607 50 235 58 7923 -*
R211.100 705.811 710.753 100 3 25 401 -*

Table 2.10: Solution overview for the R2 problems.

49
Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
C201.25 214.700 2 1 0 48 0.2
C201.50 360.200 3 1 0 103 1.1
C201.100 589.100 3 1 0 40 2.0
C202.25 214.700 2 1 0 139 1.3
C202.50 360.200 3 1 0 574 20.3
C202.100 589.100 3 1 0 32 5.7
C203.25 214.700 2 1 0 102 5.7
C203.50 359.800 3 1 0 808 203.1
C203.100 588.700 3 1 0 77 73.9
C204.25 211.004 211.042 213.100 1 13 4 364 46.0
C204.50 350.100 2 1 0 471 402.2
C204.100 100 4 1 277 -*
C205.25 212.050 214.700 2 1 3 86 0.4
C205.50 357.350 359.000 359.800 3 3 1 400 7.2
C205.100 586.400 3 1 0 47 4.6
C206.25 197.700 214.700 2 1 6 112 0.9
C206.50 344.200 359.000 359.800 3 5 5 989 33.9
C206.100 585.400 586.000 3 1 2 64 14.9
C207.25 207.981 214.400 214.500 2 7 12 208 3.7
C207.50 356.269 359.600 3 1 13 236 22.3
C207.100 581.969 585.800 3 1 6 429 74.6
C208.25 193.280 214.500 2 1 10 165 1.8
C208.50 340.425 350.500 2 1 4 190 8.4
C208.100 581.767 585.800 3 1 4 159 57.0

Table 2.11: Solution overview for the C2 problems.

50
 Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
RC201.25 356.650 360.200 3 1 1 39 0.2
RC201.50 670.150 681.983 684.900 5 15 5 285 3.0
RC201.100 1240.398 1253.430 1261.800 9 679 18 16461 1205.9
RC202.25 290.408 313.389 338.000 3 665 30 12772 222.3
RC202.50 504.140 548.224 572.352 50 705 80 42784 -*
RC202.100 1004.398 1013.888 1028.253 100 126 30 6683 -*
RC203.25 214.475 260.811 295.866 25 984 111 49000 -*
RC203.50 409.246 480.052 486.953 50 23 42 2942 -*
RC203.100 815.276 831.971 834.207 100 7 38 714 -*
RC204.25 188.593 244.810 260.494 25 127 39 8907 -*
RC204.50 50 1 0 199 -*
RC204.100 100 1 0 196 -*
RC205.25 307.600 320.788 338.000 3 23 6 791 6.9
RC205.50 541.592 579.905 617.816 50 1093 59 84589 -*
RC205.100 1056.111 1070.859 1087.132 100 162 32 10709 -*
RC206.25 250.106 288.983 324.000 3 503 39 14526 195.5
RC206.50 441.336 532.192 551.372 50 375 73 35262 -*
RC206.100 952.406 982.887 995.911 100 126 40 7947 -*
RC207.25 217.961 263.894 292.133 3 2427 98 123256 -*
RC207.50 390.837 468.865 475.274 50 15 38 1693 -*
RC207.100 866.668 877.849 888.931 100 47 36 2988 -*
RC208.25 169.671 233.078 243.122 25 56 34 4095 -*
RC208.50 50 1 21 408 -*
RC208.100 100 1 13 389 -*

Table 2.12: Solution overview for the RC2 problems.

Author R1 C1 RC1 Type 1 R2 C2 RC2 Type 2 Total

Irnich and Villeneuve [19] 3783.1 1011.8 2651.1 7446.0 5847.1 10146.9 3673.2 19667.2 27113.2
This paper 2550.0 48.2 397.1 2995.3 1981.0 1155.6 1484.5 4621.2 7616.4
Speed-up 1.5 21.0 6.7 2.5 3.0 8.8 2.5 4.3 3.6

Table 2.13: Total solution time in seconds for each problem set for the 117 problems solved by the LBCP
algorithm on the Dell Inspiron 7500 vs. the corresponding minimum solution time reported by Irnich and
Villeneuve [12].

Problem Name of Solomon problem with suffix indicating number of customers.

LB1 Value of root node before any strong valid inequalities are generated.
LB2 Value of root node after strong valid inequalities are generated.
B&B nodes Number of selected nodes in the branch-and-bound tree.
k k-cycle elimination used in obtaining minimum the solution time.
Seconds Total CPU-time in seconds to solve problem to optimality.

Table 2.14: Description of columns in Table 2.15 and 2.16.

51
 Irnich and Villeneuve [19] LBCP Algorithm
Problem LB1 LB2 B&B nodes k Seconds LB1 LB2 B&B nodes Seconds Increase
R104.50 616.500 620.758 34 2 212.6 616.500 620.758 117 406.9 194.3
R203.50 598.350 598.350 23 3 470.4 590.930 593.430 479 661.7 191.3
RC202.25 308.033 321.425 7 3 12.6 290.408 313.389 741 196.6 184.0
RC206.25 318.950 318.950 9 4 12.3 250.163 288.983 517 194.2 181.9
R209.50 599.813 599.813 7 4 255.4 582.877 588.413 531 421.1 165.7
R206.25 373.600 373.600 7 3 5.1 363.132 365.908 85 15.7 10.6
RC205.25 338.000 338.000 1 3 1.5 307.600 320.788 23 6.5 5.0
C203.25 214.700 214.700 1 2 3.7 214.700 214.700 1 4.9 1.2
RC106.25 345.500 345.500 1 3 0.4 342.829 343.200 13 1.2 0.8
R101.25 617.100 617.100 1 2 0.1 617.100 617.100 1 0.7 0.6

Table 2.15: Top 10 increases in computational time using the LBCP algorithm compared to Irnich and
Villeneuve [12].

Irnich and Villeneuve [19] LBCP Algorithm

Problem LB1 LB2 B&B nodes k Seconds LB1 LB2 B&B nodes Seconds Decrease
R201.100 1140.300 1140.300 35 4 3537.2 1136.222 1138.650 203 281.6 3255.6
RC201.100 1255.770 1256.260 39 4 3620.4 1240.398 1253.430 679 1082.7 2537.7
C208.100 581.767 585.800 2 2 2183.3 581.767 585.800 1 37.4 2145.9
C207.100 581.969 585.800 5 2 2068.8 581.969 585.800 1 52.1 2016.7
C203.100 588.700 588.700 1 2 1706.3 588.967 588.967 7 509.0 1197.3
RC105.100 1471.160 1509.800 25 2 899.0 1471.160 1509.800 41 56.4 842.6
C204.50 350.100 350.100 1 2 1159.4 350.100 350.100 1 332.8 826.6
C206.100 586.000 586.000 1 3 814.4 585.400 586.000 1 11.6 802.8
C202.100 589.100 589.100 1 2 585.6 589.100 589.100 1 6.5 579.1
R211.25 339.981 339.981 139 4 876.3 330.140 330.477 1127 335.3 541.0

Table 2.16: Top 10 decreases in computational time using the LBCP algorithm compared to Irnich and
Villeneuve [12].

Problem LB1 LB2 LB3 Vehicles B&B nodes Cuts Iterations Seconds
R1_2_1.200 4654.913 4661.033 4667.200 23 521 21 7980 1363.39
C1_2_1.200 2698.600 20 1 0 127 7.7
C1_2_2.200 2682.187 2694.300 20 75 6 2816 992.61
C1_2_5.200 2694.900 20 1 0 112 9.64
C1_2_6.200 2694.900 20 1 0 126 14.15
C1_2_7.200 2694.900 20 1 0 105 11.18
C1_2_8.200 2667.870 2668.118 2684.000 20 137 19 2409 601.15
C1_4_1.400 7138.767 7138.800 40 1 1 97 28.59
C110_1.1000 42444.400 42444.683 42444.800 100 5 3 858 1298.21

Table 2.17: Overview of the solved Homberger problems.

52
Acknowledgements The authors would like to thank Professor Kaj Madsen from the department of In-
formatics and Mathematical Modelling, Technical University of Denmark, for all the fruitful discussions
regarding the development of Algorithm 1.

References
[1] N. Ascheuer, M. Fischetti, and M. Grötschel. Solving the asymmetric travelling salesman problem
with time windows by branch-and-cut. Mathematical Programming Series A, 90:475–506, 2001.
[32]
[2] A. Chabrier. Vehicle routing problem with elementary shortest path based column generation. Tech-
nical report, ILOG, 2003. [33, 42, 44, 45]
[3] E. W. Cheney and A. A. Goldstein. Newton’s method for convex programming and tchebycheff
approximation. Numerische Mathematik, 1:253–268, 1959. [33]
[4] V. Chvátal. Linear Programming. Freeman, New York, 1983. [38]
[5] W. Cook and J. L. Rich. A parallel cutting-plane algorithm for the vehicle routing problem with
time windows. Technical report, Rice University, 1999. [33, 42, 44, 45]
[6] J.-F. Cordeau, G. Desaulniers, J. Desrosiers, M. M. Solomon, and F. Soumis. VRP with time win-
dows. In P. Toth and D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete
Mathematics and Applications, pages 157–193. SIAM, Philadelphia, 2002. [32, 36]
[7] G. B. Dantzig and P. Wolfe. A decomposition principle for linear programs. Operations Research,
8:101–111, 1960. [33]
[8] M. Desrochers. An algorithm for the shortest path problem with resource constraints. Technical
report, GERAD, 1988. [37]
[9] M. Desrochers and F. Soumis. A generalized permanent labelling algorithm for the shortest path
problem with time windows. INFOR, 26:191–211, 1988. [37]
[10] M. Desrochers, J. Desrosiers, and M. Solomon. A new optimization algorithm for the vehicle routing
problem with time windows. Operations Research, 40:342–354, 1992. [32, 36, 37, 41]
[11] M. Dror. Note on the complexity of the shortest path models for column generation in VRPTW.
Operations Research, 42:977–978, 1994. [33, 37]
[12] O. du Merle, D. Villeneuve, J. Desrosiers, and P. Hansen. Stabilized column generation. Discrete
Mathematics, 194:229–237, 1999. [33]
[13] D. Feillet, P. Dejax, M. Gendreau, and C. Gueguen. An exact algorithm for the elementary shortest
path problem with resource constraints: Application to some vehicle routing problems. Technical
report, Laboratoire d’Informatique d’Avignon, Université d’Avignon, 2004. [33]
[14] M. L. Fisher. The Lagrangian relaxation method for solving integer programming problems. Man-
agement Science, 27:1–18, 1981. [33]
[15] H. Gehring and J. Homberger. A parallel two-phase metaheuristic for routing problems with time
windows. Asia-Pacific Journal of Operational Research, 18:35–47, 2001. [33, 40, 45]

53
[16] R. E. Griffith and R. A. Stewart. A nonlinear programming technique for the optimization of con-
tinuous processing systems. Management Science, 7:379–392, 1961. [38, 39]
[17] J.-B. Hiriart-Urruty and C. Lemaréchal. Convex Analysis and Minimization Algorithms I-II.
Grundlehren der matematischen Wissenschaften 304-305. Springer-Verlag, Berlin Heidelberg,
1993. [33, 38, 40]
[18] D. J. Houck, J. C. Picard, M. Queyranne, and R. R. Vemuganti. The travelling salesman problem
as a constrained shortest path problem: Theory and computational experience. OPSEARCH, 17:
93–109, 1980. [37]
[19] S. Irnich and D. Villeneuve. The shortest path problem with k-cycle elimination (k > 3): Improving
a branch-and-price algorithm for the VRPTW. Technical report, Lehr- und Forschungsgebiet Oper-
ations Research und Logistik Management, Rheinisch-Westfälische Technische Hochschule, 2003.
[33, 42, 44, 45, 51, 52]
[20] M. Jünger and S. Thienel. The ABACUS system for branch-and-cut-and-price algorithms in integer
programming and combinatorial optimization. Software: Practice and Experience, 30:1325–1352,
2000. [40]
[21] J. E. Kelley. The cutting-plane method for solving convex programs. Journal of SIAM, 8:703–712,
1960. [33]
[22] N. Kohl. Exact methods for Time Constrained Routing and Related Scheduling Problems. PhD
thesis, Department of Mathematical Modelling, Technical University of Denmark, 1995. [40]
[23] N. Kohl and O. B. G. Madsen. An optimization algorithm for the vehicle routing problem with time
windows based on Lagrangean relaxation. Operations Research, 45:395–406, 1997. [32, 33, 37]
[24] N. Kohl, J. Desrosiers, O. B. G. Madsen, M. M. Solomon, and F. Soumis. 2-path cuts for the vehicle
routing problem with time windows. Transportation Science, 33:101–116, 1999. [33, 37]
[25] J. Larsen. Parallelization of the Vehicle Routing Problem with Time Windows. PhD thesis, Depart-
ment of Mathematical Modelling, Technical University of Denmark, 1999. [37, 40]
[26] K. Levenberg. A method for the solution of certain problems in least squares. Quarterly Applied
Mathematics, 2:164–168, 1944. [40]
[27] K. Madsen. An algorithm for minimax solution of overdetermined systems of non-linear equations.
Journal of the Institute of Mathematics and Its Applications, 16:321–328, 1975. [38]
[28] D. Marquardt. An algorithm for least squares estimation of nonlinear parameters. SIAM Journal of
Applied Mathematics, 11:431–441, 1963. [40]
[29] R. E. Marsten, W. W. Hogan, and J. W. Blankenship. The BOXSTEP method for large-scale opti-
mization. Operations Research, 23:389–405, 1975. [39]
[30] D. Naddef and G. Rinaldi. Branch-and-cut algorithms for the capacitated VRP. In P. Toth and
D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and
Applications, pages 49–78. SIAM, Philadelphia, 2002. [32]
[31] P. Neame, N. Boland, and D. Ralph. An outer approximate subdifferential method for piecewise
affine optimization. Mathematical Programming Series A, 87:57–86, 2000. [37]

54
[32] G. L. Nemhauser and L. A. Wolsey. Integer and Combinatorial Optimization. Wiley, New York,
1988. [37]
[33] M. Savelsbergh. Local search in routing problems with time windows. Annals of Operations Re-
search, 4:285–305, 1985. [32]
[34] M. M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35:254–265, 1987. [33, 40]

55
56
Chapter 3

Path inequalities for the vehicle routing

problem with time windows

Brian Kallehauge
Centre for Traffic and Transport, Technical University of Denmark

Natashia Boland
Department of Mathematics and Statistics, University of Melbourne

Oli B. G. Madsen
Centre for Traffic and Transport, Technical University of Denmark

Abstract
In this paper we introduce a new formulation of the vehicle routing problem with time windows (VRPTW)
involving only binary variables associated with the arcs in the underlying digraph. The new formulation is
based on the formulation of the asymmetric traveling salesman problem with time windows by Ascheuer
et al. [3] and has the advantage of avoiding additional variables and linking constraints. In the new
formulation of the VRPTW time windows are modeled using path inequalities. The path inequalities
eliminate time and capacity infeasible paths. We present a new class of strengthened path inequalities
based on the polyhedral results obtained by Mak [17] in the context of the asymmetric traveling salesman
problem with replenishment arcs. We study the VRPTW polytope and determine the polytope dimension.
We show that the lifted path inequalities are facet defining under certain assumptions. We also introduce
precedence constraints in the context of the VRPTW. Computational experiments are performed with
a branch-and-cut algorithm on the Solomon test problems with wide time windows. Based on results
on 25-node problems the outcome is that the algorithm shows promising results compared to leading
algorithms in the literature. In particular we report a solution to a previously unsolved 50-node Solomon
test problem R208. The conclusion is therefore that the path formulation of Desrochers et al. [9] is no
longer the unchallenged winning strategy for solving the VRPTW.

3.1 Introduction
This paper presents a polyhedral and computational study of a variant of the vehicle routing problem with
time windows.

57
 Throughout this paper, D = (V, A) will be a directed graph, where V = {0, 1, . . . , n + 1} is the set of
nodes, |V | = n + 2, and A is the set of arcs. Node 0 and n + 1 represent the start and destination depot,
respectively. The remaining n nodes represent the set of customers N = V \ {0, n + 1}. With every arc
(i, j) ∈ A a cost ci j ∈ Z+ and time ti j ∈ Z+ is associated. We assume that the triangle inequality on the
costs and travel times is satisfied, i.e. ci j 6 cih + ch j and ti j 6 tih + th j for all (i, j) ∈ A. With every node
j ∈ N we associate a demand d j ∈ Z+ , a release date a j ∈ Z+ , and a due date b j ∈ Z+ , where a j > t0 j
and b j > a j for all j ∈ N, and b j > ai + ti j for all (i, j) ∈ A. The release date ai and the due date bi are
the earliest, respectively the latest possible starting time for servicing node i. For the nodes 0 and n + 1
we assume that d0 = dn+1 = 0, a0 = an+1 = 0, and b0 = bn+1 = +∞. With the graph D we associate
a vehicle capacity q ∈ Z+ , where q > di for all i ∈ N, and q > di + d j for all (i, j) ∈ A. As usual, for
each W ⊂ V let δ − (W ) = {(i, j) ∈ A | i ∈ V \ W, j ∈ W }, δ + (W ) = {(i, j) ∈ A | i ∈ W, j ∈ V \ W }, and
A(W ) = {(i, j) ∈ A | i, j ∈ W }. We assume that A ∩ {(0, n + 1)} = 0, / |δ + (0)| = |δ − (n + 1)| = n and
δ (0) = δ (n + 1) = 0.
− + / Let m = |A(N)| so |A| = 2n + m.
A k-route consists of a partition {Ni | i = 1, . . . , k} of the set of customer nodes N into k subsets, and
an associated sequence, or route, of each subset ri = (v1 , . . . , v|Ni | ) specifying the order of service of the
customers. A route ri represents a vehicle leaving the start depot 0, servicing the set of customers Ni in the
order defined by ri and entering the destination depot n + 1. Note that a partition of the customer nodes
may correspond to several k-routes. Let RN+ be the set of nonnegative real vectors whose components
are indexed by N. With each k-route we associate a vector s ∈ RN+ , defined as follows: sv1 = av1 and
sv j = max{sv j−1 +tv j−1 v j , av j } for j = 2, . . . , |Ni | and i = 1, . . . , k. We say that a k-route is time and capacity
feasible if sv j 6 bv j for j = 1, . . . , |Ni | and ∑ j∈Ni d j 6 q for i = 1, . . . , k. We say that the partition size k,
or vehicle fleet size, is feasible, if a corresponding feasible k-route exists and we denote by k(N) the
minimum feasible fleet size. We denote by K = {k(N), . . . , |N|} the set of feasible fleet sizes. For each
k ∈ K we denote by Rk the set of feasible k-routes using exactly k vehicles. The set of feasible k-routes
for all k ∈ K is denoted RK = {∪Rk | k = k(N), . . . , |N|}. Let Rk ∈ RK denote any feasible k-route and let
c(Rk ) be the corresponding cost defined as c(Rk ) = (∑ c(ri ) | i = 1, . . . , k) where c(ri ) = c0v1 + (∑ cv j−1 v j |
j = 2, . . . , |Ni |) + cv|N | n+1 . The vehicle routing problem with time windows is
i

(VRPTW) min{c(Rk ) | Rk ∈ RK }.
The VRPTW reduces to the capacitated vehicle routing problem (CVRP) if ai = 0 and bi = +∞ for every
i ∈ N. Therefore the VRPTW is NP-hard. Indeed, it is strongly NP-complete to find a feasible solution
for the VRPTW with a fixed number of vehicles [19].
Bard et al. [6] present an integer linear programming formulation of a different variant of the VRPTW
where the objective is to minimize the number of vehicles. In this formulation binary variables are as-
sociated with arcs in the underlying digraph. Integer node variables are introduced to model the time
and capacity restrictions following the approach of Miller et al. [18]. In the context of the asymmetric
traveling salesman problem with time windows (ATSPTW) Ascheuer et al. [3, 4] modeled time window
restrictions using “infeasible path elimination” constraints and they noted that similar constraints can
also be used to model any other kind of path infeasibility. We follow this suggestion and present a new
formulation of the VRPTW involving binary arc variables only. In our formulation time and capacity
restrictions are modeled using infeasible path elimination constraints (IPECs), which we denote path in-
equalities. We present a class of strengthened path inequalities based on the polyhedral results obtained by
Mak [17] in the context of the asymmetric traveling salesman problem with replenishment arcs (RATSP).
Our formulation is exponential in size since there are exponentially many path inequalities. We present
the first polyhedral results on the VRPTW polytope by proving that the new class of strengthened path
inequalities are facet-defining for the VRPTW polytope under certain conditions. We solve the problem
by a branch-and-cut algorithm based on the new path inequalities and other classes of inequalities adopted
from the asymmetric traveling salesman problem (ATSP) and the precedence constrained ATSP [5].
The reasons for this line of research were the following. First, the polyhedral approach to the related

58
ATSPTW indicates that most instances in the range of up to 50-70 nodes can be solved to optimality
via branch-and-cut codes in a few minutes [3]. Second, we believe that the polyhedral approach may
outperform the dynamic programming approach of the path formulation when the time windows are wide.
Third, the results of Bard et al. [6] indicate that in their case instances with 50 nodes can be routinely
solved using a branch-and-cut code. Finally, no polyhedral results on the VRPTW polytope were known
when this investigation started and further research in this area could show what the relative advantages
of the different solution methods are.
This paper reports some computational results on VRPTW test problems developed by Solomon [20].
We focus on test problems having a long planning horizon and large vehicle capacity making it possible
for each vehicle to visit many customers on a route.
The paper is organized as follows. In Section 3.2 we present a new binary integer programming (BIB)
formulation of the VRPTW and determine the dimension of the associated polytope. In Section 3.3 we
prove that a new class of strengthened path inequalities define a proper face of the VRPTW polytope and
that the inequalities are facet defining under certain conditions. In Section 3.4 we describe how classes of
valid inequalities for the precedence constrained ATSP polytope can be transferred to the VRPTW. Sec-
tion 3.6 reviews preprocessing routines for the VRPTW based on time windows and their implications. In
Section 3.7 we describe the test problems providing the data for our computational study and the compu-
tational platform for performing the experiments. Section 3.8 is dedicated to the implementational details
of the branch-and-cut algorithm. In Section 3.9 we report extensive computational results following the
approach of Ascheuer et al. [4]. Finally, we present our conclusions in Section 3.10.

3.2 A BIP formulation of the VRPTW

A path in the graph D is a sequence of nodes P = (v1 , . . . , v p ) such that (vi , vi+1 ) ∈ A for all i = 1, . . . , p − 1.
Let V (P) and A(P) denote respectively the set of nodes and arcs of the path. The path is always open
and simple, i.e. |A(P)| = p − 1 and vi 6= v j for i 6= j. We say that a path P is infeasible if it does not
p
occur as a subpath in any feasible route, i.e. if either ∑i=1 dvi > q or svi > bvi , for some i ∈ {1, . . . , p}. An
infeasible path P is said to be minimal infeasible if the truncated subpaths defined by A(P) \ {(v1 , v2 )}
and A(P) \ {(v p−1, v p )} are feasible. We denote by PD the set of all minimal infeasible paths in D. A
necessary and sufficient condition for a k-route to be feasible is that no route that defines it contains any
minimal infeasible path.
Let RA+ be the set of nonnegative real vectors whose components are indexed by A; the incidence
vector of a k-route in RK is x ∈ RA+ , defined as follows: xi j = 1 if (i, j) ∈ A is used in the k-route and
xi j = 0 otherwise. For notational convenience we do not distinguish between a k-route and its incidence
vector. For any Q ⊆ A, we write x(Q) for ∑(i, j)∈Q xi j . With this notation, the set of feasible k-routes
RK for the VRPTW is the set of those k-routes for which the incidence vector x ∈ BA satisfy the degree
equations

(3.1) x(δ + (i)) = 1, ∀ i ∈ N,

(3.2) x(δ (i)) = 1,
−
∀ i ∈ N,

the subtour inequalities

(3.3) x(A(W )) 6 |W | − 1, ∀ 0/ 6= W ⊆ N,

and the path inequalities

(3.4) x(A(P)) 6 |A(P)| − 1, ∀P ∈ PD .

59
We denote by PVRPTW ⊂ RA+ the VRPTW polytope, i.e. the convex hull of the incidence vectors of all
the k-routes in RK , PVRPTW = conv{RK }. As for the dimension of the VRPTW polytope, we have
dim(PVRPTW ) = m.

Proposition 3.2.1 dim(PVRPTW ) = m.

Proof: Clearly dim(PVRPTW ) 6 |A| − 2n = m, since PVRPTW ⊂ RA and the system (3.1) and (3.2) of
degree equations has rank 2n.
Next we show that there are m+ 1 affinely independent k-routes in PVRPTW . First consider the k-route
Rn ∈ RK consisting of n routes, one for each customer

Rn = {(0, i, n + 1) | i ∈ N},

and then consider the following m k-routes denoted by Rn−1 (i, j), each of them containing an arc (i, j) in
A(N) which is not used by any other of the m − 1 k-routes or by the k-route Rn , thus proving the affine
independence.

Rn−1 (i, j) = {(0, i, j, n + 1)} ∪ {(0, i′, n + 1) | i′ ∈ N \ {i, j}}, ∀ (i, j) ∈ A(N).

This gives a total of m + 1 affinely independent k-routes and so dim(PVRPTW ) > m. Therefore, from
dim(PVRPTW ) 6 m and dim(PVRPTW ) > m it follows that dim(PVRPTW ) = m. 

3.3 Lifted path inequalities

The path inequalities (3.4) can be very weak. Mak [17] proposes ways to strengthen path inequalities
for the ATSP with Replenishment Arcs (RATSP) and introduces a class of valid inequalities she called
S1 constraints. In what follows we transfer Mak’s S1 inequalities to the VRPTW adopting the definitions
used therein.

Definition 3.3.1 Suppose we have a minimal infeasible path P = (v1 , . . . , v p ) ∈ PD . Then we define the
following arc sets.
S
1. ∆+ (P) = h=1 p−1 +
∆vh (P) is the set of non-path arcs starting at nodes in P, where ∆+
vh (P) = {(vh , j) ∈
A | j 6= vh+1 }, for h = 1, . . . , p − 1.
e+ (P) = S p−1 ∆
2. ∆ e+
h=1 vh (P) is the set of time and capacity feasible forward and escaping non-path arcs
w.r.t. the minimal infeasible path P, where ∆ e+vh (P) = {(vh , j) ∈ A(N) | j 6= v1 , . . . , vh+1 , ∑l=1 dvl +
h

d j 6 q and svh + tvh , j 6 b j } ∪ {(vh, n + 1) ∈ A}, for h = 1, . . . , p − 1.

S p−1 +
3. ∆¯ + (P) = h=1 ∆¯ vh (P) is the set of time and capacity infeasible forward and escaping non-path arcs
w.r.t. the minimal infeasible path P, where ∆¯ + vh (P) = {(vh , j) ∈ A(N) | j 6= v1 , . . . , vh+1 , ∑l=1 dvl +
h

d j > q or svh +tvh , j > b j }, for h = 1, . . . , p − 1. Since all arcs (i, j) ∈ A are time and capacity feasible
we have ∆¯ +v1 (P) = 0. /
S
4. ∆B (P) = h=1p−1 B
∆vh (P) is the set of backward arcs w.r.t. the minimal infeasible path P, where
∆vh (P) = {(vh , j) ∈ A(N) | j ∈ {v1 , . . . , vh−1 }}, for h = 2, . . . , p − 1 and ∆Bv1 (P) = 0.
B /

5. δ + (vh ) = ∆ e+ (P) ∪ ∆¯ + (P) ∪ ∆B (P) ∪ {(vh , vh+1 }) is the set of all arcs leaving node vh , for h =
vh vh vh
1, . . . , p − 1.

60
The outdegree equations (3.1) give us a means to alter the form of the path inequalities (3.4) (the "in-
side" form). By subtracting the outdegree equations x(δ + (i)) = 1 for the nodes i ∈ {v1 , . . . , v p−1 } of an
infeasible path in (3.4) and dividing the inequality by −1 we obtain the "outside" form

(3.5) x(∆+ (P)) > 1 ∀P ∈ PD .

From definition 3.3.1, we get

∆+ (P) = ∆
e+ (P) ∪ ∆¯ + (P) ∪ ∆B (P),

and therefore (3.5) is equivalent to

x(∆
e + (P)) + x(∆¯ + (P)) + x(∆B (P)) > 1 ∀P ∈ PD .

Inequalities (3.5) can be lifted to give

(3.6) x(∆
e + (P)) > 1 ∀P ∈ PD ,

and we now prove that this is a valid class of inequalities for the VRPTW.

Proposition 3.3.1 For P ∈ PD , the path inequality

(3.7) x(∆
e + (P)) > 1

is valid for the VRPTW polytope.

Proof: Assume to the contrary that there exists an x ∈ RK such that (3.7) does not hold, i.e. x(∆
e + (P)) = 0
and so x(∆vh (P)) = 0, for h = 1, . . . , p − 1. We now prove by induction that if (3.7) does not hold, then
e +

xvh ,vh+1 = 1, for h = 1, . . . , p − 1, contradicting our assumption that x ∈ RK .

Initial step. When h = 1 the outdegree constraint for node v1 gives us x(δ + (v1 )) = x(∆ v1 (P))+x(∆v1 (P))
e+ ¯+
+x(∆Bv1 (P)) + xv1 v2 = 1. By assumption we have x(∆ e+v1 (P)) = 0 and by definition we have ∆v1 (P) =
B

/ hence x(∆v1 (P)) = 0. Recall that ∆¯ v1 (P) = 0/ and therefore x(∆¯ v1 (P)) = 0. So for h = 1 xvh vh+1 = 1
0, B + +

is true.
Inductive step. Assume there is a t, for 2 6 t 6 p − 1, such that xvl vl+1 = 1 is true for l = 1, . . . ,t − 1. We
now prove that xvl vl+1 = 1 is true for l = t. When l = t the outdegree constraint for node vt gives
us x(δ + (vt )) = x(∆e + (P)) + x(∆¯ + (P)) + x(∆B (P)) + xv v = 1. Again, by assumption we have
vt vt vt t t+1
x(∆e+
vt (P)) = 0 and by the inductive assumption x vl vl+1 = 1, for l = 1, . . . ,t − 1, and therefore we must
have ∆¯ +
vt (P) = 0,
/ otherwise the solution would induce a minimal infeasible path. Now, assume there
is an l, for 1 6 l 6 t − 1, such that xvt vl = 1. Then by the inductive assumption the solution would
induce a subtour and therefore we must have x(∆Bvt (P)) = 0. Now, from the outdegree constraint
we have xvt vt+1 = 1. This completes the inductive step and hence the proposition is proved.

The inside form of the inequality (3.7) is

(3.8) x(P) + x(∆¯ + (P)) + x(∆B (P)) 6 |A(P)| − 1.

Definition 3.3.2 (Mak [17]) Given a time and capacity constrained digraph D and a minimal infeasible
path P ∈ PD we denote by S1D,P the constraint (3.8).

61
 Ascheuer et al. [3] propose ways to strengthen path inequalities for the ATSPTW and introduce a
class of lifted path inequalities they call tournament constraints. To begin we define additional arc sets.

Definition 3.3.3 Suppose we have a minimal infeasible path P = (v1 , . . . , v p ) ∈ PD . Then we define the
following arc sets.
S
1. ∆E (P) = h=1 ∆vh (P) is the set of escaping non-path arcs w.r.t. the minimal infeasible path P,
p−1 E

where ∆Evh (P) = {(vh , j) ∈ A(N) | j 6∈ {v1 , . . . , v p }} ∪ {(vh, n + 1) ∈ A}, for h = 1, . . . , p − 1

S
2. ∆F (P) = h=1p−1 F
∆vh (P) is the set of non-path forward arcs w.r.t. the minimal infeasible path P, where
∆Fvh (P) = {(vh , j) ∈ A(N) | j = vh+2 , . . . , v p }, for h = 1, . . . , p − 2 and ∆Fvp−1 (P) = 0.
/

3. δ + (vh ) = ∆Evh (P) ∪ ∆Fvh (P) ∪ ∆Bvh (P) ∪ {(vh , vh+1 )} is the set of all arcs leaving node vh , for h =
1, . . . , p − 1.
We note that the relationship between the arc sets of definition 3.3.1 and 3.3.3 is given by

(3.9) ∆
e + (P) ∪ ∆¯ + (P) = ∆E (P) ∪ ∆F (P).

Given a P ∈ PD , Ascheuer et al. [3] denote by [P] = {(vi , v j ) ∈ A | 1 6 i 6 j 6 p} the transitive closure
of P = (v1 , . . . , v p ). Ascheuer et al. [3] show that for every P ∈ PD the tournament constraint

(3.10) x([P]) 6 |A(P)| − 1

is valid for the ATSPTW polytope. Using definition 3.3.3 we can write (3.10) as

(3.11) x(P) + x(∆F (P)) 6 |A(P)| − 1.

Observe that (3.11) expands to

p−1 
(3.12) ∑ xvh vh+1 + x(∆Fvh (P)) 6 |A(P)| − 1.
h=1

By the outdegree equation, this is equivalent to

p−1 
(3.13) ∑ 1 − x(∆Evh (P)) − x(∆Bvh (P)) 6 |A(P)| − 1,
h=1

and so to
p−1 
(3.14) ∑ x(∆Evh (P)) + x(∆Bvh (P)) > 1,
h=1

and also

(3.15) x(∆E (P)) + x(∆B (P)) > 1.

We now prove that (3.15) is a valid constraint for the VRPTW.

Proposition 3.3.2 For P ∈ PD , the path inequality

(3.16) x(∆E (P)) + x(∆B (P)) > 1

is valid for the VRPTW polytope.

62
Proof: Assume to the contrary that there exists an x ∈ RK such that (3.16) does not hold, i.e. x(∆E (P)) +
x(∆B (P)) = 0 and so x(∆Evh (P)) + x(∆Bvh (P)) = 0, for h = 1, . . . , p − 1. We now prove by induction that if
(3.16) does not hold, then xvh−1 ,vh = 1, for h = p, . . . , 2, contradicting our assumption that x ∈ RK .
Initial step. When h = p the outdegree constraint for node v p−1 gives us x(δ + (v p−1 )) = x(∆Evp−1 (P)) +
x(∆Fvp−1 (P)) + x(∆Bvp−1 (P)) + xv p−1 ,v p = 1. By assumption we have x(∆Evp−1 (P)) + x(∆Bvp−1 (P)) = 0
and by definition we have ∆Fvp−1 (P) = 0, / hence x(∆Fvp−1 (P)) = 0. So for h = p xvh−1 vh = 1 is true.

Inductive step. Assume there is a t, for 2 6 t 6 p − 1, such that xvl−1 vl = 1 is true for l = p, . . . ,t + 1.
We now prove that xvl−1 vl = 1 is true for l = t. When l = t the outdegree constraint for node vt
gives us x(δ + (vt )) = x(∆Evt (P)) + x(∆Fvt (P)) + x(∆Bvt (P)) + xvt−1 ,vt = 1. Again, by assumption we
have x(∆Evt (P)) + x(∆Bvt (P)) = 0 and by the inductive assumption xvl−1 vl = 1, for l = p, . . . ,t + 1,
and the indegree constraints x(δ − (i)) = 1, for i = p, . . . ,t + 1 we must have x(∆Fvt (P)) = 0. Now,
from the outdegree constraint we have xvt−1 vt = 1. This completes the inductive step and hence the
proposition is proved.

3.3.1 Facet proof

Definition 3.3.4 Given any minimal infeasible path P ∈ PD , we define the face of PVRPTW induced by
S1D,P to be F1D,P = {x ∈ PVRPTW | x(A(P)) + x(∆¯ + (P)) + x(∆B (P)) = |A(P)| − 1}.

Lemma 1 For any time and capacity constrained digraph D = (V, A) with N = {1, . . . , n} where |N| > 5,
and any minimal infeasible path P = (v1 , . . . , v p ) ∈ PD with 3 6 p 6 |N| − 2, F1D,P is a proper face of
PVRPTW .
Proof: To show that F1D,P is a proper face of PVRPTW , we need to show that 0/ 6= F1D,P 6= PVRPTW . We
first show that 0/ 6= F1D,P by showing that there is at least one solution in RK that satisfies constraint S1D,P at
equality. Without loss of generality, we assume that P = {1, . . . , p}. Consider a feasible solution x1 ∈ RK
defined by x1i j = 1 if (i, j) ∈ {(k, k + 1) | k = 1, . . . , p − 2} ∪ {(0, j) | j > p} ∪ {( j, n + 1) | j > p} and
x1i j = 0 otherwise. Since this is clearly a feasible k-route using exactly |A(P)| − 1 arcs, constraint S1D,P is
satisfied at equality. Next we show that F1D,P 6= PVRPTW by showing that there is at least one solution in
RK that does not satisfy constraint S1D,P at equality. Consider another feasible solution x2 ∈ RK where
x2i j = 1 if (i, j) ∈ {(0, j) | j = 1, . . . , n} ∪ {( j, n + 1) | j = 1, . . . , n} and x2i j = 0 otherwise. Clearly, as none
of the arcs in A(P) ∪ ∆¯ + (P) ∪ ∆B (P) are used in the k-route, constraint S1D,P is not satisfied at equality. 

Given a path P ∈ PD let Ω(P) = ∪h=1 vh (P)}. Let ωvh

Ωvh (P), where Ωvh (P) = {k ∈ V | ( j, k) ∈ ∆¯ +
p−1

denote any arbitrary node in the set Ωvh (P).

Theorem 3.3.1 For any time and capacity constrained digraph D = (V, A) with N = {1, . . . , p + 2} and
any minimal infeasible path P = (v1 , . . . , v p ) ∈ PD , constraint S1D,P defines a facet for PVRPTW if the
following conditions hold:
1. all arcs in A are time and capacity feasible, i.e. b j > ai + ti j and q > di + d j for all (i, j) ∈ A,
2. if (i, vh ) exists in A(N) for i ∈ N \ V (P) and vh ∈ V (P) \ {v1 , v p }, then there must exist an arc
(vh , ωvh ) ∈ A(N) and (i, vh , ωvh ) must be feasible,
3. if (v1 , vh ) exists in A(N) for any h ∈ {3, . . . , p − 1}, then there must exist (vh , ωvh ) ∈ A(N) and
(vh−1 , ωvh−1 ) ∈ A(N) where ωvl is distinct for distinct l and (v1 , vh , ωvh ) and (v2 , . . . , vh−1 , ωvh−1 )
must be feasible.

63
 4. if (v1 , v p ) or (v p , v1 ) exists in A(N), then there must exist (v p−1 , ωv p−1 ) ∈ A(N) and (v2 , . . . , v p−1 ,
ωv p−1 ) must be feasible.
5. if (vh , v1 ) exists in A(N) for any h ∈ {3, . . . , p − 1}, then there must exist (vh−1 , ωvh−1 ) ∈ A(N) and
(v2 , . . . , vh−1 , ωvh−1 ) must be feasible.
6. if (v p , vh ) exists in A(N) for any h ∈ {1, . . . , p − 1}, then there must exist (vh , ωvh ) ∈ A(N) and
(v p−1 , ωv p−1 ) ∈ A(N) where ωvl is distinct for distinct l and (v p , vh , ωvh ) and (vh+1 , . . . , v p−1 , ωv p−1 )
must be feasible.
7. if (vh , v p ) exists in A(N) for any h ∈ {2, . . . , p − 2}, then there must exist (v p−1 , ωv p−1 ) ∈ A(N) and
(vh+1 , . . . , v p−1 , ωv p−1 ) must be feasible.
8. if (vi , v j ) exists in A(N), for i, j ∈ {2, . . . , p−1}, i < j −1, then there must exist (v j , ωv j ) ∈ A(N) and
(v j−1 , ωv j−1 ) ∈ A(N) where ωvl is distinct for distinct l and (v1 , . . . , vi , v j , ωv j ) and (vi+1 , . . . , v j−1 ,
ωv j−1 ) must be feasible.
9. if (vi , v j ) exists in A(N), for i, j ∈ {2, . . . , p − 1}, i > j, then there must exist (v j , ωv j ) ∈ A(N) and
(vi−1 , ωvi−1 ) ∈ A(N) where ωvl is distinct for distinct l and (vi , v j , ωv j ) and (v j+1 , . . . , vi−1 , ωvi−1 )
must be feasible.
Proof: We now show that under the conditions of the lemma, F1D,P has dimension dim(PVRPTW ) − 1.
From proposition 1 we know that F1D,P has dimension at most dim(PVRPTW ) − 1. We therefore have
to show that the dimension of F1D,P is at least dim(PVRPTW ) − 1. We show this directly by constructing
dim(PVRPTW ) affinely independent feasible k-routes that satisfy constraint S1D,P at equality. First we split
A(N) into the following disjoint sets:

A1 = A(P),
A2 = (V (P) \ {v1, v p },V \ V (P)),
A3 = (V \ V (P),V (P) \ {v1, v p }),
A4 = ({v1 , v p },V \ V (P)),
A5 = (V \ V (P), {v1 , v p }),
A6 = ({v1 },V (P) \ {v1}) \ A(P),
A7 = (V (P) \ {v1}, {v1 }),
A8 = ({v p },V (P) \ {v1, v p }),
A9 = (V (P) \ {v1, v p }, {v p}) \ A(P),
A10 = (V (P) \ {v1, v p },V (P) \ {v1, v p }) \ A(P), with i < j for (vi , v j ) ∈ A10 ,
A11 = (V (P) \ {v1, v p },V (P) \ {v1, v p }), with i > j for (vi , v j ) ∈ A11 ,
A12 = (V \ V (P),V \ V (P)).

Then for every Al , l = 1, . . . , 12, we construct a k-route for all (i, j) ∈ Al that satisfy constraint S1D,P at
equality. We also make sure that the arc (i, j) ∈ Al for which we construct a k-route is not contained in
any of the other previously introduced k-routes, thereby ensuring the affine independence. Note that the
k-routes constructed for A1 are special because we actually ensure the affine independence by the absence
of the arc (i, j) ∈ A1 in the corresponding k-route.
1. {(v1 , . . . , vh )} ∪ {(vh+1, . . . , v p )} ∪ {(vk ) | k ∈ N \ V (P)}, ∀(vh , vh+1 ) ∈ A1 .

2. (a) {(v1 , . . . , vh , j)} ∪ {(vh+1 , . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ { j}}} if (vh , j) 6∈ ∆¯ +

vh (P),

64
 (b) {(v1 , . . . , vh−1 )} ∪ {(vh, j)} ∪ {(vh+1 , . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ { j}}} otherwise,
∀(vh , j) ∈ A2 .
3. {(i, vh , ωvh )} ∪{(v1 , . . . , vh−1 )} ∪{(vh+1 , . . . , v p )} ∪{(vk ) | k ∈ N \ {V (P)∪{i, ωvh }}}, ∀(i, vh ) ∈ A3 .
4. (a) {(v1 , j)} ∪ {(v2 , . . . , v p )} ∪ {(vk ) | k ∈ N \ {V(P) ∪ { j}}} if h = 1,
(b) {(v p , j)} ∪ {(v1 , . . . , v p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ { j}}} otherwise, ∀(vh , j) ∈ A4 .
5. (a) {(i, v1 )} ∪ {(v2, . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {i}}} if h = 1,
(b) {(i, v p )} ∪ {(v1 , . . . , v p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {i}}} otherwise, ∀(i, vh ) ∈ A5 .
6. (a) {(v1 , vh , ωvh )} ∪ {(v2 , . . . , vh−1 , ωvh−1 } ∪ {(vh+1, . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P) ∪{ωvh ,
ωvh−1 }}} if h 6= p,
(b) {(v1 , v p )} ∪ {(v2 , . . . , v p−1 , ωv p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {ωv p−1 }}} otherwise, ∀(v1 , vh ) ∈
A6 .
7. (a) {(vh , v1 )} ∪ {(v2 , . . . , vh−1 , ωvh−1 )} ∪ {(vh+1 , . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {ωvh−1 }}} if
h 6= p,
(b) {(v p , v1 )} ∪ {(v2 , . . . , v p−1 , ωv p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {ωv p−1 }}} otherwise, ∀(vh , v1 ) ∈
A7 .
8. {(v p , vh , ωvh )}∪{(v1 , . . . , vh−1 )}∪{(vh+1, . . . , v p−1 , ωv p−1 )}∪{(vk ) | k ∈ N \{V (P)∪{ωvh , ωv p−1 }}},
∀(v p , vh ) ∈ A8 .
9. (a) {(vh , v p )} ∪ {(v1 , . . . , vh−1 )} ∪ {(vh+1 , . . . , v p−1 , ωv p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {ωv p−1 }}}
if (vh , v p ) ∈ ∆¯ +
vh (P),
(b) {(v1 , . . . , vh , v p )} ∪ {(vh+1, . . . , v p−1 , ωv p−1 )} ∪ {(vk ) | k ∈ N \ {V (P) ∪ {ωv p−1 }}} otherwise,
∀(vh , v p ) ∈ A9 .
10. {(v1 , . . . , vi , v j , ωv j )} ∪ {(vi+1 , . . . , v j−1 , ωv j−1 )} ∪ {(v j+1, . . . , v p )} ∪ {(vk ) | k ∈ N \ {V (P)∪
{ωv j , ωv j−1 }}}, ∀(vi , v j ) ∈ A10 .
11. {(vi , v j , ωv j )} ∪ {(v1 , . . . , v j−1 )} ∪ {(v j+1 , . . . , vi−1 , ωvi−1 )} ∪ {(vi+1 , . . . , v p )} ∪ {(vk ) |
k ∈ N \ {V(P) ∪ {ωv j , ωvi−1 }}}, ∀(vi , v j ) ∈ A11 .
12. {(i, j)} ∪ {(v1 , . . . , v p−1 )} ∪ {(vk ) | k ∈ N \ {i, j, v1 , . . . , v p−1 }}, ∀(i, j) ∈ A12 .


3.4 Precedence constraints

Time windows induce precedences among the nodes [4], that is, whenever a set of customers Nl ∈ N
are served on the same route l ∈ {1, . . . , k} and bi < a j + t ji for i, j ∈ Ni , we can conclude that i has to
precede j in any feasible route. Let i ≺ j denote the fact that i has to precede j in any feasible route
and let P = (V, R) denote an additional precedence digraph defined on the node set V = V of D. An arc
(i, j) ∈ R represents a precedence relationship i ≺ j. Note that by definition δ + (0), δ − (n + 1) ∈ R and
R ∩ {(0, n + 1)} = 0. / In the context of the ATSPTW which Ascheuer et al. [4] considers, the precedence
digraph is acyclic and transitively closed. However neither is true for the VRPTW, because in our case
the nodes i ∈ N are not required to all be visited on the same route as in the ATSPTW, e.g. if a 2-cycle
exists (i − j − i) it means that the node pair (i, j) cannot belong to the same route and we refer to it as an
incompatible pair. This term was introduced by Bard et al. [6]. Also, if i ≺ j and j ≺ k it does not mean

65
that i ≺ k because j may be served by a different vehicle than i, which means it may be possible for k to
precede i on the same route. Now, let

(3.17) π (v) = {i ∈ V | (i, v) ∈ R},

(3.18) σ (v) = { j ∈ V | (v, j) ∈ R},

represent the set of the predecessors and successors of a node v ∈ V , respectively.

Balas et al. [5] derived families of valid inequalities for the precedence constrained asymmetric trav-
eling salesman (PCATS) polytope that can be seen as strengthenings of the subtour inequalities (3.3).
These inequalities can also be written in the equivalent cut form

(3.19) x(W,W ) > 1 ∀ 0/ 6= W ⊆ N

Now we present two strengthenings of (3.19) in the VRPTW context. The results we present follow easily
from Balas et al. [5], but these strengthenings have not been considered previously in our context:

Proposition 3.4.1 For W ⊆ N and any given j ∈ W , W = V \ W , the weak predecessor-inequality (weak
π -inequality)

(3.20) x(W \ π ( j),W \ π ( j)) > 1

is valid for the VRPTW polytope.

Proof: Let r be any feasible route and let j ∈ r. Let ŵ be the last node of W served by r, and note that
ŵ ∈ W \ π ( j). The successor of ŵ in r cannot be in π ( j). 

Balas et al. [5] presented a second class of inequalities, in which the successors of W play a role
analogous to that of the predecessors of W in (3.20). The authors describe how the role of predecessor
and successor can be switched by basically replacing the digraph D with the transposed DT = (V, AT ),
where AT = {(i, j) ∈ V × V | ( j, i) ∈ A}. This immediately gives us the following result.

Proposition 3.4.2 For W ⊆ N, and any given j ∈ W , W = V \ W , the weak successor-inequality (weak
σ -inequality)

(3.21) x(W \ σ ( j),W \ σ ( j)) > 1

is valid for the VRPTW polytope.

Note that the general π - and σ −inequalities are not generally valid for the VRPTW, since they are derived
from the fact that in the PCATSP context all nodes have to be served by the same vehicle. Balas et al. [5]
also presented a predecessor-successor inequality or π − σ -inequality and a precedence cycle breaking
inequality or pcb-inequality but it is easy to show that these families of inequalities are not generally valid
for the VRPTW polytope. These inequalities are also derived from the fact that in the PCATSP context
all nodes have to be served by the same vehicle. However, we can still make use of these families of
inequalities when we are considering the VRPTW with exactly one vehicle:

(1-VRPTW) min{c(R1 ) | R1 ∈ R1 }.

The 1-VRPTW is an important subproblem in the solution of the VRPTW. The 1-VRPTW is identical
with the ATSPTW if the total demand is not greater than the vehicle capacity, i.e. ∑i∈N di 6 q. The
strengthened precedence inequalities can be used in the branching strategy where nodes with the number
of vehicles set to one is bounded using these inequalities, x(N) = 1. The inequalities can also be used

66
in the separation of the 2-path inequalities proposed by Kohl et al. [15], in which one has to determine
whether it is feasible to serve a subset of customer nodes using exactly one vehicle. It is also possible
to use these inequalities in the case constraint branching is performed on a subset of customers, S ⊂ N,
dividing the problem into (i) x(S) = 1 and (ii) x(S) > 2. In subproblem (i) the strengthened inequalities
are valid with respect to S ⊂ N, and the separation problem can be solved for this subset.

3.5 ATSP inequalities

In this section we describe the classes of inequalities from the ATSP that we use in our implementation
of the branch-and-cut algorithm:
• Odd Closed Alternating Trail (CAT) inequalities [10]
Two distinct arcs (i, j) and (u, v) are called incompatible if i = u, or j = v, or i = v and j = u;
compatible otherwise. Let (i, j) = (u, v) denote the fact that (i, j) and (u, v) are incompatible and
(i, j) ↔ (u, v) that they are compatible. An odd CAT is a sequence
(3.22) T = {a1, . . . , at } ∈ A(N), t > 3 and odd,
of t distinct arcs, such that for k = 1, . . . ,t
a0 = at
at+1 = a1
ak = ak+1
ak = ak−1
ak ↔ ai , i = 1, . . . ,t, i 6= k + 1, and i 6= k − 1.
We denote by TD the set of all CATs in the directed graph D. For T ∈ TD let
(3.23) s(N) = {i ∈ N | |δ + (i) ∩ T | = 2}
and
(3.24) t(N) = {i ∈ N | |δ − (i) ∩ T | = 2},
denote the set of source and sink nodes of N, respectively. Moreover, let
(3.25) Q = {(i, j) ∈ A(N) \ T | i ∈ s(N), j ∈ t(N)}.
For any T ∈ TD the odd CAT inequality is
|T | − 1
(3.26) x(T ∪ Q) 6 .
2
• D+
k -inequalities [10]
k k−1 k−1
(3.27) xi1 ik + ∑ xih ih−1 + 2 ∑ xi1 ih + ∑ x({i2 , . . . , ih−1 }, ih ) 6 k − 1,
h=2 h=2 h=3

where (i1 , . . . , ik ) is any sequence of k ∈ {3, . . ., n − 1} distinct nodes.

• D−
k -inequalities [10]
k k−1 k−1
(3.28) xik i1 + ∑ xih−1 ih + 2 ∑ xih i1 + ∑ x(ih , {i2 , . . . , ih−1 }) 6 k − 1,
h=2 h=2 h=3

where (i1 , . . . , ik ) is any sequence of k ∈ {3, . . ., n − 1} distinct nodes.

67
3.6 Preprocessing
For the VRPTW, preprocessing includes three main steps: tightening of the time windows, deducing from
the time windows the precedence relationships among customer nodes, and eliminating arcs.

3.6.1 Tightening of the time windows

Kontoravdis and Bard [16] presented criteria (3.29) and (3.30) for removing the time windows of the
depot nodes:

(3.29) ak = max{ak , a0 + t0k } ∀ k ∈ N,

(3.30) bk = min{bk , bn+1 − tkn+1} ∀ k ∈ N.

Desrochers et al. [9] presented criteria (3.31)-(3.32) for increasing the release time of the time windows
and criteria (3.33)-(3.34) for decreasing the due time of the time windows:

(3.31) ak = max{ak , min {ai + tik }} ∀ k ∈ N s.t. δ − (k) 6= 0,

/
(i,k)∈A

(3.32) ak = max{ak , min{bk , min {a j − tk j }}} ∀ k ∈ N s.t. δ + (k) 6= 0,

/
(k, j)∈A

(3.33) bk = min{bk , max{ak , max {bi + tik }}} ∀ k ∈ N s.t. δ − (k) 6= 0,

/
(i,k)∈A

(3.34) bk = min{bk , max {b j − tk j }} ∀ k ∈ N s.t. δ + (k) 6= 0.

/
(k, j)∈A

In a fully connected directed graph it is obvious that criterion (3.29) is equivalent to (3.31), i.e. {(0, k)} ∈
Argmin(i,k)∈A {ai + tik }, and criterion (3.30) is equivalent to (3.34), i.e. {(k, n + 1)} ∈ Argmax(k, j)∈A {b j −
tk j }. However, if we do not have a graph with all possible arcs with head in 0 or tail in n + 1 we need
criterion (3.31) and (3.34). This is for example the case if we branch on the arc variables, i.e. we apply
criterion (3.31) on the subproblem having x0 j = 0, j ∈ N, and (3.34) is used when xin+1 = 0, i ∈ N.

3.6.2 Precedence relationships and elimination of arcs

In Section 3.4 we described how the time windows implied certain precedence relationships. The con-
struction of precedence relationships in the VRPTW follows the description in Ascheuer et al. [4] except
that the precedence digraph in our case is neither acyclic nor transitively closed. The time windows
of the VRPTW and the construction of precedence relationships moves our problem from a fully con-
nected graph to a graph where certain arcs are eliminated. By construction, if (i, j) is in the set R of the
precedence digraph, the arc ( j, i) cannot be contained in any feasible solution to the VRPTW.

3.7 Test problems and computational platform

The test problems that we use in this paper were developed by Solomon [20] and are divided into two
classes, these are denoted by the number 1 and 2. For each of the classes he generated three subclasses,
these are denoted by the letter R, C, and RC. We present an overview of the test problems in Table 3.1. The

68
number of problems #P in class 1 and 2 is respectively 29 and 27; all problems contain n = 100 customer
nodes denoted by 1, . . . , n and a single depot node denoted by 0. We enlarged the set of test problems by
only considering the first 25 and 50 customer nodes of each original problem. The nodes are specified by
integer coordinates (x′i , y′i ) in the Euclidean [0, 100]2 plane and the vehicle capacity by an integer q. All
test problems have an upper bound on the number of vehicles of 25. For each node i = 0, . . . , n a demand
di , start time a′i , due time b′i , and service time sti′ is specified. For the R, C, and RC subclasses the service
times are respectively 10, 90, and 10 for all i ∈ N, 0 otherwise. All node parameters are integer values.
The depot start time a′i is 0 and all customer due times b′i , i ∈ N, are less than the depot due time b0 . In
order to fulfill the assumptions stated in Section 3.1 regarding the model parameters of the VRPTW we
perform the following transformations of the test problem data: Step 1. Create a copy of node 0 and call it
n + 1, set b0 = 0. Step ′ ′ ′ ′ ′
p 2. For i = 0, . . . , n + 1 set xi = 10 xi , yi = 10 yi , ai = 10 ai , bi = 10 bi , and sti = 10 sti .
2 2
Step 3. Set ci j = ⌊ (xi − x j ) + (yi − y j ) ⌋ and ti j = sti + ci j for i, j = 0, . . . , n + 1, i 6= j. Step 4. Add 1
to all ci j for i 6= 0 in order to fulfill the triangle inequality; the service time is positive hence the triangle
inequality on the travel times is also satisfied. Step 5. Apply the preprocessing steps of Section 3.6. The
solution value for the original problem is calculated as (c(R̂k ) − n)/10, where c(R̂k ) denotes the solution
value for the transformed problem.
Subclasses R, C, and RC consists of problems where the distribution of nodes in the plane is respec-
tively random, clustered, and semi-clustered. Problem class 1 consists of problems with a relatively small
vehicle capacity q compared to the total customer demand; in problem class 2 the vehicle capacity is rela-
tively large. Now we want to assess whether the length of the planning horizon [0, bn+1 ] is a constraining
factor. As noted in Section 3.1 this can be seen as another constraint associated with the vehicles. We
consider the complete graph DC = (AC ,V ) on the n + 2 nodes in V . We compute the Hamilton path in
DC starting in 0 and ending in n + 1. Note that concerning subclasses R and RC the nodes are identically
distributed in class 1 and 2, while this is not the case for subclass C. This means that in relation to sub-
classes R and RC the Hamilton path solution is identical for class 1 and 2; this is not true for subclass
C. In Table 3.1 we present the depot due time bn+1 and the duration of the Hamilton path, i.e. the time
the vehicle visits the destination depot n + 1. Problem class 1 consists of test problems with a relatively
short planning horizon compared to the duration of the Hamilton path while class 2 consists of problems
with a relatively long planning horizon. In class 2 the 25- and 50-customer test problems in subclasses
R and RC are not constrained by the vehicle capacity nor the length of the planning horizon, whereas in
subclass C this is only true for the 25-customer problems. In class 1 all test problems are constrained by
vehicle capacity and the length of the planning horizon.
Next we turn to the individual time constraints associated with the customers. Solomon designed two
different methods for assigning time windows to customers. The first method was designed for the random
generation of time windows and used in the subclasses R and RC. The second method was designed for
assigning time windows in a structured way to the clustered problems in subclass C. In method 2 Solomon
used a 3-opt method on each cluster to create routes and then generated time windows by choosing the
center as the arrival time at each customer. The time windows generated in method 2 therefore allows for
cluster-by-cluster solutions to the C-problems. In terms of the number of customers that received a time
window smaller than the depot time window, Solomon created problems where 25, 50, 75, and 100% of
the customers received such a time window. However, in the measurements we made the test problems
R110, R111, R112, RC105, RC107, RC108, and R210 had a time window density of 86-99%. In Table
3.2 we present a grouping of the test problems based on their time windows density. Note that we have
included the test problems in the range 86-99% in the 100% category.
The computational experiments were conducted on two different machines. The hardware and soft-
ware configuration of the machines is given in Table 3.3. The branch-and-cut algorithm presented in this
paper is implemented using the ILOG CPLEX Concert Technology for C++ and the ILOG CPLEX 9.1
Mixed Integer Optimizer [11]. In Section 3.8 we will describe implementational details of the branch-
and-cut algorithm.

69
 Subclass #P q ∑1≤i≤n di bn+1 Hamilton path duration
n = 25 n = 50 n = 100 n = 25 n = 50 n = 100
Problem class 1
R 12 200 332 721 1458 230 562.3 959.7 1635.9
C 9 200 460 860 1810 1236 2381.8 4739.5 9501.1
RC 8 200 540 970 1724 240 475.6 869.6 1640.7
Total 29
Problem class 2
R 11 1000 332 721 1458 1000 562.3 959.7 1635.9
C 8 700 460 860 1810 3390 2445.5 4825.2 9542.3
RC 8 1000 540 970 1724 960 475.6 869.6 1640.7
Total 27

Table 3.1: Solomon’s test problems

R C RC
Problem class 1
100% 1, 5, 9, 10, 11, 12 1, 5, 6, 7, 8, 9 1, 5, 6, 7, 8
75% 2, 6 2 2
50% 3, 7 3 3
25% 4, 8 4 4
Problem class 2
100% 1, 5, 9, 10, 11 1, 5, 6, 7, 8 1, 5, 6, 7, 8
75% 2, 6 2 2
50% 3, 7 3 3
25% 4, 8 4 4

Table 3.2: Time window density of Solomon’s 100-customer test problems

Machine Sun Fire 15K Dell Inspiron 7500

CPU UltraSPARC III Cu 900MHz Intel Pentium III 600MHz
RAM 384Gb 256Mb
Operating system Solaris Microsoft Windows XP
Compiler Sun Studio CC 9 Microsoft Visual C++ .NET 7.1
Compiler options -fast -xarch=v8plusb /O2 /MD
ILOG CPLEX version 9.1.0 9.1.0

Table 3.3: Hardware and software configuration for the computational experiments

70
3.8 The branch-and-cut algorithm
According to the ILOG CPLEX 9.1 User’s Manual [11] the Mixed Integer Optimizer of CPLEX solves
MIP problems using a general branch-and-cut algorithm. We therefore assume that the CPLEX algorithm
follows the basic steps of a branch-and-cut algorithm described in e.g. Wolsey [21, Fig. 9.5]. In this
section we describe the parts of the branch-and-cut algorithm which are specific for our approach to
solving the VRPTW:
• Formulation of the initial binary integer program
• Separation routines
• Branching
• Enumeration strategy
The separation routines, branching, and enumeration strategy is implemented as three CPLEX callback
functions, which are executed respectively in the CUT, BRANCHING, and NODE step of the branch-
and-cut algorithm described in Wolsey [21, Fig. 9.5]. We have turned off the presolve routines and the
general valid inequalities of CPLEX, i.e. clique cuts, Gomory fractional cuts etc. We also set CPLEX
to emphasize optimality over feasibility. For all other parameters in CPLEX we use the default value.
Note that our approach for solving the VRPTW could have been implemented in any branch-and-cut
framework available, e.g. the open-source framework ABACUS [13]. The result of our implementational
work is a basic Dantzig-Fulkerson-Johnson branch-and-cut (DFJBC) algorithm [2] for the VRPTW.

3.8.1 Formulation of the initial binary integer program

We generate all the variables xi j ∈ A, the integer constraints x ∈ BA , and the degree constraints (3.1)
and (3.2). Finally, we add a lower bound on the number of vehicles x(N) > ⌈∑i∈N di /q⌉. A common
approach to solving large-scale integer linear programs by branch-and-cut is to combine cutting-planes
with the dual concept of column generation [8]. In CPLEX 9.1 it is not possible to combine cutting-planes
and column generation, however, this is possible in ABACUS and other open-source frameworks.

3.8.2 Separation routines

In this section we describe the separation procedures for the class of inequalities used in our branch-and-
cut algorithm.
• Subtour, π -, σ -, (π , σ )-inequalities: For the subtour and π -inequalities we use the algorithm
proposed by Balas et al. [5] that simultaneously solves the separation problem for both the subtour
inequalities and the (weak) π -inequalities. The maxflow step of the algorithm is implemented
using the ILOG CPLEX 9.1 Network Optimizer, which solves a minimum cost flow problem using
a network simplex method. The maxflow problem is therefore formulated as a minimum cost flow
problem [1]. The overall complexity of the algorithm is O(n4 ). For the separation of σ - and
(π , σ )-inequalities we also use the algorithms proposed by Balas et al. [5] with complexities of
respectively min{n, |R|} · O(n3) and O(|R| · n3).
• Lifted path inequalities: For the tournament constraints we use the enumeration procedure pro-
posed in Ascheuer et al. [4]. We adopt this procedure for the separation of S1 -inequalities. We
detect paths where x∗ (∆e + (P)) is less than 1. Given a path P from the start depot 0 to a node i ∈ N
we backtrack if x∗ (∆
e + (P)) > 1, otherwise we extend the path to all i ∈ N \ {V (P)}. If an infeasible
path is detected and it is minimal we have found a violated inequality; otherwise we backtrack
again.

71
 • Odd CAT-inequalities: We use the separation routine described in [10].
• Dk -inequalities: We use the separation routine described in [10].

Separation order Suppose we are considering a node in the branch-and-bound tree with the branching
constraints k 6 x(N) 6 k̄. The separation routines are called in the following order:
1. Subtour and π -inequalities (if x(N) 6 1 we lift the π -inequalities)
2. σ -inequalities (if x(N) 6 1 we lift the σ -inequalities)
3. If x(N) 6 1 separation of (π , σ )-inequalities is performed
4. Tournament inequalities
5. S1 -inequalities
6. Odd CAT inequalities
7. D+
k -inequalities

8. D−
k -inequalities

We skip all the subsequent routines whenever one of the procedures generate an inequality.

3.8.3 Branching
We have used two branching strategies: branching on the number of vehicles and branching on arcs:
• Branching on vehicles
Given a point x∗ ∈ RA let k = ⌊x∗ (N)⌋. We create two subproblems if x∗ (N) is fractional: one by
adding the constraint x(N) 6 k and the other by adding the constraint x(N) > k + 1.
• Branching on arcs
If it is not possible to branch on the number of vehicles we branch on the x-variables. We use the
default variable selection strategy of CPLEX [11].

3.8.4 Enumeration strategy

We have used best bound search as the enumeration strategy, which means that the node with the lowest
objective function value will be selected.

3.9 Computational results

The DFJBC algorithm was tested on the class 2 Solomon test problems. The total number of test problems
in this class is 81 when we include the test problems created by only considering the first 25 and 50
customers of each original 100 customer problem.

72
 Key to Table 3.4: (preprocessing results)

ITER : Number of preprocessing loops

TW2 . . . #TW : Number of release time adjustments according to (3.32)
TIME : Total increase in release times
TW3 . . . #TW : Number of release time adjustments according to (3.33)
TIME : Total decrease in due times
REDUCTION : Total time of adjustments in percent; calculated by (TW2.TIME + TW3.TIME) / ∑i∈N (bi −
max{ai ,t0i })

TW2 TW3
ITER #TW TIME #TW TIME REDUCTION
25 customers
R101 4 7 32.0 3 21.8 21.5
C101 4 2 11.0 3 13.8 1.6
C106 4 2 17.5 2 21.2 2.1
RC101 3 1 9.8 0 0.0 1.3
R201 3 1 10.9 3 49.9 2.1
C201 11 14 940.1 16 2556.7 87.4
C205 3 1 320.0 2 0.8 4.0
C206 3 0 0.0 1 70.9 0.6
C208 3 0 0.0 1 47.2 0.3
RC201 3 0 0.0 1 17.9 0.6
50 customers
R101 3 6 15.0 3 5.8 4.2
RC101 3 1 9.0 0 0.0 0.6
C201 6 7 5.0 8 4.7 0.1
C205 3 0 0.0 1 0.4 0.0
RC201 3 0 0.0 1 17.9 0.3
100 customers
R101 3 3 8.9 1 7.0 1.6

Table 3.4: Preprocessing results

3.9.1 Preprocessing
First we present the results of the preprocessing routines. Only in 16 out of the 168 test problems do
we observe adjustments according to criteria (3.32)-(3.33). The criterion (3.34) (or (3.30)) has no effect
on the test problems. We do not include the adjustments of criterion (3.31) (or (3.29)); an additional
151 problems have adjustments only related to this criterion because of release times of zero. Note that
there are also some minor adjustments due to the distance function we use. The number of remaining arc
variables |A| after preprocessing is presented in Section 3.9.2. The test problems would have (n + 1)n arcs
in case the graphs were fully connected, i.e. the 25, 50, and 100 customer test problems would have 650,
2 550, and 10 100 arcs (variables), respectively. The number of remaining arcs in a test problem is the
difference between the number of arcs in the complete graph and the number of precedence relationships
|R| induced by the time windows, i.e. |A| = (n + 1)n − |R|.

73
3.9.2 Solomon’s test problems
In this section we present the results for the class 2 Solomon test problems. Table 3.5 gives an overview
of the results compared to algorithms in the literature. The algorithm of Chabrier [7] is an elementary
shortest path decomposition of the VRPTW. In Irnich and Villeneuve [12] non-elementary paths are
allowed but the formulation is strengthened by eliminating k-cycles (k > 3). In addition 1-path and 2-path
inequalities are generated at the root node of the branch-and-bound tree. Kallehauge et al. [14] use a non-
elementary shortest path with 2-cycle elimination but the master algorithm is accelerated by stabilizing
the dual multipliers. Also, 2-path inequalities are generated in the root node and 1-path inequalities in all
nodes of the branch-and-bound tree. It is clear that our cutting plane algorithm can be combined with all
path pricing algorithms above. The question is where the relative advantages of the different algorithms
lie. The results presented in this paper show that the polyhedral approach in certain cases outperform the
shortest path decomposition of the VRPTW when the time windows are wide and therefore shows that
the shortest path decomposition method is no longer the unchallenged winning strategy. Including arc
pricing in our method would further improve the computational efficiency [2].
The new test problem reported solved in Table 3.5 by the DFJBC algorithm was R208.50. The solution
consists of two routes with a total length of 487.7. In order to solve R208.50 we removed the 1 h CPU
time limit. The solution time was 53815.2 s on the Sun Fire 15K. This brings the total number of solved
test problems up to 63 out of 81. It is clear that the branch-and-cut code presented in this paper overall is
not competitive, however, based on results on the 25-node problems the DFJBC algorithm shows that the
polyhedral approach is promising for problems with wide time windows. We solved 44 problems within
the 1 h CPU time limit, i.e. 24, 14, and 6 test problems out of the 27 25-node, 50-node, and 100-node test
problems, respectively.
Now we compare our computational times for the class 2 Solomon test problems with Irnich and
Villeneuve [12], Kallehauge et al. [14], and Chabrier [7]. Note that [7] only reports solution times for
new solutions found and it is therefore only possible to compare solutions for 6 test problems.
Irnich and Villeneuve [12] performed all their computational experiments on a Pentium III 600MHz
machine with 512Mb RAM. The Dell Inspiron 7500 machine is similar except that it only has 256Mb
RAM. We should therefore be able to compare computational times for experiments performed on the
Dell Inspiron 7500 with the results reported in Irnich and Villeneuve [12]. We tried to solve the 44
problems on the Dell Inspiron 7500 machine. However because of lack of memory we solved only 43 out
of the 44 problems (RC203.25 was not solved). Irnich and Villeneuve [12] report solution times using
2-, 3-, and 4-cycle elimination. We compare our method with the k-cycle elimination method with the
minimum solution time. In Table 3.6 we compare the total solution times per subclass and number of
customers. In Table 3.7 we show the 10 problems for which we observed respectively the largest increase
and decrease in the solution time by our method compared to the minimal solution time reported in Irnich
and Villeneuve [12]. We describe the associated columns in Table 3.9.2. The results demonstrate that our
method reduces the computational times significantly for a number of the 25-node problems, however,
our method experienced problems with the R211.25 test problem solved by [12]. We will comment on
this in Section 3.9.2. The DFJBC algorithm is also competitive with respect to the 50-node C-problems,
however, for the remaining 50- and 100-node problems our method is not completive compared to [12].
Kallehauge et al. [14] also performed computational experiments on the Sun Fire 15K machine. In Table
3.8 and 3.9 we compare our results on the Sun Fire 15K machine with the results in [14]. Also in
this comparison the DFJBC algorithm performs significantly better on the 25-node problems than the
algorithm of [14]. The only exception is again the R211.25 test problem. Finally, in the comparison with
Chabrier [7] in Table 3.10 we note that for the problems with randomly distributed customers and wide
time windows (R208.25 and R204.25) the DFJBC algorithm performs well. However, it is also clear
that our algorithm is not competitive in relation to the semi-clustered RC-problems nor 50- and 100-node
problems. The results of Chabrier [7] has been obtained using a 1.5GHz Pentium IV with 256Mb RAM

74
 Problem C IV KLM This paper Total solved
Subclass R
1 3 3 3 2 3
2 2 2 2 1 2
3 2 2 2 1 2
4 1 (1) 2 1 1 2
5 2 2 2 2 2
6 2 2 1 1 2
7 1 1 1 1 1
8 1 1 1 (1) 2 2
9 2 2 2 2 2
10 2 2 1 1 2
11 1 (1) 2 1 0 2
Total 19 (2) 21 17 14 22
Subclass C
1 3 3 3 3 3
2 3 3 3 3 3
3 3 3 3 2 3
4 3 3 2 2 3
5 3 3 3 3 3
6 3 3 3 3 3
7 3 3 3 3 3
8 3 3 3 3 3
Total 24 24 23 22 24
Subclass RC
1 3 3 3 2 3
2 3 3 1 1 3
3 2 2 0 1 2
4 (1) 1 0 0 0 1
5 3 3 1 2 3
6 2 2 1 2 2
7 (1) 2 1 0 1 2
8 (1) 1 0 0 0 1
Total (3) 17 14 6 9 17
Grand total (3) 60 (2) 59 46 (1) 45 63

Table 3.5: The number of Solomon problems solved in Chabrier [7] (C), Irnich and Villeneuve [12] (IV),
Kallehauge et al. [14] (KLM), and this paper. 1 = 25 customer problem solved, 2 = 25- and 50-customer
problem solved, 3 = 25-, 50-, and 100-customer problem solved. The numbers in parenthesis indicate
solutions not reported elsewhere

75
 Key to Table 3.7, 3.9, and 3.10

k : k-cycle elimination used in obtaining minimum solution time

RLB1 : Value of root node before any strong valid inequalities are generated
RLB2 : Value of root node after strong valid inequalities are generated
RLB : Value of root node
#N : Number of processed nodes in the branch-and-bound tree
CPU(TOTAL) : Total CPU-time in seconds to solve problem to optimality

Subclass #P IV This paper

25 customers
R 10 684.10 206.74
C 8 329.90 49.28
RC 5 1420.80 200.04
Total 23 2434.80 456.06
50 customers
R 3 1353.20 5046.78
C 8 2136.90 1108.80
RC 3 1043.00 1462.51
Total 14 4533.10 7618.09
100 customers
C 6 5973.80 6749.14
Total 6 5973.80 6749.14
Grand total 43 12941.70 14823.29

Table 3.6: Total solution time in seconds for the 43 class 2 test problems solved by the DFJBC algorithm
on the Dell Inspiron 7500 vs. the corresponding minimum solution time reported by Irnich and Villeneuve
[12] (IV)

and using the Java version of ILOG CPLEX 7.5. We have used the computational times obtained on the
Dell Inspiron for comparison.

Results for 25-node problems

In this section we demonstrate in detail the performance of the branch-and-cut code on the class 2
Solomon test problems with 25 customer nodes. The results are summarized in Tables 3.11-3.14. Ta-
ble 3.11 provides an overview of the results, Table 3.12 presents the number of generated cutting planes,
Table 3.13 gives information on the branch-and-bound tree, and Table 3.14 summarizes the percentages
of computing time spent in the various parts of the algorithm.
The branch-and-cut code shows promising results on the 25-node problems. However, three 25-node
problems could not be solved within 1 h (R211.25, RC204.25, and RC208). For example the R204.25
and R208.25 problems are solved significantly faster by the branch-and-cut code than any other algorithm
in the literature. It should be noted that these two problems have an optimal solution value close to the
Hamilton path length of 312.3, see Table 3.1. This is not surprisingly an indication that the code will
work well on problems where the capacity and time constraints are not very tight, i.e. with a solution
structure close to the traveling salesman problem. It is interesting to note the column |A| in Table 3.11.
We can see that RC208.25 is a "pure" time-constrained problem, i.e. the time windows do not imply any
precedence relationships. Also R211.25 has only 3 precedence relationships. These two problems are
special because the time windows share a common band, which makes the problems very hard compared

76
 Irnich and Villeneuve [12] This paper
k RLB1 RLB2 #N CPU(TOTAL) RLB #N CPU(TOTAL) DELTA
Top 10 increases
R205.50 4 682.8 682.8 136 1091.40 662.5 2479 3722.51 2631.11
R209.50 4 599.8 599.8 6 255.40 582.4 420 1302.49 1047.09
C205.100 3 586.4 586.4 0 221.90 586.4 0 980.35 758.45
C206.100 3 586.0 586.0 0 814.40 586.0 0 1315.11 500.71
RC205.50 4 621.6 630.2 4 82.40 621.0 106 495.95 413.55
RC201.50 4 683.1 683.1 4 25.70 680.1 11 103.11 77.41
C202.100 2 589.1 589.1 0 585.60 589.1 0 659.13 73.53
R202.25 3 410.5 410.5 0 0.80 393.1 66 30.69 29.89
R203.25 4 391.4 391.4 0 5.80 373.0 111 28.38 22.58
R210.25 3 403.6 403.6 2 8.00 389.0 65 30.58 22.58
Top 10 decreases
RC207.25 4 264.6 280.4 133 1393.70 268.3 2455 186.98 1206.72
C204.50 2 350.1 350.1 0 1159.40 347.8 70 648.20 511.20
R208.25 3 323.3 323.4 15 363.50 319.8 29 16.23 347.27
C208.100 2 581.8 585.8 1 2183.30 585.8 0 1884.58 298.72
C204.25 2 211.0 211.0 10 279.80 204.9 283 37.72 242.08
R204.25 4 349.1 349.1 34 231.70 332.2 83 29.70 202.00
C207.50 2 356.3 359.6 5 274.00 359.4 2 74.50 199.50
C207.100 2 582.0 585.8 4 2068.80 585.6 2 1887.45 181.35
C208.50 2 340.4 350.5 1 138.50 350.5 0 28.98 109.52
C202.50 2 360.2 360.2 0 196.80 354.1 12 92.00 104.80

Table 3.7: Top 10 increases and decreases in computational time using the DFJBC algorithm compared
to Irnich and Villeneuve [19]

Subclass #P KLM This paper

25 customers
R 10 314.62 138.33
C 8 59.96 23.43
RC 4 424.84 13.49
Total 22 799.42 175.25
50 customers
R 3 3089.91 2359.87
C 8 698.36 433.51
RC 1 3.04 42.30
Total 12 3791.31 2835.68
100 customers
C 6 158.82 2065.03
Total 6 158.82 2065.03
Grand total 40 4749.55 5075.96

Table 3.8: Total solution time in seconds for the 40 class 2 test problems solved by both the LBCP
algorithm in Kallehauge et al. [14] (KLM) and the DFJBC algorithm in this paper. All test problems are
solved on the Sun Fire 15K machine

77
 Kallehauge et al. [14] This paper
RLB1 RLB2 #N CPU(TOTAL) RLB #N CPU(TOTAL) DELTA
Top 10 increases
C208.100 581.8 585.8 0 57.02 585.8 0 569.74 512.72
C207.100 582.0 585.8 0 74.58 585.6 1 527.52 452.94
C206.100 585.4 586.0 0 14.90 586.0 0 418.99 404.09
C205.100 586.4 586.4 0 4.62 586.4 0 334.44 329.82
R209.50 582.9 588.4 524 530.41 582.4 626 816.31 285.90
C202.100 589.1 589.1 0 5.74 589.1 0 202.86 197.12
C202.50 360.2 360.2 0 20.34 354.1 13 62.55 42.21
RC201.50 670.2 682.0 14 3.04 680.1 8 42.30 39.26
R202.25 406.4 408.4 4 0.77 393.1 85 23.89 23.12
R203.25 379.9 381.6 36 5.41 373.0 111 24.13 18.72
Top 10 decreases
R205.50 666.6 672.4 5254 2558.66 662.5 2320 1527.47 1031.19
RC202.25 290.4 313.4 664 222.25 311.8 56 10.13 212.12
RC206.25 250.1 289.0 502 195.49 324.0 0 1.55 193.94
C204.50 350.1 350.1 0 402.18 347.8 113 214.02 188.16
R204.25 333.1 335.4 776 190.64 332.2 68 17.60 173.04
C203.50 359.8 359.8 0 203.11 352.3 23 75.28 127.83
R208.25 318.1 318.9 74 58.23 319.9 17 7.33 50.90
C204.25 211.0 211.0 12 46.00 204.9 155 15.64 30.36
C206.50 344.2 359.0 4 33.88 359.8 0 15.22 18.66
R207.25 347.6 349.7 124 30.20 347.4 54 15.33 14.87

Table 3.9: Top 10 increases and decreases in computational time using the DFJBC algorithm compared
to Kallehauge et al. [20]

Chabrier [7] This paper

RLB #N CPU(TOTAL) RLB #N CPU(TOTAL) DELTA
Increases
R205.50 682.9 94 531.00 662.5 2479 3722.51 3191.51
R209.50 599.8 4 195.40 582.4 420 1302.49 1107.09
RC206.50 610.0 0 9.40 594.1 411 863.45 854.05
RC205.50 630.2 0 10.60 621.0 106 495.95 485.35
Decreases
R208.25 328.2 0 741.50 319.8 29 16.23 725.27
R204.25 350.5 16 171.60 332.2 83 29.70 141.90

Table 3.10: Comparing the DFJBC algorithm to solutions reported in Chabrier [18]

78
Key to Table 3.11, 3.15, and 3.18:

|A| : Number of arcs after preprocessing

SOLUTION . . . k : Number of vehicles in optimal solution
OPT : Value of an optimal solution. If the instance is not solved to optimality, the global lower bound glb
and global upper bound gub are given in the form [glb, gub]
GAP : Optimality gap in percent; calculated by (gub − glb)/glb · 100
ROOT . . . BOUNDS : Lower bound rlb and upper bound rub at the root LP
GAP : Optimality gap at the root node of the branch-and-cut (BC) tree in percent; calculated by (rub −
rlb)/rlb · 100
QUAL : Quality of lower bound at the root of the BC tree in percent; calculated by 100 − (gub − rlb)/rlb · 100
BC-TREE . . . #N : Number of processed nodes in the branch-and-cut tree
LEV : Depth of the BC tree
#CUTS : Number of generated cutting planes
#LP : Number of linear programs solved
CPU : Total CPU-time in seconds to solve problem to optimality. If the problem could not be solved within
a certain time limit we give the maximum CPU-time for computations before termination

Key to Table 3.12, 3.16, and 3.19:

SEC/ π : Number of generated SEC / Number of generated π inequalities / Number of calls of combined SEC
and π separation routine
Lifted π : Number of generated strong π inequalities in BC-nodes where lower and upper bound on number of
vehicles are both 1
σ : Number of generated σ inequalities / Number of calls of σ separation routine
Lifted σ : Number of generated strong σ inequalities in BC-nodes where lower and upper bound on number of
vehicles are both 1
(π , σ ) : Number of generated (π , σ ) inequalities / Number of calls of (π , σ ) separation routine
TOURN : Number of generated tournament constraints / Number of calls of tournament separation routine
S1 : Number of generated S1 inequalities / Number of calls of S1 separation routine
CAT : Number of generated odd CAT inequalities / Number of calls of CAT separation routine
D+
k : Number of generated D+ +
k inequalities / Number of calls of Dk separation routine
D−
k : Number of generated D− −
k inequalities / Number of calls of Dk separation routine

Key to Table 3.13:

#N : Total number of processed nodes / Total number of generated nodes

#NVEH : Number of processed vehicle branching nodes / Number of generated vehicle branching nodes
#NARC : Number of processed arc branching nodes / Number of generated arc branching nodes
k : Number of vehicles in optimal solution (or best incumbent when algorithm is terminated)
BC-TREE : In each generated interval of the lower and upper bound on the number of vehicles [lbk, lbk] we show
the total number of processed nodes vs. total number of generated nodes

Key to Table 3.14, 3.17, and 3.20:

INIT : Computing time spent in initialization phase (in %)

LP : Computing time spent in LP solver (in %)
SEP : Computing time spent in separation routines (in %)
MISC : Computing time spent in other parts of the program (in %)
TOTAL : Total CPU time in seconds

79
to R208.25 and R204.25 which both also have very few precedence relationships. Actually RC203.25
and RC204.25 have less precedence relationships than R208.25 but the semi-clustered distribution of the
customers make these problems very hard. The conclusion is therefore that the code performs better on
randomly distributed problems than semi-clustered problems and that problems with the special band-
structure of time windows are much harder than problems where the time windows are non-overlapping.
Table 3.12 shows that the precedence constraints not surprisingly are important for problems with
precedence constraints. In problems with little or no precedence structure the path inequalities are critical
in achieving a tight lower bound (R211.25, RC208.25) and improvements are needed here. The lifted
precedence constraints are seldomly generated and the CAT and Dk -constraints are mainly found in the
R-subclass.
In our algorithm we simultaneously determine the number of vehicles and the design of the corre-
sponding number of routes. In branch-and-cut algorithms for the CVRP it is standard to consider a fixed
number of vehicles. The purpose of Table 3.13 is to provide an overview of the computational work of
the separation of inequalities relative to the bounds on the number of vehicles. We note that for the class
2 problems with 25 customers a large proportion of the computational work is related to branch-and-
bound nodes where the number of vehicles leaving the depot is one, i.e. we are in fact trying to solve
an ATSPTW. We have taken advantage of this observation by implementing a separation routine for the
(π , σ )-inequalities, but for the hard (and unsolved) instances further work is required in the case where
we are considering one vehicle. We would also like to note that the optimal number of vehicles is always
greater than one (except for R208.25) so the purpose of solving nodes with x(N) = 1 efficiently is being
able to quickly prune that part of the search tree.
Table 3.14 gives the percentage of computing time spent in different parts of the algorithm. We can
conclude that the majority of computing time is spent in the separation routines. However, for the hard
problems (e.g. R211.25, RC204.25) a significant part of the computing time is spent in the LP solver.
The reason for this is simply that as the number of generated cutting planes increases the LP problems
become larger and more difficult.

Results for 50-node problems

In this section we present similar information as in the section with the 25-node results except we do not
show details on the branch-and-bound tree. For the 50-node problems the algorithm is only effective in
relation to the C-problems. However, the algorithm is overall an improvement compared to the algorithm
of Kallehauge et al. [14] but especially compared to the algorithm of Chabrier [7] it is clear that the bounds
are not tight enough for the R- and RC-problems compared to the bounds provided by the elementary
shortest path decomposition. It is not possible to compare in details the results for the C-problems with
[7], because the author does not provide computational times for this subclass. It is also clear that even for
these relatively small problems the separation of subtour and precedence constraints with a complexity of
O(n4 ) is a bottle-neck. Heuristic separation algorithms should be considered following the approach of
e.g. Bard et al. [6].

Results for 100-node problems

In this section we provide similar information as in the section with the results for the 50-node problems.
It is clear that the algorithm breaks down for the R- and RC-problems. In many cases it is not possible
to finish the root node within the 1 h limit. For the 100-node problems the algorithm is also inefficient in
relation to the C-problems compared to the algorithm of Kallehauge et al. [14]. However the bounds are
still tight for the C-problems and therefore we believe it is possible to achieve an acceptable performance
by introducing arc pricing.

80
 SOLUTION ROOT BC-TREE
|A| k OPT GAP BOUNDS GAP QUAL #N LEV #CUTS #LP CPU
R201 397 4 463.3 0.00 463.3 0.00 100.00 0 0 11 24 0.64
R202 518 4 410.5 0.00 [ 393.1, 1244.6] 216.60 95.58 85 17 308 589 23.89
R203 596 3 391.4 0.00 [ 373.0, 1244.6] 233.69 95.06 111 21 347 691 24.13
R204 627 2 355.0 0.00 [ 332.2, 1244.6] 274.64 93.14 68 10 309 576 17.60
R205 488 3 393.0 0.00 [ 390.6, 393.0] 0.61 99.39 2 1 26 93 1.20
R206 565 3 374.4 0.00 [ 356.6, 405.4] 13.70 94.99 82 12 302 597 19.29
R207 609 3 361.6 0.00 [ 347.4, 392.9] 13.09 95.92 54 16 284 543 15.33
R208 631 1 328.2 0.00 [ 319.9, 336.9] 5.33 97.39 17 9 161 324 7.33
R209 545 2 370.7 0.00 [ 364.2, 372.6] 2.30 98.22 17 10 120 289 7.32
R210 560 3 404.6 0.00 [ 389.0, 448.4] 15.27 95.99 64 17 309 571 21.60
R211 647 2 [ 347.7, 352.7] 1.44 [ 312.7, 1244.6] 297.96 87.22 12792 255 6704 22279 __*
C201 353 2 214.7 0.00 214.7 0.00 100.00 0 0 0 1 0.15
C202 498 2 214.7 0.00 [ 209.8, 214.7] 2.36 97.64 2 1 22 52 1.04
C203 584 2 214.7 0.00 [ 209.8, 214.7] 2.36 97.64 2 1 54 158 2.12
C204 622 1 213.1 0.00 [ 204.9, 227.8] 11.19 95.98 155 22 279 605 15.64
C205 391 2 214.7 0.00 214.7 0.00 100.00 0 0 7 8 0.35
C206 415 2 214.7 0.00 214.7 0.00 100.00 0 0 22 47 0.94
C207 463 2 214.5 0.00 [ 214.3, 277.0] 29.26 99.91 1 1 42 90 1.81
C208 438 2 214.5 0.00 214.5 0.00 100.00 0 0 35 75 1.38
RC201 401 3 360.2 0.00 360.2 0.00 100.00 0 0 10 21 0.51
RC202 522 3 338.0 0.00 [ 311.8, 372.8] 19.56 91.60 56 29 181 403 10.13
RC203 596 3 326.9 0.00 [ 264.0, 986.3] 273.62 76.17 14498 68 4629 21250 1834.85
RC204 627 3 [ 276.7, 300.3] 8.54 [ 246.7, 368.1] 49.19 78.29 15101 81 13723 31972 __*
RC205 483 3 338.0 0.00 338 0.00 100.00 0 0 33 120 1.30
RC206 492 3 324.0 0.00 324 0.00 100.00 0 0 39 134 1.55
RC207 561 3 298.3 0.00 [ 268.3, 1884.4] 602.28 88.83 765 35 474 1593 50.55
RC208 650 3 [ 230.8, 334.7] 45.02 [ 226.5, 1884.4] 732.12 52.20 11856 327 8836 23865 __*
__* : time limit of 1 CPU hour exceeded

Table 3.11: Computational results for the DFJBC algorithm (n = 25)

81
 Table 3.12: Number of generated cuts / Number of calls of separation routine (n = 25)
SEC/ π Lifted π σ Lifted σ (π , σ ) TOURN S1 CAT D+
k D−
k

R201 0/ 9/ 14 0/ 0 1/ 5 0/ 0 0/ 0 0/ 4 0/ 4 1/ 4 0/ 3 0/ 3
R202 2/ 129/ 379 5/ 49 85/ 191 0/ 3 52/ 69 10/ 106 16/ 96 9/ 80 0/ 71 0/ 71
R203 7/ 146/ 421 2/ 30 92/ 195 0/ 15 71/ 84 1/ 103 11/ 102 12/ 91 3/ 79 2/ 76
R204 16/ 141/ 362 0/ 28 58/ 148 0/ 17 57/ 83 1/ 90 13/ 89 13/ 76 7/ 63 3/ 56
R205 3/ 20/ 31 0/ 0 1/ 8 0/ 0 0/ 0 2/ 7 0/ 5 0/ 5 0/ 5 0/ 5
R206 5/ 108/ 369 2/ 22 105/ 221 0/ 0 33/ 35 6/ 116 10/ 110 25/ 100 6/ 75 2/ 69
R207 12/ 111/ 334 1/ 9 83/ 176 5/ 10 34/ 51 6/ 88 5/ 82 24/ 77 2/ 53 1/ 51
R208 20/ 76/ 183 0/ 3 30/ 85 0/ 0 2/ 2 0/ 55 5/ 55 16/ 50 8/ 34 4/ 26
R209 5/ 47/ 139 0/ 15 24/ 72 0/ 13 15/ 30 6/ 48 13/ 42 5/ 29 3/ 24 2/ 21
R210 9/ 122/ 371 1/ 39 71/ 181 0/ 12 58/ 90 8/ 110 20/ 102 17/ 82 1/ 65 2/ 64
R211 277/ 439/ 17650 0/ 166 398/ 16422 0/ 63 512/ 5171 1133/ 16024 3092/ 14891 640/ 11799 111/ 11159 102/ 11048
C201 0/ 0/ 1 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
C202 1/ 16/ 26 0/ 0 5/ 9 0/ 0 0/ 0 0/ 4 0/ 4 0/ 4 0/ 4 0/ 4
82

C203 7/ 40/ 61 0/ 1 5/ 13 0/ 0 1/ 1 1/ 8 0/ 7 0/ 7 0/ 7 0/ 7
C204 33/ 84/ 386 0/ 17 29/ 266 0/ 0 3/ 43 80/ 237 18/ 157 13/ 139 10/ 126 9/ 116
C205 0/ 7/ 8 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
C206 0/ 22/ 25 0/ 0 0/ 3 0/ 0 0/ 0 0/ 3 0/ 3 0/ 3 0/ 3 0/ 3
C207 0/ 30/ 48 0/ 0 0/ 18 0/ 0 0/ 0 9/ 18 2/ 9 0/ 7 0/ 7 1/ 7
C208 2/ 31/ 38 0/ 0 1/ 5 0/ 0 0/ 0 1/ 4 0/ 3 0/ 3 0/ 3 0/ 3
RC201 0/ 9/ 12 0/ 0 1/ 3 0/ 0 0/ 0 0/ 2 0/ 2 0/ 2 0/ 2 0/ 2
RC202 3/ 92/ 229 2/ 29 49/ 125 0/ 0 7/ 7 4/ 76 16/ 72 5/ 56 2/ 51 1/ 49
RC203 27/ 1238/ 12810 65/ 1563 446/ 8896 0/ 9 2584/ 2590 2/ 8450 141/ 8448 101/ 8307 11/ 8206 14/ 8195
RC204 89/ 3008/ 27272 324/ 4911 1242/ 15447 0/ 71 8404/ 8762 28/ 14205 370/ 14177 230/ 13807 14/ 13577 14/ 13563
RC205 2/ 25/ 34 0/ 0 6/ 7 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
RC206 5/ 30/ 41 0/ 0 4/ 6 0/ 0 0/ 0 0/ 2 0/ 2 0/ 2 0/ 2 0/ 2
RC207 23/ 155/ 913 3/ 41 131/ 692 1/ 13 40/ 60 18/ 560 34/ 542 43/ 508 14/ 465 12/ 451
RC208 536/ 0/ 20687 0/ 0 0/ 20151 0/ 0 0/ 2658 2806/ 20151 4376/ 17345 850/ 12969 134/ 12119 134/ 11985
Table 3.13: Number of processed branch-and-bound nodes / Number of generated nodes (n = 25)

#N #NVEH #NARC k BC-TREE

R202 85/ 134 9/ 12 76/ 122 4 [1,1] 1 / 3 [2,2] 33 / 51 [2,25] 16 / 19 [3,3] 17 / 30 [3,25] 17 / 27 [4,25] 1 / 4
R203 111/ 144 8/ 8 103/ 136 3 [1,1] 1 / 3 [2,2] 43 / 61 [2,25] 56 / 67 [3,25] 11 / 13 0 0
R204 68/ 96 2/ 2 65/ 94 2 [1,1] 9 / 13 [2,25] 58 / 83 0 0 0 0
R205 2/ 2 0/ 0 2/ 2 3 [1,25] 2 / 2 0 0 0 0 0
R206 82/ 124 10/ 10 72/ 114 3 [1,1] 1 / 3 [2,2] 13 / 22 [2,25] 55 / 81 [3,25] 13 / 18 0 0
R207 54/ 84 2/ 2 52/ 82 3 [1,1] 1 / 3 [2,25] 53 / 81 0 0 0 0
R208 17/ 28 2/ 2 15/ 26 1 [1,1] 1 / 1 [1,25] 14 / 22 [2,25] 2 / 5 0 0 0
R209 17/ 22 2/ 2 14/ 20 2 [1,1] 1 / 1 [2,25] 15 / 21 0 0 0 0
R210 64/ 114 9/ 14 55/ 100 3 [1,1] 1 / 3 [2,2] 14 / 24 [2,25] 47 / 79 [3,25] 2 / 8 0 0
83

R211 12792/ 21880 304/ 656 12488/ 21224 2 [1,1] 2000 / 3394 [1,25] 10607 / 17922 [2,25] 185 / 564 0 0 0
C202 2/ 2 2/ 2 0/ 0 2 [1,1] 1 / 1 [2,25] 1 / 1 0 0 0 0
C203 2/ 2 2/ 2 0/ 0 2 [1,1] 1 / 1 [2,25] 1 / 1 0 0 0 0
C204 155/ 204 11/ 12 144/ 192 1 [1,1] 25 / 32 [1,25] 125 / 166 [2,25] 5 / 6 0 0 0
C207 1/ 4 0/ 0 1/ 4 2 [1,25] 1 / 4 0 0 0 0 0
RC202 56/ 86 7/ 8 49/ 78 3 [1,1] 1 / 1 [2,2] 36 / 55 [2,25] 17 / 27 [3,25] 2 / 3 0 0
RC203 14498/ 16326 1547/ 1826 12951/ 14500 3 [1,1] 463 / 486 [1,2] 1 / 1 [1,25] 1211 / 1238 [2,2] 685 / 924 [2,25] 11848 / 13238 [3,25] 290 / 439
RC204 15101/ 27076 1632/ 6496 13469/ 20580 3 [1,1] 925 / 3413 [1,25] 12060 / 16350 [2,2] 30 / 95 [2,25] 2083 / 7177 [3,25] 3 / 41 0
RC207 765/ 870 40/ 40 725/ 830 3 [1,1] 2 / 4 [1,25] 2 / 2 [2,2] 156 / 196 [2,25] 587 / 644 [3,25] 18 / 24 0
RC208 11856/ 23692 21/ 42 11835/ 23650 3 [1,1] 947 / 1905 [1,25] 10909 / 21766 [2,25] 0 / 21 0 0 0
 INIT LP SEP MISC TOTAL
R201 14.1 4.7 73.4 7.8 0.64
R202 0.3 6.7 83.6 9.3 23.89
R203 0.3 9.4 70.2 20.1 24.13
R204 0.5 13.4 69.6 16.5 17.60
R205 6.7 8.3 78.3 6.7 1.20
R206 0.3 6.9 78.7 14.1 19.29
R207 0.5 8.3 74.2 17.0 15.33
R208 1.1 4.9 79.7 14.3 7.33
R209 1.0 3.4 86.3 9.3 7.32
R210 0.3 7.0 78.9 13.8 21.60
R211 0.0 58.2 39.1 2.7 3652.21
C201 53.3 0.0 33.3 13.3 0.15
C202 6.7 5.8 76.9 10.6 1.04
C203 3.8 5.2 82.1 9.0 2.12
C204 0.4 6.2 86.1 7.3 15.64
C205 20.0 0.0 74.3 5.7 0.35
C206 7.4 4.3 83.0 5.3 0.94
C207 3.9 3.9 85.6 6.6 1.81
C208 5.1 2.9 86.2 5.8 1.38
RC201 15.7 2.0 74.5 7.8 0.51
RC202 0.7 4.7 86.1 8.5 10.13
RC203 0.0 63.5 34.7 1.8 1834.85
RC204 0.0 60.0 37.4 2.6 3691.98
RC205 6.2 5.4 77.7 10.8 1.30
RC206 4.5 4.5 79.4 11.6 1.55
RC207 0.1 13.2 79.5 7.1 50.55
RC208 0.0 50.3 46.2 3.5 3643.42

Table 3.14: Percentage of computing time spent in different parts of the DFJBC algorithm (n = 25)

84
 SOLUTION ROOT BC-TREE
|A| k OPT GAP BOUNDS GAP QUAL #N LEV #CUTS #LP CPU
R201 1518 6 791.9 0.00 791.9 0.00 100.00 0 0 32 90 16.09
R202 1984 50 [ 638.8, 2620.8] 310.25 [ 635.8, 2620.8] 312.24 -212.24 1927 617 2879 6362 __*
R203 2295 50 [ 552.5, 2620.8] 374.35 [ 541.1, 2620.8] 384.31 -284.31 1893 507 3477 6698 __*
R204 2514 4 [ 483.5, 543.2] 12.34 [ 471.8, 2620.8] 455.48 84.87 2570 921 3604 7548 __*
R205 1878 4 690.1 0.00 [ 662.5, 707.6] 6.81 95.83 2320 32 772 4502 1527.47
R206 2183 4 [ 596.4, 722.4] 21.13 [ 572.1, 2620.8] 358.09 73.73 2209 510 3007 7310 __*
R207 2380 3 [ 532.8, 610.1] 14.52 [ 520.9, 2620.8] 403.13 82.88 2419 873 3199 6846 __*
R208 2525 3 [ 471.6, 534.4] 13.31 [ 464.1, 2620.8] 464.65 84.86 3514 500 3070 8167 __*
R209 2150 4 600.6 0.00 [ 582.4, 2620.8] 349.98 96.88 626 124 883 2253 816.31
R210 2188 4 [ 594.0, 670.9] 12.95 [ 582.7, 2620.8] 349.79 84.86 2001 517 3221 6592 __*
R211 2523 50 [ 467.8, 2620.8] 460.26 [ 464.6, 2620.8] 464.12 -364.12 2466 854 3351 6962 __*
C201 1339 3 360.2 0.00 360.2 0.00 100.00 0 0 0 1 0.83
C202 1890 3 360.2 0.00 [ 354.1, 382.1] 7.92 98.26 13 6 117 174 62.55
C203 2253 3 359.8 0.00 [ 352.3, 359.8] 2.14 97.86 23 9 159 354 75.28
C204 2505 2 350.1 0.00 [ 347.8, 363.1] 4.40 99.34 113 63 411 759 214.02
C205 1476 3 359.8 0.00 359.8 0.00 100.00 0 0 22 65 11.11
C206 1589 3 359.8 0.00 359.8 0.00 100.00 0 0 33 118 15.22
C207 1814 3 359.6 0.00 [ 359.4, 366.2] 1.89 99.94 2 2 75 208 37.48
C208 1691 2 350.5 0.00 350.5 0.00 100.00 0 0 34 129 17.02
RC201 1495 5 684.8 0.00 [ 680.1, 775.6] 14.05 99.30 8 3 67 166 42.30
RC202 1971 5 [ 591.2, 621.9] 5.20 [ 515.2, 872.8] 69.42 79.28 2657 265 2399 6536 __*
RC203 2291 50 [ 480.2, 4056.0] 744.65 [ 446.4, 4056.0] 808.52 -708.52 2168 792 3156 6634 __*
RC204 2513 3 [ 396.9, 499.3] 25.80 [ 389.6, 4056.0] 940.97 71.86 3958 790 2713 8250 __*
RC205 1834 5 630.2 0.00 [ 621.0, 738.6] 18.94 98.51 90 13 212 547 149.71
RC206 1860 5 610.0 0.00 [ 594.1, 4056.0] 582.69 97.33 470 22 298 1151 329.28
RC207 2173 50 [ 500.7, 4056.0] 710.09 [ 495.2, 4056.0] 719.11 -619.11 2459 963 2268 6062 __*
RC208 2548 50 [ 373.6, 4056.0] 985.76 [ 372.4, 4056.0] 989.07 -889.07 3079 785 3161 7373 __*
__* : time limit of 1 CPU hour exceeded

Table 3.15: Computational results for the DFJBC algorithm (n = 50)

85
 Table 3.16: Number of generated cuts / Number of calls of separation routine (n = 50)
SEC/ π Lifted π σ Lifted σ (π , σ ) TOURN S1 CAT D+
k D−
k

R201 2/ 24/ 34 0/ 0 5/ 8 0/ 0 0/ 0 1/ 3 0/ 2 0/ 2 0/ 2 0/ 2
R202 7/ 1407/ 4802 25/ 786 974/ 3101 2/ 25 262/ 290 32/ 2125 120/ 2093 37/ 1973 9/ 1936 4/ 1927
R203 15/ 1819/ 5358 13/ 486 1286/ 3306 0/ 13 205/ 221 32/ 2020 65/ 1988 26/ 1923 10/ 1897 6/ 1887
R204 106/ 1421/ 6173 31/ 545 893/ 3921 0/ 57 694/ 803 16/ 3028 38/ 3012 318/ 2974 50/ 2656 37/ 2606
R205 9/ 230/ 2173 0/ 0 141/ 1934 0/ 0 0/ 0 69/ 1793 166/ 1724 130/ 1558 17/ 1428 10/ 1411
R206 17/ 1265/ 5210 8/ 170 1108/ 3828 0/ 10 92/ 108 62/ 2720 310/ 2658 116/ 2348 18/ 2232 11/ 2214
R207 47/ 1682/ 5609 9/ 68 1034/ 3820 0/ 12 51/ 59 67/ 2786 196/ 2719 81/ 2523 19/ 2442 13/ 2423
R208 175/ 1146/ 6583 6/ 448 502/ 4833 3/ 89 423/ 538 121/ 4328 149/ 4207 439/ 4058 58/ 3619 48/ 3561
R209 23/ 355/ 1295 0/ 0 226/ 917 0/ 0 0/ 0 44/ 691 130/ 647 70/ 517 18/ 447 17/ 429
R210 21/ 1544/ 5215 5/ 108 887/ 3508 1/ 48 137/ 162 132/ 2620 421/ 2488 63/ 2067 6/ 2004 4/ 1998
R211 120/ 736/ 5811 394/ 1297 301/ 4414 288/ 560 147/ 4940 365/ 3825 790/ 3460 131/ 2670 52/ 2539 27/ 2487
C201 0/ 0/ 1 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
C202 2/ 68/ 131 0/ 0 28/ 61 0/ 0 0/ 0 6/ 33 11/ 27 0/ 16 2/ 16 0/ 14
86

C203 14/ 101/ 178 0/ 0 34/ 63 0/ 0 0/ 0 3/ 29 6/ 26 1/ 20 0/ 19 0/ 19

C204 37/ 137/ 515 0/ 0 87/ 341 0/ 0 0/ 0 39/ 254 30/ 215 39/ 185 24/ 146 18/ 122
C205 0/ 21/ 24 0/ 0 0/ 3 0/ 0 0/ 0 1/ 3 0/ 2 0/ 2 0/ 2 0/ 2
C206 4/ 29/ 34 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
C207 3/ 62/ 83 0/ 0 1/ 18 0/ 0 0/ 0 4/ 17 2/ 13 2/ 11 1/ 9 0/ 8
C208 4/ 27/ 37 0/ 0 3/ 6 0/ 0 0/ 0 0/ 3 0/ 3 0/ 3 0/ 3 0/ 3
RC201 2/ 40/ 83 0/ 0 2/ 41 0/ 0 0/ 0 12/ 39 8/ 27 3/ 19 0/ 16 0/ 16
RC202 16/ 1112/ 4916 36/ 824 762/ 3692 0/ 0 60/ 60 28/ 2930 342/ 2902 26/ 2560 10/ 2534 7/ 2524
RC203 32/ 1526/ 5319 28/ 1025 819/ 3309 0/ 5 424/ 435 61/ 2490 168/ 2429 65/ 2261 22/ 2196 11/ 2174
RC204 117/ 1004/ 6672 22/ 371 638/ 5103 3/ 72 426/ 477 46/ 4462 118/ 4416 244/ 4298 56/ 4054 39/ 3998
RC205 6/ 112/ 292 0/ 0 43/ 174 0/ 0 0/ 0 24/ 131 20/ 107 6/ 87 0/ 81 1/ 81
RC206 15/ 126/ 586 0/ 0 77/ 445 0/ 0 0/ 0 12/ 368 34/ 356 22/ 322 5/ 300 7/ 295
RC207 22/ 761/ 4724 15/ 169 651/ 3563 3/ 45 363/ 808 34/ 2909 299/ 2875 67/ 2576 29/ 2509 24/ 2480
RC208 308/ 169/ 6228 31/ 184 121/ 5602 17/ 85 118/ 6670 744/ 5464 1344/ 4720 181/ 3376 83/ 3195 45/ 3112
 INIT LP SEP MISC TOTAL
R201 1.0 1.5 96.0 1.6 16.09
R202 0.0 5.2 89.2 5.6 3710.71
R203 0.0 11.4 79.3 9.3 3675.22
R204 0.0 15.9 76.5 7.5 3660.86
R205 0.0 10.6 86.9 2.5 1527.47
R206 0.0 13.4 81.9 4.7 3672.94
R207 0.0 10.7 77.7 11.6 3657.81
R208 0.0 12.4 83.6 4.0 3662.96
R209 0.0 9.3 85.3 5.3 816.31
R210 0.0 9.3 81.9 8.8 3678.79
R211 0.0 8.4 81.5 10.1 3709.52
C201 20.5 1.2 72.3 6.0 0.83
C202 0.3 0.7 97.5 1.5 62.55
C203 0.2 1.2 96.2 2.3 75.28
C204 0.1 2.0 94.5 3.4 214.02
C205 1.6 1.0 95.3 2.1 11.11
C206 1.1 1.2 95.0 2.8 15.22
C207 0.4 1.1 96.7 1.7 37.48
C208 1.0 1.1 95.4 2.5 17.02
RC201 0.4 0.7 97.5 1.3 42.30
RC202 0.0 17.3 80.4 2.4 3647.85
RC203 0.0 7.5 84.1 8.4 3682.08
RC204 0.0 9.0 84.7 6.3 3653.89
RC205 0.1 1.2 97.0 1.7 149.71
RC206 0.1 2.0 95.5 2.4 329.28
RC207 0.0 4.5 89.5 6.0 3759.56
RC208 0.0 9.8 82.0 8.2 3639.20

Table 3.17: Percentage of computing time spent in different parts of the DFJBC algorithm (n = 50)

87
 SOLUTION ROOT BC-TREE
|A| k OPT GAP BOUNDS GAP QUAL #N LEV #CUTS #LP CPU
R201 5917 9 [ 1132.7, 1155.6] 2.03 [ 1123.6, 4980.0] 343.23 97.15 72 26 295 666 __*
R202 7718 100 [ 888.6, 4980.0] 460.42 [ 888.6, 4980.0] 460.42 -360.42 0 0 464 1165 __*
R203 9050 100 [ 748.1, 4980.0] 565.70 [ 748.1, 4980.0] 565.70 -465.70 0 0 538 1219 __*
R204 9845 100 [ 661.9, 4980.0] 652.39 [ 661.9, 4980.0] 652.41 -552.41 14 11 563 1070 __*
R205 7327 100 [ 900.0, 4980.0] 453.33 [ 899.7, 4980.0] 453.53 -353.53 7 4 402 865 __*
R206 8521 100 [ 783.6, 4980.0] 535.52 [ 783.6, 4980.0] 535.53 -435.53 2 2 463 1108 __*
R207 9378 100 [ 714.8, 4980.0] 596.69 [ 714.8, 4980.0] 596.70 -496.70 2 2 543 1231 __*
R208 9936 100 [ 651.8, 4980.0] 664.02 [ 651.6, 4980.0] 664.29 -564.29 21 16 571 1065 __*
R209 8518 100 [ 785.8, 4980.0] 533.79 [ 785.2, 4980.0] 534.22 -434.22 8 5 476 1010 __*
R210 8568 100 [ 798.3, 4980.0] 523.85 [ 798.2, 4980.0] 523.87 -423.87 5 4 456 1074 __*
R211 9997 100 [ 645.1, 4980.0] 671.99 [ 645.1, 4980.0] 672.01 -572.01 37 31 553 958 __*
C201 5221 3 589.1 0.00 589.1 0.00 100.00 0 0 0 1 11.48
C202 7350 3 589.1 0.00 589.1 0.00 100.00 0 0 26 70 202.86
C203 8866 4 [ 586.0, 632.3] 7.90 [ 584.4, 643.8] 10.17 91.79 66 19 447 821 __*
C204 9789 3 [ 584.4, 597.1] 2.18 [ 583.5, 5934.4] 917.06 97.67 95 47 509 929 __*
C205 5698 3 586.4 0.00 586.4 0.00 100.00 0 0 38 85 334.44
C206 6212 3 586.0 0.00 586 0.00 100.00 0 0 50 158 418.99
C207 6578 3 585.8 0.00 [ 585.6, 585.8] 0.03 99.97 1 1 57 212 527.52
C208 6665 3 585.8 0.00 585.8 0.00 100.00 0 0 65 206 569.74
RC201 5918 7 [ 1250.1, 1288.2] 3.05 [ 1249.2, 6609.4] 429.09 96.88 61 48 309 693 __*
RC202 7752 100 [ 940.1, 6609.4] 603.03 [ 940.1, 6609.4] 603.03 -503.03 0 0 467 1071 __*
RC203 9056 100 [ 781.6, 6609.4] 745.67 [ 781.6, 6609.4] 745.67 -645.67 1 1 522 1239 __*
RC204 9854 100 [ 692.7, 6609.4] 854.19 [ 692.7, 6609.4] 854.21 -754.21 18 13 562 1043 __*
RC205 7173 100 [ 1081.7, 6609.4] 511.00 [ 1081.5, 6609.4] 511.12 -411.12 7 5 414 893 __*
RC206 7366 100 [ 974.8, 6609.4] 578.01 [ 974.8, 6609.4] 578.01 -478.01 1 1 416 955 __*
RC207 8619 100 [ 832.4, 6609.4] 694.03 [ 832.4, 6609.4] 694.03 -594.03 0 0 480 1095 __*
RC208 10091 100 [ 647.7, 6609.4] 920.43 [ 647.7, 6609.4] 920.44 -820.44 226 221 372 907 __*
__* : time limit of 1 CPU hour exceeded

Table 3.18: Computational results for the DFJBC algorithm (n = 100)

88
 Table 3.19: Number of generated cuts / Number of calls of separation routine (n = 100)
SEC/ π Lifted π σ Lifted σ (π , σ ) TOURN S1 CAT D+
k D−
k

R201 5/ 120/ 367 0/ 0 60/ 242 0/ 0 0/ 0 35/ 182 37/ 147 34/ 110 3/ 76 1/ 73
R202 5/ 383/ 464 0/ 0 76/ 76 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0
R203 13/ 439/ 538 0/ 0 86/ 86 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0
R204 33/ 357/ 578 0/ 0 127/ 188 0/ 0 0/ 0 3/ 61 8/ 58 18/ 50 9/ 32 8/ 23
R205 14/ 221/ 411 0/ 0 104/ 176 0/ 0 0/ 0 5/ 72 33/ 67 21/ 34 2/ 13 2/ 11
R206 9/ 314/ 467 0/ 0 118/ 144 0/ 0 0/ 0 2/ 26 9/ 24 7/ 15 1/ 8 3/ 7
R207 23/ 369/ 547 0/ 0 126/ 155 0/ 0 0/ 0 0/ 29 7/ 29 11/ 22 5/ 11 2/ 6
R208 50/ 294/ 594 0/ 0 143/ 250 0/ 0 0/ 0 0/ 107 7/ 107 52/ 100 19/ 48 6/ 29
R209 22/ 251/ 486 0/ 0 171/ 213 0/ 0 0/ 0 5/ 42 20/ 37 6/ 17 0/ 11 1/ 11
R210 9/ 273/ 463 0/ 0 156/ 181 0/ 0 0/ 0 2/ 25 12/ 23 4/ 11 0/ 7 0/ 7
R211 62/ 223/ 592 0/ 0 79/ 307 0/ 0 0/ 0 20/ 228 77/ 208 48/ 131 26/ 83 18/ 57
C201 0/ 0/ 1 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
C202 0/ 26/ 27 0/ 0 0/ 1 0/ 0 0/ 0 0/ 1 0/ 1 0/ 1 0/ 1 0/ 1
89

C203 28/ 228/ 509 0/ 0 44/ 253 0/ 0 0/ 0 98/ 209 37/ 111 8/ 74 2/ 66 2/ 64
C204 53/ 225/ 603 0/ 0 60/ 325 0/ 0 0/ 0 73/ 265 52/ 192 23/ 140 17/ 117 6/ 100
C205 1/ 35/ 41 0/ 0 0/ 5 0/ 0 0/ 0 1/ 5 1/ 4 0/ 3 0/ 3 0/ 3
C206 5/ 42/ 52 0/ 0 2/ 5 0/ 0 0/ 0 0/ 3 1/ 3 0/ 2 0/ 2 0/ 2
C207 1/ 47/ 63 0/ 0 3/ 15 0/ 0 0/ 0 2/ 12 3/ 10 0/ 7 0/ 7 1/ 7
C208 9/ 48/ 70 0/ 0 3/ 13 0/ 0 0/ 0 2/ 10 1/ 8 2/ 7 0/ 5 0/ 5
RC201 3/ 138/ 372 0/ 0 51/ 231 0/ 0 0/ 0 35/ 180 47/ 145 26/ 98 5/ 72 4/ 67
RC202 12/ 373/ 467 0/ 0 82/ 82 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0 0/ 0
RC203 12/ 371/ 525 0/ 0 115/ 142 0/ 0 0/ 0 2/ 27 14/ 25 3/ 11 1/ 8 4/ 7
RC204 30/ 314/ 582 0/ 0 150/ 238 0/ 0 0/ 0 1/ 88 28/ 87 21/ 59 9/ 38 9/ 29
RC205 2/ 263/ 423 0/ 0 103/ 158 0/ 0 0/ 0 5/ 55 34/ 50 4/ 16 3/ 12 0/ 9
RC206 17/ 243/ 419 0/ 0 98/ 159 0/ 0 0/ 0 7/ 61 45/ 54 2/ 9 4/ 7 0/ 3
RC207 16/ 303/ 481 0/ 0 134/ 162 0/ 0 0/ 0 3/ 28 15/ 25 3/ 10 2/ 7 4/ 5
RC208 115/ 43/ 600 0/ 0 25/ 442 0/ 0 0/ 0 23/ 417 84/ 394 41/ 310 22/ 269 19/ 247
 INIT LP SEP MISC TOTAL
R201 0.0 0.2 99.1 0.7 3616.18
R202 0.0 0.9 98.6 0.5 3615.42
R203 0.0 1.6 97.8 0.6 3623.32
R204 0.0 1.9 93.0 5.0 3616.80
R205 0.0 0.7 96.1 3.3 3655.33
R206 0.0 1.2 94.1 4.7 3655.12
R207 0.0 1.9 92.5 5.6 3736.64
R208 0.0 1.6 93.9 4.5 3635.19
R209 0.0 1.0 95.1 3.9 3872.42
R210 0.0 1.2 94.2 4.6 3650.14
R211 0.0 1.5 94.3 4.2 3621.49
C201 4.4 0.3 94.3 1.0 11.48
C202 0.3 0.3 99.0 0.5 202.86
C203 0.0 0.3 99.2 0.5 3605.29
C204 0.0 0.6 98.1 1.3 3605.84
C205 0.2 0.2 99.4 0.3 334.44
C206 0.1 0.2 99.1 0.5 418.99
C207 0.1 0.2 99.1 0.6 527.52
C208 0.1 0.2 99.2 0.5 569.74
RC201 0.0 0.2 99.1 0.7 3617.12
RC202 0.0 0.9 98.7 0.4 3614.09
RC203 0.0 1.7 93.4 4.9 3740.49
RC204 0.0 1.8 92.8 5.3 3634.98
RC205 0.0 0.5 97.3 2.2 3665.94
RC206 0.0 0.7 96.4 2.9 3626.73
RC207 0.0 1.2 96.5 2.2 3615.31
RC208 0.0 0.8 95.9 3.4 3614.08

Table 3.20: Percentage of computing time spent in different parts of the DFJBC algorithm (n = 100)

90
3.10 Conclusions
In this paper we presented a new formulation of the VRPTW involving only binary arc variables. The
new formulation is based on the formulation of the ATSPTW by Ascheuer et al. [3] and has the advantage
of avoiding additional variables and linking constraints. In the new formulation of the VRPTW time win-
dows are modeled using path inequalities. A path inequality eliminates a path that is infeasible because of
some deadline or vehicle capacity is violated. We presented a new class of strengthened path inequalities
based on the polyhedral results obtained by Mak [17] in the context of the TSP with replenishment arcs.
We studied the VRPTW polytope and determined the polytope dimension. We shoved that the new class
of path inequalities is facet defining under reasonable assumptions. These are the first polyhedral results
for the VRPTW. We introduced precedence constraints in the context of the VRPTW. We designed a
branch-and-cut algorithm for the exact solution of the VRPTW and evaluated the computational perfor-
mance on the long-horizon Solomon test problems. The outcome is based on 25-node problems that the
algorithm shows promising results compared to leading algorithms in the literature. In particular we re-
port a solution to a previously unsolved 50-node Solomon test problem R208. The conclusion is therefore
that the path pricing algorithm is no longer the unchallenged winning strategy for solving the VRPTW.

Acknowledgements We would like to thank Dr. Vicky Mak of Deakin University, Australia, for all the
fruitful discussions regarding this paper. We would also like to thank Dr. Irina Dumitrescu of HEC Mon-
treal, Canada, for giving corrections and suggestions improving this paper. The computational platform
for carrying out the tests on the Sun Fire 15K machine were provided by The Danish Center for Scientific
Computing.

References
[1] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network Flows. Prentice Hall, New Jersey, 1993. [71]
[2] D. Applegate, R. Bixby, V. Chvátal, and W. Cook. Implementing the Dantzig-Fulkerson-Johnson
algorithm for large traveling salesman problems. Mathematical Programming Series B, 97:91–153,
2003. [71, 74]
[3] N. Ascheuer, M. Fischetti, and M. Grötschel. A polyhedral study of the asymmetric traveling sales-
man problem with time windows. Networks, 36:69–79, 2000. [57, 58, 59, 62, 91]
[4] N. Ascheuer, M. Fischetti, and M. Grötschel. Solving the asymmetric travelling salesman problem
with time windows by branch-and-cut. Mathematical Programming Series A, 90:475–506, 2001.
[58, 59, 65, 68, 71]
[5] E. Balas, M. Fischetti, and W. R. Pulleyblank. The precedence-constrained asymmetric traveling
salesman polytope. Mathematical Programming, 68:241–265, 1995. [58, 66, 71]
[6] J. F. Bard, G. Kontoravdis, and G. Yu. A branch-and-cut procedure for the vehicle routing problem
with time windows. Transportation Science, 36:250–269, 2002. [58, 59, 65, 80]
[7] A. Chabrier. Vehicle routing problem with elementary shortest path based column generation. Tech-
nical report, ILOG, 2003. [74, 75, 78, 80]
[8] V. Chvátal. Linear Programming. Freeman, New York, 1983. [71]
[9] M. Desrochers, J. Desrosiers, and M. Solomon. A new optimization algorithm for the vehicle routing
problem with time windows. Operations Research, 40:342–354, 1992. [57, 68]

91
[10] M. Fischetti and P. Toth. A polyhedral approach to the asymmetric traveling salesman problem.
Management Science, 43:1520–1536, 1997. [67, 72]
[11] ILOG. ILOG CPLEX 9.1 User’s Manual, 2005. [69, 71, 72]
[12] S. Irnich and D. Villeneuve. The shortest path problem with k-cycle elimination (k > 3): Improving
a branch-and-price algorithm for the VRPTW. Technical report, Lehr- und Forschungsgebiet Oper-
ations Research und Logistik Management, Rheinisch-Westfälische Technische Hochschule, 2003.
[74, 75, 76, 77]
[13] M. Jünger and S. Thienel. The ABACUS system for branch-and-cut-and-price algorithms in integer
programming and combinatorial optimization. Software: Practice and Experience, 30:1325–1352,
2000. [71]
[14] B. Kallehauge, J. Larsen, and O. B. G. Madsen. Lagrangian duality applied to the vehicle routing
problem with time windows. Computers & Operations Research, 33:1464–1487, 2006. [74, 75, 77,
78, 80]
[15] N. Kohl, J. Desrosiers, O. B. G. Madsen, M. M. Solomon, and F. Soumis. 2-path cuts for the vehicle
routing problem with time windows. Transportation Science, 33:101–116, 1999. [67]
[16] G. Kontoravdis and J. F. Bard. A GRASP for the vehicle routing problem with time windows. ORSA
Journal on Computing, 7:10–23, 1995. [68]
[17] V. H. Mak. On the Asymmetric Travelling Salesman Problem with Replenishment Arcs. PhD thesis,
Department of Mathematics and Statistics, The University of Melbourne, 2001. [57, 58, 60, 61, 91]
[18] C. E. Miller, A. W. Tucker, and R. A. Zemlin. Integer programming formulation of traveling sales-
man problems. Journal of the Association for Computing Machinery, 7:326–329, 1960. [58]
[19] M. W. Savelsbergh. Local search in routing problems with time windows. Annals of Operations
Research, 4:285–305, 1985. [58]

[20] M. M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35:254–65, 1987. [59, 68]
[21] L. A. Wolsey. Integer Programming. John Wiley & Sons, New York, 1998. [71]

92
Chapter 4

Vehicle routing problem with time

windows

Brian Kallehauge
Centre for Traffic and Transport, Technical University of Denmark

Jesper Larsen
Informatics and Mathematical Modelling, Technical University of Denmark

Oli B. G. Madsen
Centre for Traffic and Transport, Technical University of Denmark

Marius M. Solomon
College of Business Administration, Northeastern University and GERAD

Abstract
In this chapter we discuss the vehicle routing problem with time windows in terms of its mathemati-
cal modeling, its structure and decomposition alternatives. We then present the master problem and the
subproblem for the column generation approach, respectively. Next, we illustrate a branch-and-bound
framework and address acceleration strategies used to increase the efficiency of branch-and-price meth-
ods. Then, we describe generalizations of the problem and report computational results for the classic
Solomon test sets. Finally, we present our conclusions and discuss some open problems.

4.1 Introduction
The vehicle routing problem (VRP) involves finding a set of routes, starting and ending at a depot, that
together cover a set of customers. Each customer has a given demand, and no vehicle can service more
customers than its capacity permits. The objective is to minimize the total distance traveled or the number
of vehicles used, or a combination of these. In this chapter, we consider the vehicle routing problem with
time windows (VRPTW), which is a generalization of the VRP where the service at any customer starts
within a given time interval, called a time window. Time windows are called soft when they can be
considered non-biding for a penalty cost. They are hard when they cannot be violated, i.e., if a vehicle

93
arrives too early at a customer, it must wait until the time window opens; and it is not allowed to arrive
late. This is the case we consider here.
The remarkable advances in information technology have enabled companies to focus on efficiency
and timeliness throughout the supply chain. In turn, the VRPTW has increasingly become an invaluable
tool in modeling a variety of aspects of supply chain design and operation. Important VRPTW applica-
tions include deliveries to supermarkets, bank and postal deliveries, industrial refuse collection, school
bus routing, security patrol service, and urban newspaper distribution. Its increased practical visibility has
evolved in parallel with the development of broader and deeper research directed at its solution. Signifi-
cant progress has been made in both the design of heuristics and the development of optimal approaches.
In this chapter we will concentrate on exact methods for the VRPTW based on column generation.
These date back to Desrochers, Desrosiers, and Solomon [12] who used column generation in a Dantzig-
Wolfe decomposition framework and Halse [20] who implemented a decomposition based on variable
splitting (also known as Lagrangean decomposition). Later, Kohl and Madsen [29] developed an al-
gorithm exploiting Lagrangean relaxation. Then, Kohl, Desrosiers, Madsen, Solomon, and Soumis [30],
Larsen [35], and Cook and Rich [7] extended the previous approaches by developing Dantzig-Wolfe based
decomposition algorithms involving cutting planes and/or parallel platforms. Kallehauge [25] suggested
a hybrid algorithm based on a combination of Lagrangean relaxation and Dantzig-Wolfe decomposition.
Recently, Chabrier [5], Chabrier, Danna, and Le Pape [6], Feillet, Dejax, Gendreau, and Gueguen [18],
Irnich and Villeneuve [23], and Rousseau, Gendreau, and Pesant [38] have proposed algorithms based on
enhanced subproblem methodology. Advancements in master problem approaches have been made by
Danna and Le Pape [10] and Larsen [34].
This chapter has the following organization. In section 4.2 we describe the mathematical model of
the VRPTW and in section 4.3 we discuss the structure of the problem and decomposition alternatives.
Next, sections 4.4 and 4.5 present the master problem and the subproblem for the column generation ap-
proach, respectively. Section 4.6 illustrates the branch-and-bound framework, while section 4.7 addresses
acceleration strategies used to increase the efficiency of branch-and-price methods. Then, we describe
generalizations of the VRPTW in section 4.8 and report computational results for the classic Solomon
test sets in section 4.9. Finally we present our conclusions and discuss some open problems in 4.10.

4.2 The model

The VRPTW is defined by a fleet of vehicles, V , a set of customers, C , and a directed graph G . Typically
the fleet is considered to be homogeneous, that is, all vehicles are identical. The graph consists of |C | + 2
vertices, where the customers are denoted 1, 2, . . . , n and the depot is represented by the vertices 0 (“the
starting depot”) and n + 1 (“the returning depot”). The set of all vertices, that is, 0, 1, . . . , n + 1 is denoted
N . The set of arcs, A , represents direct connections between the depot and the customers and among
the customers. There are no arcs ending at vertex 0 or originating from vertex n + 1. With each arc (i, j),
where i 6= j, we associate a cost ci j and a time ti j , which may include service time at customer i.
Each vehicle has a capacity q and each customer i a demand di . Each customer i has a time window
[ai , bi ] and a vehicle must arrive at the customer before bi . If it arrives before the time window opens, it
has to wait until ai to service the customer. The time windows for both depots are assumed to be identical
to [a0 , b0 ] which represents the scheduling horizon. The vehicles may not leave the depot before a0 and
must return at the latest at time bn+1 .
It is assumed that q, ai , bi , di , ci j are non-negative integers and ti j are positive integers. Note that this
assumption is necessary to develop an algorithm for the shortest path with resource constraints used in
the column generation approach presented later. Furthermore it is assumed that the triangle inequality is
satisfied for both ci j and ti j .
The model contains two sets of decision variables x and s. For each arc (i, j), where i 6= j, i 6= n+1, j 6=

94
0, and each vehicle k we define xi jk as

1, if vehicle k drives directly from vertex i to vertex j
xi jk =
0, otherwise
The decision variable sik is defined for each vertex i and each vehicle k and denotes the time vehicle
k starts to service customer i. In case vehicle k does not service customer i, sik has no meaning and
consequently it’s value is considered irrelevant. We assume a0 = 0 and therefore s0k = 0, for all k.
The goal is to design a set of routes that minimizes total cost, such that
• each customer is serviced exactly once,
• every route originates at vertex 0 and ends at vertex n + 1, and
• the time windows of the customers and capacity constraints of the vehicles are observed.
This informal VRPTW description can be stated mathematically as a multicommodity network flow prob-
lem with time windows and capacity constraints:
(4.1) min ∑ ∑ ∑ ci j xi jk s.t.
k∈V i∈N j∈N

(4.2) ∑ ∑ xi jk = 1 ∀i ∈ C
k∈V j∈N

(4.3) ∑ di ∑ xi jk ≤ q ∀k ∈ V
i∈C j∈N

(4.4) ∑ x0 jk = 1 ∀k ∈ V
j∈N

(4.5) ∑ xihk − ∑ xh jk = 0 ∀h ∈ C , ∀k ∈ V
i∈N j∈N

(4.6) ∑ xi,n+1,k = 1 ∀k ∈ V
i∈N
xi jk (sik + ti j − s jk ) ≤ 0
(4.7) ∀i, j ∈ N , ∀k ∈ V
(4.8) ai ≤ sik ≤ bi ∀i ∈ N , ∀k ∈ V
(4.9) xi jk ∈ {0, 1} ∀i, j ∈ N , ∀k ∈ V
The objective function (4.1) minimizes the total travel cost. The constraints (4.2) ensure that each
customer is visited exactly once, and (4.3) state that a vehicle can only be loaded up to it’s capacity. Next,
equations (4.4), (4.5) and (4.6) indicate that each vehicle must leave the depot 0; after a vehicle arrives at
a customer it has to leave for another destination; and finally, all vehicles must arrive at the depot n + 1.
The inequalities (4.7) establish the relationship between the vehicle departure time from a customer and
its immediate successor. Finally constraints (4.8) affirm that the time windows are observed, and (4.9) are
the integrality constraints. Note that an unused vehicle is modeled by driving the “empty” route (0, n + 1).
The model can also incorporate a constraint giving an upper bound on the number of vehicles, as is
the case in Desrosiers, Dumas, Solomon, and Soumis [14]:
(4.10) ∑ ∑ x0 jk ≤|V | ∀k ∈ V , ∀ j ∈ N
k∈V j∈N

Note also that the nonlinear restrictions (4.7) can be linearized as:
(4.11) sik + ti j − Mi j (1 − xi jk )≤s jk ∀i, j ∈ N , ∀k ∈ V

95
The large constants Mi j can be decreased to max{bi + ti j − a j }, (i, j) ∈ A.
For each vehicle, the service start variables impose a unique route direction thereby eliminating any
subtours. Hence, the classical VRP subtour elimination constraints become redundant. Finally, the ob-
jective function (4.1) has been universally used when solving the VRPTW to optimality. In the research
on heuristics it has been common to minimize the number of vehicles which may lead to additional travel
cost.
The VRPTW is a generalization of both the traveling salesman problem (TSP) and the VRP. When
the time constraints (4.7) and (4.8)) are not binding the problem relaxes to a VRP. This can be modeled
by setting ai = 0 and bi = M, where M is a large scalar, for all customers i. If only one vehicle is available
the problem becomes a TSP. If several vehicles are available and the cost structure is: c0 j = 1, j ∈ C
and ci j = 0, otherwise, we obtain the bin-packing problem. Since trips between customers are “free”,
the order in which these are visited becomes unimportant and the objective turns to “squeezing” as much
demand as possible into as few vehicles (bins) as possible. In case the capacity constraints (4.2) are not
binding the problem becomes a m-TSPTW, or, if only one vehicle is available, a TSPTW.

4.3 Structure and decomposition

A closer look at the above model reveals that only the assignment constraints (4.2) are coupling the
vehicles while the remaining constraints are dealing with each vehicle separately. This strongly suggests
the use of Lagrangean relaxation (LR) or decomposition, for example Dantzig-Wolfe (DWD), to break up
the overall problem into a subproblem for each vehicle and a master problem. To date, the most successful
decomposition approaches for the VRPTW cast the subproblem as a constrained shortest path structure.
The master problem is an integer program whose solution cannot be obtained directly, so its LP relaxation
is solved. The column generation process alternates between solving this linear master problem and the
subproblem. The former finds new multipliers to send to the latter which uses this information to find new
columns to send back. A lower bound on the optimal integer solution of the VRPTW model is obtained at
the end of this back and forth process. This is then used within a branch-and-bound framework to obtain
the optimal VRPTW solution. If the vehicles are identical, as we have assumed here, all subproblems will
be equivalent and therefore it is necessary to only solve one. The master problem and the subproblem
will be discussed in more detail in sections 4.4 and 4.5, respectively. The complete column generation
process is described in Chapter 1, while the subproblem forms the subject of Chapter 2.
In addition, other LRs are possible but not promising. One may consider relaxing the time and
capacity constraints (4.3), (4.7) and (4.8). This yields a linear network flow problem which possesses
the integrality property. The corresponding bound can be calculated very fast, but is not likely to be
very strong unless capacity is not binding and time windows are very narrow (see Desrosiers, Dumas,
Solomon, and Soumis [14]). Relaxing only the capacity or time window constraints also does not seem
sensible since the relaxed problem is not generally easier to solve than the original.
Desrochers, Desrosiers, and Solomon [12] were the first to apply DWD with a free number of vehicles.
The assignment constraints were considered the coupling constraints, while the subproblem was a shortest
path problem with resource constraints. Relaxing the same constraint set and applying LR was first
proposed by Kohl and Madsen [29]. Desrosiers, Sauvé, and Soumis [13] have used a similar relaxation
to calculate a lower bound for the minimum fleet size for the m-TSPTW.
Jörnsten, Madsen, and Sørensen [24] suggested solving the VRPTW by variable splitting (later called
Lagrangean decomposition, or LD). In follow-up work, Halse [20] described three different variable
splitting methods where ∑ j xi jk was replaced by yik in constraint set (4.2) and possibly (4.3). In turn,
the constraint yik = ∑ j xi jk was introduced and Lagrangean relaxed. The problem decomposes into two
problems, one in the x- and s-variables and the other in the y-variables. The former problem is further
decomposed by vehicle and it is a shortest path problem with resource constraints. The latter is an

96
assignment-type problem. Specifically, the approaches are:

• VS1: Keep constraints (4.2) and (4.3) in the y-problem. This represents a generalized assignment
problem (GAP) and the x/s-problem becomes a shortest path problem with time windows (SPPTW).
The GAP has the special structure where all right hand sides in (4.3) are identical and di does not
depend on k.
• VS2: Keep constraints (4.2) in the y-problem. The y-problem becomes a "Semi assignment" prob-
lem (SAP) consisting of constraints (4.2) only. The x/s-problem is equivalent to a shortest path
problem with time windows and capacity constraints (SPPTWCC). The SAP is easily solvable and
possesses the integrality property.
• VS3: Keep constraints (4.2) in the y-problem and constraints (4.3) in both the y- and the x/s-
problem. The y-problem is a GAP and the x/s-problem constitutes a SPPTWCC.

In the LD master problem, whose role is to find multipliers to the relaxed equation relating x and
y, the number of multipliers is larger than in the LR considered above. This clearly makes the master
problem more difficult. Also the subproblems are no longer identical since the LD multipliers depend on
both customer and vehicle. Note that only VS1 and VS2 have been implemented.
We now define LB(VS1), LB(VS2) and LB(VS3) as the best lower bounds obtainable from the three
variable splitting approaches, respectively. It can be shown that the previous LR and the DWD yield
the same lower bound LB(LR/DWD). Provided that the vehicles are identical, Kohl [28] has derived the
following results:

LB(VS3) ≥ LB(VS1)
LB(VS3) ≥ LB(VS2)
LB(LR/DWD) = LB(VS2)

There exist instances for which LB(VS3)>LB(VS1). He further showed that LB(VS2)=LB(VS3) under
some weak supplementary conditions. This is surprising because it implies there is no additional gain
to be derived from solving two hard integer problems (the SPPTWCC and GAP) instead of just one
(the SPPTWCC). However, in the more general case where vehicles have different capacities it might be
possible that the VS3 model yields a better bound than VS2.
To conclude, in VRPTW case, the variable splitting methods mentioned above generally provide
similar lower bounds to those obtained from the ordinary LR or DWD.

4.4 The master problem

The column generation methodology has been successfully applied to the VRPTW by numerous re-
searchers. It represents a generalization of the linear DWD since the master problem and the subproblem
are integer and mixed-integer programs, respectively. Often the master problem is simply stated as a
set partitioning problem on which column generation is applied, thereby avoiding the description of the
DWD on which it is based. To gain an appreciation for different cutting and branching opportunities
compatible with column generation, here we present the master problem by going through the steps of
the DWD based on the multicommodity network flow formulation (4.1) - (4.9).
The column generation approach exploits the fact that only constraint set (4.2) links the vehicles
together. Hence, the integer master problem is defined through (4.1) - (4.2) and (4.9), that is, it contains
the objective function, the assignment of customers to exactly one vehicle and the binary requirement on
the flow variables. The rest of the constraints and (4.9) are part of the subproblem which has a modified
objective function that decomposes into |V | independent subproblems, one for each vehicle. In the rest

97
of this section we will focus on the linear master problem (4.1) - (4.2). Branching, necessary to solve the
integer master problem, will be discussed in section 4.6.
Let P k be the set of feasible paths for vehicle k, k ∈ V . Hence, p ∈ P k corresponds to an elementary
path which can also be described by using the binary values xkij p , where xkij p = 1, if vehicle k goes directly
from vertex i to vertex j on path p, and xkij p = 0, otherwise. Any solution xkij to the master problem (4.1)
- (4.2) can be written as a non-negative convex combination of a finite number of elementary paths, i.e.,

(4.12) xkij = ∑ xkij p ykp ∀k ∈ V , ∀(i, j) ∈ A

p∈P k

(4.13) ∑ ykp = 1 ∀k ∈ V
p∈P k

(4.14) ykp ≥ 0 ∀k ∈ V , ∀p ∈ P k

Using xkij p we can define the cost of a path, ckp , and the number of times a customer i is visited by vehicle
k, aki , as:

ckp = ∑(i, j)∈A ckij xkij p ∀k ∈ V , ∀p ∈ P k

akip = ∑ j∈N ∪{n+1} xkij p ∀k ∈ V , ∀i ∈ N , ∀p ∈ P k

Now we can substitute these values into (4.1) - (4.2) and arrive at the revised formulation of the master
problem:

(4.15) min ∑ ∑ ckp ykp s.t.

k∈V p∈P k

(4.16) ∑ ∑ akip ykp = 1 ∀i ∈ C

k∈V p∈P k

(4.17) ∑ ykp = 1 ∀k ∈ V
p∈P k

(4.18) ykp ≥ 0 ∀k ∈ V , ∀p ∈ P k

The mathematical formulation (4.15) - (4.18) is then the linear relaxation of a set partitioning type prob-
lem with an additional constraint on the total number of vehicles and a set of convex combination con-
straints.
In the usual case of a single depot and a homogeneous fleet of vehicles with the same initial condi-
tions for all vehicles, all P k are identical, that is, P k = P, k ∈ V . Furthermore, the networks for the
subproblems are also identical. Therefore constraints (4.17) can be aggregated. By letting y p = ∑k∈V ykp ,
the index k can be eliminated from the formulation (4.15) - (4.18). The resulting model given below is
the classical linear relaxation of the set partitioning formulation:

(4.19) min ∑ c py p s.t.

p∈P

(4.20) ∑ aip y p = 1 ∀i ∈ C
p∈P

(4.21) yp ≥ 0 ∀p ∈ P

In the column generation methodology, the set of columns in the linear master problem is limited to only
those that have already been generated, hence the term restricted master problem. It consists of finding
a set of minimum cost paths among all paths presently in the master problem. The restricted master

98
problem can mathematically be stated as:

(4.22) min ∑ c p y p s.t.

p∈P ′

(4.23) ∑ aip y p = 1 ∀i ∈ C
p∈P ′

(4.24) yp ≥ 0 ∀p ∈ P ′

Each decision variable y p counts the number of times path p is used. This is not necessarily integer,
but can be any real number in the interval [0; 1]. The set P ′ contains all the paths generated, aip denotes
the number of times customer i is serviced on path p, and, c p is the cost of the path. The parameter aip
should in principle be either 0 or 1, but since the subproblem is relaxed (see section 4.5) it can take larger
integer values.
Solving the restricted master problem yields a solution y = (y1 , y2 , . . . , y|P ′ | ) which might be integer
but this is not guaranteed. If it is integer, a feasible but not necessarily optimal solution to the VRPTW
has been found. In addition to the primal solution, a dual solution φ = (φ1 , φ2 , . . . , φ|C| ) is also obtained.
An initial start for the restricted master problem is often the set of routes visiting a single customer,
that is, routes of the type depot-i-depot (cf. section 4.8). When the optimal solution to the restricted
master problem is found, the simplex algorithm asks for a new variable (i.e. a column/path p ∈ P \ P ′ )
with negative reduced cost. Such a column is found by solving a subproblem, sometimes called the
pricing problem. For the VRPTW, the subproblem should solve the problem “Find the path with minimal
reduced cost.” Solving the subproblem is in fact an implicit enumeration of all feasible paths, and the
process terminates when the optimal objective of the subproblem is non-negative (it will actually be 0).
It is not surprising that the behavior of the dual variables plays a pivotal role in the overall performance
of the column generation principle for the VRPTW. It has been observed by Kallehauge [25] that dual
variables do not converge smoothly to their respective optima. Assume that the paths (0, i, n + 1) are
used to initialize the algorithm. Figure 4.1 illustrates the instability of the column generation algorithm
compared to the stabilized cutting-plane algorithm presented in the above paper. Furthermore, Figure
4.2 illustrates the effect of the size of the multipliers on the computational difficulty of the SPPTWCC
subproblems. Whereas the multipliers are large in the Dantzig-Wolfe process, they are small in the
cutting-plane approach. This problem originates in the coordination between the master problem and the
subproblem.

Labels Stabilized Cutting-Plane Algorithm

100,000
Column Generation Algorithm
80,000

60,000

40,000

20,000

0 u
0 20 40 60 80 100 120

Figure 4.1: Number of labels generated in the subproblem with respect to the iteration number for
the Dantzig-Wolfe method and the bundle method on the Solomon instance R104 with 100 customers
(from Kallehauge [25]).

99
 Distance
125,000
100,000
75,000
50,000
25,000
0 Seconds
0 200 400 600 800
(a) Column Generation Algorithm

Distance

1,000

750

500

250

0 Seconds
0 10 20 30 40 50 60 70
(b) Stabilized Cutting-Plane Algorithm

Figure 4.2: The Euclidian distance between the current dual variables and the optimum dual variables.
Observe the different scales.

This corresponds to the principle of stabilization of column generation as discussed in du Merle,

Villeneuve, Desrosiers, and Hansen [16]. Stabilization in the VRPTW context is reported by Kallehauge,
Larsen, and Madsen [27]. Here a speedup factor of 6 is reported for the root node of all R1 instances.
Finally, in many routing problems the optimal solution remains unchanged even if overcovering rather
than exact covering of customers is allowed. Due to the triangle inequality in the VRPTW, overcovering
will always be more expensive than just covering and therefore an optimal solution will always be one
where each customer is visited exactly once. The advantage of allowing overcovering is that the linear
relaxation of the Set Covering Problem is easier to solve than that of the Set Partitioning Problem, and
this will in turn lead to the computation of good estimates of the dual variables.

4.5 The subproblem

In the column generation approach for the VRPTW, the subproblem decomposes into |V | identical prob-
lems, each one being a shortest path problem with resource constraints (time windows and vehicle ca-
pacity). More specifically, the subproblem is an Elementary Shortest Path Problem with Time Windows
and Capacity Constraints (ESPPTWCC), where elementary means that each customer can appear at most

100
once in the shortest path. It can be formulated as:

(4.25) min ∑ ∑ ĉi j xi j , s.t.

i∈N j∈N

(4.26) ∑ di ∑ xi j ≤ q
i∈C j∈N

(4.27) ∑ x0 j = 1
j∈N

(4.28) ∑ xih − ∑ xh j = 0 ∀h ∈ C
i∈N j∈N

(4.29) ∑ xi,n+1 = 1
i∈N
(4.30) si + ti j − Mi j (1 − xi j ) ≤ sj ∀i, j ∈ N
(4.31) ai ≤ si ≤ bi ∀i ∈ N
(4.32) xi j ∈ {0, 1} ∀i, j ∈ N

Constraint (4.26) is the capacity constraint, constrains (4.30) and (4.31) are time constraints, while
constraint (4.32) ensures integrality. The constraints (4.27), (4.28) and (4.29) are flow constraints result-
ing in a path from the depot 0 to the depot n + 1. When solving the ESPPTWCC as the subproblem
in the VRPTW, ĉi j is the modified cost of using arc (i, j), where ĉi j = ci j − πi . Note that while ci j is a
non-negative integer, ĉi j can be any real number.
This subproblem does not posses the integrality property, and therefore solving it as a linear mixed-
integer programming problem will potentially result in a reduction of the integrality gap between the
optimal solution of the LP-relaxed version of the VRPTW and the optimal integer solution to the problem.
Since the ESPPTWCC is NP-hard in the strong sense (see Dror [15] and Kohl [28]), the usual ap-
proach has been to slightly alter the problem by relaxing some of the constraints. In particular, allowing
cycles changes the problem to the Shortest Path Problem with Time Windows and Capacity Constraints
(SPPTWCC). Since arcs can now be used more than once (and customers may therefore be visited more
than once), the decision variables xi j and si are replaced by xli j and sli . The variable xli j is set to 1 if the
arc (i, j) is used as the l’th arc on the shortest path, and 0 otherwise, and the variable sli is set to the
bn+1
start of service at customer i as customer number l, where l ∈ L = {1, 2, . . ., |L |}, |L | = ⌊ mint ij
⌋. The

101
SPPTWCC can now be described by the following mathematical model:
(4.33) min ∑ ∑ ∑ ĉi j xli j , s.t.
l∈L i∈N j∈N

(4.34) ∑ ∑ x1i j = 1
i∈N j∈N

(4.35) ∑ ∑ xli j − ∑ ∑ xl−1

ij ≤ 0 ∀l ∈ L − {1}
i∈N j∈N i∈N j∈N

(4.36) ∑ di ∑ ∑ xli j ≤ q
i∈C l∈L j∈N

(4.37) ∑ x10 j = 1
j∈N

(4.38) ih − ∑
∑ xl−1 xlh j = 0 ∀h ∈ C ∀l ∈ L − {1}
i∈N j∈N

(4.39) ∑ ∑ xli,n+1 = 1
l∈L i∈N
sli + ti j − K(1 − xli j ) ≤ slj
(4.40) ∀i, j ∈ N ∀l ∈ L − {1}
(4.41) ai ≤ sli ≤ bi ∀i ∈ N
(4.42) xli j ∈ {0, 1} ∀i, j ∈ N
In this formulation, (4.34) forces the first arc to be used only once, while (4.35) states that arc l can
only be used provided that arc l − 1 is used. The remaining constraints are the original constraints (4.3)
to (4.9) extended to include the additional superscript l and the changes related to its inclusion. Note
that (4.34) is redundant as it is covered by (4.37), but it has been kept in the model as to indicate the
origin node.
This problem can be solved by a pseudo-polynomial algorithm described in Desrochers, Desrosiers,
and Solomon [12]. This and all other current approaches are based on dynamic programming. Even
though negative cycles are possible, the time windows and the capacity constraints prohibits infinite
cycling. Note that capacity is accumulated every time a customer is serviced in a cycle. If the distance
used to compute the cost of routes satisfies the triangle inequality, the optimal solution contains only
elementary routes. Solving the SPPTWCC instead of the ESPPTWCC augments the size of the set of
admissible columns generated for the master problem. Consequently the lower bound on the master
problem may decrease. A slight improvement can be obtained by implementing 2-cycle elimination in
the solution process which dates back to Kolen, Rinnooy Kan, and Trienekens [31].
While the SPPTWCC relaxation was at the time a computational necessity, the ESPPTWCC has re-
cently been tackled directly. Work on this problem and k-cycle elimination, where k ≥ 3, proved very
successful in expanding the scope of the VRPTW problems solved. Even though the ESPPTWCC con-
tinues to be regarded as difficult to solve when time windows are wide, two research groups have recently
used it directly in VRPTW optimal algorithms. Chabrier [5] and Chabrier, Danna, and Le Pape [6],
and independently Feillet, Dejax, Gendreau, and Gueguen [18] have extended the dynamic programming
approach of Desrochers, Desrosiers, and Solomon [12] to the ESPPTWCC by adapting the path domi-
nance rule. They then incorporated several heuristic modifications to make the algorithm much faster.
Chabrier [5] and Chabrier, Danna, and Le Pape [6] obtained lower bounds superior to those based on
the SPPTWCC resulting in excellent computational results to be described in section 4.9. A different
approach that has not yet been tried on the VRPTW is presented in Dumitrescu and Boland [17]. The
authors compare three scaling techniques and a standard label-setting method. They show that integrating
preprocessing information within the label-setting method can be very beneficial in terms of both memory

102
and run time. Further improvements of the label-setting method can be obtained by using Lagrangean
relaxation.
Instead of dealing with the computational burden of the ESPPTWCC or the weakened lower bound
provided by the SPPTWCC, one could consider a middle of the road approach. That is, disallow cycles of
small length. As discussed above, cycle elimination corresponding to k = 2 has been a common technique.
In the SPPTWCC-k-cyc, paths with cycles of length of at most k are eliminated. The case k ≥ 3 has been
considered by Irnich and Villeneuve [23] with encouraging results presented in section 4.9. Recently
Rousseau, Gendreau, and Pesant [38] have presented results where Constraint Programming is used to
solve the subproblem. Taking into account the difference in computer power, the authors conclude that
their approach is not any faster than that of Desrochers, Desrosiers, and Solomon [12].

4.6 Branch-and-bound
The column generation approach does not automatically guarantee integer solutions and often the solu-
tions obtained will indeed be fractional. Therefore a branch-and-bound framework has to be established.
The calculations are organized in a branching tree. For the VRPTW only binary strategies have been
proposed in the literature although it should be noted that it is generally not difficult to come up with
non-binary branching trees for the problem. The branching decisions are generally based on considera-
tions related to the original 3-index flow formulation (4.2) - (4.9). The column generation process is then
repeated at each node in the branch-and-bound tree.

4.6.1 Branching on the number of vehicles

Branching on the number of vehicles was originally proposed by Desrochers, Desrosiers, and Solomon
[12]. If the number of vehicles is fractional we introduce a bound on the number of vehicles. Note that this
branching strategy does not require that the flow and time variables of the original model be computed.
This branching rule can be implemented fairly easily and only concerns the master problem. We
denote the flow over an arc by fi j and this is the sum of all flows over that arc, that is fi j = ∑k∈V xi jk .
The fi j values can easily be derived from the solution of the master problem. When we branch on the
number of vehicles, two child nodes are created, one imposing on the master problem parent node the
additional constraint ∑ j∈C f0 j ≥ ⌈l⌉ while the other forcing ∑ j∈C f0 j ≤ ⌊l⌋, where l is the fractional sum
of all variables in the master problem.
Note that branching on the number of vehicles is not necessarily enough to obtain an integer solution
as it is possible to derive solutions where the sum of the vehicles is integer, but yet there are fractional
vehicles driving around the network.

4.6.2 Branching on flow variables

Branching on a single variable xi jk is possible only if each vehicle can be distinguished. In column gen-
eration this can be achieved by solving the subproblem for each vehicle individually and by introducing
an additional constraint in the master problem

∑ yp = 1 ∀k ∈ V
p∈Pk

where Pk is the set of routes generated for each vehicle k and y p is the binary variable indicating whether
route p is used.
Since most cases described in the literature assume a homogeneous fleet, it doesn’t make sense to
branch on individual vehicles. Instead, branching can be done on sums of flows, that is either on ∑ j xi jk

103
or on ∑k xi jk (equivalent to fi j ). Branching on ∑ j xi jk results in a different subproblem for each vehicle,
even though the vehicles are identical. That is because imposing ∑ j xi jk = 1 forces customer i to be visited
by vehicle k, while ∑ j xi jk = 0 implies that customer i is assigned to any vehicle but k.
The standard practice has been to branch on ∑k xi jk since the branching decision can easily be trans-
ferred to the master problem and subproblem. This was proposed independently by Halse [20] and
Desrochers, Desrosiers, and Solomon [12]. When ∑k xi jk = 1, customer j succeeds customer i on the
same route, while if ∑k xi jk = 0, customer i does not immediately precede j. If there is more than one
candidate for branching, that is, there are several fractional variables, we would generally like to choose
a candidate that is not close to either 0 or 1 in order to make an impact. When selecting among the nodes
to branch on, a common heuristic is to branch on the variable maximizing ci j (min{xi jk , 1 − xi jk }) using
a best-first strategy In order to create more complex strategies the branching schemes can be applied
hierarchically, such as first branching on the number of vehicles and then on ∑k xi jk , or mixed.

4.6.3 Branching on resource windows

Branching on resource windows was first proposed by Gélinas, Desrochers, Desrosiers, and Solomon
[19] and is presently the only alternative to branching on flow variables. In the VRPTW model resource
windows can be interpreted as either the time windows or the capacity constraints. We will only discuss
branching on time windows, as capacity is significantly less constraining in many cases. In Gélinas,
Desrochers, Desrosiers, and Solomon [19] only branching on time windows was used.
Branching on time windows results in splitting a time window into two smaller ones. Branching has
to be done in such a way that at least one route is infeasible in each of the two sub-windows.
In order to branch on time windows three decisions have to be taken:
1. How should the node for branching be chosen?
2. Which time window should be divided?
3. Where should the partition point be?
In order to decide on the above issues, we define feasibility intervals [lir , uri ] for all vertices i ∈ N and
all routes r with fractional flow. lir is the earliest time that service can start at vertex i on route r, and uri
is the latest time that service can start, that is, [lir , uri ] is the time interval during which route r must visit
vertex i to remain feasible.
The intervals can easily be computed by a recursive formula. Additionally we define
(4.43) Li = max {lir }, i ∈ N
fractional routes r
(4.44) Ui = min {uri }, i ∈ N
fractional routes r
If Li > Ui at least two routes (or two visits by the same route) have disjoint feasibility intervals, i.e., the
vertex is a candidate for branching on time windows. We can branch on a candidate vertex i by dividing
the time windows [ai , bi ] at any integer value in the open interval [Ui , Li [. It should be noted that situations
can arise where there are no candidates for branching on time windows, but the solution is not feasible.
Three different strategies were proposed by Gélinas, Desrochers, Desrosiers, and Solomon [19] aim-
ing at the elimination of cycles, the minimization of the number of visits to a customer i and the balancing
of flow in the two branch-and-bound nodes.
After having chosen the candidate vertex i for branching, an integer t ∈ [Ui , Li [ has to be selected in
order to determine the division. Here t is chosen in order to divide the time window of the customer such
that 1) the flow is balanced and 2) the time window is divided as evenly as possible.

104
4.7 Acceleration strategies
4.7.1 Preprocessing
The aim of preprocessing is to narrow the solution space by tightening the formulation before the actual
optimization is started. This can be done by fixing some variables, reducing the interval of values a
variable can take and so on. In the VRPTW, the time windows can be narrowed if the triangle inequality
holds. Accordingly, Kontoravdis and Bard [32] propose the following scheme. The earliest time a vehicle
can arrive at a customer is by arriving straight from the depot and the latest time it can leave is by going
directly back to the depot. Hence, for each customer i, its time window can be strengthened from [ai , bi ]
to [max{a0 + t0i , ai }, min{bn+1 − ti,n+1 , bi }].
A further reduction of the time windows can be achieved by the method developed by Desrochers,
Desrosiers, and Solomon [12]. The time windows are reduced by applying the following four rules in a
cyclic manner. The process is stopped when one whole cycle is performed without changing any of the
time windows. The four rules are:
1. Minimal arrival time from predecessors:
al = max{al , min{bl , min(i,l) {ai + til }}}
2. Minimal arrival time to successors:
al = max{al , min{bl , min(l, j) {a j − tl j }}}
3. Maximal departure time from predecessors:
bl = min{bl , max{al , max(i,l) {bi + til }}}
4. Maximal departure time to successors:
bl = min{bl , max{al , max(l, j) {b j − tl j }}}
The first rule adjusts the start of the time window to the earliest time a vehicle can arrive coming
straight from any possible predecessor. In a similar fashion, the second rule modifies the start of the time
window in order to minimize the excess time spent before the time windows of all possible successors
open if the vehicle continues to a successor as quickly as possible. The two remaining rules use the same
principles to adjust the closing of the time window. With respect to capacity, an arc (i, j) can obviously
be removed if di + d j > q.

4.7.2 Subproblem strategies

A well known strategy for accelerating column generation is to return many negative marginal cost
columns to the master problem. Even though in principle only one needs to be returned, several can
be if they are available. Computational tests conducted by Kohl [28] and Larsen [35] confirm the benefits
of this approach.

4.7.3 Master problem strategies

Along with the novel perspectives on the subproblem solution described in 4.5, master problem ac-
celeration strategies have been key to the evolution of VRPTW approaches over the last few years. One
approach is to accelerate the solution at the root node of the branch-and-bound tree by using a local search
method to generate a set of initial columns. This helps the column generation process get a fast increase
in the quality of the dual variables. It has been implemented by numerous researchers and has finally been
discussed in the literature by Danna and Le Pape [10]. The authors use a local search method based on
the savings algorithm incorporating time windows which produces a set of routes better than the trivial

105
depot-customer-depot ones. Furthermore, local search is used along with a MIP solver throughout the
branch-and-price process to generate good integer solutions fast. Two different heuristics, a local search
method based on large neighborhood search and a guided tabu search, were tested and proved beneficial,
especially on Solomon’s R1 and RC1 problem classes.
Two new approaches have been suggested by Larsen [34] and Larsen [35]. First, during the execution
of the branch-and-price a large number of columns are generated and many of these only participate in
a few computations and will not be used afterwards. If kept, each column will increase computing time
when solving the relaxed set partitioning problem and when adjusting the upper bounds on variables due
to branching decisions. Therefore Larsen [34] suggests to keep track of how long a column is part of a
basis. If it does not participate in a basis for a given number of branch-and-bound nodes it is removed
from the model. This was also suggested by Desaulniers, Desrosiers, and Solomon [11] where it was also
noted that a certain number of nonbasic columns should remain in the problem. Larsen [34] reports that
deleting columns that have not been part of the basis for the last 20 branch-and-bound nodes outperforms
the code without column deletion by a factor of 2.5 aggregated over 27 instances.
The second acceleration approach is to stop the algorithm for the SPPTWCC before it completes.
Computations can be stopped as soon as at least one route with negative cost has been generated. This
approach is denoted “forced early stop” in Larsen [35] and results in dramatic running time reductions,
especially for problems with large time windows. For these, the values of the dual variables at the begin-
ning of the procedure will however be of poor quality. Only when the subproblem proves optimality it
cannot be stopped prematurely.

4.7.4 Cutting planes

The barebone column generation methodology for solving the VRPTW is part of the popular approach
for solving difficult integer programming problems by relaxing the integrality constraints of the original
problem. Typically, the optimal solution to the relaxed problem is not feasible for the original problem
and branch-and-bound is used in order to get integer solutions.
Cutting planes has been proposed to improve the polyhedral description of the relaxed problem in
order to get an integer solution or at least narrow the integrality gap. Kohl, Desrosiers, Madsen, Solomon,
and Soumis [30] suggested three cuts in order to tighten the LP formulation of the VRPTW problem. As
these cuts are only introduced at the root node, this is not a branch-and-cut approach, where cuts can be
introduced at any node of the search tree.
The method is based on subtour elimination constraints and comb inequalities transferred from the
TSP, and 2-path cuts. To detect subtour elimination constraints, a separation algorithm by Crowder and
Padberg [9] was implemented. With respect to the comb inequalities, only combs with 3 teeth and 2
nodes were detected. The separation algorithm was a primitive enumeration scheme. Neither of these
constraints had a large impact on tightening the bound.
A new idea introduced by Kohl, Desrosiers, Madsen, Solomon, and Soumis [30] was the inclusion
of 2-path cuts. The basis of this set of cuts is the subtour elimination inequality in the strong form:
x(S) ≥ k(S), ∀S ⊆ C , where x(S) is the flow leaving the set S, and k(S) is the minimum number of
vehicles needed to service the customers in S. Determining k(S) is not an easy task, but using the triangle
inequality on the travel times we have that S1 ⊂ S2 ⇒ k(S1 ) ≤ k(S2 ). Sets S that satisfy x(S) < 2 and
k(S) > 1 must now be found. As k(S) is an integer, k(S) > 1 implies k(S) ≥ 2. So we need to identify
sets S that require at least two vehicles to be serviced, but are currently serviced by less than two.
For a set S, two checks have to be performed: 1) k(S) > 1 and 2) can the customers be serviced by
a single vehicle? The first check is easy, but the second requires the solution of the TSPTW feasibility
problem. Since this problem is NP-hard the separation algorithm can only be applied to small sets. This
is done heuristically using a simple greedy algorithm based on Laporte, Nobert, and Desrochers [33].
The 2-path cuts outperformed the branch-and-price method without 2-path cuts. The proportion of

106
the integrality gap closed by the 2-path cuts varies from 100% to 10% in a few cases. Overall 12 new
unsolved Solomon instances were closed.
Cook and Rich [7] extended the above 2-path cut approach to k-path cuts involving the solution of a
VRPTW with (k − 1) customers as part of the separation algorithm. The authors performed experiments
with k up to 6. For larger k, the percentage of the integrality gap that is closed is of course larger, but the
separation algorithm requires substantially more time and therefore it is not evident that it is preferable to
use k larger than 2.
Recently, Bard, Kontoravdis, and Yu [4] have proposed a branch-and-cut algorithm for the arc formu-
lation of the VRPTW. This development parallels the initial uses of this technique for the VRP (Naddef
and Rinaldi [37]). Based on the results obtained by Mak [36], a new arc formulation of the VRPTW is
presented in Kallehauge and Boland [26]. In this formulation the time and capacity restrictions are mod-
eled using infeasible path elimination constraints (IPECS). This new class of inequalities can be viewed
as a strengthening of the IPECS described in Ascheuer, Fischetti, and Grötschel [1], Ascheuer, Fischetti,
and Grötschel [2], and Bard, Kontoravdis, and Yu [4] and can also be incorporated at the master problem
level in the path formulation considered in this chapter.
Another line of research involves valid inequalities derived from the precedence relationships estab-
lished by the time windows. That is, if a set of customers is served by the same vehicle, the associated
time windows create a precedence structure among the corresponding nodes (Ascheuer, Fischetti, and
Grötschel [2]). In Kallehauge and Boland [26], two classes of valid inequalities for the precedence-
constrained asymmetric traveling salesman polytope (Balas, Fischetti, and Pulleyblank [3]) are trans-
ferred to the VRPTW.

4.8 Generalizations of the VRPTW model

The methods considered in this chapter can be generalized and applied to a number of related problems
as discussed by Desrosiers, Dumas, Solomon, and Soumis [14]. Here we will concentrate on routing
generalizations and show how a number of more complex routing problems can be modeled based on the
framework introduced in the previous sections.

4.8.1 Non-identical vehicles

In the general case vehicles may differ with respect to travel time, travel costs, capacity and possibly
other characteristics. We define a class of vehicles as a set of identical vehicles. There may be a cost
associated with the vehicles of a particular class, and there may be bounds on their availability as well.
These bounds are modeled in to the master problem as supplementary constraints.
The subproblem must be solved separately for each class of vehicles. The marginal costs of the arcs
originating at the depot of the subproblem for a particular vehicle class must be modified by the simplex
multiplier of the constraints on the availability of this class in the master problem. One can chose to solve
one or more of the subproblems between each master iteration. The LP optimality criterion is that no
subproblem generates columns with negative reduced costs. It is likely to be efficient to branch on the
number of vehicles of a particular class if this number is fractional.
A special case occurs if vehicles do not differ with respect to traveling time, travel cost and time
windows, but only have different capacities and possible availability and fixed costs. This problem is
clearly solvable as described above, but it can also be transformed into the identical vehicle problem
described earlier in this chapter. The advantage of this transformation is that only one subproblem must
be solved at each iteration. To illustrate how the transformation works consider a problem with two
classes of vehicles, with vehicle capacities q1 and q2 respectively, where q1 < q2 . The fixed costs of
using the vehicles are c1 and c2 , respectively. Two extra nodes are inserted in parallel between the depot

107
and the customers and any path must go through exactly one of these nodes. The two arcs from the depot
to the new nodes are priced c1 and c2 , respectively. If node 1 is chosen, the capacity is reduced by q2 − q1
since the resource window of node 1 starts at this quantity. Since the resource window of the depot is
[0, q2 ], a path going through node 1 cannot service customers with accumulated demand of more than
q2 − (q2 − q1 ) = q1 . If there are bounds on the availability of the vehicles, these are inserted in the master
problem and the simplex multipliers modify the cost of the two new arcs between the depot and the new
nodes.

4.8.2 Multiple depots

If the vehicles are based at different depots, one subproblem must be solved for each depot. Constraints
on the availability of vehicles at a particular depot are kept in the master problem, and the associated
simplex multiplier modifies the cost of arcs originating at the depot. This is equivalent to the general
non-identical vehicle case discussed above.
One may assume that the vehicles are allowed to finish their routes at a depot different from the one
the vehicles started, but that the number of vehicles starting and ending at any depot remains constant. In
this particular case it is sufficient to solve one subproblem. One extra node per depot is created "before"
the customers and one "after" the customers. For each depot there will be a constraint r in the master
problem requiring the number of vehicles housed at that depot be kept constant. The right hand side
will be zero, and the left hand side coefficient (r, p) will be 1 if route p starts at the depot associated
with constraint r and ends at another depot, −1 if the route starts at another depot and ends at the depot
associated with constraint r, and zero otherwise. The corresponding simplex multipliers modify the cost
of arcs originating at the depot (with opposite sign). It is also easy to introduce different fixed costs
associated with the vehicles housed at the depots.

4.8.3 Multiple or soft time windows

Customers may have several (disjoint) time intervals in which they can be serviced. A vehicle arriving
between two time windows must wait until the beginning of the next time window. This doesn’t truly
complicate the problem since the usual dominance criterion in the subproblem remains valid. A vehicle
arriving at a particular node at time t1 can do everything a vehicle arriving at time t2 can, provided that
t1 < t2 .
If there exist a cost c(si ) dependent on the time si service at customer i begins, the time window is said
to be soft. If the cost is non-decreasing with increasing time this is not problematic, since the dominance
criteria remain valid. The most general case where c(si ) is a general function is not efficiently solvable.
Ioachim, Gélinas, Desrosiers, and Soumis [22] present an algorithm for the linear case.

4.9 Computational experiments

Almost from the first computational experiments, a set of problems became the test-bed for both heuristic
and exact investigations of the VRPTW. Solomon [39] proposed a set of 168 instances that have remained
the leading test set ever since. For the researchers working on heuristic algorithms for the VRPTW a need
for bigger problems made Homberger and Gehring [21] propose a series of extended Solomon problems.
These larger problems have as many as 1000 customers and several have been solved by exact methods.

4.9.1 The Solomon instances

The test sets reflect several structural factors in vehicle routing and scheduling such as geographical
data, number of customers serviced by a single vehicle and the characteristics of the time windows (e.g.,

108
tightness, positioning and the fraction of time-constrained customers in the instances). Customers are
distributed within a [0, 100]2 square.
The instances are divided into 6 groups (test-sets) denoted R1, R2, C1, C2, RC1 and RC2. Each of the
test sets contain between 8 and 12 instances. In R1 and R2 the geographical data is randomly generated by
a random uniform distribution. In the test sets C1 and C2 the customers are placed in clusters, and finally
in the RC1 and RC2 test-sets some customers are placed in clusters while others are placed randomly. In
the test sets R1, C1 and RC1 the scheduling horizon is short permitting approximately 5 to 10 customers
to be serviced on each route. The R2, C2 and RC2 problems have a long scheduling horizon allowing
routes with more than 30 customers to be feasible. This makes the problems very hard to solve exactly
and they have not been used until recently to test exact methods. The time windows for the test sets C1
and C2 are generated to permit good, maybe even optimal, cluster-by-cluster solutions. For each class
of problems the geographical position of the customers is the same in all instances whereas the time
windows are changed.
Each instance has 100 customers, but by considering only the first 25 or 50 customers, smaller in-
stances can easily be generated. It should be noted that for the RC-sets this results in the customers
being clustered since the clustered customers appear at the beginning of the file. Travel time between two
customers is usually assumed to be equal to the travel distance plus the service time at the predecessor
customer.

4.9.2 Computational results

This section reviews the results obtained by the best exact algorithms for the VRPTW. All are based on
the column generation approach. The tables 4.1 through 4.6 present the solutions for the six different sets
of the Solomon instances that have been solved to optimality. Column K indicates the number of vehicles
used in the optimal solution while the column “Authors” give reference to the first publication(s) of the op-
timal solution for the problem: Kohl, Desrosiers, Madsen, Solomon, and Soumis [30] (KDMSS), Larsen
[35] (L), Kallehauge, Larsen, and Madsen [27] (KLM), Cook and Rich [7] (CR), Irnich and Villeneuve
[23] (IV), Chabrier [5] (C), and Danna and Le Pape [10] (DLP). It should be noted that Desrochers,
Desrosiers, and Solomon [12] prior to Kohl, Desrosiers, Madsen, Solomon, and Soumis [30] solved 50
of the 87 Solomon problems with narrow time windows, but with different travel times. Whereas all the
above mentioned papers compute the travel times using one decimal point precision and truncation, time
and cost is computed differently in Desrochers, Desrosiers, and Solomon [12]. Furthermore, solutions
to all C1 instances were reported for the first time by Kohl and Madsen [29], who used a Lagrangean
relaxation approach.
As discussed in Cordeau, Desaulniers, Desrosiers, Solomon, and Soumis [8], the optimal algorithm
of Kohl, Desrosiers, Madsen, Solomon, and Soumis [30] solved 69 of the 87 Solomon benchmark short
horizon problems to optimality. Eleven additional problems were solved by Larsen [35], Cook and Rich
[7], and Kallehauge, Larsen, and Madsen [27]. Recently, Irnich and Villeneuve [23] were successful in
closing three additional instances. Four 100-customer instances are still open.
As also reported in Cordeau, Desaulniers, Desrosiers, Solomon, and Soumis [8], Larsen [35], Cook
and Rich [7], and Kallehauge, Larsen, and Madsen [27] also provided exact solutions to 42 of the 81
Solomon long horizon problems. Since then, Irnich and Villeneuve [23], Chabrier [5] and Danna and Le
Pape [10] have solved an additional 21 instances, leaving 18 problems still unsolved.

4.10 Conclusions
In this chapter we have highlighted the noteworthy developments for optimal column generation ap-
proaches to the VRPTW. To date, such methods incorporating branching and cutting on solutions ob-

109
 Table 4.1: Optimal solutions for the R1 instances
Problem K Dist. Authors Problem K Dist. Authors
R101.25 8 617.1 KDMSS R107.25 4 424.3 KDMSS
R101.50 12 1044 KDMSS R107.50 7 711.1 KDMSS
R101.100 20 1637.7 KDMSS R107.100 11 1064.6 CR+KLM
R102.25 7 547.1 KDMSS R108.25 4 397.3 KDMSS
R102.50 11 909 KDMSS R108.50 6 617.7 CR+KLM
R102.100 18 1466.6 KDMSS R108.100
R103.25 5 454.6 KDMSS R109.25 5 441.3 KDMSS
R103.50 9 772.9 KDMSS R109.50 8 786.8 KDMSS
R103.100 14 1208.7 CR+L R109.100 13 1146.9 CR+KLM
R104.25 4 416.9 KDMSS R110.25 5 444.1 KDMSS
R104.50 6 625.4 KDMSS R110.50 7 697 KDMSS
R104.100 11 971.5 IV R110.100 12 1068 CR+KLM
R105.25 6 530.5 KDMSS R111.25 4 428.8 KDMSS
R105.50 9 899.3 KDMSS R111.50 7 707.2 CR+KLM
R105.100 15 1355.3 KDMSS R111.100 12 1048.7 CR+KLM
R106.25 5 465.4 KDMSS R112.25 4 393 KDMSS
R106.50 8 793 KDMSS R112.50 6 630.2 CR+KLM
R106.100 13 1234.6 CR+KLM R112.100

tained through Dantzig-Wolfe decomposition are the best performing algorithms. Valid inequalities have
proved an invaluable tool in strengthening the LP relaxation for this class of problems.
Recent advances have stemmed from work on parallel implementations of the overall approach, accel-
eration strategies, primarily at the master problem level, and the subproblem. Solving the subproblem as
a ESPPTWCC or a SPPTWCC-k-cyc has shown to be very beneficial. Nevertheless, 25% of Solomon’s
problems are still unsolved. Additional research in each of these areas should lead to further advances.
We expect that the further study of polyhedral structures, parallelism, acceleration strategies, and the sub-
problem will constitute the backbone of research in this area for the next several years. Master problem
acceleration methods relying on local search heuristics is just beginning.
Decomposition algorithms are also easily adaptable to other settings. This is because they comprise
modules, such as dynamic programming, that can handle a variety of objectives. Lateness, for one, is be-
coming an increasingly important benchmark in today’s supply chains that emphasize on time deliveries.
Moreover, they can be run as optimization-based heuristics by means of early stopping criteria.
We hope that this chapter has shed sufficient light on current developments to lead to exciting further
research.

Acknowledgments The research of Marius M. Solomon was partially supported by the Patrick F. and
Helen C. Walsh Research Professorship.

References
[1] N. Ascheuer, M. Fischetti, and M. Grötschel. A polyhedral study of the asymmetric traveling sales-
man problem with time windows. Networks, 36:69–79, 2000. [107]
[2] N. Ascheuer, M. Fischetti, and M. Grötschel. Solving the asymmetric travelling salesman problem

110
 Table 4.2: Optimal solutions for the C1 instances
Problem K Dist. Authors Problem K Dist. Authors
C101.25 3 191.3 KDMSS C106.25 3 191.3 KDMSS
C101.50 5 362.4 KDMSS C106.50 5 362.4 KDMSS
C101.100 10 827.3 KDMSS C106.100 10 827.3 KDMSS
C102.25 3 190.3 KDMSS C107.25 3 191.3 KDMSS
C102.50 5 361.4 KDMSS C107.50 5 362.4 KDMSS
C102.100 10 827.3 KDMSS C107.100 10 827.3 KDMSS
C103.25 3 190.3 KDMSS C108.25 3 191.3 KDMSS
C103.50 5 361.4 KDMSS C108.50 5 362.4 KDMSS
C103.100 10 826.3 KDMSS C108.100 10 827.3 KDMSS
C104.25 3 186.9 KDMSS C109.25 3 191.3 KDMSS
C104.50 5 358 KDMSS C109.50 5 362.4 KDMSS
C104.100 10 822.9 KDMSS C109.100 10 827.3 KDMSS
C105.25 3 191.3 KDMSS
C105.50 5 362.4 KDMSS
C105.100 10 827.3 KDMSS

Table 4.3: Optimal solutions for the RC1 instances

Problem K Dist. Authors Problem K Dist. Authors
RC101.25 4 461.1 KDMSS RC105.25 4 411.3 KDMSS
RC101.50 8 944 KDMSS RC105.50 8 855.3 KDMSS
RC101.100 15 1619.8 KDMSS RC105.100 15 1513.7 KDMSS
RC102.25 3 351.8 KDMSS RC106.25 3 345.5 KDMSS
RC102.50 7 822.5 KDMSS RC106.50 6 723.2 KDMSS
RC102.100 14 1457.4 CR+KLM RC106.100
RC103.25 3 332.8 KDMSS RC107.25 3 298.3 KDMSS
RC103.50 6 710.9 KDMSS RC107.50 6 642.7 KDMSS
RC103.100 11 1258 CR+KLM RC107.100 12 1207.8 IV
RC104.25 3 306.6 KDMSS RC108.25 3 294.5 KDMSS
RC104.50 5 545.8 KDMSS RC108.50 6 598.1 KDMSS
RC104.100 RC108.100 11 1114.2 IV

with time windows by branch-and-cut. Mathematical Programming Series A, 90:475–506, 2001.

[107]

[3] E. Balas, M. Fischetti, and W. R. Pulleyblank. The precedence-constrained asymmetric traveling

salesman polytope. Mathematical Programming, 68:241–265, 1995. [107]
[4] J. F. Bard, G. Kontoravdis, and G. Yu. A branch-and-cut procedure for the vehicle routing problem
with time windows. Transportation Science, 36:250–269, 2002. [107]
[5] A. Chabrier. Vehicle routing problem with elementary shortest path based column generation. To
appear in Computers & Operations Research. [94, 102, 109]

[6] A. Chabrier, E. Danna, and C. Le Pape. Cooperation entre generation de colonnes avec tournees
sans cycle et recherche locale appliquee au routage de vehicules. Huitiemes Journees Nationales sur
la resolution de Problemes NP-Complets (JNPC 2002). [94, 102]

111
 Table 4.4: Optimal solutions for the R2 instances
Problem K Dist. Authors Problem K Dist. Authors
R201.25 4 463.3 CR+KLM R207.25 3 361.6 KLM
R201.50 6 791.9 CR+KLM R207.50
R201.100 8 1143.2 KLM R207.100
R202.25 4 410.5 CR+KLM R208.25 1 328.2 IV+C
R202.50 5 698.5 CR+KLM R208.50
R202.100 R208.100
R203.25 3 391.4 CR+KLM R209.25 2 370.7 KLM
R203.50 5 605.3 IV+C R209.50 4 600.6 IV+C
R203.100 R209.100
R204.25 2 355 IV+C R210.25 3 404.6 CR+KLM
R204.50 2 506.4 IV R210.50 4 645.6 IV+C
R204.100 R210.100
R205.25 3 393 CR+KLM R211.25 2 350.9 KLM
R205.50 4 690.1 IV+C R211.50 3 535.5 IV+DLP
R205.100 R211.100
R206.25 3 374.4 CR+KLM
R206.50 4 632.4 IV+C
R206.100

[7] W. Cook and J. L. Rich. A parallel cutting-plane algorithm for the vehicle routing problem with time
windows. Technical Report TR99-04, Departement of Computational and Applied Mathematics,
Rice University, 1999. [94, 107, 109]
[8] J.-F. Cordeau, G. Desaulniers, J. Desrosiers, M. M. Solomon, and F. Soumis. VRP with time win-
dows. In P. Toth and D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete
Mathematics and Applications, pages 157–193. SIAM, Philadelphia, 2002. [109]
[9] H. Crowder and M. Padberg. Solving large-scale symmetric travelling salesman problems to opti-
mality. Management Science, 26:495–509, 1980. [106]
[10] E. Danna and C. Le Pape. Branch-and-price heuristics: A case study on the vehicle routing prob-
lem with time windows. In G. Desaulniers, J. Desrosiers, and M. M. Solomon, editors, Column
generation, GERAD 25th Anniversary Series, pages 99–129. Springer, New York, 2005. [94, 105,
109]
[11] G. Desaulniers, J. Desrosiers, and M. M. Solomon. Accelerating strategies in column generation
methods for vehicle routing and crew scheduliung problems. In C. Ribeiro and P. Hansen, editors,
Essays and Surveys in Metaheuristics, pages 309–324. Kluwer Academic Publishers, 2002. [106]
[12] M. Desrochers, J. Desrosiers, and M. Solomon. A new optimization algorithm for the vehicle routing
problem with time windows. Operations Research, 40:342–354, 1992. [94, 96, 102, 103, 104, 105,
109]
[13] J. Desrosiers, M. Sauvé, and F. Soumis. Lagrangean relaxation methods for solving the minimum
fleet size multiple travelling-salesman problem with time windows. Management Science, 34:1005–
1022, 1988. [96]
[14] J. Desrosiers, Y. Dumas, M. M. Solomon, and F. Soumis. Time constrained routing and scheduling.
In M. Ball, T. Magnanti, C. Monma, and G. L. Nemhauser, editors, Network Routing, volume 8

112
 Table 4.5: Optimal solutions for the C2 instances
Problem K Dist. Authors Problem K Dist. Authors
C201.25 2 214.7 CR+L C205.25 2 214.7 CR+L
C201.50 3 360.2 CR+L C205.50 3 359.8 CR+KLM
C201.100 3 589.1 CR+KLM C205.100 3 586.4 CR+KLM
C202.25 2 214.7 CR+L C206.25 2 214.7 CR+L
C202.50 3 360.2 CR+KLM C206.50 3 359.8 CR+KLM
C202.100 3 589.1 CR+KLM C206.100 3 586 CR+KLM
C203.25 2 214.7 CR+L C207.25 2 214.5 CR+L
C203.50 3 359.8 CR+KLM C207.50 3 359.6 CR+KLM
C203.100 3 588.7 KLM C207.100 3 585.8 CR+KLM
C204.25 1 213.1 CR+KLM C208.25 2 214.5 CR+L
C204.50 2 350.1 KLM C208.50 2 350.5 CR+KLM
C204.100 3 588.1 IV C208.100 3 585.8 KLM

Table 4.6: Optimal solutions for the RC2 instances

Problem K Dist. Authors Problem K Dist. Authors
RC201.25 3 360.2 CR+L RC205.25 3 338 L+KLM
RC201.50 5 684.8 L+KLM RC205.50 5 630.2 IV+C
RC201.100 9 1261.8 KLM RC205.100 7 1154 IV+C
RC202.25 3 338 CR+KLM RC206.25 3 324 KLM
RC202.50 5 613.6 IV+C RC206.50 5 610 IV+C
RC202.100 8 1092.3 IV+C RC206.100
RC203.25 2 326.9 IV+C RC207.25 3 298.3 KLM
RC203.50 4 490.122 IV+C RC207.50 4 558.6 C
RC203.100 RC207.100
RC204.25 3 299.7 C RC208.25 2 269.1 C
RC204.50 3 444.2 DLP RC208.50
RC204.100 RC208.100

of Handbooks in Operations Research and Management Science, pages 35–139. Elsevier Science
Publishers B.V. (North-Holland), Amsterdam, The Netherlands, 1995. [95, 96, 107]
[15] M. Dror. Note on the complexity of the shortest path models for column generation in VRPTW.
Operations Research, 42:977–978, 1994. [101]
[16] O. du Merle, D. Villeneuve, J. Desrosiers, and P. Hansen. Stabilized column generation. Discrete
Mathematics, 194:229–237, 1999. [100]
[17] I. Dumitrescu and N. Boland. Improved preprocessing, labeling and scaling algorithms for the
weight-constrained shortest path problem. Networks, 42:135–153, 2003. [102]
[18] D. Feillet, P. Dejax, M. Gendreau, and C. Gueguen. An exact algorithm for the elementary shortest
path problem with resource constraints: Application to some vehicle routing problems. Networks,
44:216–229, 2004. [94, 102]
[19] S. Gélinas, M. Desrochers, J. Desrosiers, and M. M. Solomon. A new branching strategy for time
constrained routing problems with application to backhauling. Annals of Operations Research, 61:
91–109, 1995. [104]

113
[20] K. Halse. Modeling and Solving Complex Vehicle Routing Problems. PhD thesis, Technical Univer-
sity of Denmark, 1992. [94, 96, 104]
[21] J. Homberger and H. Gehring. Two evolutionary metaheuristics for the vehicle routing problem
with time windows. INFOR, 37:297–318, 1999. [108]
[22] I. Ioachim, S. Gélinas, J. Desrosiers, and F. Soumis. A dynamic programming algorithm for the
shortest path problem with time windows and linear node costs. Networks, 31:193–204, 1998.
[108]
[23] S. Irnich and D. Villeneuve. The shortest-path problem with resource constraints and k-cycle elim-
ination for k ≥ 3. To appear in INFORMS Journal on Computing. [94, 103, 109]
[24] K. Jörnsten, O. B. G. Madsen, and B. Sørensen. Exact solution of the vehicle routing and scheduling
problem with time windows by variable splitting. Technical Report 5/1986, Institute of Mathemati-
cal Statistics and Operations Research, Technical University of Denmark, 1986. [96]
[25] B. Kallehauge. Lagrangian duality and non-differentiable optimization – applied on vehicle routing.
Master’s thesis, Technical University of Denmark, 2000. In Danish. [94, 99]
[26] B. Kallehauge and N. Boland. Path inequalities for the vehicle routing problem with time windows.
Technical Report 2005-2, Centre for Traffic and Transport, Technical University of Denmark, 2005.
[107]
[27] B. Kallehauge, J. Larsen, and O. B. G. Madsen. Lagrangean duality and non-differentiable optimiza-
tion applied on routing with time windows – experimental results. Technical Report IMM-REP-
2000-8, Department of Mathematical Modelling, Technical University of Denmark, 2000. [100,
109]
[28] N. Kohl. Exact Methods for Time Constrained Routing and Related Scheduling Problems. PhD
thesis, Technical University of Denmark, 1995. [97, 101, 105]
[29] N. Kohl and O. B. G. Madsen. An optimization algorithm for the vehicle routing problem with time
windows based on Lagrangean relaxation. Operations Research, 45:395–406, 1997. [94, 96, 109]
[30] N. Kohl, J. Desrosiers, O. B. G. Madsen, M. M. Solomon, and F. Soumis. 2-path cuts for the vehicle
routing problem with time windows. Transportation Science, 33:101–116, 1999. [94, 106, 109]
[31] A. W. J. Kolen, A. H. G. Rinnooy Kan, and H. W. J. M. Trienekens. Vehicle routing with time
windows. Operations Research, 35:266–273, 1987. [102]
[32] G. Kontoravdis and J. F. Bard. A grasp for the vehicle routing problem with time windows. ORSA
Journal on Computing, 7:10–23, 1995. [105]
[33] G. Laporte, Y. Nobert, and M. Desrochers. Optimal routing under capacity and distance restrictions.
Operations Research, 33:1050–1073, 1985. [106]
[34] J. Larsen. Refinements of the column generation process for the vehicle routing problem with time
windows. Journal of Systems Science and Systems Engineering, 13:326–341, 2004. [94, 106]
[35] J. Larsen. Parallelization of the Vehicle Routing Problem with Time Windows. PhD thesis, Technical
University of Denmark, 1999. [94, 105, 106, 109]
[36] V. H. Mak. On the Asymmetric Travelling Salesman Problem with Replenishment Arcs. PhD thesis,
Department of Mathematics and Statistics, The University of Melbourne, 2001. [107]

114
[37] D. Naddef and G. Rinaldi. Branch-and-cut algorithms for the capacitated VRP. In P. Toth and
D. Vigo, editors, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and
Applications, pages 49–78. SIAM, Philadelphia, 2002. [107]
[38] L.-M. Rousseau, M. Gendreau, and G. Pesant. Solving VRPTWs with constraint programming
based column generation. Annals of Operations Research, 130:199–216, 2004. [94, 103]
[39] M. M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35:254–65, 1987. [108]

115
View publication stats