This article was downloaded by: [128.197.229.

194] On: 15 January 2020, At: 17:00

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

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

Slack Induction by String Removals for Vehicle Routing

Problems
Jan Christiaens, Greet Vanden Berghe

To cite this article:

Jan Christiaens, Greet Vanden Berghe (2020) Slack Induction by String Removals for Vehicle Routing Problems. Transportation
Science

Published online in Articles in Advance 15 Jan 2020

. https://doi.org/10.1287/trsc.2019.0914

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

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

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

Copyright © 2020, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 TRANSPORTATION SCIENCE
Articles in Advance, pp. 1–17
http://pubsonline.informs.org/journal/trsc ISSN 0041-1655 (print), ISSN 1526-5447 (online)

Slack Induction by String Removals for Vehicle Routing Problems

Jan Christiaens,a Greet Vanden Berghea
a
CODeS, imec, Department of Computer Science, KU Leuven, 9000 Gent, Belgium
Contact: [email protected], https://orcid.org/0000-0003-1204-904X (JC); [email protected],
https://orcid.org/0000-0002-0275-5568 (GVB)

Received: June 20, 2017 Abstract. Dedicated algorithm and modeling improvements continue to advance the
Revised: May 7, 2018; December 21, 2018 state of the art with respect to vehicle routing problems (VRPs). Despite these academic
Accepted: February 5, 2019 achievements, solving large VRP instances sufficiently fast for real-life applicability remains
Published Online in Articles in Advance: challenging. By exploiting VRP solution characteristics in an effective manner, this paper
January 15, 2020 arrives at a powerful and fast optimization heuristic. Its primary contributions are threefold:
https://doi.org/10.1287/trsc.2019.0914 a ruin method, a recreate method, and a fleet minimization procedure. The ruin method
functions via adjacent string removal, introducing with it a novel property regarding vehicle
Copyright: © 2020 INFORMS routing problems that we term spatial slack, whereas the recreate method is categorized as
greedy insertion with blinks. Combining these results in slack induction by string removals
(SISRs), a powerful ruin and recreate approach. The fleet minimization procedure, mean-
while, introduces an absences-based acceptance criterion that serves as a complementary
optimization component for when minimizing the number of vehicles constitutes the
primary VRP objective. Together these three elements provide a suite of simple, powerful,
and easily reproducible algorithmic methods that are successfully applied not only to the
capacitated VRP but also to a wide range of related problems such as pickup and delivery
problems and others that include time windows. SISRs serves to strip back the layers of
complexity and specialization synonymous with the trend of algorithmic development
throughout recent decades. Moreover, such simplicity and reproducibility are shown
to not necessarily come at the expense of solution quality, with SISRs consistently out-
performing alternative general approaches as well as dedicated single-purpose methods.
Finally, aside from performance-related criteria, SISRs also serves to showcase a fresh per-
spective with respect to VRPs more generally, introducing a range of new terminology and
procedures that, it is hoped, will invigorate further research and innovation.

Funding: This work was partially supported by the Institute for the Promotion of Innovation through
Science and Technology in Flanders (IWT-Vlaanderen) in cooperation with Conundra [Grant
130855].
Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2019.0914.

Keywords: vehicle routing • large-scale instances • string removal • spatial slack • capacity slack • fleet minimization • ruin and recreate

1. Introduction branch-and-cut-and-price algorithm by Fukasawa et al.

The capacitated vehicle routing problem (CVRP) was (2006) improved on previous branch-and-cut algo-
introduced by Dantzig and Ramser (1959) as the truck rithms. Several improvements such as that by Baldacci,
dispatching problem, a generalization of the traveling Christofides, and Mingozzi (2008) have been combined
salesman problem (TSP). Despite multiple decades of with new elements (Pecin et al. 2014) and have proved
research (Laporte 2009), the CVRP remains a highly capable of solving CVRPs of up to 350 customers and
active field of research, and it continues to represent a even 650 customers in some cases. Computation
significant computational challenge. Research concern- times may, however, be as long as five days, which makes
ing the problem mostly falls into one of two categories: such solvers impractical for many real-world purposes.
exact algorithms or heuristics. Heuristics, by contrast, are characterized as fast and
Developing exact algorithms is an interesting ac- adaptable methods. Absolute optimality is, however,
tivity in itself, while also providing objective function not ensured—a necessary trade-off in terms of accom-
bounds that enable accurate heuristic quality assess- modating many real-world situations that rely on the
ment. Toth and Vigo (2002) provide a comprehensive presence of timely solutions. Heuristics generally em-
overview of exact algorithms to solve CVRPs with ei- ploy specific neighborhoods within a metaheuristic to
ther symmetric or asymmetric distance matrices. They gradually improve candidate solutions. Prior to the latest
primarily focus on branch-and-bound algorithms and improvements, classical neighborhoods such as 2-opt*
review a set of original powerful approaches. The (Potvin and Rousseau 1995) and CROSS exchange
1
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
2 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

(Taillard et al. 1997) have been explored in single- paper ultimately ends with a conclusion that summarizes
solution-based heuristics. Larger neighborhoods were the paper’s contribution while also delineating the scope
employed by Shaw (1998) and Pisinger and Røpke and possibilities for future research.
(2007) in a ruin and recreate (R&R) framework. Prins
(2004) contributed a hybrid genetic algorithm, the first 2. Problem Description
population-based framework for vehicle routing prob- The CVRP primarily considered by this paper is de-
lems (VRPs), which improved significantly on the prior fined as follows. Let &
{9, !} be a complete undi-
approaches. This approach was refined by Vidal et al. rected graph in which 9 is the set of vertices and !
(2015), thereby realizing the state-of-the-art VRP heuristic. the set of arcs connecting the vertices. The vertices
The vast majority of VRPs consider minimizing the vi ∈ 9 for i ∈ {0, . . . , n} represent locations in a two-
total travel distance as their objective. However, a much dimensional space where v0 corresponds to the depot
smaller number of variants require a hierarchical ob- and the other n vertices to customers having a de-
jective to be handled, which is primarily the minimi- mand qi . Each arc (i, j) in !
{(i, j) : i, j ∈ 9, i 
j} is
zation of the number of vehicles and secondarily the associated with a cost cij . An unlimited homogeneous
minimization of the total distance. In such problems, fleet of vehicles with capacity Q is situated at the
solutions associated with smaller fleet sizes are pre- depot. The CVRP consists of designing vehicle tours at
ferred over solutions whose total travel distance is minimum cost such that each customer is served exactly
smaller but that require additional vehicles. Although once by a single vehicle, and each tour starts at the depot
this objective might appear artificial, it reflects the at- v0 , serves customers without exceeding the vehicle’s
titude of real-world transportation companies who far capacity Q, and finally ends at the same depot v0 .
prefer to be able to accommodate all requests with their Although the CVRP constitutes the primary prob-
own fleet rather than having to outsource certain re- lem addressed by this paper, the general applicability
quests to other companies—the cost of which is pro- of SISRs is demonstrated by way of solving several other
hibitively high. Solution methods that address this associated problems within the domain of vehicle rout-
hierarchical objective are generally two-stage approaches ing. Key references for papers that provide descrip-
wherein the number of vehicles is minimized in the first tions of individual problems are supplied in Table 1.
stage and the distance in the second stage, such as the The remainder of this paper refers to a string as a
methods by Pisinger and Røpke (2007) and Nagata, sequence of consecutive locations in a tour. Strings
Bräysy, and Dullaert (2010). Hybrid methods, for ex- may contain at most all customers served by the tour.
ample, by Vidal et al. (2013), are less common but also The number of customers included in a string is re-
possible. ferred to as its cardinality, with tour t’s cardinality
Although the initial ambition of heuristics is noble, denoted by |t|.
they have gradually become increasingly complicated.
This paper seeks to remedy this by introducing a low- 3. Ruin and Recreate (R&R) Strategies
level yet simultaneously powerful and fast heuristic Classical neighborhoods for VRP heuristics—such as,
that is sufficiently adaptable and may be easily incor- 2-opt (Croes 1958); 2-opt* (Potvin and Rousseau
porated into any current or future VRP approach. This 1995); k-opt (Lin 1965); Or-opt (Or 1976); exchange,
ruin and recreate approach’s improvements to the cross, and relocate (Savelsbergh 1988); and CROSS
CVRP state of the art and its not only successful but exchange (Taillard et al. 1997)—are all defined via slight
also competitive application to a host of related modifications to the incumbent solution. They therefore
VRPs will be demonstrated and documented. exhibit limited neighborhood size complexity, imply-
Structurally, this paper begins by first offering a de- ing that a full neighborhood 1(s) of a solution s may
tailed problem definition of the CVRP, followed by an be evaluated within a reasonable amount of com-
overview of the other VRPs addressed during experi- putational time. Very large neighborhoods often oc-
mentation. Next, Section 3 provides an introduction to cur when a significant part of the solution is modified,
ruin and recreate, whereas Section 4 provides a brief making a full neighborhood exploration impractical.
description of simulated annealing (SA), the metaheu- Therefore, neighbors may be selected from a reduced
ristic employed for guiding the local search. Section 5 candidate set #(s) ⊆ 1(s), or a single neighbor may be
introduces slack induction by string removals (SISRs), generated by first ruining the current solution before
this paper’s novel R&R contribution. Following this, recreating it into a feasible one. Such a method may be
the new absences-based acceptance criterion for fleet defined by a ruin method 5− and recreate method 5+ .
minimization is detailed in Section 6. Parameters asso- This underlying concept is present in a variety of re-
ciated with this work are detailed in Section 7. Com- search papers. To our knowledge, it was established
putational results for the CVRP and a variety of asso- for the first time by Dees and Smith (1981) in their
ciated problems are provided in Section 8, and the rip-up and reroute strategies for wiring point-to-point
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 3

Table 1. CVRP and Associated Problems

CVRP

The capacitated VRP, also known as the classical VRP, was introduced by Dantzig and Ramser (1959) as the truck dispatching problem. The
algorithms by Subramanian, Uchoa, and Ochi (2013) and Vidal et al. (2014) are considered as the current best approaches.
Multidepot VRP

Vehicles are available at a fixed number of depots. Each vehicle is assumed to return to the depot where its tour starts. This problem was
introduced by Gillett and Johnson (1976). The algorithms by Pisinger and Røpke (2007); Subramanian, Uchoa, and Ochi (2013); and Vidal et al.
(2012) are considered as the current best approaches.
VRP with backhauls (VRPB)

The vehicle routing problem with backhauls, introduced by Goetschalckx and Jacobs-Blecha (1989), considers serving two sets of customers:
delivery (line haul) and pickup (backhaul) customers. Each tour must serve at least one delivery customer, and all delivery customers must be
served before any pickup customers are served. A tour’s total delivery and pickup demand should not exceed its vehicle’s capacity. The
algorithms by Gajpal and Abad (2009), Zachariadis and Kiranoudis (2012), and Vidal et al. (2014) represent the best solution approaches for
this problem.
VRP with mixed delivery and pickup (VRPMDP)

This problem, introduced by Min (1989), relaxes the VRPB by enabling any order of delivery and pickup customers within a tour. This implies
that a vehicle’s used capacity may increase and decrease during the tour, but should never exceed the vehicle’s capacity. Best solutions are
obtained by Subramanian, Uchoa, and Ochi (2013) and Vidal et al. (2014).
VRP with simultaneous delivery (VRPSDP)

Unlike the VRPMDP this problem, also introduced by Min (1989), does not require customers to be associated exclusively with either delivery
or pickup demand. This implies that a single customer can have both delivery and pickup requests. Best solutions are obtained by the same
authors as for the VRPMDP.
Multidepot VRP with mixed delivery and pickup

Min (1989) generalizes the VRPSDP by addition of multiple depots into the multidepot VRPMDP. Best approaches are those of Røpke and
Pisinger (2006b) and Subramanian, Uchoa, and Ochi (2013).
Open VRP

In contrast to the CVRP, vehicles must not return to the depot after serving the last customer in the open VRP, introduced by Sariklis and
Powell (2000). This reflects situations in which customers are assigned to external carriers. Current best approaches are those proposed by
Subramanian, Uchoa, and Ochi (2013) and Vidal et al. (2014).
Cumulative CVRP

The cumulative CVRP (Ngueveu, Prins, and Wolfler Calvo 2010) requires, similar to “minavg routing” (Campbell, Vandenbussche, and
Hermann 2008), the sum of arrival times at customers to be minimized rather than the total traveled distance. The approach of Vidal et al.
(2014) obtained the current best known solutions for this problem.
VRPTW

The VRP with time windows, introduced sometime in the late 1970s or early 1980s, extends the CVRP with temporal constraints. Customers
are associated with individual time windows within which they prefer to be served. Vehicles start their tour in the depot and are assumed to
arrive back at the depot without exceeding the maximum tour duration as defined in Gehring and Homberger (1999). The current best
approaches are those proposed by Repoussis, Tarantilis, and Ioannou (2009); Nagata, Bräysy, and Dullaert (2010); and Vidal et al. (2013).
PDPTW

Literature on the pickup and delivery problem also dates back to the same time frame as the VRPTW. As defined by Dumas, Desrosiers, and
Soumis (1991), goods require transportation between pickup and delivery locations. Similar to the VRPTW, temporal constraints are present.
The PDPTW should not be confused with delivery and pickup problems where goods are transported to or from the depot, rather than
between the customers themselves. Røpke and Pisinger (2006a), Nagata and Kobayashi (2010), and Curtois et al. (2018) have contributed the
current best solution approaches.
Dial-a-ride problem (DARP)

In the dial-a-ride problem, which was first examined by Wilson et al. (1971), transportation of customers between specified pickup and
delivery locations are considered. The most recent definition of the DARP (Cordeau and Laporte 2003) is in line with the work of Jaw et al.
(1986). The DARP restricts the PDPTW by imposing a maximum ride time in order to manage user inconvenience. Best approaches are those of
Chassaing, Duhamel, and Lacomme (2016); Masmoudi et al. (2017); and Gschwind and Drexl (2019).
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
4 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

connections in electronic design automation. Shaw 4. Local Search Metaheuristic

(1998) introduced large neighborhood search (LNS), The SISRs neighborhood search, detailed in Section 5,
wherein the ruin phase is implemented as the removal is guided by simulated annealing. This well-known
of related customers (related in terms of time, distance, metaheuristic, introduced by Kirkpatrick, Gelatt, and
and being served by the same vehicle). A branch-and- Vecchi (1983), is based on the statistical model concern-
bound technique optimally inserts the removed cus- ing energy changes in annealing systems (Metropolis
tomers in the recreate phase. Shaw (1998) applied LNS et al. 1953). In this study, the system’s initial tem-
to both the CVRP and the vehicle routing problem with perature 70 is reduced by the exponential cooling
time windows (VRPTW). The term ruin and recreate schedule (Equation (1)) to its final temperature 7f in f
was introduced by Schrimpf et al. (2000, p. 139), who iterations. Cooling constant c is obtained by way of
applied the technique to a number of prominent prob- Equation (2):
lems, including the TSP and VRPTW. They ruined so-
lutions by removing randomly selected customers, 7k+1
c 7k 0 ≤ k <f, (1)
customers within a certain radius, or one single string. ( )
7f 1/f
The solution is recreated by greedily reinserting the re- where c
70 > 7f > 0. (2)
70
moved customers in a random order at minimum cost.
Røpke and Pisinger (2006a) defined several 5− and The probability h of accepting a neighbor solution s∗
several 5+ methods that compete to modify the so- depends on the current temperature 7 and the dif-
lution in their adaptive large neighborhood search ference in energy ΔE
dist(s∗ ) − dist(s) relative to the
(ALNS) framework and applied their framework to current solution s (Equation (3)). Neighbor solution s∗
the pickup and delivery problem with time windows is accepted if a random selected value from the con-
(PDPTW). They implemented seven different strategies tinuous uniform distribution U(0, 1) is less than the
for selecting customers for removal: random, worst, probability h(ΔE, 7). Therefore, the acceptance cri-
related, cluster, time oriented, historical node pair, and terion is expressed by way of Equation (4):
historical request pair. The solution is recreated by
employing either greedy or regret insertion (Potvin h(ΔE, 7)
exp(−ΔE/7), (3)
∗
and Rousseau 1993) with or without a noise function. dist(s ) < dist(s) − 7 ln(U(0, 1)). (4)
Pisinger and Røpke (2007) applied ALNS to the rich
pickup and delivery problem with time windows. A solution s is represented by two sets s
{T, A}. Set T
Recent heuristic development based on the R&R is the set of tours, each of which serves at least one
principle builds on the general ALNS framework by customer, whereas set A is the set of absent customers,
enlarging the set of 5− and 5+ methods. Examples of all of whom are not served by any tour t ∈ T. The ini-
such studies include, but are not limited to, the pa- tial solution contains an equal number of tours as
pers by Laporte, Musmanno, and Vocaturo (2010) and there are customers defined in the problem; therefore,
Ribeiro and Laporte (2012). In distinct contrast to the only one customer is served in each tour. This initial
recent trend of introducing more and more 5− and 5+ solution is passed to the local search procedure, de-
methods, this paper introduces a streamlined R&R tailed by way of Algorithm 1. First, the initial solu-
implementation: slack induction by string removals tion is saved as the best solution sbest found thus far
(SISRs; phonetically, “scissors”). This proposed ap- (line 2). Temperature 7 is set to the initial temperature
proach employs a single 5− method and a single 5+ 70 (Line 3). Next, f improvement iterations are per-
method: adjacent string removal and greedy insertion formed (lines 4–10). During each iteration, a neighbor
with blinks, respectively, with the significant addi- solution s∗ is generated from current solution s by
tional benefit of being highly reproducible. An over- employing SISRs’ R&R procedure (line 5). This neigh-
view of these R&R methodological implementations is bor is accepted as the new current solution s if SA’s ac-
provided in Table 2. ceptance criterion is satisfied (lines 6 and 7) and saved if

Table 2. Overview of R&R Methodologies

Method 5− 5+

Shaw (1998) (LNS) related removal optimal insertion

Schrimpf et al. (2000) random, radial, or string removal greedy insertion
Pisinger and Røpke (2007) (ALNS) random, worst, related, cluster, greedy or regret insertion with
time-oriented, or historical removal or without noise function
This study’s approach (SISRs) adjacent string removal greedy insertion with blinks
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 5

it improves on the best solution (lines 8 and 9). Finally, potential failures of alternative 5− methods within
the system is cooled down (line 10) before proceeding the context of slack induction.
onto the next iteration.
Premise 1 (Remove a “Sufficient” Number of Customers).
Removing a small number of customers (such as in Figure 1(a1))
Algorithm 1 (Local Search Metaheuristic)
may fail to introduce sufficient capacity slack in the vehi-
Input: Initial solution s
{T, A}, where T is the set of
cles for serving customers of higher demand. Although it is
tours and A the set of absent customers visually clear (Figure 1(a2)) by which vehicles the removed
1: procedure LOCAL-SEARCH(s)
customers should be served, it may be impossible if the ve-
2: sbest ← s
hicles gained insufficient capacity slack.
3: 7 ← 70
4: for f iterations do Premise 2 (Remove “Adjacent” Customers). Removing
5: s∗ ← SISRs-RUIN-RECREATE(s) customers who are not geographically close (not adjacent) to
6: if dist(s∗ ) < dist(s) − 7 ln(U(0, 1)) then each other (such as randomly selected customers in Figure 1(b1))
7: s ← s∗ generally introduces capacity and spatial slack scattered
8: if dist(s∗ ) < dist(sbest ) then across many tours (Figure 1(b2)). Such slack is unlikely to
9: sbest ← s∗ be beneficial when recreating the solution. The recreated so-
10: 7 ← c7 lution may be comparable to the result of multiple R&R it-
11: end procedure erations wherein a small number of customers is removed each
time.
5. Slack Induction by String Removals Premise 3 (Remove Strings of Customers). Removing
This section provides a detailed description of the SISRs customers of which their selection is only based on adjacency
method. The concepts of capacity slack and spatial slack (as in the case of radial removal, Figure 1(c1)) may often fail
and three premises are introduced in Section 5.1, to capture complete detours that remain present in the ruined
which lays out the rationale for the proposed method. state (Figure 1(c2)). This may be remedied by increasing the
Detailed descriptions of SISRs’ 5− and 5+ methods radius; however, many more customers than strictly nec-
may be found in Sections 5.2 and 5.3, respectively. essary may be removed as well. Therefore, removing strings
may be much more effective.
5.1. Rationale If one can remove multiple strings that are geo-
When customers are removed from a tour’s sequence, graphically close to each other—removal of adjacent
several types of slack may be introduced. First, the most strings—a very effective yet simple of origin 5− method
apparent type of slack may be capacity slack, referring may be obtained. An example of such adjacent
to the vehicle’s capacity that is set free. Second, spa- string removals is depicted in Figure 2. Only two ad-
tial slack may be identified if one imagines that the jacent strings are selected for removal in Figure 2(a),
vehicle’s travel distance is limited. Given this limited and the ruined state of the solution is illustrated
distance, the vehicle’s reachable area may be defined. in Figure 2(b). Notice that (1) spatial slack is greatly
This area appears as an elliptical region around introduced in only two tours, and (2) the spatial slacks
each edge, where each edge’s end points are the focus of both overlap significantly. Ideally, the ruined state
points for its corresponding ellipse. This area increases
obtained by the 5− method (in Figure 2(b)) is recre-
when the locations of the removed customers are no
ated by 5+ in the solution illustrated in Figure 2(c).
longer visited, and we refer to this increase as spatial
slack. Although the reduced travel distance contributes
the most to gaining spatial slack, it is not a necessary 5.2. Ruin Method—Adjacent String Removal
condition. For example, removing the intermediate of The proposed 5− method is based on Premises 1 to 3
three collinear locations does not reduce the travel as follows. The number of removed customers is man-
distance, but the reachable area (spatial slack) does aged by means of the average number of removed
increase given the larger focal distance. Although the customers c̄. In addition, the 5− method may be set
vehicle’s travel distance is not limited within the CVRP for removing many strings of small cardinality or few
definition, one may express spatial slack relative to the strings of high cardinality by means of the maximum
vehicle’s original tour as the additional reachable area cardinality of the removed strings Lmax . Let s be the
delimited by its original travel distance. solution being ruined. A valid number of removed
An effective 5− method should introduce both strings ks and cardinality lt for every string (each of
types of slack into the solution such that the possi- which are removed from a different tour t) is selected
bilities for improving the solution are greatly en- as follows. The maximum string cardinality lmax
s is ini-
hanced. The following premises serve as the basis tially set to the minimum of Lmax and the solution’s
for such an 5− method. Each premise addresses the average tour cardinality |t ∈ T| (Equation (5)) before
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
6 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

Figure 1. Examples of Ruin Methods That Introduce Insufficient Slack

Notes. The figure shows the removal of too few customers (panels (a1) and (a2)), randomly selected customers (panels (b1) and (b2)), and radially
selected customers (panels (c1) and (c2)). Panels (a1), (b1), and (c1) illustrate the partial solution, wherein the customers selected for removal are
colored white and outlined by dashed lines. Panels (a2), (b2), and (c2) illustrate the ruined state, wherein spatial slack relative to the original state
is shaded gray.

the maximum number of strings ksmax is derived made. First, for each customer c, an adjacency list adj(c)
(Equation (6)): is assumed to be available for identifying adjacent
strings. The list contains all customers ordered by
lmax
s
min{ Lmax , |t ∈ T|} lmax
s ∈ R+ , (5) increasing distance from c, which includes c as its first
4 c̄ element. Second, each string being removed originates
ksmax
−1 ksmax ∈ R+ , (6)
1 + lmax
s
from a different tour such that spatial slack, as well as
ks
U( 1, ksmax + 1 ) ks ∈ N+ . (7) capacity slack, is spread across a number of tours, con-
trolled by ks . Therefore, a tour may be ruined at most
The number of strings ks to be removed is obtained once, and removing ks strings implies ruining ks tours.
by rounding down a randomly selected real value from
the continuous uniform distribution U(1, ksmax + 1) by Algorithm 2 (SISRs’ Ruin Method)
way of Equation (7). The maximum cardinality of Input: Solution s
{T, A}, where T is the set of tours
each string is limited by the initial limit lmax
s and the and A the set of absent customers
tour’s cardinality |t| from which the string is to be re- 1: procedure RUIN(s)
s , ks , ks ← calculate(Equations(5), (6), and (7).
lmax
moved (Equation (8)). Finally, the cardinality for the max
2:
string to be removed from tour t is obtained by round- 3: cseed
s ← randomCustomer(s)
ing down a real value selected from U(Lmin , lmaxt + 1) 4: R←∅
(Equation (9)): 5: s ) and |R| < ks do
for c ∈ adj(cseed
6: if c ∈/ A and t ∈/ R then
lmax
t
min{ |t|, lmax
s } lmax
t ∈ R+ , (8) 7: c∗t ← c
+
lt
U( 1, lmax
t +1 ) lt ∈ N . (9) 8: t , lt ← calculate(Equations (8) and (9).
lmax
9: A ← A ∪ removeSelected(t, lt , c∗t )
Before the procedural steps of the ruin phase 5− are 10: R ← R ∪ {t}
detailed by way of Algorithm 2, two assumptions are 11: end procedure

Figure 2. An Example of Adjacent String Removals The procedural steps of the ruin phase 5− are given
by Algorithm 2. First, values for lmax max
s , ks , and k
are derived by Equations (1)–(3) (line 2). A randomly
selected customer serves as the seed customer, cseed s ,
from which string removals are initiated (line 3). The
auxiliary set R, for storing ruined tours, is initially
empty (line 4). The adjacency list from cseed
s is iterated
over from the first customer to the last until set R
contains ks ruined tours, indicating that ks strings have
been removed (lines 5–10). In each iteration, if the
customer c has not been removed (c ∈/ A) and its
serving tour t has not been ruined (t ∈/ R; line 6), a
string is removed from this tour (lines 7–10). Such a
customer is special in the sense that it is the closest
Note. The figure shows (a) selection of two adjacent strings, (b) a
ruined state with significant spatial slack induced, and (c) an improved customer to cseed
s out of all customers served by t,
recreated state. denoted by c∗t (line 7). The maximum cardinality lmax t
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 7

and final cardinality lt of the string to be removed Figure 4. Example of Split String Removal
from tour t are obtained by way of Equations (4) and
(5) (line 8). The string is removed via the “string” or
“split string” procedure based on c∗t and lt . In either
case, lt customers are selected and removed from t be-
fore adding them to set A (line 9), and the ruined tour
is added to set R (line 10).
The string procedure removes a randomly selected
string of cardinality lt in tour t, which includes cus-
tomer c∗t . By removing a string that includes c∗t , other
customers inside this string are also likely to be lo- Note. The figure shows (a) random selection of a string when l
5
cated close to cseed . Therefore, every call to this pro- and m
2 and (b) preserving a substring of m
2 customers.
cedure will result in a string removal of customers
much greater probability, indicating that the customers
close to the seed customer or, simply, strings adjacent
removed are close to the depot (see Figure 5(b2)).
to each other. An example where lt
3 is illustrated in
Figure 6 presents an example of the ruin phase
Figure 3. One of the three possible strings is randomly
where two strings (ks
2) are removed from solution s.
selected for removal.
First, the seed customer cseed
s is randomly selected (see
The split string procedure begins much like the
string procedure, by randomly selecting a string of car- s ), is iter-
Figure 6(a)). Following this, the list adj(cseed
ated over. Because cseed is always the first element in
dinality l + m that includes customer c∗t (see Figure 4(a)). s
the list and no tours are ruined, c∗t1
cseed
s . A string of
However, the ruin phase bypasses and preserves
length lt1
4 is removed by the string procedure, which
a random substring of m intervening customers, as
includes c∗t1 (see Figure 6, (b) and (c)). In Figure 6(d),
shown in Figure 4(b). The number of preserved cus- ∗
tomers m is determined as follows. Initially m
1, and s ) is iterated over until the next customer (ct2 )
adj(cseed
is found in a not-yet-ruined tour. The second string
its current value is maintained either if a random num-
is also removed by the string method, and lt2 is, co-
ber U(0, 1) is larger than β, which is referred to as the
incidentally, again 4 (see Figure 6(e)). The final ruined
split depth, or once the maximum value for m is reached
state is illustrated in Figure 6(f).
(m
mmax
|t| − l). If neither of these conditions is sat-
isfied, m is incremented (m
m + 1), and the process 5.3. Recreate Method—Greedy Insertion with Blinks
repeats. Although basic greedy methods insert each customer
The probability of executing the split string is re- sequentially at the best position, one may deviate
ferred to as the split rate α. Examples are illustrated in slightly from the best position. Pisinger and Røpke
Figure 5 for string (panel (a)) and split string (panels (2007) apply a greedy insertion with noise function by
(b1) and (b2)) removals with small and large m-values. adding a randomized noise term to the insertion cost.
Customers being removed are colored white. The This paper, by contrast, introduces greedy insertion
string procedure removes a string, introducing spa- with blinks (Algorithm 3).
tial slack (see Figure 5(a)). In the split string removal First, the set of absent customers A is sorted accord-
procedure, a small value for m is obtained with a very ing to one of the following orders: random, demand,
low probability. When this occurs, customers close far, or close (line 2). Random enables the set’s inser-
to c∗t are removed while remaining spatially bonded tion without any ordering. Demand sorts customers
to the same regions as the preserved customers (see by demand, placing those with the largest demand
Figure 5(b1)). Otherwise, m
mmax is obtained with first. Far inserts the most distant customers from the
depot first. Finally, close inserts customers closest to the
Figure 3. Example of String Removal
depot first. Set A is sorted by random, demand, far,
and close by weights equal to four, four, two, and one,
respectively. Each customer c ∈ A (line 3) is inserted into
solution s at the “best” position P as follows. Initially,
P is set to null, indicating that no position for c is se-
lected (line 4). All current tours that are part of the so-
lution (t ∈ T) are iterated over in a random order (line 5).
When a tour t has sufficient capacity slack to serve c, all
positions Pt inside this tour are iterated over (line 6).
Each position is evaluated with a probability of 1 − γ,
Note. There are three possible string selections, which include c∗t from otherwise skipping the position as if the algorithm
tour t when lt
3. blinks, and therefore, γ
0 represents the blink rate
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
8 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

Figure 5. Examples of String and Split String Removals

Note. The figure shows (a) string removal and split string removal when m equals the (b1) lowest or (b2) highest possible value.

(line 7). If a position Pt is found for which the cost of Of all feasible options for inserting a customer, one
inserting c is lower than the current best position P, Pt may make a random choice or use a heuristic to pick
becomes the new best position (lines 8 and 9). If no the best option (greedy insertion). The effect of blinking
position is found in existing tours, a new empty tour is is reflected by the probability p of selecting a specific
created to serve c (lines 10–12). Finally, c is inserted option based on its rank r. A blink rate of γ implies
at P and removed from the set of absent customers that the best option, which is ranked first (r
1), is
(lines 13 and 14). Blink rate γ
0 denotes that cus- selected with probability p(1)
(1 − γ). If the best-
tomers are always inserted at the best position, a ranked option is blinked over, the probability of select-
strategy proven to be suboptimal after experimenta- ing the second-best option is p(2)
(1 − γ) γ, and the
tion. This may appear somewhat counterintuitive third-ranked option equals p(3)
(1 − γ) γ2 . The se-
given that it is more logical to assume that consistently lection probability for each rank is expressed by the
inserting each customer at the best possible position exponential function
would result in the highest-quality solution.
p(r)
(1 − γ) γ(r−1) r ∈ {1, . . . , ∞}. (10)
Algorithm 3 (SISRs’ Recreate Method)
Input: Ruined solution s
{T, A} where T is the set of Notice how the blinking algorithm itself is unaware
tours and A the set of absent customers of each option’s rank; it blinks only with probability
1: procedure RECREATE(s) γ while iterating over all options. Thus, blinking re-
2: sort(A) sults in rank-based selection probabilities without re-
3: for c ∈ A do quiring one to rank the options first, in contrast to
4: P ← null heuristic-biased stochastic sampling (HBSS; Bresina 1996),
5: for t ∈ T (which can serve c) do a closely related selection algorithm. A comparison
6: for Pt in t do between HBSS and the blinking algorithm can be
7: if U(0, 1) < 1 − γ then found in Online Appendix A.
8: if P
null or costAt(Pt ) < costAt(P)
then 6. Fleet Minimization
9: P ← Pt This paper introduces a two-stage approach to accom-
10: if P
null then modate a hierarchical objective that is primarily the
11: T ← T ∪ {new tour t} minimization of the number of vehicles. The newly de-
12: P ← position in t veloped absences-based neighbor acceptance criterion
13: insert c at P is employed during the vehicle minimization stage, de-
14: A ← A \ {c} tailed by way of Algorithm 4. The solution that serves
15: end procedure all customers with the lowest number of vehicles is

Figure 6. An Example of SISRs’ Ruin Method

Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 9

represented by sbest (line 2). For each customer c, the the proposed algorithm and are therefore employed
number of solutions where c was not served by any throughout the experiments reported in Section 8.
tour (absent from the solution) is tracked by its as-
sociated absence counter absc . This counter is initially 7.1. Parameters for Distance Minimization
set to 0 for all customers (lines 3 and 4). In each it- The benchmark sets employed throughout this analysis
eration, the following steps are performed. First, a were introduced by Uchoa et al. (2017) for the CVRP,
neighbor solution s∗ is obtained by SISRs’ 5− and 5+ Gehring and Homberger (2001) for the VRPTW, and
procedures (line 6). This neighbor is accepted if (1) the Li and Lim (2001) for the PDPTW. Whereas results for
number of absent customers decreased or (2) the sum the CVRP can be directly compared by the solution’s
of absence counters associated with the absent cus- distance, results for the VRPTW and PDPTW are not
∑
tomers sumAbs(A)
c∈A (absc ) decreased (lines 7 and directly comparable when the number of utilized ve-
8). If the current solution is feasible (A∗
∅), a copy is hicles differs. Solutions employing fewer vehicles usu-
saved as sbest . Afterward, the tour, whose served cus- ally result in much longer travel distances. Therefore, the
tomers represent the lowest sum of absence counters, is number of utilized vehicles is fixed to the most fre-
removed from the current solution (lines 9–11). Finally, quently obtained number per instance, thereby ensur-
each absent customer in the neighbor solution has ing a fair comparison. For each evaluated parameter
its associated absence counter incremented (lines 12 setting, the complete benchmark set is solved once.
and 13). Results for all settings are compared by way of their
average ranks, presented throughout Tables 4, 5, and 6.
For each benchmark instance, every parameter set-
Algorithm 4 (Fleet Minimization) ting is ranked by its obtained result with respect to
Input: Solution s
{T, A} where T is the set of tours those obtained by all other settings. The average rank
and A the set of absent customers for each setting is the average of all its obtained ranks.
1: procedure FLEET-MINIMIZATION(s) Low average ranks represent the best-performing set-
2: sbest ← s tings and are colored dark gray. Meanwhile, the highest
3: for c defined in the problem do average ranks correspond to the worst-performing set-
4: absc ← 0 tings and are colored light gray.
5: while minimizing tours do Because maximum string cardinality Lmax and the
6: s∗ ← SISRs-RUIN-RECREATE(s) average number of removed customers c̄ both influ-
7: if |A∗ | < |A| or sumAbs(A∗ ) < sumAbs(A) then ence the maximum number of strings Kmax , the values
8: s ← s∗ of both these parameters are varied simultaneously
9: if A∗
∅ then in order to determine which combination performs
10: sbest ← s∗ best. Given that average computational run time is
11: remove t ∈ T with the lowest sumAbs(t) proportional to c̄, changes to this value will result in
12: for c ∈ A∗ do varying run times. This is compensated for by adjusting
13: absc ← absc + 1 the number of iterations inversely proportional to c̄.
14: end procedure This enables a comparison of different settings for c̄
In essence, neighbor s∗ is accepted only if the number while maintaining equal computational effort or
of absent customers |A∗ | decreased or if the sum of calculation time. Average ranks are presented in
absence counters associated with the absent customers Table 4. Note that not all combinations are feasible.
sumAbs(A∗ ) decreased. This absences-based acceptance Given that at least one string is removed in each iter-
criterion for minimizing fleet size proved to be very ation, it is impossible to obtain an average number of
efficient without complicating the overall method. In removed customers less than (1 + Lmax )/2. For example,
fact, it requires only the replacement of SA’s distance- when Lmax is set to 5, c̄ should be equal to (1 + 5)/2
3
based acceptance criterion by this absences-based ac- or greater. Therefore, smaller values for c̄ such as 2 have
ceptance criterion during the fleet minimization phase. not been considered. It is observed that for both the
CVRP and PDPTW, the best parameter combinations are
identical (c̄
10, Lmax
10), whereas slightly larger
7. Parameters values (c̄
15 and Lmax
20) perform better when
Parameter variations for both the 5− and 5+ methods solving the VRPTW.
are evaluated for the CVRP, VRPTW, and PDPTW. At Next, the split rate α and split depth β are varied
most, two parameters are varied simultaneously to simultaneously. The average rank obtained by each
determine their influence on the algorithm’s perfor- combination is presented in Table 5. In general, the
mance while the values of all other parameters remain best average ranks are obtained when split rate α is set
fixed. These predefined values, presented in Table 3, within the range 0.25–0.75 in combination with high
were observed as high performing while developing values for split depth β such as 0.9 or 0.99. Whereas this
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
10 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

Table 3. Parameter Settings for SISRs and SA

SISRs’ 5−

c̄
10 An average of 10 removed customers has proved a good ruin parameter. When c̄ is equal to 10, it
implies the removal of at least one and at most 19 customers. Decreasing this parameter, for
example, to five, was observed to be insufficient for diversification. By contrast, increasing c̄, for
example, to 20, results in extensively ruined solutions. Such solutions cannot easily be recreated
into improving solutions and most likely result in highly divergent neighbors, which disables
intensification of potentially good solutions.
Lmax
10 The cardinality of removed strings is limited to 10 because removing longer strings fails to improve
final solution quality. This confirms the observations made by Fahrion and Wrede (1990), whose
chain-exchange procedure limits string cardinality to half the average tour length. By employing
Equation (7) the number of removed strings ks will range from 1 to 3 for solutions with large tours
(|t ∈ T| ≥ Lmax ) and 1 to 19 for solutions whose tours all have a cardinality of 1 (|t ∈ T|
1). This
situation almost always only occurs during the first iteration of the search (see Section 4).
α
0.5 By setting the split rate to 50% (α
0.5), both string and split string procedures are executed with
equal probability.
β
0.01 This setting results in values for the number of preserved customers m
1, 2, 3, . . . , mmax having
respective probabilities
p
1.00%, 0.99%, 0.9801%, . . . , (100 − pm
1 − pm
2 − pm
3 − · · · − pm
mmax −1 )%. As such, it is very
likely that customers removed within a split string removal are those at the start or end in the tour’s
sequence.
SISRs’ 5+

γ
0.01 It was observed that a blink rate of 1% (γ
0.01) was the most effective in terms of final solution
quality.
Fleet minimization

iterations
10% of all iterations Generally the minimum number of vehicles is found within three minutes. Therefore, dedicating 10%
of all iterations to the fleet minimization is considered to be sufficient.
SA

70
100, 7f
1 The search begins at an initial temperature 70
100 and ends at the final temperature 7f
1.
iterations
it(v) The number of iterations it is determined as a function of problem size v by linear interpolation. The
minimum and maximum problem sizes are 100 and 1,000, respectively, as per Uchoa et al. (2017).
The present study set it(100)
3 × 107 and it(1, 000)
3 × 108 , thus enabling the direct comparison
of SISRs’ calculation times against those from the aforementioned paper.

well-performing region is less pronounced for the the large instances. For both the VRPTW and PDPTW,
PDPTW, it is distinctive for both the CVRP and the the best value tends toward γ
0.05.
VRPTW.
Finally, blink rate γ is varied while all other pa- 7.2. Parameters for Fleet Minimization
rameters remain fixed. Average ranks obtained for The performance of Algorithm 4 is evaluated on the 1,000-
the CVRP exhibit two peaks. Further analysis of customer benchmark set of Gehring and Homberger
these results showed that the peak at γ
0 is in- (2001) by taking the best of five runs per instance. The
troduced by results on the small instances, whereas the cumulative number of vehicles (CNV) for each set is
second peak at γ
0.05 emerges because of those for compared against the (best of five) CNV obtained

Table 4. Maximum String Cardinality and Average Number of Removed Locations

CVRP VRPTW PDPTW

c̄ c̄ c̄

Lmax 2 5 10 15 30 2 5 10 15 30 2 5 10 15 30

2 12.2 8.1 8.9 11.0 12.6 13.4 10.9 11.1 11.7 12.7 10.1 8.6 9.2 9.3 10.8
5 4.7 4.7 5.4 8.9 7.2 6.3 6.1 8.2 5.3 5.5 6.3 7.6
10 4.1 5.5 7.8 5.0 4.1 5.7 4.8 5.2 6.6
20 5.3 7.0 3.6 4.6 6.1 6.6
50 8.0 4.9 7.1

Notes. The average ranks are shaded in proportion to their values. Low average ranks are shaded dark
gray and correspond to the best performing settings. Highest average ranks are shaded light gray and
correspond to the worst performing settings.
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 11

Table 5. Split String Parameters α and β

CVRP VRPTW PDPTW

α α α

β 0.0 0.25 0.50 0.75 1.0 0.0 0.25 0.50 0.75 1.0 0.0 0.25 0.50 0.75 1.0

0.00 14.2 12.3 12.5 13.8 13.4 12.7 12.1 12.3 12.1 13.6 11.0 10.5 11.1 11.3 11.3
0.50 14.3 13.2 12.2 12.8 11.7 12.8 12.4 11.2 12.5 12.6 10.6 10.6 9.8 11.4 11.1
0.90 14.5 10.5 10.4 8.6 9.7 13.0 10.5 10.8 10.3 12.3 12.0 10.0 9.2 10.2 9.9
0.99 14.1 9.4 9.0 10.7 12.5 13.6 10.1 10.3 10.9 14.2 11.2 10.6 10.6 11.3 13.2
1.00 13.2 11.2 10.0 10.1 13.7 12.9 10.1 11.6 12.4 17.2 11.5 11.0 10.9 12.5 18.0

Notes. The average ranks are shaded in proportion to their values. Low average ranks are shaded dark
gray and correspond to the best performing settings. Highest average ranks are shaded light gray and
correspond to the worst performing settings.

Table 6. Blink Rate Parameter γ

CVRP VRPTW PDPTW

γ α α

0.0 0.001 0.01 0.05 0.10 0.0 0.001 0.01 0.05 0.10 0.0 0.001 0.01 0.05 0.10

2.57 2.76 2.87 2.53 3.12 2.91 2.89 2.75 2.72 3.06 3.00 2.79 2.72 2.43 2.63

Notes. The average ranks are shaded in proportion to their values. Low average ranks are shaded dark
gray and correspond to the best performing settings. Highest average ranks are shaded light gray and
correspond to the worst performing settings.

by EAMA(100) (Nagata, Bräysy, and Dullaert 2010) observed. Consequently, the parameter settings pre-
and HGSADC (Vidal et al. 2013) in Table 7. In ad- sented in Table 3 are employed during the fleet min-
dition, results of SISRs with different Lmax settings imization phase when solving the VRPTW and
are reported. The best results are marked dark gray, PDPTW.
whereas results of SISRs are marked light gray when
better than the competing algorithms but not the best 8. Computational Results
of the different Lmax settings. The average calculation SISRs’ computational results are documented in this
time T is specified in minutes, which is at most 10 section. First, SISRs’ results for the most relevant CVRP
minutes for EAMA(100) and unknown for HGSADC. benchmark set are compared against those obtained
These results indicate the best settings may vary for by current state-of-the-art methods in Section 8.1. Fol-
different sets. For example, removing more strings lowing this, although the CVRP constitutes this work’s
of lower cardinality (Lmax
2) enables us to find the core focus, SISRs’ general applicability is verified by
best CNV for set C2, whereas removing strings of in- way of solving numerous CVRP variants. The very
termediate cardinality (Lmax
5) obtains a better CNV minor adjustments necessary to enable SISRs to solve
than EAMA(100) and HGSADC for set C1. However, each of these problems are detailed in Section 8.2.
after conducting a parameter analysis similar to that
performed in Section 7.1, no superior regions were 8.1. CVRP
Uchoa et al. (2017, p. 845) introduced a new bench-
Table 7. Fleet Minimization with Different Lmax Settings mark set because “[t]he existing sets suffer from at
least one of the following drawbacks: (i) became too
SISRs
easy for current algorithms; (ii) are too artificial;
EAMA(100) HGSADC Lmax
10 Lmax
5 Lmax
2 (iii) are too homogeneous, not covering the wide range
Set n CNV (T) CNV CNV (T) CNV (T) CNV (T)
of characteristics found in real applications.” This
recent benchmark set contains 100 instances whose
R1 1,000 919 (≤ 10) 919 919 (0.2) 919 (0.2) 919 (0.1) sizes range from 100 to 1,000 customers and that cover
R2 1,000 190 (≤ 10) 190 190 (0.0) 190 (0.0) 190 (0.0) a wide variety of characteristics. Uchoa et al. (2017)
C1 1,000 941 (≤ 10) 941 939 (2.8) 938 (2.4) 939 (2.7)
C2 1,000 291 (≤ 10) 288 291 (9.1) 292 (3.7) 288 (2.5)
report the detailed results obtained by two state-of-
RC1 1,000 900 (≤ 10) 900 900 (0.1) 900 (0.0) 900 (0.0) the-art heuristics, both conducted on a Xeon central
RC2 1,000 183 (≤ 10) 182 182 (0.3) 182 (0.3) 188 (0.4) processing unit (CPU) at 3.07 GHz: iterated local
Notes. Results are ranked first (best), second, and third by way of dark
search with set partitioning (ILS-SP; Subramanian,
gray, light gray, and no shading, respectively. T, Average calculation Uchoa, and Ochi 2013) and unified hybrid genetic
time in minutes. search (UHGS; Vidal et al. 2014).
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
12 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

Table 8. Benchmark Subsets which are highlighted in dark and light gray, respec-
n Instances No. of instances
tively, in the “New BKS” columns.
To provide a clear overview, average calculation times
Small 100–250 X-n101-k25–X-n247-k47 32 and gaps to the best-known solution in percentage
Medium 250–500 X-n251-k28–X-n491-k59 36 terms (100(result − BKS)/BKS) for the average and best
Large 500–1,000 X-n502-k39–X-n1001-k43 32
All 100–1,000 X-n101-k25–X-n1001-k43 100
results are summarized by way of the minimum, max-
imum, average, and median values in Table 9. These
values were taken directly from the bottom of Tables B1,
B2, and B3 in Online Appendix B.
This study’s experiments were performed on a Xeon UHGS outperforms the other methods for small
E5-2650 v2 CPU at 2.60 GHz and were compared against instances. For the medium set, UHGS and SISRs
the results of ILS-SP and UHGS, which were taken from perform equally well insofar as obtaining the best
Uchoa et al. (2017). No reimplementation of these solutions (“Best” in Table 9). Additionally, it was
methods was therefore required. The full bench- observed that SISRs was the most robust method.
mark set was divided into three sets for comparison Meanwhile, for the large set, SISRs outperforms both
based on problem size n: small, medium, and large heuristics on all aspects—it is the most robust method
(Table 8). that finds the best solutions in the least amount of
The complete results obtained by each heuristic are computation time.
documented in Online Appendix B, where Tables B1, The average solution quality (“Avg” columns in
B2, and B3 report the results for the small, medium, Tables B1, B2, and B3 in Online Appendix B) expressed
and large instances, respectively. The average dis- as gaps has also been visualized by way of box plots for
tance (“Avg”), the best distance (“Best”), and the each heuristic and the problem set in Figure 7. These
average calculation time in minutes (“Time”) obtained box plots demonstrate SISRs’ efficiency and robustness
after 50 runs per instance are presented in these ta- by exhibiting both a small median and spread of av-
bles. Heuristics are ranked first (best), second, and third erage gaps.
by way of dark gray, gray, and light gray, respec- Finally, average solution quality is analyzed by conduc-
tively, for the columns “Avg,” “Best,” and “Time.” ting a one-tailed Wilcoxon signed-rank test (Wilcoxon
The distance and number of vehicles of the best known 1945), which is a nonparametric test for comparing two
solution (BKS) are also reported. Throughout the en- independent groups and performed by the software
tire experimental campaign, 40 BKSs were improved environment R (R Core Team 2017). The only assump-
by SISRs and 40 solutions of equal quality were found, tion required for this test is that observations are

Table 9. Summarized Computational Results

ILS-SP UHGS SISRs

Avg Best Time Avg Best Time Avg Best Time

Small (100–250)
Min 0.00 0.00 0.1 0.00 0.00 1.4 0.00 0.00 0.2
Max 2.50 1.19 17.8 0.30 0.06 20.4 0.50 0.16 18.4
Avg 0.31 0.12 2.4 0.07 0.00 6.0 0.11 0.01 5.2
Median 0.19 0.00 1.6 0.03 0.00 5.4 0.05 0.00 5.0
Medium (250–500)
Min 0.00 0.00 2.0 0.00 0.00 6.5 0.01 −0.18 9.8
Max 1.54 0.73 60.6 0.49 0.10 86.7 1.35 0.28 58.4
Avg 0.54 0.17 23.1 0.24 0.02 30.3 0.20 0.02 27.0
Median 0.42 0.12 16.8 0.26 0.00 22.4 0.15 0.01 22.4
Large (500–1,000)
Min 0.00 0.00 27.3 0.11 0.00 33.1 −0.26 −0.42 60.9
Max 1.74 1.45 792.8 0.68 0.33 560.8 0.38 0.15 412.7
Avg 0.76 0.49 195.7 0.29 0.04 268.6 −0.01 −0.13 152.0
Median 0.75 0.39 144.7 0.26 0.02 258.6 0.01 −0.14 144.8
All (100–1,000)
Min 0.00 0.00 0.1 0.00 0.00 1.4 −0.26 −0.42 0.2
Max 2.50 1.45 792.8 0.68 0.33 560.8 1.35 0.28 412.7
Avg 0.53 0.26 71.7 0.20 0.02 98.8 0.10 −0.03 60.0
Median 0.40 0.11 17.7 0.20 0.00 22.4 0.07 0.00 22.4

Notes. Results are ranked first (best), second, and third place by use of the colors dark gray, gray, and light
gray respectively. Time, average calculation time in minutes; Min, minimum; Max, maximum; Avg, average.
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 13

Figure 7. Gaps of the Average Results

Note. Gaps of the average results to the best-known solutions (Uchoa et al. 2017) are visualized by way of box plots.

independent. The following null hypothesis H0 and than this level, the hypothesis is rejected, and the cell
alternative hypothesis H1 are considered, where X rep- is colored gray. The null hypothesis H0 and alterna-
resents ILS-SP or UHGS: tive hypothesis H1 are both rejected when considering
medium, large, and all benchmark sets, indicating
H0 : Avg(SISRs)
Avg(X); that SISRs significantly differs from both other heu-
H1 : Avg(SISRs) > Avg(X). ristics and that its average results are significantly
better.
The average results obtained by SISRs are compared In addition to the performance analysis employing
with the average results of ILS-SP and UHGS. Rejecting the benchmark set introduced by Uchoa et al. (2017),
H0 implies that the average results obtained by SISRs SISRs is applied to the classical benchmarks listed
are different from those obtained by alternative heuris- in Table 11. Detailed results obtained by SISRs with
tic X. Rejecting H1 implies that the values of the av- respect to these multiple benchmarks are docu-
erage results obtained by SISRs are lower than those mented in Online Appendix C. These experiments are
obtained by X or that SISRs’ average solution quality performed only in the interests of completeness.
is better than X’s. The obtained p-values are presented Going forward, we highly recommend that further
in Table 10. research use the most recent benchmark set of Uchoa
Although the overall significance level α is set to et al. (2017). The reasons why immediately become
2.5%, Bonferroni correction (Dunn 1961) is applied in clear on providing a brief comparative overview of
order to compensate for the family-wise error rate. the various benchmarks.
Bonferroni correction sets the significance level for The composition of each set is summarized in
testing individual hypotheses to α/m, where m is the Table 12: the number of instances (“#Inst”), a box plot
number of tests conducted using the same data. that represents problem size distribution, the distribu-
Because four tests are performed using the small, tion of customers (“Cust”), and the cumulative distance
medium, and large data sets and four again using (CD) obtained by SISRs. Customer distribution is cate-
all data, a total of eight hypotheses are tested each gorized as random (R), clustered (C), real geographic
time with the same data. Therefore p-values are coordinates (G), or artificially distributed along geomet-
assessed with respect to the Bonferroni-adjusted level ric shapes such as concentric circles or squares (A).
0.025/8
0.003125, or 0.3125%. If the p-value is lower All benchmark sets prior to R95 (Rochat and Taillard

Table 10. p-Values of the Wilcoxon Rank-Sum Test

Small Medium Large All

ILS-SP UHGS ILS-SP UHGS ILS-SP UHGS ILS-SP UHGS

H0 0.029239 0.022526 0.000013 0.002624 0.000000 0.000000 0.000000 0.00000

H1 0.014620 0.989361 0.000007 0.001312 0.000000 0.000000 0.000000 0.00000

Note. p-Values are shaded gray when the corresponding hypothesis is rejected.
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
14 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

Table 11. CVRP Benchmark Sets benchmarks. Therefore, the number of executed it-
erations was divided by 10 for this benchmark.
Acronym Authors

C69 Christofides and Eilon (1969)

8.2. CVRP Variants
C79 Christofides, Mingozzi, and Toth (1979)
F94 Fisher (1994) Various sets of benchmark instances from the lit-
A95 Augerat et al. (1995) erature were employed to evaluate SISRs’ perfor-
R95 Rochat and Taillard (1995) mance. Minimal adjustments were necessary to
G98 Golden et al. (1998) enable SISR to solve the CVRP variants addressed,
L05 Li, Golden, and Wasil (2005)
and therefore, a structured overview for the ad-
U17 Uchoa et al. (2017)
justments related to each problem attribute is pro-
vided by way of Table 13. If more than one change
was necessary, the changes are given in a numbered
1995) contain very few instances and are of very small list. The computational results were then compared
size. The first medium-sized instances were introduced against state-of-the-art methods for each CVRP vari-
in G98 (Golden et al. 1998), wherein customers are ant. A detailed breakdown of the results obtained
distributed along geometric shapes that are very un- by both SISRs and these state-of-the-art methods
likely to appear in real-world environments. L05 (Li, for each individual instance is presented in Online
Golden, and Wasil 2005) contains large instances that Appendix D. A causal consultation of these results
are based on those of G98 with similarly unrealistic cus- immediately reveals SISRs’ impressive dominance.
tomer distributions. U17 (Uchoa et al. 2017) is the only Despite the various state-of-the-art methods it is com-
benchmark set that exhibits a wide range of problem pared against, being general frameworks or dedicated
sizes with relevant customer distributions, vehicle single-problem heuristics, SISRs proves highly com-
capacities, and many other realistic characteristics. petitive, very often matching or improving the best-
Interested readers are referred to Uchoa et al. (2017) known solutions across all variants. Even more im-
for full details. SISRs’ results for each benchmark are pressively, for large-scale instances, which are more
summarized by the average gap values (“Gap”) in representative of real-world environments, SISRs
Table 12. produces many new best-known solutions. This is
Average gaps obtained by SISRs are at most 0.13% especially true for the PDPTW, for which SISRs
for all benchmark sets except for G98 and L05, where improves more than half the considered benchmark
average gaps were 0.44% and 0.82%, respectively. set’s instances (178 of 298, with 58 of these solutions
Note that initial calculation times for L05 were con- using fewer vehicles). Full details concerning these
siderably longer than for other instances of compa- improved solutions can be found in Table D11 in
rable size. This is very likely due to the high num- Online Appendix D, where improved solutions are
ber of customers per tour compared with all other colored dark gray and also marked with an asterisk

Table 12. Results for the Classical Benchmark Sets

Note. Set, the acronym of the benchmark set; #Inst, the number of instances; Cust, the distribution of customers which can be categorized as random
(R), clustered (C), real geographic coordinates (G), or artificially distributed (A); Gap, the average gap of 10 runs per instance obtained by SISRs.
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 15

Table 13. Adjustments Accommodating VRP Variants

Multidepot

(1) (5+ ) While sorting customers by far from or close to the depot (see Section 5.3), the distance to the closest depot is employed. (2) (5+ ) When
a new vehicle must be used to include customer c, the closest depot to c that has an unused vehicle is selected.
Backhauls

(1) (5− ) If a tour contains only backhaul customers after a string is removed, all remaining customers are removed. (2) (SA) A solution
associated with a higher number of tours that exclusively contain backhaul customers is never accepted.
Mixed/simultaneous deliveries and pickups

For each tour, location two additional values are maintained, which correspond to the maximum capacity consumption in all previous/
following locations. This preprocessing requires O(|t|) time and enables a capacity check in O(1) time (5+ ).
Open

A dummy location to which all other locations have a distance equal to zero is introduced. This location represents the end depot in each tour.
Cumulative

(5+ ) The delta evaluation method, which is employed to select the best insertion position, is modified to represent the cumulative objective.
Service time

(5+ ) An additional ordering in terms of decreasing service time is employed when each customer has an individual service time.
Time windows

(1) Feasibility checks concerning time windows are based on the “push forward” global variables (Savelsbergh 1990), which require O(|t|) time
for preprocessing. These variables enable time window feasibility checks in O(1) time. (2) (5+ ) In addition to the sorting orders introduced in
Section 5.3 (random, demand, far, and close), customers are sorted with respect to increasing time window length, increasing time window
start, and decreasing time window end.
Maximum ride times

The procedure for deriving the global variables concerning time windows is adapted for accommodating maximum ride times,a increasing its
complexity to O(|t|Q). In addition, this procedure is employed for deriving the maximum time span between one location and all successive
location. Because it takes |t| executions for deriving the maximum time span between all pairs, preprocessing of a tour requires O(|t|2 Q) time,
which again enables feasibility checks in O(1). A similar approach was reported by Gschwind (2019).
Pickup and deliveries

(1) (5− ) Strings are removed as described in Section 5.2. In addition, the associated pickup/delivery locations that were not included by the
string are also removed. (2) (5+ ) Greedy insertion with blinks (Algorithm 3) is modified to iterate over feasible combinations for pickup and
delivery insertion positions.
Fleet minimization

(1) Minimization of fleet size is performed by employing the absences-based acceptance criterion during the fleet minimization phase, which is
executed during the first 10% of iterations. Algorithm 4 provides details concerning this procedure. (2) A solution with an increased number of
tours is never accepted in the distance minimization phase.
a
This corresponds to the procedure by Tang et al. (2010), who assume a worst-case complexity of O(|t|2 ). Nevertheless, if |t|/2 > Q, a worst-case
complexity of O(|t|Q) may be assumed because a service’s timing is adjusted at most Q times.

when the number of vehicles has been reduced. In addition to these components, this paper also
Meanwhile, solutions of equal quality are colored contributes a fleet minimization approach to VRP
light gray. variants focusing primarily on minimizing their num-
ber of vehicles employed.
9. Conclusions In many respects, Laporte (2009, p. 415) provided
This paper introduced SISRs, a heuristic procedure ca- much of the motivation behind this paper’s original
pable of achieving highly competitive and new best contribution when writing of how “[t]here is . . . a
results within short computation times for both the sense that several of the most successful metaheuristics
CVRP and a wide range of associated problems such are over-engineered and one should now attempt to
as the VRP with time windows, the pickup and de- produce simple and flexible algorithms capable of
livery problem, and the dial-a-ride problem. SISRs handling a larger variety of constraints, even if this were
is composed of a ruin method that generates spatial to translate into a small loss in accuracy.” Whereas SISRs
slack, followed by a greedy recreate heuristic that exhibits the simplicity and flexibility Laporte (2009)
leaves sufficient room for solution space exploration. recommends, it also serves as a counterargument that
 Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
16 Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS

this necessitates a loss in accuracy. As demonstrated Fukasawa R, Longo H, Lysgaard J, Poggi M, Reis M, Uchoa E,
by this paper, no such trade-off occurs. Werneck RF (2006) Robust branch-and-cut-and-price for the
capacitated vehicle routing problem. Math. Programming 106(3):
Furthermore, performance with respect to large-
491–511.
scale academic instances was particularly impres- Gajpal Y, Abad P (2009) An ant colony system (ACS) for vehicle
sive and exceeded expectations. Therefore, there is routing problem with simultaneous delivery and pickup. Com-
also room for further research in terms of even larger put. Oper. Res. 36(12):3215–3223.
instances that are more representative of real-world Gehring H, Homberger J (1999) A parallel hybrid evolutionary met-
conditions, the results for which should be broadly aheuristic for the vehicle routing problem with time windows.
Miettinen K, Mäkelä M, Toivanen J, eds. Proc. EUROGEN99,
similar to those obtained for the largest instances
Jyväskylä, Finland, 57–64.
addressed within the present study. Finally, all the Gehring H, Homberger J (2001) A parallel two-phase metaheuristic
newly introduced heuristic components demonstrate for routing problems with time windows. Asia-Pacific J. Oper.
an impressive convergence speed which engenders Res. 18(1):35–47.
an array of future real-world applications, providing Gillett BE, Johnson JG (1976) Multi-terminal vehicle-dispatch algo-
a multitude of exciting avenues for future research. rithm. Omega 4(6):711–718.
Goetschalckx M, Jacobs-Blecha C (1989) The vehicle routing problem
with backhauls. Eur. J. Oper. Res. 42(1):39–51.
Acknowledgments
Golden BL, Wasil EA, Kelly JP, Chao IM (1998) The Impact of Meta-
The authors acknowledge the statistical advice provided by
heuristics on Solving the Vehicle Routing Problem: Algorithms, Problem
Wim Vancroonenburg (KU Leuven) and the editorial con- Sets, and Computational Results (Springer, Boston), 33–56.
sultation provided by Luke Connolly (KU Leuven). Gschwind T (2019) Route feasibility testing and forward time slack
for the synchronized pickup and delivery problem. OR Spectrum
References 41(2):491–512.
Augerat P, Belenguer JM, Benavent E, Corberan A, Naddef D, Gschwind T, Drexl M (2019) Adaptive large neighborhood search
Rinaldi G (1995) Computational results with a branch-and-cut with a constant-time feasibility test for the dial-a-ride problem.
code for the capacitated vehicle routing problem. Working pa- Transportation Sci. 53(2):480–491.
per, Université Joseph Fourier, Saint-Martin-d’Hères, France. Jaw JJ, Odoni AR, Psaraftis HN, Wilson NHM (1986) A heuristic
Baldacci R, Christofides N, Mingozzi A (2008) An exact algorithm for algorithm for the multi-vehicle advance request dial-a-ride prob-
the vehicle routing problem based on the set partitioning formu- lem with time windows. Transportation Res. Part B: Methodological
lation with additional cuts. Math. Programming 115(2):351–385. 20(3):243–257.
Bresina JL (1996) Heuristic-biased stochastic sampling. Proc. AAAI Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simu-
Conf. Artificial Intelligence, Portland OR, 271–278. lated annealing. Science 220(4598):671–680.
Campbell AM, Vandenbussche D, Hermann W (2008) Routing for Laporte G (2009) Fifty years of vehicle routing. Transportation Sci.
relief efforts. Transportation Sci. 42(2):127–145. 43(4):408–416.
Chassaing M, Duhamel C, Lacomme P (2016) An ELS-based ap- Laporte G, Musmanno R, Vocaturo F (2010) An adaptive large
proach with dynamic probabilities management in local search neighbourhood search heuristic for the capacitated arc-routing
for the dial-a-ride problem. Engrg. Appl. Artificial Intelligence problem with stochastic demands. Transportation Sci. 44(1):
48(February):119–133. 125–135.
Christofides N, Eilon S (1969) An algorithm for the vehicle-
Li F, Golden B, Wasil E (2005) Very large-scale vehicle routing: new
dispatching problem. J. Oper. Res. Soc. 20(3):309–318.
test problems, algorithms, and results. Comput. Oper. Res. 32(5):
Christofides N, Mingozzi A, Toth P (1979) The vehicle routing
1165–1179.
problem. Christofides N, ed. Combinatorial Optimization (Wiley,
Li H, Lim A (2001) A metaheuristic for the pickup and delivery
Chichester), 315–338.
problem with time windows. Proc. 13th IEEE Internat. Conf. Tools
Cordeau JF, Laporte G (2003) A tabu search heuristic for the static
Artificial Intelligence (Institute of Electrical and Electronics En-
multi-vehicle dial-a-ride problem. Transportation Res. Part B:
gineers, Piscataway, NJ), 160–167.
Methodological 37(6):579–594.
Lin S (1965) Computer solutions of the traveling salesman problem.
Croes GA (1958) A method for solving traveling-salesman problems.
Bell System Tech. J. 44(10):2245–2269.
Oper. Res. 6(6):791–812.
Curtois T, Landa-Silva D, Qu Y, Laesanklang W (2018) Large Masmoudi MA, Braekers K, Masmoudi M, Dammak A (2017) A
neighbourhood search with adaptive guided ejection search hybrid genetic algorithm for the heterogeneous dial-a-ride prob-
for the pickup and delivery problem with time windows. EURO lem. Comput. Oper. Res. 81(May):1–13.
J. Transportation Logist. 7(2):151–192. Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E
Dantzig GB, Ramser JH (1959) The truck dispatching prob- (1953) Equation of state calculations by fast computing machines.
lem. Management Sci. 6(1):80–91. J. Chemical Phys. 21(6):1087–1092.
Dees WA, Smith RJ (1981) Performance of interconnection rip-up and Min H (1989) The multiple vehicle routing problem with simulta-
reroute strategies. 18th Design Automation Conf. (IEEE, Piscat- neous delivery and pick-up points. Transportation Res. Part A:
away, NJ), 382–390. General 23(5):377–386.
Dumas Y, Desrosiers J, Soumis F (1991) The pickup and delivery Nagata Y, Kobayashi S (2010) A memetic algorithm for the pickup
problem with time windows. Eur. J. Oper. Res. 54(1):7–22. and delivery problem with time windows using selective route
Dunn OJ (1961) Multiple comparisons among means. J. Amer. Statist. exchange crossover. Schaefer R, Cotta C, Kołodziej J, Rudolph G,
Assoc. 56(293):52–64. eds. Parallel Problem Solving from Nature, PPSN XI, Lecture Notes
Fahrion R, Wrede M (1990) On a principle of chain-exchange for in Computer Science, vol. 6238 (Springer, Berlin), 536–545.
vehicle-routeing problems (1-vrp). J. Oper. Res. Soc. 41(9):821–827. Nagata Y, Bräysy O, Dullaert W (2010) A penalty-based edge as-
Fisher ML (1994) Optimal solution of vehicle routing problems using sembly memetic algorithm for the vehicle routing problem with
minimum k-trees. Oper. Res. 42(4):626–642. time windows. Comput. Oper. Res. 37(4):724–737.
Christiaens and Vanden Berghe: SISRs for Vehicle Routing Problems
Transportation Science, Articles in Advance, pp. 1–17, © 2020 INFORMS 17

Ngueveu SU, Prins C, Wolfler Calvo R (2010) An effective memetic Schrimpf G, Schneider J, Stamm-Wilbrandt H, Dueck G (2000) Record
algorithm for the cumulative capacitated vehicle routing prob- breaking optimization results using the ruin and recreate prin-
lem. Comput. Oper. Res. 37(11):1877–1885. ciple. J. Comput. Phys. 159(2):139–171.
Or I (1976) Traveling salesman-type combinatorial problems and Shaw P (1998) Using constraint programming and local search
their relation to the logistics of regional blood banking. Thesis, methods to solve vehicle routing problems. Proc. 4th Internat.
Xerox University Microfilms, Ann Arbor, MI. Conf. Principles Practice Constraint Programming (Springer-Verlag,
Pecin D, Pessoa A, Poggi M, Uchoa E (2014) Improved branch-cut- London), 417–431.
and-price for capacitated vehicle routing. Lee J, Vygen J, eds. Subramanian A, Uchoa E, Ochi LS (2013) A hybrid algorithm for a
Integer Programming and Combinatorial Optimization, Lecture Notes class of vehicle routing problems. Comput. Oper. Res. 40(10):
in Computer Science, vol. 8494 (Springer International Publish- 2519–2531.
ing, Cham, Switzerland), 393–403. Taillard E, Badeau P, Gendreau M, Guertin F, Potvin JY (1997) A tabu
Pisinger D, Røpke S (2007) A general heuristic for vehicle routing search heuristic for the vehicle routing problem with soft time
problems. Comput. Oper. Res. 34(8):2403–2435. windows. Transportation Sci. 31(2):170–186.
Potvin JY, Rousseau JM (1993) A parallel route building algorithm for Tang J, Kong Y, Lau H, Ip AW (2010) A note on efficient feasibility
the vehicle routing and scheduling problem with time windows. testing for dial-a-ride problems. Oper. Res. Lett. 38(5):405–407.
Eur. J. Oper. Res. 66(3):331–340. Toth P, Vigo D (2002) Models, relaxations and exact approaches for
Potvin JY, Rousseau JM (1995) An exchange heuristic for routeing the capacitated vehicle routing problem. Discrete Appl. Math.
problems with time windows. J. Oper. Res. Soc. 46(12):1433–1446. 123(1):487–512.
Prins C (2004) A simple and effective evolutionary algorithm for the Uchoa E, Pecin D, Pessoa A, Poggi M, Vidal T, Subramanian A (2017)
vehicle routing problem. Comput. Oper. Res. 31(12):1985–2002. New benchmark instances for the capacitated vehicle routing
R Core Team (2017) R: A language and environment for statistical problem. Eur. J. Oper. Res. 257(3):845–858.
computing. R Foundation for Statistical Computing, Vienna. Vidal T, Crainic TG, Gendreau M, Prins C (2013) A hybrid genetic
Repoussis PP, Tarantilis CD, Ioannou G (2009) Arc-guided evolu- algorithm with adaptive diversity management for a large class
tionary algorithm for the vehicle routing problem with time of vehicle routing problems with time-windows. Comput. Oper.
windows. IEEE Trans. Evolutionary Comput. 13(3):624–647. Res. 40(1):475–489.
Ribeiro GM, Laporte G (2012) An adaptive large neighborhood Vidal T, Crainic TG, Gendreau M, Prins C (2014) A unified solution
search heuristic for the cumulative capacitated vehicle routing framework for multi-attribute vehicle routing problems. Eur.
problem. Comput. Oper. Res. 39(3):728–735. J. Oper. Res. 234(3):658–673.
Rochat Y, Taillard ÉD (1995) Probabilistic diversification and intensifi- Vidal T, Crainic TG, Gendreau M, Prins C (2015) Time-window
cation in local search for vehicle routing. J. Heuristics 1(1):147–167. relaxations in vehicle routing heuristics. J. Heuristics 21(3):
Røpke S, Pisinger D (2006a) An adaptive large neighborhood search 329–358.
heuristic for the pickup and delivery problem with time win- Vidal T, Crainic TG, Gendreau M, Lahrichi N, Rei W (2012) A hybrid
dows. Transportation Sci. 40(4):455–472. genetic algorithm for multidepot and periodic vehicle routing
Røpke S, Pisinger D (2006b) A unified heuristic for a large class of problems. Oper. Res. 60(3):611–624.
vehicle routing problems with backhauls. Eur. J. Oper. Res. Wilcoxon F (1945) Individual comparisons by ranking methods.
171(3):750–775. Biometrics Bull. 1(6):80–83.
Sariklis D, Powell S (2000) A heuristic method for the open vehicle Wilson NHM, Sussman JM, Wang HK, Higonnet BT (1971) Sched-
routing problem. J. Oper. Res. Soc. 51(5):564–573. uling algorithms for a dial-a-ride system. Technical Report USL
Savelsbergh M (1990) An efficient implementation of local search TR-70-13, Massachusetts Institute of Technology, Cambridge,
algorithms for constrained routing problems. Eur. J. Oper. Res. MA.
47(1):75–85. Zachariadis EE, Kiranoudis CT (2012) An effective local search ap-
Savelsbergh MWP (1988) Computer aided routing. Doctoral disser- proach for the vehicle routing problem with backhauls. Expert
tation, Centrum voor Wiskunde en Informatica, Amsterdam. Systems Appl. 39(3):3174–3184.