A general heuristic for vehicle routing problems

David Pisinger ∗
Stefan Ropke ∗
25th February 2005

Abstract
We present a unified heuristic, which is able to solve five different variants of the vehicle routing
problem: the vehicle routing problem with time windows (VRPTW), the capacitated vehicle routing
problem (CVRP), the multi-depot vehicle routing problem (MDVRP), the site dependent vehicle routing
problem (SDVRP) and the open vehicle routing problem (OVRP).
All problem variants are transformed to a rich pickup and delivery model and solved using the Adap-
tive Large Neighborhood Search (ALNS) framework presented in Ropke and Pisinger (2004). The ALNS
framework is an extension of the Large Neighborhood Search framework by Shaw (1998) with an adap-
tive layer. This layer adaptively chooses among a number of insertion and removal heuristics, to intensify
and diversify the search. The presented approach has a number of advantages: ALNS provides solutions
of very high quality, the algorithm is robust, and to some extent self-calibrating. Moreover, the uni-
fied model allows the dispatcher to mix various variants of VRP problems for individual customers or
vehicles.
As we believe that the ALNS framework can be applied to a large number of tightly constrained
optimization problems, a general description of the framework is given, and it is discussed how the
various components can be designed in a particular setting.
The paper is concluded with a computational study, in which the five different variants of the vehicle
routing problem are considered on standard benchmark tests from the literature. The outcome of the
tests is promising as the algorithm is able to improve 183 best known solutions out of 486 benchmark
tests. The heuristic has also shown promising results for a large class of vehicle routing problems with
backhauls, as demonstrated in Ropke and Pisinger (2005).

Keywords: metaheuristics, large neighborhood search, vehicle routing problem

1 Introduction
Most scientific papers in the area of heuristic solution methods for vehicle routing problems target a specific
vehicle routing problem, for example vehicle routing problems with time windows (VRPTW). In such
papers a heuristic is designed, implemented and fine-tuned to fit this particular problem type. Only a few
papers (see e.g. Cordeau et al. [17, 19]) consider heuristics that “out-of-the-box” can be used to solve
several problem types. We believe that general vehicle routing heuristics are an important research area
as such heuristics are needed for real life problems, where the transportation needs of different companies
often are different and thus call for various types of vehicle routing problems.
The heuristic in this paper is applied to five different problems: the vehicle routing problem with
time windows (VRPTW), the capacitated vehicle routing problem (CVRP), the multi-depot vehicle routing
problem (MDVRP), the site dependent vehicle routing problem (SDVRP) and the open vehicle routing
problem (OVRP). In the CVRP one has to deliver goods to a set of customers with known demands on
minimum-cost vehicle routes originating and terminating at a depot. The vehicles are assumed to be homo-
geneous and having a certain capacity. In some versions of the CVRP one also has to obey a route duration
∗ DIKU - Department of Computer Science, University of Copenhagen, Universitetsparken 1, DK-2100 Copenhagen Ø, Denmark.

E-mail: {pisinger, sropke}@diku.dk

1
constraint that limits the lengths of the feasible routes. The VRPTW extends the CVRP by associating
time windows with the customers. The time window defines an interval where in the customer should be
supplied. The OVRP is closely related to the CVRP, but contrary to the CVRP a route ends as soon as the
last customer has been served as the vehicles do not need to return to the depot. The MDVRP extends the
CVRP by allowing multiple depots. The SDVRP is another generalization of the CVRP. In the SDVRP
one can specify that certain customers only can be served by a subset of the vehicles, furthermore, vehicles
can have different capacities in the SDVRP. In the CVRP, MDVRP and SDVRP one seeks to minimize the
total traveled distance whereas in the OVRP and VRPTW, the first priority is to minimize the number of
vehicles and minimizing the traveled distance is the second priority. It should be mentioned that most exact
methods and some metaheuristics for the VRPTW minimize total traveled distance instead of minimizing
number of vehicles used.
All problem types are transformed to a Rich Pickup and Delivery problem with time windows (RPDPTW)
and are solved using the Adaptive Large Neighborhood Search (ALNS) framework presented in [49].
In the RPDPTW we have a number of requests to be carried out by a fixed set of vehicles. Each request
consists of picking up a quantity of goods at one location and delivering it to another location. The objective
of the problem is to find a feasible set of routes for the vehicles so that all requests are serviced, and such
that the overall travel distance is minimized. A feasible route of a vehicle should start at a given location,
service a number of requests such that the capacity of the vehicle is not exceeded, and finally end at a given
location. A pickup or delivery should take place within a given time window. Each request furthermore
has an associated pickup precedence number, and a delivery precedence number. A vehicle must visit the
locations in nondecreasing order of precedences (see e.g. Sigurd et al. [54] for various applications of
precedence constraints). Since not all vehicles may be able to service all requests (e.g. due to their physical
size or the absence of some cooling compartments) we need to ensure that every request is serviced by
a given subset of vehicles. Between any two locations we have an associated, nonnegative distance and
travel time. It is assumed that travel times satisfy the triangle inequality. This assumption implies that
any removal of requests from a feasible route will keep the route feasible with respect to the imposed time
windows.
The five vehicle routing problems considered in the present paper have all been intensively studied in
the literature. The two best known problems are the VRPTW and the CVRP. The VRPTW has been the
target of extensive research and almost any type of metaheuristic has been applied to the problem. For
recent surveys on the state of art in VRPTW research we recommend the survey by Cordeau et al [15] that
describes both exact and heuristic methods, and the survey by Bräysy and Gendreau [8] that focuses on
metaheuristics. It is hard to single out a few VRPTW metaheuristics as the amount of proposed heuristics
are enormous, and no heuristic dominates all the other heuristics in all aspects. We would, however, like to
mention the metaheuristic by Mester and Bräysy [42] as it has shown outstanding results on larger VRPTW
instances with between 200 and 1000 customers. For the smaller VRPTW instances like the Solomon data
set, some of the best heuristics in terms of solution quality achieved are the Large Neighborhood Search by
Bent and Van Hentenryck [2] and the Hybrid Genetic Algorithm by Homberger and Gehring [32].
Solving the VRPTW to optimality has also received much attention. The current state of the art exact
methods are proposed by Kallehauge et al [35], Irnich and Villeneuve [34] and Chabrier [10], and all follow
the branch-and-price framework. The two first mentioned approaches furthermore strengthen the obtained
lower bound by adding valid inequalities to the master problem. The size of the instances that consistently
can be solved to optimality is rather limited as unsolved instances with 50 customers exist, but some large
scale instances can be solved. For example, Kallehauge et al. [35] report that a 1000 customer instance has
been solved. Solving problems of this size is only possible by current techniques if the instance has a certain
structure and the time constraints are very tight. These observations justify the research into heuristics for
the VRPTW as industrial routing problems demand robust algorithms for large-sized instances.
The CVRP literature is also vast, classical heuristics for the problem have been surveyed by Laporte
and Semet [38], and metaheuristics have been surveyed by Gendreau et al. [29] and more recently by
Cordeau et al. [16]. CVRP heuristics have typically been tested on 14 instances containing between 50 and
199 customers. In the early ’90s very good metaheuristics for the CVRP were developed like the parallel
tabu search by Taillard [23]. Most of the solutions to the 14 classical instances found back then have still
not been improved. More recently, some larger instances have been introduced containing between 240
and 1200 customers by Golden et al. [30] and Li et al. [40]. These new instances seem to have spurred a

2
new interest into metaheuristics for the CVRP as indicated in the survey by Cordeau et al. [16].
Until recently, exact methods for the CVRP were dominated by branch-and-cut methods. One of the
best branch-and-cut algorithms for the CVRP was developed by Lysgaard et al. [41]. Recent research re-
sults indicate that branch-and-cut-and-price algorithms are a more promising approach as shown by Fuka-
sawa et al. [26]. For the CVRP, the largest problem that has been solved to optimality contains 135
customers.
The OVRP is a variant of the CVRP that has received less attention. The problem appears in various
distribution problems, where the vehicle simply stops after the last delivery. The problem was introduced
by Sariklis and Powell [51] where a two-phase cluster first-route second heuristic was proposed. Recently
tabu search heuristics have been proposed by Fu et al. [25] and Brandão [6].
Tabu search heuristics for the MDVRP have been proposed by Renaud et al. [48] and Cordeau et al.
[17]. The last paper deserves special attention as it describes a general heuristic that also solve periodic
vehicle routing problems (PVRP) and periodic traveling salesman problems. Earlier, Chao et al. [12]
proposed a record-to-record improvement heuristic for the MDVRP.
The SDVRP was first studied by Nag et al. [43] who developed several simple heuristics for the
problem. Chao et al. [11] developed a more advanced heuristic and constructed several new test instances.
Cordeau and Laporte [18] showed that the problem could be seen as a special case of the PVRP and they
presented computational results obtained by solving the problem using their PVRP tabu search heuristic.

The main contribution of this paper is to describe a general ALNS heuristic that is able to solve all the
above variants of the VRP problem. The computational results are promising as we, for the large scale
VRPTW instances suggested by Gehring and Homberger [27], on average are able to decrease the number
of vehicles used, and the method becomes more attractive compared to other heuristics as the problem
size increases. For the OVRP, MDVRP and SDVRP we are able to improve a large number of best known
solutions. The results for the CVRP problems are fair as the ALNS heuristic is comparable to most recently
proposed heuristics but it is surpassed by the very best heuristic for the problem type.
Due to the promising results of ALNS, we give a general description of the paradigm, to make it easier
to adapt the framework to other problem types. Various strategies for designing construction and removal
heuristics are discussed, and the overall framework is presented in a general form.
In the following Section 2 we give a formal mathematical definition of the RPDPTW and in Section 3
we describe how the considered variants of the problem are transformed to the RPDPTW. In Section 4 we
give a general presentation of the ALNS algorithm forming the core of our solution approach. Section 5
describes how the general framework has been adapted to solve the RPDPTW. Section 6 presents a number
of computational experiments which document that the proposed heuristic performs not worse than state-
of-art heuristics specialized to solve the considered variants of the problem. The paper is concluded in
Section 7.

2 Formal problem definition

We now present a mathematical formulation of the RPDPTW problem. The mathematical model is used
to describe the heuristic in details in later sections and to describe how the considered VRP variants are
transformed to the RPDPTW.
Following the terminology of Desaulniers et al. [21], a problem instance of the pickup and delivery
problem contains n requests and m vehicles. The problem is defined on a graph where P = {1, . . . , n} is
the set of pickup nodes, and D = {n + 1, . . . , 2n} is the set of delivery nodes. Request i is represented
by node i and i + n. K = {1, . . . , m} is the set of all vehicles; one vehicle might not be able to serve all
requests. These kinds of limitations are modeled by letting Pk ⊆ P and Dk ⊆ D be the set of pickups
and deliveries that can be served by vehicle k. Since every request is serviced by the same vehicle we may
assume that i ∈ Pk ⇔ i + n ∈ Dk , i.e. that both the pickup and delivery can be serviced by vehicle k.
Define N = P ∪ D and Nk = Pk ∪ Dk . Let τk = 2n + k, k ∈ K and τk0 = 2n + m + k, k ∈ K be
the nodes that represent the start and end terminals of vehicle k. The directed graph G = (V, A) consists
of the nodes V = N ∪ {τ1 , . . . , τm } ∪ {τ10 , . . . , τm
0
} and the arcs A = V × V . For each vehicle we have
a subgraph Gk = (Vk , Ak ), where Vk = Nk ∪ {τk } ∪ {τk0 } and Ak = Vk × Vk . For each edge (i, j) ∈ A

3
we assign a distance dij ≥ 0 and a travel time tij ≥ 0. It is assumed that the travel times satisfy the
triangle inequality i.e. tij ≤ til + tlj for all i, j, l ∈ V . We assign a service time si and a time window
[ai , bi ] to each node i∈ V . The service time represents the time needed for loading and unloading and the
time window indicates when the visit at the particular site should start; a visit to node i can only take place
between time ai and bi . A vehicle is allowed to arrive to a site before the start of the time window but it
has to wait until the start of the time window before the visit can be performed. For each node i ∈ N we
define li to be the amount of goods that should be loaded onto the vehicle at the particular node. We have
that li ≥ 0 for i ∈ P and li = −li−n for i ∈ D. Each vehicle k ∈ K has room for a certain amount of
goods, this capacity is given by Ck . Each node has assigned a precedence or priority Πi . Nodes with low
precedence must always be visited before nodes with higher precedence.
Each vehicle k should follow a legal route from its start terminal τk to its destination terminal τk0 . A
legal route r is a simple (loop-free) path

r = (τk = v1 , v2 , . . . , vh = τk0 ) (1)

satisfying the precedences and time windows at the customers, the capacity of the vehicle, and ensuring
that a pickup takes place before a delivery, and that only requests serviceable by vehicle k are carried out.
More formally, we demand that the vehicle only visits nodes that can be serviced by the vehicle, i.e.

vi ∈ Nk , i = 2, . . . , h − 1 (2)

A pickup-delivery pair should be served by the same vehicle, and the pickup should take place before the
delivery, hence we have

i ≤ j, vi ∈ Pk , vj ∈ Dk , vj = vi + n (3)

Precedences should be obeyed along the route, this is ensured by the constraints

i ≤ j, Πvi ≤ Πvj (4)

To ensure that time windows are satisfied, we introduce Si ∈ R+

0 to denote when the vehicle starts the
service at site vi . We then have the constraints
avi ≤ Si ≤ bvi i = 1, . . . , h (5)
Si+1 ≥ Si + si + tvi ,vi+1 i = 1, . . . , h − 1 (6)
aτ k ≤ S 1 ≤ b τ k (7)
aτk0 ≤ Sh ≤ bτk0 (8)

where [aτk , bτk ] is the time window of terminal τk and [aτk0 , bτk0 ] is the time window of terminal τk0 . Finally,
the capacity of the vehicle should be respected throughout the path. For this purpose we introduce Li ∈ R+ 0
to denote the load of the vehicle at node i after serving node i. Then we have

Li ≤ Ck i = 1, . . . , h (9)
Li+1 = Li + li+1 i = 1, . . . , h − 1 (10)
L1 = 0 (11)
Lh = 0 (12)

The travel cost of a given route r is

h−1
X
cr = dvi ,vi+1 (13)
i=1

Situations may occur where not all requests can be serviced by the available vehicles. To model this
situation we create n dummy routes, consisting of a single request. These routes do not make use of any
vehicles but they have a large cost, denoted Γ. Requests that are not served by a vehicle are said to be
located in the request bank.

4
 The whole problem can now be formulated as follows: let R be the set of all feasible routes. The
boolean matrix (αjr ) for r ∈ R and j = 1, . . . , n is used to indicate whether request j is serviced using
route r. The boolean matrix (βkr ) for r ∈ R and k = 1, . . . , m is used to indicate whether the route r is
carried out by vehicle k. Using binary variables xr to indicate whether route r is used in the solution we
get the following model
X
min f (x) = cr xr (14)
X r∈R
s.t. αjr xr = 1 j = 1, . . . , n (15)
r∈R
X
βkr xr = 1 k = 1, . . . , m (16)
r∈R
xr ∈ {0, 1} r∈R (17)

Note that a dummy route is not assigned to any vehicle, that is for any dummy route r we have that
βkr = 0, ∀k = 1, . . . , m

3 Problem transformations
The heuristic in this paper is applied to five different problems — VRPTW, CVRP, OVRP, MDVRP, SD-
VRP — which all are transformed to a RPDPTW. The conversions which will be described in the following
paragraphs are extensions of the transformations presented by Ropke and Pisinger [50] for solving VRP
problems with Backhauls.

3.1 Vehicle Routing Problem with Time Windows

In order to transform a VRPTW instance to a RPDPTW instance we map every customer in the VRPTW to
a request in the RPDPTW. Such a request consists of a pick up at the depot and delivery at the customer site.
The amount of goods that should be carried by the requests is equal to the demand of the corresponding
customer. The time window of the pickup is set to [ad , ad ] where ad is the start of the time window of the
depot in the VRPTW and its service time is set to zero. The time window and service time of the delivery
are copied from the corresponding customer in the VRPTW. In order to avoid routes that return to the depot
for restocking we let all pickups and deliveries have precedence zero and one respectively. All vehicles in
the RPDPTW have the same start and end terminals corresponding to the depot in the VRPTW. Distances
and travel times in the RPDPTW are set in the natural way.

3.2 Capacitated Vehicle Routing Problem

A CVRP instance can easily be transformed to a VRPTW instance. This can for example be done by setting
all travel and service times to zero and all time windows to [0,0]. If the CVRP contains a route duration
constraint then travel times and durations should be set as in the CVRP. All time windows (including the
ones at the end terminals) should be set to [0, D] where D is the route duration. The VRPTW is transformed
to a RPDPTW as described in Section 3.1.

3.3 Site Dependent Vehicle Routing Problem

In the SDVRP a customer may only be serviced by a given subset of the vehicles, typically because the
access paths to the node do not allow given vehicles to pass, or because specific facilities are demanded in
the vehicle (e.g. a freezing compartment).
The SDVRP is easily modeled as a RPDPTW by using the transformation from CVRP to RPDPTW
and noting that the RPDPTW allows us to specify the pickups Pk and deliveries Dk that can be carried out
by vehicle k.

5
3.4 Open Vehicle Routing Problem
The OVRP is very close to the CVRP. The difference between the two problems is that in the OVRP the
vehicles do not have to return to the depot. Thus an OVRP can be solved as an asymmetric CVRP by
setting distances and travel times from every customer to the depot to zero.
The travel times in the resulting RPDPTW do not satisfy the triangle inequality, but our method is
able to handle the problems anyway since tij ≤ til + tlj only is violated when l is an end terminal. Our
only reason for assuming that the triangle inequality is satisfied for the travel times is that we have to
avoid situations where the removal of one or more requests causes the travel time to increase. As the node
sequence i → l → j where l ∈ {τ10 , . . . , τm
0
} never occurs in a valid route this violation of the triangle
inequality does not cause problems.

3.5 Multi-Depot Vehicle Routing Problem

Even though the underlying RPDPTW model supports multiple depots, we cannot use these depots to
model the MDVRP. The problem is that we must assign each pickup to a depot and we do not know which
depot is going to serve a given request.
Instead we create a dummy base location where all routes start and end and where all ordinary requests
are picked up. We also create a dummy request for each vehicle k in the problem. The pickup and delivery
locations of these requests are located at the depot of the corresponding vehicle. A dummy request has
demand zero, it does not have any service time and it can be served at any time. The set Nk of each vehicle
k contains all ordinary requests and the dummy request corresponding to the vehicle. In this way we ensure
that each vehicle will carry precisely one dummy request.
The precedences Πi of a pickup and delivery corresponding to an ordinary request are set to zero and
two respectively, the precedence of the pickup and delivery of the dummy requests are set to one and three
respectively. This ensures that all ordinary deliveries will be surrounded by the pickup and delivery of
a dummy request. The distance and travel time between a pickup of an ordinary request and any other
location is set to zero. All other distances and travel times are set as defined by the original MDVRP.
In a solution to the RPDPTW that serves all requests we know that each route will start at a start
terminal, located at the dummy base location and perform a number of pickups, then go to the pickup of
the dummy request. Next, the ordinary deliveries will be served and the vehicle will return to the delivery
of the dummy request and then to the end terminal of the route. Before starting the pickup of the dummy
request and after the delivery of it all travel times and distances will be zero. Furthermore travel times and
distances are accumulated correctly while carrying the dummy request.
While solving MDVRP problems the cost of dummy routes Γ must be set to a sufficiently high number
such that it never will be profitable to leave a dummy request in the request bank.

VNS ALNS
Nk N5 N4

N3 N3
N2
N1
N2

N1
N∗

Figure 1: Illustration of neighborhoods used by VNS and ALNS. VNS typically operates on one
type of neighborhood with variable depth while ALNS operates on structurally different neighborhoods
N1 , . . . , Nk defined by the corresponding search heuristics. All neighborhoods N1 , . . . , Nk in ALNS are
a subset of the neighborhood N ∗ defined by modifying q variables.

6
4 Adaptive Large Neighborhood Search
We will now describe the Adaptive Large Neighborhood Search (ALNS) framework used in the present
paper. We believe that ALNS can be applied to a large class of difficult optimization problems, hence in
the following we consider an optimization problem in the general IP form:

min{f (x) : Ax ≤ b, x ∈ Zn } (18)

ALNS is a local search framework in which a number of simple algorithms compete to modify the current
solution. In each iteration an algorithm is chosen to destroy the current solution, and an algorithm is chosen
to repair the solution. The new solution is accepted if it satisfies some criteria defined by the local search
framework applied at the master level.
To be more formal, we extend the domain of each variable xi to Z ∪ {⊥}, where ⊥ means undefined.
A destroy heuristic chooses at most q variables which are assigned the value ⊥. A repair heuristic assigns
feasible values xi ∈ Z to the q variables.
The ALNS framework is an extension of the Large Neighborhood Search presented by Shaw [53],
where a large collection of variables are modified in each iteration. In ALNS the neighborhoods are
searched by simple and fast heuristics. ALNS is also based on the Ruin and Recreate paradigm presented
by Schrimpf et al. [52], or the Ripup and Reroute paradigm applied in [59]. In each iteration the current
solution is partially destroyed and then repaired using some heuristics. ALNS also has similarities with
Very Large Neighborhoods Search (VLNS) presented by Ahuja et al. [1]. In VLNS the algorithm operates
on very large neighborhoods chosen in a way so that they still can be searched efficiently.
Variable Neighborhood Search (VNS) was presented by Hansen and Mladenovic [31]. VNS makes
use of a parameterized family of neighborhoods, typically obtained by using a given neighborhood with
variable depth. When the algorithm reaches a local minimum using one of the neighborhoods, it proceeds
with a larger neighborhood from the parameterized family. When the VNS algorithm gets out of the local
minimum it proceeds with the smaller neighborhood. On the contrary, ALNS operates on a predefined
set of large neighborhoods corresponding to the destroy (removal) and repair (insertion) heuristics. The
neighborhoods are not necessarily well-defined in a formal mathematical sense — they are rather defined
by the corresponding heuristic algorithm. The difference between VNS and ALNS is graphically illustrated
in Figure 1. In the sections that follow, we will identify a neighborhood with the heuristic searching it.
Instead of viewing the ALNS heuristic as a sequence of destroy and repair operations one can alterna-
tively see it as a sequence of fix and optimize operations. The fix operation selects a number of variables
that are fixed at their current value; the optimize operation seeks to find a near-optimal solution that respects
the fixed variables, that is, only non-fixed variables can be changed. After the optimization operation, all
variables are unlocked again. The fix operation is analogous to the destroy operation and the optimize
operation is analogous to the repair operation. The fix/optimize view might be helpful when applying the
heuristic to problems where the destroy and repair operations do not seem intuitive.

4.1 Outline of algorithm

ALNS can be based on any local search framework, e.g. simulated annealing, tabu search, guided local
search. The general framework is outlined in Figure 2, where lines 2–8 form the main loop of the local
search framework at the master level. Implementing a simulated annealing algorithm is straight forward as
one solution is sampled in each iteration of the ALNS. A simple tabu search could for example be imple-
mented by randomly sampling a number of candidate solutions and choosing the best non tabu solution.
In each iteration of the main loop we choose one destroy and one repair neighborhood in line 3. An
adaptive layer stochastically controls which neighborhoods to choose according to their past performance
(score). The more a neighborhood Ni has contributed to the solution process, the larger score πi it gets,
and hence it has a larger probability of being chosen.
The adaptive layer uses roulette wheel selection for choosing a destroy and a repair neighborhood. If the
past score of a neighborhood i is denoted πi and we have m neighborhoods, then we choose neighborhood
Nj with probability
πj
Pm
i=1 πi

7
 Adaptive Large Neighborhood Search
1 Construct a feasible solution x; set x∗ := x
2 Repeat
3 Choose a destroy neighborhood N − and a repair neighborhood
N + using roulette wheel selection based on previously
obtained scores {πj }
4 Generate a new solution x0 from x using the heuristics
corresponding to the chosen destroy and repair neighborhoods
5 If x0 can be accepted then set x := x0
6 Update scores πj of N − and N +
7 If f (x) < f (x∗ ) set x∗ := x
8 Until stop criteria is met
9 Return x∗

Figure 2: Outline of the ALNS framework

Notice that the destroy and repair neighborhoods are selected independently, and hence two separate
roulette wheel selections are performed.
In most applications the neighborhoods are searched by fast heuristics, and hence we can assume that
they are equally fast. But if some heuristics are significantly slower than others, one may normalize the
score πi of a neighborhood with a measure of the time consumption ti of the corresponding heuristic. This
ensures a proper trade-off between time consumption and solution quality.
In line 4 of the ALNS-algorithm, we first destroy the current solution x using a heuristic searching the
neighborhood N − and then repair the solution using the heuristic corresponding to neighborhood N + . It
can be advantageous to use noising or randomization in the destroy and repair heuristics to obtain a proper
diversification. In traditional local search heuristics the diversification is controlled implicitly by the local
search paradigm (accept ratio, tabu list, etc.), but since we use large neighborhoods which are searched by
simple heuristics, it is not sufficient to have a diversification operator at the master level. We also need a
diversification operator at the sub-level to avoid stagnating search processes where the destroy and repair
neighborhoods keep performing the same modifications to a solution.
Finally, in line 6 we update the scores πi of the neighborhoods. A number of criteria can be used to
measure how much a neighborhood contributes to the solution process: new best solutions are obviously
given a large score, but also not previously visited solutions are given a score. Depending on the local
search framework used on the master level, one may also give specific scores to accepted solutions e.g. in
a simulated annealing framework. Since each step of the ALNS heuristic involves two neighborhoods (a
destroy and a repair neighborhood), the score obtained in a given iteration is divided equally between them.
Every M iterations of the ALNS algorithm, the scores πi are reset, and the probabilities for choosing the
neighborhoods are recalculated. Each neighborhood is assigned a minimum probability for being chosen
to ensure that statistical information about its performance can be collected. The probabilities for choosing
a neighborhood can also be a weighted sum of the score during the last M iterations, and the overall score
since the beginning of the algorithm.

4.2 Designing an ALNS algorithm

In order to design an ALNS algorithm for a given optimization problem one needs to
• Choose a number of fast construction heuristics which are able to respect a partial solution
• Choose a number of destroy heuristics supporting the chosen construction heuristics
• Choose a local search framework at the master level.
In each iteration the heuristic corresponding to a destroy neighborhood should remove a given number
q of variables. The destroy neighborhoods (N − ) should be a proper mix of neighborhoods which can

8
intensify and diversify the search. To diversify the search, one may randomly choose q decision variables,
i.e. using a random removal neighborhood. To intensify the search one may try to remove q “critical”
variables, i.e. variables having a large cost or variables spoiling the current structure of the solution (e.g.
edges crossing each other in a Euclidean traveling salesman problem). This is known as worst removal
or critical removal. Concrete examples on random removal and worst removal neighborhoods in a VRP
context are given in Sections 5.1.1–5.1.2.
One may also choose a number of related variables that are easy to interchange while maintaining
feasibility of the solution. This related removal neighborhood was introduced by Shaw [53]. More formally
we can measure the relatedness rij of two variables xi and xj by the deviation of the corresponding
coefficients in the constraint matrix A in (18). The smaller rij the more related are variables xi and
xj . How exactly rij should be defined depends on the concrete problem at hand, and one may even have
several simultaneous neighborhoods defined by various choices of the relatedness measure (rij ). In order
to choose the q most related variables, one needs to solve the NP-hard dispersion-sum problem given by
n X
X n
minimize rij xi xj
i=1 j=1
n
X (19)
subject to xj = q
j=1
xj ∈ {0, 1}, j = 1, . . . , n

A greedy heuristic for this problem running in O(n3 ) was presented in [44] together with a more time-
consuming exact algorithm. If n is large, it may be too time-consuming even to compute the whole matrix
(rij ) and one will instead choose related variables according to some heuristics. Shaw [53] presented an
algorithm running in O(qn) time by initially selecting a variable at random, and then repeatedly selecting
an already selected variable i and finding a variable j which minimizes rij and adding j to the set of chosen
variables. An alternative heuristic is based on a modified Kruskal’s algorithm for the minimum spanning
tree problem, using rij as edge weights, which stops when a connected component with q or more elements
has been constructed. The variables in this component are set to ⊥. The worst-case running time of this
algorithm is O(n2 log n) as we have n2 edges in Kruskal’s algorithm. Ropke and Pisinger [49] used a
modified version of this algorithm in the VRP for splitting requests on a route into two strongly connected
subsets. It should be noted that solving the dispersion sum problem (19) to optimality seldom would be
a good idea even if it could be done in a very short time. If rij is independent of the current solution the
destroy neighborhood obtained by solving the dispersion sum problem to optimality would always assign
⊥ to the same set of variables. Concrete examples on various related removal neighborhoods are given in
Sections 5.1.3–5.1.5.
Following the same idea as in related removal one may choose a number of variables having small co-
efficients in the resource constraints in (18), as these generally are easy to interchange and loosely speaking
can fill up unused resource constraints. We denote this strategy small removal.
Finally, one may use history based removal where the q variables are chosen according to some histor-
ical information as presented in [49]. The historical information could for example count how often setting
a given variable (or set of variables) to a specific value leads to a bad solution. One may then try to remove
variables that currently are assigned an improper value, based on the historical information. Variants of the
history based removal neighborhood are discussed in Sections 5.1.6–5.1.7.

Recreate neighborhoods (N + ) are typically based on concrete well-performing heuristics for the given
problem. These heuristics can make use of variants of the greedy paradigm, e.g. performing the locally best
choice in each step, or performing the least bad choice in each step. An alternative variant of the greedy
paradigm is to set all variables to their upper bound in (18), and repeatedly decrease the most expensive
variable until a feasible solution is obtained. The repair heuristics can also be based on approximation
algorithms or on exact algorithms that in some sense have been truncated to obtain faster solution times.
Shaw [53] and Bent and Van Hentenryck [2] proposed more expensive algorithms like searching N + based
on truncated branch-and-bound methods. Although ALNS mainly is intended to use cheap heuristics, more
expensive search methods can be used if the scores of the corresponding neighborhoods are normalized with

9
respect to the time consumption. In the context of VRP problems, recreate neighborhoods are considered
in more details in Section 5.2 discussing both simple greedy approaches and variants of regret heuristics.
Some optimization problems can naturally be split into a number of sub-problems, where each sub-
problem can be solved individually. Such problems include the Bin Packing Problem in which a number
of bins are to be filled, or the Vehicle Routing Problem in which a number of routes are to be constructed.
For such problems one should decide whether the subproblems should be solved one by one (sequential
heuristics) or all subproblems should be solved at the same time (parallel heuristics). Sequential heuristics
are easier to implement but may have the disadvantage that the last subproblem solved will be left with
variables that do not fit well together. This is in some extent avoided in parallel heuristics.

A natural extension to the ALNS framework is to have coupled neighborhoods. In principle one may,
for each destroy neighborhood Ni− , define a subset Ki ⊆ {N + } of repair neighborhoods that can be used
with Ni− . The roulette wheel selection of repair neighborhoods will then only choose a neighborhood in
Ki if Ni− was chosen.
As a special case, one may have Ki = ∅ meaning that the neighborhood Ni− takes care of both the
destroy and repair steps. One could use an ordinary local search heuristic to compete with the other destroy
and repair neighborhoods, ensuring that a thorough investigation of the solution space close to the current
solution is made from time to time.

For some problems it may be sufficient to have a number of destroy and repair heuristics that are
selected randomly with equal probability, that is without the adaptive layer. We will denote such a heuristic
a Large Multiple-Neighborhood Search (LMNS). The LMNS heuristics share the robustness of the ALNS
heuristics, while having considerably fewer parameters to calibrate.

4.3 Properties of the ALNS framework

The ALNS framework has several advantages. For most optimization problems we already know a number
of well-performing heuristics which can form the core of an ALNS algorithm. Due to the large neigh-
borhoods and diversity of the neighborhoods, the ALNS algorithm will explore large parts of the solution
space in a structured way. The resulting algorithm becomes very robust, as it is able to adapt to various
characteristics of the individual instances, and seldom is trapped in a local minima.
ALNS is particularly well suited for tightly constrained problems, where small neighborhoods are
not sufficient to escape a local minima or certain areas of the solution space. In such problems, the large
neighborhood search makes it possible to change many variables each time to reach new, feasible solutions.
The calibration of the ALNS algorithm is quite limited as the adaptive layer automatically adjusts the
influence of each neighborhood used. It is still necessary to calibrate the individual sub-heuristics used for
searching the destroy and repair neighborhoods, but one may calibrate these individually or even use the
parameters used in existing algorithms.
In the design of most local search algorithms the researcher has to choose between a number of possible
neighborhoods. In ALNS the question is not “either-or” but rather “both-and”. As a matter of fact, our
experience is that the more (reasonable) neighborhoods the ALNS heuristic makes use of, the better it
performs [49].

5 ALNS applied to the RPDPTW

We will now describe how the general ALNS framework has been adapted to the RPDPTW problem. The
“variables” in the ALNS framework correspond to requests in the RPDPTW. A destroy neighborhood N −
consists of the removal of q requests from the existing routes, and assigning them to the request bank.
We will describe the various heuristics associated with the destroy neighborhoods in Section 5.1. A repair
neighborhood N + inserts requests from the request bank into one or more legal routes. The associated
insertion heuristics are described in Section 5.2. The local search framework used at the master level
is simulated annealing to be described in Section 5.3. Section 5.4 describes the noising method used to
diversify the search of the heuristics. Finally, the scheme used for adjusting the weights in the roulette
wheel selection is described in Section 5.5.

10
5.1 Request removal
The ALNS heuristic for the RPDPTW makes use of seven different removal heuristics, each searching a
given removal neighborhood N − . The heuristics take as input a given solution x and outputs q requests
that have been removed from the routes.

5.1.1 Random removal

The simplest removal heuristic, random removal, selects q requests at random and removes them from
the solution. This obviously has the effect of diversifying the search.

5.1.2 Worst removal

The purpose of the worst removal heuristic is to choose a number of requests that are very expensive,
or which somehow spoil the structure of the current solution. In the RPDPTW it seems reasonable to try to
remove requests with high cost and insert them at another place in the solution to obtain a better solution
value.
Given a request i served by some vehicle in a solution x we define the cost of the request ∆f−i as
the difference between the value of f (x) and the cost of solution x where request i is removed completely
from the problem.
The worst removal heuristic now repeatedly chooses a new request i, having the largest cost ∆f−i
until q requests have been removed. The removal heuristic is randomized, the randomization is controlled
by the parameter p. If p is small, the most expensive request is selected, while less expensive requests may
be chosen for larger values of p with a probability that decreases with the cost ∆f−i . We refer to [49] for
additional details.

5.1.3 Related removal

The purpose of the related removal heuristic is to remove a set of requests that in some sense are
related and hence easy to interchange. For the RPDPTW we define the relatedness rij of two orders i and j
solely by the distance between the requests, as introduced by Ropke and Pisinger [49]. Since each request
i consists of a pickup node i and a delivery node i + n we get the expression
1 0
rij = (d (i, j) + d0 (i, j + n) + d0 (i + n, j) + d0 (i + n, j + n)) (20)
D
where the distance measure d0 (u, v) between two nodes in this context is defined as

0 duv if u and v are not located at a terminal
d (u, v) = (21)
0 if u or v is located at a terminal

The motivation for neglecting the distance from a terminal is that the terminal will be visited in any case,
and hence should not contribute to the relatedness measure of two requests.
The denominator D is set to the number of nonzero terms in (20), i.e. the number of pickup and
deliveries taking place at a site different from a terminal. Hence if all nodes are different from a terminal,
D := 4 while if both requests have a pickup at a terminal D := 1.
The relatedness measure is used to remove customers as described in Shaw [53]. The algorithm initially
selects a request i by random. Then it repeatedly chooses an already selected request j and selects a new
request which is most related to j. The algorithm stops when q requests have been chosen. Like in the
worst removal heuristic (Section 5.1.2) the process is controlled by a randomization parameter p. If
p is zero, the most related request is always chosen in the inner loop. If p > 0 a less related request may
be chosen, where the probability of choosing a request decreases with the relatedness measure rij and
increases with p. The algorithm is described in more details in [49].

11
5.1.4 Cluster removal
The cluster removal heuristic is a variant of the related removal heuristic in which we try
to remove clusters of related requests from a few routes. As a motivation, consider a route where the
requests are grouped into two geographical clusters. When removing requests from such a route it is often
important to remove one of these clusters entirely as the insertion methods otherwise would be prone to
insert the removed requests back into the route. The related removal heuristic from Section (5.1.3)
has a tendency to leave requests from such a cluster on the original route so therefore we propose a heuristic
that seeks to remove an entire cluster at once.
Although we could use the same algorithm as above for selecting related requests — just restricted to
a single route — we have chosen to use a heuristic based on strongly connected components, as described
in Section 4.2. We simply run Kruskal’s algorithm for the minimum spanning tree problem (using rij for
the edge distances) and terminate the algorithm when two connected components remain. One of these
clusters is chosen at random and the requests from the chosen cluster are removed. If less than q requests
have been selected, we randomly pick a removed request and choose a request from a different route, that
is most related to the given request. The route of the new request is partitioned into two clusters and so the
process continues until the desired number of requests has been removed. We refer the reader to Ropke
and Pisinger [50] for more details.

5.1.5 Time-oriented removal

The time oriented removal is another variant of the related removal heuristic. In this heuris-
tic we try to remove requests that are served at roughly the same time as we hope that these requests are
easy to interchange.
The heuristic works as follows. A request r̃ is chosen at random and the B requests that are closest to
r̃ (according to the distance rij defined in (20)) are marked. We define a time-oriented distance between
two requests as
∆tij = |tpi − tpj | + |tdi − tdj | (22)
where tpi and tdi are the time when the pickup and the delivery of request i are served in the current
solution. Among the B marked requests we select the q − 1 that are closest to r̃ according to ∆tij . The
process is controlled by a randomization parameter p like in the related removal heuristic described
in Section 5.1.3. These requests are removed together with r̃.
Before running the removal heuristic we first select a subset of all requests that are geographically
close to the chosen request, as we observed that this selection made the heuristic perform better on large
instances. The reason for this is that if the heuristic only considered requests that are close to the chosen
request time-wise, then only one or two requests would be removed from each route in the larger problems,
and this makes it hard to make any major improvements to the solution.

5.1.6 Historical node-pair removal

It is well-known from several metaheuristics that using historical information from the past local search
(e.g. the long term memory or the aspiration level in tabu search) may improve the performance of a local
search algorithm. In the present heuristic we look at the historical success of visiting two nodes right after
each other in a route, while the heuristic in Section 5.1.7 looks at the historical success of servicing two
requests by the same vehicle.
The historical node-pair removal heuristic (denoted the neighbor graph removal heuristic
in [50]) makes use of both historical information and the present solution when removing the requests. With
each pair of nodes (u, v) ∈ A we associate a weight f(u,v)∗
which indicates the best solution value found
∗
so far, in a solution which used edge (u, v). Initially f(u,v) is set to infinity, and each time a new solution
∗
is found in the ALNS algorithm, we update the weights f(u,v) of all edges used in the given solution, for
which the edge weight can be improved.
∗
We may use the edge weights f(u,v) to remove requests that seem to be misplaced. The removal
heuristic simply calculates the cost of a request (i, i + n) in the current solution by summing the weights
of edges incident to i and i + n. The most costly request is removed, and the process is repeated until q

12
requests have been extracted. To ensure some variation in the extracted requests, randomness is introduced
in the removal process.

5.1.7 Historical request-pair removal

An alternative history-based removal heuristic can make use of the historical success of placing pairs of
requests in the same route. We will call this approach historical request-pair removal (de-
noted request graph removal in [50]).
For this purpose we introduce the weight h(a,b) with each pair of requests (a, b) ∈ {1, . . . , n} ×
{1, . . . , n}. The weight h(a,b) denotes the number of times the two requests a and b have been served by the
same vehicle in the B best unique solutions observed so far in the search. Initially h(a,b) is set to zero, and
each time a new unique top-B solution is observed, the weights are incremented and decremented according
to the solutions entering and leaving the top-B solutions. A proper value of B was experimentally found
to be B = 100.
The weights h(a,b) could be used in a similar way as in the historical node-pair removal heuristic
described above, but initial experiments indicated that this was an unpromising approach. Instead, the
graph is used to define the relatedness between two requests, such that two requests are considered to be
related if the weight of the corresponding edge in the request graph is high. This relatedness measure is
used as in the related removal heuristic described in Section 5.1.3.

5.2 Inserting requests

The considered insertion heuristics all construct a number of routes for the vehicles. As each route can be
considered as an individual sub-problem the heuristics can build the routes sequentially or in parallel as
discussed in Section 4.2. The first-mentioned heuristics build one route at a time while parallel heuristics
construct several routes at the same time. The heuristics presented in this paper are all parallel, as they are
used in a context where a number of partial routes k ∈ R are given, and a number of unplaced requests U
should be inserted from the request bank.

5.2.1 Basic greedy heuristic

A simple greedy approach is to repeatedly insert a request in the cheapest possible route. More formally,
let ∆fi,k denote the change in the objective value incurred by inserting request i at the cheapest position
in route k. We set ∆fi,k = ∞ if request i cannot be inserted in route k. Following the greedy approach we
calculate
(i, k) := arg min ∆fi,k (23)
i∈U,k∈R

and insert request i in route k at its minimum cost position. This process continues until all requests have
been inserted or no more requests are feasible. The time complexity of this basic greedy heuristic is
decreased by tabulating all values of ∆fi,k and noting that only one route is changed in each iteration.

5.2.2 Regret heuristics

An obvious problem with the basic greedy heuristic is that it often postpones the placement of difficult
requests to the last iterations where we do not have much freedom of action. The regret heuristic tries
to circumvent the problem by incorporating a kind of look-ahead information when selecting the request
to insert. Regret heuristics have been used by Potvin and Rousseau [45] for the VRPTW and in the context
of the Generalized Assignment Problem by Trick [58].
Let ∆fiq denote the change in the objective value incurred by inserting request i into its best position
in the qth cheapest route for request i. For example ∆fi2 denotes the change in the objective value by
inserting request i in the route where the request can be inserted second cheapest. In each iteration, the
regret heuristic chooses to insert the request i that maximizes

i := arg max ∆fi2 − ∆fi1

(24)
i∈U

13
The request is inserted in the best possible route, at the minimum cost position. In other words, we maxi-
mize the difference of cost of inserting the request i in its best route and its second best route. We repeat
the process until no more requests can be inserted.
The heuristic can be extended in a natural way to define a class of regret heuristics: the regret-q
heuristic is the construction heuristic that in each construction step chooses to insert the request i that
maximizes
q
!
X
h 1
i := arg max ∆fi − ∆fi (25)
i∈U
h=2

Ties are broken by selecting the request with smallest insertion cost. The request i is inserted at its minimum
cost position, in its best route.
The regret heuristic based on criteria (24) is obviously a regret-2 heuristic and the basic
greedy heuristic from Section 5.2.1 is a regret-1 heuristic due to the tie-breaking rules. Informally
speaking, heuristics with q > 2 investigate the cost of inserting a request on the q best routes and chooses
to insert the request whose cost difference between inserting it into the best route and the q − 1 best routes
is largest. Compared to a regret-2 heuristic, regret-q heuristics with large values of q discover earlier
when the possibilities for inserting a request at a favorable place becomes limited.

5.3 Master local search framework

At the master level we have chosen to use simulated annealing as our local search framework. Our ac-
ceptance criteria in Line 5 of the main algorithm depicted in Figure 2 thus becomes to accept a candidate
solution x0 given the current solution x with probability
f (x0 )−f (x)
e− T , (26)

where T > 0 is the temperature. We use a standard exponential cooling rate, starting from the temperature
Tstart and decreasing T according to the expression T = T · c, where 0 < c < 1 is the cooling rate. We
calculate Tstart by inspecting our initial solution. The following method was developed in [50] and works
well when the number of requests in the problems to be solved is relatively constant. First the cost z 0 of the
initial solution is calculated using a modified objective function. In the modified objective function, Γ (cost
of having requests in the request bank) is set to zero. The start temperature is now set such that a solution
that is w percent worse than the current solution is accepted with probability 0.5. The reason for setting Γ
to zero is that this parameter typically is large and could cause us to set the starting temperature to a too
large number if the initial solution had some requests in the request bank. Now w is a parameter that has
to be set. We denote this parameter the start temperature control parameter. We have observed that this
approach is able to cope with instances of different sizes better if we divide the start temperature found by
the number n of requests in the instance.

5.4 Applying noise to the objective function

As mentioned in Section 4.1 it can be necessary to use noising or randomization in the destroy and repair
heuristics, as a diversification operator at the master level is not sufficient.
For the RPDPTW problem we have chosen to add a noise term to the objective function of the insertion
heuristics. Every time we calculate the cost C of a request insertion into a route, we add some noise δ and
calculate a modified insertion cost C 0 = max{0, C + δ}. The noise δ is chosen as a random number in the
interval [−Nmax , Nmax ], where Nmax = η · maxi,j∈V {dij }, and η is a parameter that controls the amount
of noise. We use the maximum distance to make the noise level proportional to the objective value. The
distances form part of the objective function in all problems considered, hence the noise level is somehow
proportional to the objective function.
Every insertion heuristic is split into two heuristics — one using noise, and one using the original
objective function only. After selecting which removal and insertion heuristic to use, it is decided if the
noise applied insertion heuristic should be used. This is again done using the roulette wheel selection
principle as we keep track of how well the insertion heuristics with and without noise have been performing

14
recently. Notice that we do not keep track of how well each individual insertion heuristic is performing
with and without noise, but only the insertion heuristics in general.

5.5 Adaptive weights adjustment

The roulette wheel selection mechanism in the ALNS framework presented in Section 4.1 is based on the
scores πi of the respective heuristics. A high score corresponds to a successful heuristic, and hence the
heuristic should be chosen with larger probability.
The scores are collected during some small time segments, defined as 100 iterations. The observed
score π i,j of a heuristic i in time segment j is incremented with the following values depending on the new
solution x0 :
σ1 The last remove-insert operation resulted in a new global best solution x0 .
σ2 The last remove-insert operation resulted in a solution x0 that has not been accepted before, and the
cost of the new solution is better than the cost of current solution.
σ3 The last remove-insert operation resulted in a solution x0 that has not been accepted before. The cost
of the new solution is worse than the cost of current solution, but the solution was accepted.
We distinguish between the two latter situations since we prefer heuristics that are able to improve on the
solution, but we also want to reward heuristics that can diversify the search to some extent. We keep track
of visited solutions by assigning a hash key to each solution and storing the key in a hash table.
At the end of each segment we calculate the smoothened scores to be used in the roulette wheel selection
as
π i,j
πi,j+1 = ρ + (1 − ρ)πi,j (27)
ai
where ai is the number of times the heuristic has been called in the time segment. The reaction factor ρ
controls how quickly the weight adjustment algorithm reacts to changes in the scores. If ρ = 1 then the
roulette wheel selection is only based on the scores in the most recent segment, while if ρ < 1 the scores
of past segments is also taken into account. For an illustration that shows how the scores evolve during a
search we refer the reader to [49].

5.6 Minimizing the number of vehicles used

The presented heuristic minimizes the travel costs, hence in order to minimize the number of vehicles also,
we use a two-stage approach.
Starting from a heuristic solution which makes use of m vehicles, we repeatedly remove one route and
place the corresponding requests in the request bank. If the ALNS heuristic is able to find a solution that
serves all requests we proceed with a lower number of routes. We assign a large cost Γ to requests in the
request bank to encourage solutions with all requests serviced.
If the ALNS heuristic fails to find a solution with all requests serviced, the algorithm steps back to the
last feasible solution encountered and proceeds with the second stage of the algorithm which consists of
the ordinary ALNS heuristic with the last found feasible solution as a starting point. For additional detail
on the two-stage algorithm see [49].
A different two-stage approach was used by Bent and Van Hentenryck [2], in which two distinct neigh-
borhoods and metaheuristics were used for the two stages.

6 Computational experiments
6.1 Parameter tuning
In order to keep the parameter tuning to a minimum we have used almost the same parameter setting as
determined in [49], with the exception of the cooling rate c and the start temperature control parameter
w. These were calibrated by selecting 5 reasonable values for each parameter and testing the 25 possible

15
combinations on 8 VRPTW instances with between 100 and 1000 customers. This was done separately for
both the vehicle minimizing ALNS and the ordinary distance minimizing ALNS, so different values for c
and w are used when trying to find a feasible solution and when minimizing the distance.

6.1.1 Selecting the number of requests to remove

In our past work [49, 50] we have removed up to 100 requests in each iteration. Experiments indicated
that we seldom accepted the moves resulting from such removals as the insertion heuristics are too weak.
Consequently the maximum number of requests that can be removed in a single iteration has been reduced
to 60. It was also observed that moves resulting from removing a small number of requests often were
accepted, but seldom lead to any major improvements of the solution. Therefore we now remove at least
0.1n requests in each iteration. To be precise, the number of requests to remove is found as a random
number between min {0.1n, 30} and min {0.4n, 60}. That is, for small instances the number of requests
to remove will be in the interval [0.1n, 0.4n] while for larger instances the interval is [30, 60].

6.2 Analysis of typical search

In order to illustrate how the present ALNS heuristic works, we have produced a number of figures by
running the heuristic on a 200 customer VRPTW instance and minimizing the traveled distance. All figures
are from the same search.
Figure 3 shows the cost of the accepted solutions and the best known solution as a function of the iter-
ation count. The figure is very typical for a Simulated Annealing metaheuristic. Initially very poor moves
are accepted and consequently the graph of accepted solutions is fluctuating wildly. As the temperature
is decreased the fluctuations become smaller and they eventually nearly die out such that only improving
solutions or very mild deteriorating solutions are accepted.
The next sequence of figures all show the distance between selected solutions. We have chosen to define
the distance between two solutions x and x0 as the Hamming distance between the corresponding binary
edge-variables. Figure 4 (left) shows the distance between each new accepted solution and the previously
accepted solution (the current solution). The figure shows that the ALNS in the first half of the search
can make huge changes to the solution in a single move as discussed in Section 4.3. In the last half of
the search only smaller moves are accepted. Figure 4 (right) depicts the difference between each proposed
solution and the last accepted solution. The figure shows that large moves are proposed throughout the
search process, but toward the end of the search these large moves are not accepted.
Figure 4 suggests some possible improvements to the algorithm: (1) Towards the end of the search it
seems to be beneficial to reduce the number of requests q that are removed in each iteration as the simulated
annealing framework generally only will accept minor changes. This could speed up the algorithm or
allow us to perform more iterations within the same amount of time. (2) Several moves have distance zero,
meaning that no changes were made to the solution vector. Obviously, such moves should be avoided,
possibly by incorporating a tabu-like principle in the insertion heuristics.
Figure 5 (top left) shows the Hamming distance between the accepted solutions and the previously best
known solution. Every time the distance reaches zero we have most likely found a new best solution (or
we have returned to the previously best known solution). It is interesting to see how quickly the search
moves away from the currently best known solution. This behavior is contrary to some of the ideas behind
the Variable Neighborhood metaheuristics and the Noising Method, where one tries to stick around the
currently best known solution or return to it if the current search direction seems fruitless. Also notice that
we move very far away from the best solutions. This can be seen as the number of edges in a solution is
equal to 2n + m. The maximum Hamming distance between two solutions is therefore 2(2n + m). In
the instance studied in this section n = 200 and m = 20, thus the maximum hamming distance for this
instance is 840.
Figure 5 (top right) shows the Hamming distance from each accepted solution and the best solution
found throughout the search. It is interesting to see that this plot is much more steady compared to the plot
in Figure 5 (top left) and that even though we are moving very far away from the previously best known
solution, the distance to the overall best solution (which of course is unknown early in the search) remains
roughly stable.

16
 5400

5200

5000

4800

4600

4400

4200

4000
0 5000 10000 15000 20000 25000

Figure 3: Solution cost as function of iteration count. Along the x-axis we show the iteration count while
the y-axis shows solution cost. The upper graph shows the cost of the accepted solutions while the lower
graph shows the cost of currently best known solution.
250 300

250
200

200
150

150

100
100

50
50

0 0
0 5000 10000 15000 20000 25000 0 5000 10000 15000 20000 25000

Figure 4: Left: Difference between accepted solutions. The figure shows the Hamming distance between
an accepted solution and the last accepted solution. Right: Difference between proposed solution and last
accepted solution. The figure shows the Hamming distance between each proposed solution and the last
accepted solution. The x-axis shows iteration count and the y-axis shows solution distance.

Figure 5 (bottom) combines the two previous plots. The upper contours of the two plots fit each other
surprisingly well. This indicates that the ALNS heuristic quickly moves away from the currently best
known solution until the distance to the currently best known solution is roughly the same as the distance
to the final best known solution. The search then visits solutions where the two distances are roughly the
same until a new best solution is found. We believe that the Simulated Annealing framework is responsible
for this behavior.

6.3 Application of the heuristic to standard benchmark problems

This section examines how the proposed heuristic performs on standard benchmark instances for the five
problem types considered in this paper. In order to investigate how much influence the number of LNS
iterations has on the solution quality, we have tested two configurations of our algorithm. One version
(ALNS-25K) that does 25000 iterations while minimizing the total traveled distance and one that does
50000 iterations (ALNS-50K). Both configurations use up to 25000 iterations in the vehicle minimization
stage. The cooling rate c in the simulated annealing algorithm described in Section 5.3 was adjusted such
that both configurations go through the same temperature span.
We have applied the heuristic to each instance 5 or 10 times, depending on the instance size. We report

17
 700 800
600 700

500 600
500
400
400
300
300
200 200
100 100
0 0
0 5000 10000 15000 20000 25000 0 5000 10000 15000 20000 25000

800
700
600
500
400
300
200
100
0
0 5000 10000 15000 20000 25000

Figure 5: Top left: Hamming distance between accepted solutions and the currently best known solution.
Top right: Hamming distance between accepted solutions and the best solution found during the search.
Bottom: The two plots showed in the same diagram. The x-axis shows iteration count and the y-axis
shows Hamming distance.

18
the best solution value out of the 5 or 10 experiments as well as the average solution value.
All experiments were performed on a 3GHz Pentium 4 computer. Detailed results from the experiments
can be found in the appendix. As mentioned before, the same parameter configuration has been used for
all experiments.

6.3.1 Vehicle routing problems with time windows (VRPTW)

An large number of metaheuristics have been proposed for solving the VRPTW. Bräysy and Gendreau
[8] have surveyed most of these approaches, and their survey contains 47 metaheuristics. Most of these
metaheuristics have been applied to the Solomon data set [55]. The Solomon data set contains 56 VRPTW
instances that all contain 100 customers. The instances contain a variety of customer and time window dis-
tributions and have proved to be a challenge for both heuristics and exact methods since their introduction.
Most of the proposed metaheuristics use vehicle minimization as primary objective and travel distance min-
imization as secondary objective, we prioritize our objectives in the same way. In this section we compare
the ALNS heuristic to the “best” of the previously proposed metaheuristics. It is hard to decide which of
the previously proposed metaheuristics are the best, as several criteria for comparing the heuristics could be
used. In this paper we have selected the metaheuristics that have been able to reach the minimum number
of total vehicles used for all of the instances in the Solomon data set, as these in a certain sense can be
regarded as the best heuristics in terms of solution quality. Table 1 summarizes this comparison.
The table shows that the ALNS heuristic is able to compete with the best heuristics for the VRPTW
when considering the moderate sized Solomon instances, even though it was not specifically designed for
this problem type. The heuristics by Homberger and Gehring [32] and Bent and Van Hentenryck [2] obtain
slightly better results compared to the best solutions obtained by ALNS-25K, but the papers do not state
how many experiments that were performed to reach these results. On the other hand, ALNS-25K reaches
slightly better solutions than the three remaining heuristics and the computational time is reasonable. The
column showing the average performance of ALNS-25K indicates that one run of the heuristic can be
performed quite fast but then one should not expect to reach the minimum number of vehicles. It does
not seem worthwhile to spend 50000 iteration instead of 25000 for these rather small problems. During
the calibration of the algorithm we discovered a new best solution to problem R207. This solution can be
found in the Appendix.
When the VRPTW has been solved by exact methods in the literature one has usually considered
minimizing the traveled distance without putting any limits on the number of vehicles. Furthermore all
distances are usually truncated to one decimal (see for example the work by Larsen [39]). In Table 2 we
summarize the result of applying the ALNS-25K heuristic to the Solomon VRPTW instances using the
same objective and rounding criteria as the exact methods. The heuristic has been applied to each instance
10 times and the table reports the best and average performances. The table shows that the heuristic is
able to find solutions that are very close to the optimal solutions and in many cases the heuristic is able to
identify the optimal solution in at least one of the test runs.
The optimal solutions have been reported in Chabrier [10], Cook and Rich [14], Danna and Le Pape
[20], Feilet et al. [24], Irnich and Villeneuve [34], Kallehauge et al. [35], Kohl et al. [36] and Larsen [39].
Larger VRPTW instances have been proposed by Gehring and Homberger [27]. The Gehring/Homberger
data set contains 300 instances with between 200 and 1000 customers. In Tables 3–7 we compare the
ALNS heuristic to the best heuristics that have been applied to these problems. The two heuristics that
reach the best solution quality is the heuristic by Mester and Bräysy [42] and the ALNS heuristic. Overall
the ALNS heuristic is better at minimizing the number of vehicles which is the primary objective of these
problems. The heuristic of Mester and Bräysy is very good at minimizing the traveled distance though.
The experiments show that the time used by the ALNS heuristic scales quite well with the problem size.
The 50000 iteration ALNS configuration becomes worthwhile for the larger problems. For problems with
600 customers or more the difference in total traveled distance obtained by the ALNS-25K and ALNS-50K
configurations become quite large, as the simulated annealing metaheuristic needs more iterations to obtain
a good solution for large problems.
The ALNS heuristic has been able to improve the best known solution for 122 out of the 300 large
scale VRPTW instances. The best solutions for the large VRPTW instances obtained by the ALNS-25K
and ALNS-50K configurations are shown in Table 8.

19
 BBB HG B BH IIKMUY ALNS 25K ALNS 50K
R1 11.92 11.92 11.92 11.92 11.92 11.92 12.03 11.92 12.03
1221.10 1212.73 1222.12 1211.10 1217.40 1213.39 1216.93 1212.39 1215.16
R2 2.73 2.73 2.73 2.73 2.73 2.73 2.75 2.73 2.75
975.43 955.03 975.12 954.27 959.11 958.60 968.01 957.72 965.94
C1 10.00 10.00 10.00 10.00 10.00 10.00 10.00 10.00 10.00
828.48 828.38 828.38 828.38 828.38 828.38 828.38 828.38 828.38
C2 3.00 3.00 3.00 3.00 3.00 3.00 3.00 3.00 3.00
589.93 589.86 589.86 589.86 589.86 589.86 589.86 589.86 589.86
RC1 11.50 11.50 11.50 11.50 11.50 11.50 11.60 11.50 11.60
1389.89 1386.44 1389.58 1384.17 1391.03 1385.39 1386.91 1385.78 1385.56
RC2 3.25 3.25 3.25 3.25 3.25 3.25 3.25 3.25 3.25
1159.37 1108.52 1128.38 1124.46 1122.79 1124.77 1140.06 1123.49 1135.46
CNV 405 405 405 405 405 405 407.5 405 407.5
CTD 57952 57192 57710 57273 57444 57360 57641 57332 57550
CPU P-400 Mhz P-400 Mhz P-200Mhz SU 10 P3 1Ghz P4 3Ghz P4 3Ghz P4 3Ghz P4 3Ghz
T. (s) 1800 N/A 4950 7200 15000 86 86 146 146
Exp. 3 N/A 1 >5 1 10 1 10 1

Table 1: Solomon instances with 100 customers. The table compares the ALNS heuristic to the heuristics
by Berger et al. (BBB) [4], Homberger and Gehring (HG) [32], Bräysy (B) [7], Bent and Van Hentenryck
(BH) [2] and Ibaraki et al. (IIKMUY) [33]. The data set is divided into six groups: R1, R2, C1, C2, RC1,
RC2. For each group we report two numbers per heuristic. The top number is the number of vehicles used
and the bottom number is the distance traveled. These numbers have been averaged over all the instance
in the given group. The rows named CNV and CTD show the cumulative number of vehicles and distances
respectively. The row CPU shows the computer used in the experiment, the row T. (s) shows the number
of CPU seconds used for finding the solutions. The last row shows the number of experiments that were
performed in order to obtain the results presented in the table (if multiple experiments were performed, the
table shows the best results obtained). The two columns for the ALNS heuristic show the results obtained
with the 25000 iteration configuration and the 50000 iteration configuration. For each configuration we
show two columns. The first column shows the best result out of ten experiments, and the second column
show the average solution quality (averaged over the ten experiments). Bold entries mark the best solution
quality obtained among the heuristics in the comparison.

exact methods ALNS

Customers Instances Solved to Optimums Avg. gap Avg. gap Avg. time
optimality found all (%) opt.(%) (s)
25 56 56 56 0.02 0.02 5
50 56 53 48 0.19 0.13 15
100 56 37 27 0.36 0.26 47

Table 2: Comparison of ALNS to exact methods. The columns should be interpreted as follows: Customers
— the number of customers in the test set, Instances — the number of instances in the test set, Solved to
optimality — the number of instances that has been solved to optimality in the literature, Optimums found
— the number of optimal solutions that were found by the heuristic, Avg. gap all (%) — the average gap
over all instances, Avg. gap opt. (%) — the average gap over instances solved to optimality in the literature,
Avg. time (s) — the average time in seconds spent on performing one experiment.

20
 GH99 GH01 BH LL LC BHD MB ALNS 25K ALNS 50K
R1 18.2 18.2 18.2 18.3 18.2 18.2 18.2 18.2 18.20 18.2 18.20
3705.00 3855.03 3677.96 3736.20 3676.95 3718.30 3618.68 3635.94 3664.648 3631.226 3652.747
R2 4.0 4.0 4.1 4.1 4.0 4.0 4.0 4.0 4.05 4.0 4.05
3055.00 3032.49 3023.62 3023.00 2986.01 3014.28 2942.92 2950.30 2950.04 2949.368 2942.594
C1 18.9 18.9 18.9 19.1 18.9 18.9 18.8 18.9 18.90 18.9 18.90
2782.00 2842.08 2726.63 2728.60 2743.66 2749.83 2717.21 2723.10 2732.458 2721.522 2728.382
C2 6.0 6.0 6.0 6.0 6.0 6.0 6.0 6.0 6.00 6.0 6.00
1846.00 1856.99 1860.17 1854.90 1836.10 1842.65 1833.57 1833.33 1836.4 1832.947 1834.675
RC1 18.0 18.1 18.0 18.3 18.0 18.0 18.0 18.0 18.00 18.0 18.00
3555.00 3674.91 3279.99 3385.80 3449.71 3329.62 3221.34 3233.76 3282.989 3212.282 3257.168
RC2 4.3 4.4 4.5 4.9 4.3 4.4 4.4 4.3 4.33 4.3 4.33
2675.00 2671.34 2603.08 2518.70 2613.75 2585.89 2519.79 2560.59 2592.39 2556.874 2578.575
CNV 694 696 697 707 694 695 694 694 694.8 694 694.8
CTD 176180 179328 171715 172472 173061 172406 168573 169370 170589 169042 169941
CPU P-200Mhz P-400Mhz SU 10 P-545Mhz P-933Mhz A-700Mhz P4 2Ghz P4-3Ghz P4-3Ghz P4-3Ghz P4-3Ghz
T. (min) 4x10 4x2.1 n/a 182.1 5x10 2.4 8 4.3 4.3 7.7 7.7
Exp. 1 3 n/a 3 1 3 1 10 1 10 1

Table 3: Gehring/Homberger VRPTW instances with 200 customers. The table compares the ALNS heuris-
tic to the heuristics by Gehring and Homberger (GH99) [27] and (GH01) [28], Bent and Van Hentenryck
(BH) [2], Le Bouthillier and Cranic (LC) [5], Bräysy et al (BHD) [9] and Mester and Bräysy (MB) [42].
The table should be interpreted like Table 1. Notice that computing times are reported in minutes. Entries
of the form x × y appearing in the T. (min) row indicate that the experiment was run for y minutes on a
parallel computer with x processors.

GH99 GH01 BH LL LC BHD MB ALNS 25K ALNS 50K

R1 36.4 36.4 36.4 36.6 36.5 36.4 36.3 36.4 36.40 36.4 36.40
8925.00 9478.22 8713.37 8912.40 8839.28 8692.17 8530.03 8609.38 8663.57 8540.04 8589.90
R2 8.0 8.0 8.0 8.0 8.0 8.0 8.0 8.0 8.00 8.0 8.00
6502.00 6650.28 6959.75 6610.60 6437.68 6382.63 6209.94 6252.01 6309.84 6241.72 6277.07
C1 38.0 38.0 38.0 38.7 37.9 37.9 37.9 37.6 37.62 37.6 37.62
7584.00 7855.82 7220.96 7181.40 7447.09 7230.48 7148.27 7369.88 7450.84 7290.16 7372.49
C2 12.0 12.0 12.0 12.1 12.0 12.0 12.0 12.0 12.00 12.0 12.00
3935.00 3940.19 4154.40 4017.10 3940.87 3894.48 3840.85 3849.27 3884.44 3844.69 3875.95
RC1 36.1 36.1 36.1 36.5 36.0 36.0 36.0 36.0 36.00 36.0 36.00
8763.00 9294.99 8330.98 8377.90 8652.01 8305.55 8066.44 8149.61 8240.28 8069.30 8148.81
RC2 8.6 8.8 8.9 9.5 8.6 8.9 8.8 8.5 8.64 8.5 8.64
5518.00 5629.43 5631.70 5466.20 5511.22 5407.87 5243.06 5366.82 5388.76 5335.09 5351.56
CNV 1390 1392 1393 1414 1390 1391 1389 1385 1386.6 1385 1386.6
CTD 412270 428489 410112 405656 408281 399132 390386 395970 399377 393210 396158
CPU P-200Mhz P-400Mhz SU 10 P-545Mhz P-933Mhz A-700Mhz P4 2Ghz P4-3Ghz P4-3Ghz P4-3Ghz P4-3Ghz
T. (min) 4x20 4x7.1 n/a 359 5x20 7.9 17 9.7 9.7 15.8 15.8
Exp. 1 3 n/a 3 1 3 1 5 1 5 1

Table 4: Gehring/Homberger VRPTW instances with 400 customers

21
 GH99 GH01 BH LL LC BHD MB ALNS 25K ALNS 50K
R1 54.5 54.5 55.0 55.2 54.8 54.5 54.5 54.5 54.50 54.5 54.50
20854.00 21864.47 19308.62 19744.80 19869.82 19081.18 18358.68 19370.04 19562.34 18888.52 19048.49
R2 11.0 11.0 11.0 11.1 11.2 11.0 11.0 11.0 11.00 11.0 11.00
13335.00 13656.15 14855.43 13592.40 13093.97 13054.83 12703.52 12729.51 12826.39 12619.26 12721.32
C1 57.9 57.7 57.8 58.2 57.9 57.8 57.8 57.5 57.56 57.5 57.56
14792.00 14817.25 14357.11 14267.30 14205.58 14165.90 14003.09 14125.94 14212.71 14065.89 14098.04
C2 17.9 17.8 17.8 18.2 17.9 18.0 17.8 17.5 17.80 17.5 17.80
7787.00 7889.96 8259.04 8202.60 7743.92 7528.73 7455.83 7891.70 7834.72 7801.296 7682.61
RC1 55.1 55.0 55.1 55.5 55.2 55.0 55.0 55.0 55.00 55.0 55.00
18411.00 19114.02 17035.91 17320.00 17678.13 16994.22 16418.63 16846.71 17006.94 16594.94 16722.51
RC2 11.8 11.9 12.4 13.0 11.8 12.1 12.1 11.6 11.78 11.6 11.78
11522.00 11670.29 11987.89 11204.90 11034.71 11212.36 10677.46 10922.44 10938.30 10777.12 10828.45
CNV 2082 2079 2091 2112 2088 2084 2082 2071 2076.4 2071 2076.4
CTD 867010 890121 858040 843320 836261 820372 796172 818863 823814 807470 811014
CPU P-200Mhz P-400Mhz SU 10 P-545Mhz P-933Mhz A-700Mhz P4 2Ghz P4-3Ghz P4-3Ghz P4-3Ghz P4-3Ghz
T. (min) 4x30 4x12.9 n/a 399.8 5x30 16.2 40 10.5 10.5 18.3 18.3
Exp. 1 3 n/a 3 1 3 1 5 1 5 1

Table 5: Gehring/Homberger VRPTW instances with 600 customers.

GH99 GH01 BH LL LC BHD MB ALNS 25K ALNS 50K

R1 72.8 72.8 72.7 73.0 73.1 72.8 72.8 72.8 72.80 72.8 72.80
34586.00 34653.88 33337.91 33806.34 33552.40 32748.06 31918.47 32697.85 32905.52 32316.79 32528.76
R2 15.0 15.0 15.0 15.1 15.0 15.0 15.0 15.0 15.00 15.0 15.00
21697.00 21672.85 24554.63 21709.39 21157.56 21170.15 20295.28 20477.77 20627.40 20353.51 20499.72
C1 76.7 76.1 76.1 77.4 76.3 76.3 76.2 75.6 75.66 75.6 75.66
26528.00 26936.68 25391.67 25337.02 25668.82 25170.88 25132.27 25365.59 25547.82 25193.13 25269.64
C2 24.0 23.7 24.4 24.4 24.1 24.2 23.7 23.7 23.98 23.7 23.94
12451.00 11847.92 14253.83 11956.60 11985.11 11648.92 11352.29 11985.80 11999.28 11725.46 11741.73
RC1 72.4 72.3 73.0 73.2 72.3 73.0 73.0 73.0 73.00 73.0 73.00
38509.00 40532.35 30500.15 31282.54 37722.62 30005.95 30731.07 29864.06 30016.05 29478.3 29625.04
RC2 16.1 16.1 16.6 17.1 15.8 16.3 15.8 15.7 15.82 15.7 15.82
17741.00 17941.23 18940.84 17561.22 17441.60 17686.65 16729.18 16870.87 17022.33 16761.95 16852.95
CNV 2770 2760 2778 2802 2766 2776 2765 2758 2762.6 2758 2762.2
CTD 1515120 1535849 1469790 1416531 1475281 1384306 1361586 1372619 1381184 1358291 1365178
CPU P-200Mhz P-400Mhz SU 10 P-545Mhz P-933Mhz A-700Mhz P4-2Ghz P4-3Ghz P4-3Ghz P4-3Ghz P4-3Ghz
T. (min) 4x40 4x23.2 n/a 512.9 5x40 26.2 145 13.5 13.5 22.7 22.7
Exp. 1 3 n/a 3 1 3 1 5 1 5 1

Table 6: Gehring/Homberger VRPTW instances with 800 customers

GH99 GH01 BH LL LC BHD MB ALNS 25K ALNS 50K

R1 91.9 91.9 92.8 92.7 92.2 92.1 92.1 92.2 92.30 92.2 92.30
57186.00 58069.61 51193.47 50990.80 55176.95 50025.64 49281.48 52131.96 51900.53 50751.25 50584.55
R2 19.0 19.0 19.0 19.0 19.2 19.0 19.0 19.0 19.00 19.0 19.00
31930.00 31873.62 36736.97 31990.90 30919.77 31458.23 29860.32 30108.84 30327.34 29780.82 30016.08
C1 96.0 95.4 95.1 96.3 95.3 95.8 95.1 94.6 94.72 94.6 94.72
43273.00 43392.59 42505.35 42428.50 43283.92 42086.77 41569.67 42123.87 42266.42 41877.00 42034.65
C2 30.2 29.7 30.3 30.8 29.9 30.6 29.7 29.7 29.90 29.7 29.86
17570.00 17574.72 18546.13 17294.90 17443.50 17035.88 16639.54 17307.16 17589.70 16840.37 17052.62
RC1 90.0 90.1 90.2 90.4 90.0 90.0 90.0 90.0 90.00 90.0 90.00
50668.00 50950.14 48634.15 48892.40 49711.36 46736.92 45396.41 47735.43 48168.74 46752.15 47081.64
RC2 19.0 18.5 19.4 19.8 18.5 19.0 18.7 18.3 18.46 18.3 18.46
27012.00 27175.98 29079.78 26042.30 26001.11 25994.12 25063.51 25267.93 25466.13 25090.88 25185.45
CNV 3461 3446 3468 3490 3451 3465 3446 3438 3443.8 3438 3443.4
CTD 2276390 2290367 2266959 2176398 2225366 2133376 2078110 2146752 2157189 2110925 2119550
CPU P-200Mhz P-400Mhz SU 10 P-545Mhz P-933Mhz A-700Mhz P4 2Ghz P4-3Ghz P4-3Ghz P4-3Ghz P4-3Ghz
T. (min) 4x50 4x30.1 n/a 606.3 5x50 39.6 600 16 16 26.6 26.6
Exp. 1 3 n/a 3 1 3 1 5 1 5 1

Table 7: Gehring/Homberger VRPTW instances with 1000 customers.

22
 R1 R2 C1 C2 RC1 RC2
# Veh. Dist. Veh. Dist. Veh. Dist. Veh. Dist. Veh. Dist. Veh. Dist.
200 customers
1 20 4785.96 4 4563.55 20 2704.57 6 1931.44 18 3647.56 6 3126.03
2 18 4059.57 4 3650.54 18 2943.83 6 1863.16 18 3269.91 5 2828.39
3 18 3387.64 4 2892.07 18 2710.21 6 1776.96 18 3034.45 4 2613.12
4 18 3086.11 4 1981.30 18 2644.92 6 1713.46 18 2869.74 4 2052.74
5 18 4125.19 4 3377.18 20 2702.05 6 1878.85 18 3430.03 4 2912.13
6 18 3586.80 4 2929.72 20 2701.04 6 1857.35 18 3357.90 4 2975.13
7 18 3160.44 4 2456.71 20 2701.04 6 1849.46 18 3233.29 4 2539.85
8 18 2971.66 4 1849.87 19 2775.48 6 1820.53 18 3110.46 4 2314.61
9 18 3802.55 4 3113.74 18 2687.83 6 1830.05 18 3114.02 4 2175.98
10 18 3312.44 4 2666.10 18 2644.25 6 1808.21 18 3020.24 4 2015.61
400 customers
1 40 10432.30 8 9338.49 40 7152.06 12 4116.33 36 8813.43 11 6834.02
2 36 9115.68 8 7649.87 36 7733.55 12 3930.05 36 8118.43 9 6355.59
3 36 7988.22 8 5998.04 36 7082.13 12 3775.32 36 7663.73 8 5055.02
4 36 7415.81 8 4326.48 36 6816.17 12 3543.60 36 7368.47 8 3647.39
5 36 9479.10 8 7252.64 40 7152.06 12 3946.14 36 8426.57 9 6119.44
6 36 8556.38 8 6212.37 40 7153.45 12 3875.94 36 8390.24 8 5997.24
7 36 7725.97 8 5136.74 39 7546.78 12 3894.98 36 8223.65 8 5476.57
8 36 7390.76 8 4055.22 37 7546.32 12 3796.00 36 7922.67 8 4877.39
9 36 8970.98 8 6507.40 36 7573.18 12 3881.21 36 7953.20 8 4601.30
10 36 8325.16 8 5894.40 36 7145.92 12 3687.13 36 7774.83 8 4355.52
600 customers
1 59 21677.41 11 18837.28 60 14095.64 18 7780.84 55 17751.33 15 13163.03
2 54 20045.49 11 15069.24 56 14174.12 17 8799.38 55 16548.43 12 11853.72
3 54 17733.91 11 11291.52 56 13803.50 17 7604.00 55 15499.02 11 9863.35
4 54 16374.29 11 8163.24 56 13578.66 17 6993.77 55 15072.90 11 7231.64
5 54 21243.24 11 15418.00 60 14085.72 18 7578.12 55 17401.34 12 12560.43
6 54 18948.53 11 12936.28 60 14089.66 18 7554.61 55 17355.10 11 12282.52
7 54 17438.28 11 10269.96 58 15017.03 18 7520.34 55 17058.40 11 11052.49
8 54 16146.17 11 7752.78 57 14343.05 17 8696.15 55 16510.65 11 10488.75
9 54 20375.70 11 13885.52 56 13767.45 18 7356.19 55 16435.71 11 9882.71
10 54 18902.19 11 12568.79 56 13688.57 17 7938.94 55 16316.51 11 9340.06
800 customers
1 80 37492.04 15 28822.48 80 25184.38 24 11664.00 73 31275.38 19 20954.95
2 72 33816.69 15 23274.22 74 25536.76 24 11428.07 73 29172.08 17 18032.89
3 72 30317.49 15 18078.82 72 24629.86 24 11184.67 73 28164.66 15 14800.78
4 72 28568.78 15 13413.79 72 23938.33 23 10999.42 73 27201.39 15 11368.19
5 72 35503.63 15 25077.09 80 25166.28 24 11451.57 73 30548.23 16 19180.13
6 72 32360.07 15 20969.81 80 25160.85 24 11403.57 73 30511.07 15 19075.89
7 72 29979.63 15 16977.49 79 25425.92 24 11412.08 73 30007.82 15 17329.32
8 72 28341.21 15 12945.52 75 25450.99 23 13878.40 73 29547.96 15 16226.78
9 72 34218.41 15 22877.21 72 25737.46 24 11650.10 73 29360.93 15 15687.20
10 72 32569.97 15 21092.27 72 25697.68 23 12103.56 73 28993.52 15 14944.14
1000 customers
1 100 54720.19 19 43264.68 100 42478.95 30 16879.24 90 48933.68 21 30396.13
2 91 55428.79 19 34417.47 91 42249.60 29 17563.06 90 46165.33 18 27552.05
3 91 49634.84 19 25400.16 90 40376.43 30 16109.71 90 44014.81 18 20811.18
4 91 45303.47 19 18332.77 90 39980.07 29 16011.30 90 42607.34 18 16007.59
5 92 53089.15 19 37746.01 100 42469.18 30 16596.69 90 48934.53 18 28368.48
6 91 54555.32 19 30778.85 100 42471.29 30 16369.10 90 48766.98 18 28746.61
7 91 48141.47 19 23991.71 99 42673.51 31 16590.48 90 48005.94 18 26765.43
8 91 44853.70 19 17844.36 95 42359.27 29 18407.27 90 47122.61 18 24961.29
9 92 52015.72 19 34349.70 91 41482.00 30 16294.72 90 46889.79 18 24113.72
10 92 49769.85 19 31682.52 90 42214.60 29 17582.15 90 46080.51 18 23056.75

Table 8: The table shows the best solutions to large VRPTW instances identified by the ALNS heuristic.
The first column shows the problem number. The columns veh. and dist. show the number of vehicles and
total distance traveled in the best solution found. The table is grouped by instance type and instance size.
Bold entries indicate a best solution (either a tie with one of the heuristics from the literature or a new best
solution).

23
6.3.2 Multi depot vehicle routing problem (MDVRP)
Table 9 shows the results obtained on 33 MDVRP instances from the literature. Both ALNS configurations
have been applied 10 times to each instance. The results obtained by the ALNS heuristic are compared to
the best results obtained by heuristics proposed by Chao et al. [12], Renaud et al. [48] and Cordeau et al.
[17]. The heuristic that previously has achieved the best solution quality is the one proposed by Cordeau et
al. The cost of a solution is defined as the total distance traveled by the vehicles. The table shows that the
ALNS heuristic has been able to improve upon the best solution for a considerable number of instances.
Each configuration has found 14 new best solutions, but as most of these overlap, the total number of new
best solutions is 15. The individual improvements are typically rather small though. The table also shows
that the ALNS heuristic is quite stable as the average gap from the best known solution never surpasses
2% and 1% in the ALNS-25K and ALNS-50K configurations, respectively. It should be mentioned that
the ALNS heuristic is slower than the previously proposed heuristics. The ALNS-25K and ALNS-50K
configurations use on average two and four minutes respectively to perform one experiment on a 3GHz
Pentium 4. The heuristic by Cordeau et al. on average used 11.7 minutes to perform one experiment on a
Sun Sparcstation 10 which is considerably slower than our computer.

6.3.3 Site dependent vehicle routing problem (SDVRP)

The results obtained on the SDVRP instances are summarized in Table 10. The results are promising as the
average solution quality of ALNS-25K overall is better than results previously published. Also the sum of
the costs of the best known solutions found by the ALNS-50K configuration is more than 2% better than
the previous best known solution and the best known solution was improved for 30 out of the 35 instances.
The computational time for performing one experiment with the ALNS-25K configuration seems to be
roughly comparable with the time needed for performing one experiment with the heuristic proposed by
Cordeau and Laporte [18]. The ALNS-25K configuration spends on average 1.4 minutes to perform one
experiment while the heuristic by Cordeau and Laporte spent around 12 minutes to perform the same task
on a Sun Ultra 2, 300 MHz. It should be mentioned that the problem PR02 caused the ALNS heuristic some
difficulties, as it was only able to find a feasible solution in one out of ten experiments for the ALNS-25K
configuration and three out of ten experiments for the ALNS-50K configuration.

6.3.4 Capacitated vehicle routing problem (CVRP)

For the CVRP we have chosen to test the ALNS heuristic on three datasets. The first dataset was proposed
by Christofides et al. [13] and contains instances with between 50 and 200 customers, the second dataset
was proposed by Golden et al. [30] and contains instances with up to 483 customers. The last dataset was
proposed by Li et al. [40] and contains instances with up to 1200 customers. These are the so-far largest
instances that the ALNS heuristic has been applied to. Table 11 summarizes these experiments. Notice
that we only compare the ALNS heuristic to a subset of all the CVRP heuristics that have been proposed
in the literature. The heuristics used for benchmarking are the most recent heuristics that were surveyed by
Cordeau et al. [16].
The table shows that the ALNS heuristic cannot compete with the well-performing heuristic by Mester
and Bräysy [42], but its performance is comparable to the rest of the heuristics. For the last dataset, the
heuristic proposed by Li et al. must be considered to be the best as it is very fast compared to the ALNS
heuristic although the ALNS heuristic overall is able to reach better solutions. We discovered one new best
solution for the Golden et al. dataset and three new best solutions for the Li et al. dataset.

6.3.5 Open vehicle routing problem (OVRP)

The results on the OVRP are summarized in Table 12. The heuristic was tested on the same 16 instances
that were used by Brandão [6] and Fu et al. [25]. The primary objective considered was to minimize
the number of vehicles used, while the secondary objective was to minimize the traveled distance. The
solutions obtained by the ALNS heuristic are promising as the best known solution to 11 out of the 16
instances has been improved. The running time of the ALNS heuristic is quite respectable compared to the
two other heuristics. The configuration of Brandão’s heuristic that obtains the best results spends on average

24
9.6 minutes to solve an instance on a 500MHz Pentium III. In the paper by Fu et al. two configurations
of their heuristic are tested. These configurations spend on average 6.6 and 13.9 minutes respectively to
solve an instance on a 600 MHz Pentium II. The ALNS-25K and ALNS-50K configurations use 1.4 and
2.3 minutes respectively to solve an instance on a 3GHz Pentium IV.

6.3.6 Computational results conclusion

The computational results presented in this section are very encouraging. The results show that the general
ALNS heuristic is on par with the best specialized heuristics for the VRPTW and that the heuristic currently
is the best when it comes to minimizing the number of vehicles in large VRPTW instances. One should
keep in mind that numerous specialized heuristics have been proposed for the VRPTW making it difficult
for a general heuristic to compete on these instances.
For the MDVRP, SDVRP and OVRP the ALNS heuristic has been able to find many new best solutions
and the results on the SDVRP are especially promising. For the CVRP the proposed heuristic is able to
compete with many of the most recent heuristics, but it is outperformed by a more specialized heuristic
for this problem. Nevertheless, a couple of new best solutions were found for this problem type also. One
should also keep in mind that the heuristic was not tuned for each problem type, but a general parameter
setting was used for all experiments.
The comparison between the fast and the slow version of the ALNS heuristic showed that it did not
pay off to use the ALNS-50K variant for the smaller instances, while for instances with around 400 to
600 or more customers it seemed worthwhile to use the ALNS-50K configuration. Consequently, it might
be useful to use a variable number of iterations I which depends on the number n of requests as e.g.
I := 20000 + 50n.

7 Conclusion
A new, general heuristic framework, denoted Adaptive Large Neighborhood Search has been presented.
The framework has been used to several variants of vehicle routing problems in the present paper as well
as in [49, 50]. This includes the vehicle routing problem with time windows (VRPTW), the capacitated
vehicle routing problem (CVRP), the multi-depot vehicle routing problem (MDVRP), the site dependent
vehicle routing problem (SDVRP), the open vehicle routing problem (OVRP), the pickup and delivery
problem with time windows (PDPTW), the vehicle routing problem with backhauls (VRPB), the mixed
vehicle routing problem with backhauls (MVRPB), the multi-depot mixed vehicle routing problem with
backhauls (MDMVRPB), the vehicle routing problem with backhauls and time windows (VRPBTW), the
mixed vehicle routing problem with backhauls and time windows (MVRPBTW) and the vehicle routing
problem with simultaneous deliveries and pickups (VRPSDP).
Due to the generality of the ALNS framework, and the encouraging results demonstrated for a wide
spectrum of VRP problems, we believe that ALNS should be considered as one of the standard frameworks
for solving large-sized optimization problems.
Supply chain management is a research area getting increasing attention [37]. By co-ordinating activ-
ities in the supply chain, companies can rationalize the process resulting in mutual gains. If the involved
companies co-ordinate their transportation activities we will see a need for solving mixed transportation
problems, where the instances for example consist of a mixture of PDPTW, MDRP and SDVRP problems.
In order to handle future changes in the distribution structure, these algorithms need to be stable for various
input types, and should not need to be tuned for particular problem characteristics. It should be clear that
the ALNS framework may play an important role.
In conclusion we may add a philosophical observation: We have seen that a mixture of good and
less good heuristics lead to better solutions than using good heuristics solely. It is however necessary
to hierarchically controls the search, such that well-performing heuristics are given most influence, but
such that all heuristics participate to the solution process. Using this principle one gets a robust and well-
performing solution approach.

25
8 Appendix
New best solution to Solomon R207 instance
Route Length Visit sequence
1 437.339 42 92 45 46 36 64 11 62 88 30 20 65 71 9 81 34 78 79 3 76 28 53 40 2 87 57
41 22 73 21 72 74 75 56 4 25 55 54 80 68 77 12 26 58 13 97 37 100 98 93 59
95 94
2 453.269 27 1 69 50 33 29 24 39 67 23 15 43 14 44 38 86 16 61 91 85 99 96 6 84 8 82 7
48 47 49 19 10 63 90 32 66 35 51 70 31 52 18 83 17 5 60 89

Total length 890.61

Full VRPTW tables

The full tables documenting the VRPTW experiments described in section 6.3.1 can be found in Tables 13
– 18.
Tables 19 – 21 contain detailed result from the experiment comparing the ALNS heuristic to exact
methods.

Full CVRP tables

The detailed results for the CVRP experiments described in section 6.3.4 can be found in Tables 22 – 24.

References
[1] Ravindra K. Ahuja, Özlem Ergun, James b. Orlin, and Abraham P. Punnen. A survey of very large-
scale neighborhood search techniques. Discrete Applied Mathematics, 123:72–102, 2002.
[2] Russel Bent and Pascal Van Hentenryck. A two-stage hybrid local search for the vehicle routing
problem with time windows, 2004. To appear in Transportation Science.
[3] J. Berger and M. Barkaoui. A new hybrid genetic algorithm for the capacitated vehicle routing prob-
lem. Journal of the Operational Research Society, 54:1254–1262, 2003.
[4] Jean Berger, Mohamed Barkaoui, and Olli Bräysy. A route hybrid genetic approach for the vehicle
routing problem with time windows. INFOR, 41(2), 2003.
[5] Alexandre Le Bouthillier and Teodor Gabriel Crainic. A cooperative parallel meta-heuristic for the
vehicle routing problem with time windows. Computers and Operations Research, 2004. to appear.
[6] José Brandão. A tabu search algorithm for the open vehicle routing problem. European Journal of
Operational Research, 157(3):552–564, 2004.
[7] Olli Bräysy. A reactive variable neighborhood search for the vehicle-routing problem with time
windows. INFORMS Journal on Computing, 15(4):347–368, 2003.
[8] Olli Bräysy and Michel Gendreau. Vehicle routing problem with time windows, part ii: Metaheuris-
tics. Transportation Science. To appear.
[9] Olli Bräysy, Geir Hasle, and Wout Dullaert. A multi-start local search algorithm for the vehicle
routing problem with time windows. European Journal of Operational Research, 2003. to appear.
[10] Alain Chabrier. Vehicle routing problem with elementary shortest path based column generation.
Working Paper, ILOG, Madrid, 2003, 2003.
[11] I-Ming Chao, Bruce Golden, and Edward Wasil. A computational study of a new heuristic for the
site-dependent vehicle routing problem. INFOR, 37(3):319–336, 1999.

26
[12] I.M. Chao, B. L. Golden, and E.A. Wasil. A new heuristic for the multi-depot vehicle routing problem
that improves upon best-known solutions. Am. J. Math. Mgmt. Sci., 13:371–406, 1993.
[13] Nicos Christofides, Aristide Mingozzi, and Paolo Toth. The vehicle routing problem. In Nicos
Christofides, Aristide Mingozzi, Paolo Toth, and Claudio Sandi, editors, Combinatorial Optimization,
chapter 11, pages 315 – 338. John Wiley & Sons, 1979.
[14] William Cook and Jennifer L. Rich. A parallel cutting-plane algorithm for the vehicle routing prob-
lem with time windows. Technical Report TR99-04, Departement of Computational and Applied
Mathematics, Rice University, 1999.
[15] Jean-François Cordeau, Guy Desaulniers, Jacques Desrosiers, Marius M. Solomon, and François
Soumis. Vrp with time windows. In Paulo Toth and Daniele Vigo, editors, The Vehicle Routing
Problem, volume 9 of SIAM Monographs on Discrete Mathematics and Applications, chapter 7, pages
157–193. SIAM, Philidelphia, 2002.
[16] Jean-François Cordeau, Michel Gendreau, Alain Hertz, Gilbert Laporte, and Jean-Sylvain Sormany.
New heuristics for the vehicle routing problem. Technical Report G-2004-33, GERAD, Montreal,
Canada, 2004.
[17] Jean-François Cordeau, Michel Gendreau, and Gilbert Laporte. A tabu search heuristic for periodic
and multi-depot vehicle routing problems. Networks, 30:105–119, 1997.
[18] Jean-François Cordeau and Gilbert Laporte. A tabu search algorithm for the site dependent vehicle
routing problem with time windows. INFOR, 39:292–298, 2001.
[19] Jean-François Cordeau, Gilbert Laporte, and Anne Mercier. A unified tabu search heuristic for vehicle
routing problems with time windows. Journal of the Operational Research Society, 52:928–936,
2000.
[20] Emilie Danna and Claude Le Pape. Accelerating branch-and-price with local search: A case study on
the vehicle routing problem with time windows. Technical Report 03-006, ILOG, ILOG S.A, 9, rue
de Verdun, F-94253 Gentilly Cédex, Septemper 2003.
[21] Guy Desaulniers, Jacques Desrosiers, Andreas Erdmann, Marius M. Solomon, and François Soumis.
Vrp with pickup and delivery. In Paulo Toth and Daniele Vigo, editors, The Vehicle Routing Problem,
volume 9 of SIAM Monographs on Discrete Mathematics and Applications, chapter 9, pages 225–242.
SIAM, Philidelphia, 2002.
[22] Ozlem Ergun, James B. Orlin, and Abran Steele-Feldman. Creating very large scale neighborhoods
out of smaller ones by compounding moves: a study on the vehicle routing problem. Working Paper,
Massachusetts Institute of Technology, 2003.
[23] Éric Taillard. Parallel iterative search methods for vehicle routing problems. Networks, 23:661–673,
1993.

[24] Dominique Feillet, Pierre Dejax, Michel Gendreau, and Cyrille Gueguen. An exact algorithm for
the elementary shortest path problem with resource constraints: Application to some vehicle routing
problems. Networks, 44(3):216–229, 2004.
[25] Zhuo Fu, Richard W. Eglese, and Leon Li. A tabu search heuristic for the open vehicle routing
problem. Technical Report 2003/042, Lancaster University Management School, Lancaster, United
Kingdom, 2003.
[26] Ricardo Fukasawa, Jens Lysgaard, Marcus Poggi de Aragão, Marcelo Reis, Eduardo Uchoa, and
Renato F. Werneck. Robust branch-and-cut-and-price for the capacitated vehicle routing problem. In
Proceedings of IPCO X. Columbia University, 2004.

27
[27] Hermann Gehring and Jörg Homberger. A parallel hybrid evolutionary metaheuristic for the vehicle
routing problem with time windows. In K. Miettinen, M. Mäkelä, and J. Toivanen, editors, Pro-
ceedings of EUROGEN99 – Short Course on Evolutionary Algorithms in Engineering and Computer
Science, pages 57–64. University of Jyväskylä, 1999.
[28] Hermann Gehring and Jörg Homberger. A parallel two-phase metaheuristic for routing problems with
time windows. Asia-Pacific Journal of Operational Research, 18:35–47, 2001.
[29] Michel Gendreau, Gilbert Laporte, and Jean-Yves Potvin. Metaheuristics for the capacitated vrp. In
Paulo Toth and Daniele Vigo, editors, The Vehicle Routing Problem, volume 9 of SIAM Monographs
on Discrete Mathematics and Applications, chapter 6, pages 129–154. SIAM, Philidelphia, 2002.
[30] Bruce L. Golden, Edward A. Wasil, James P. Kelly, and I-Ming Chao. Metaheuristics in vehicle
routing. In T.G. Crainic and G. Laporte, editors, Fleet Management and Logistics, pages 33–56.
Kluwer, Boston (MA), 1998.
[31] Pierre Hansen and Nenad Mladenovic. An introduction to variable neighborhood search. In Ste-
fan Voss et al., editor, Metaheuristics, Advances and Trends in Local Search Paradigms for Optimiza-
tion, pages 433–458. Kluwer Academic Publishers, Dordrecht, 1999.
[32] Jörg Homberger and Hermann Gehring. A two-phase hybrid metaheuristic for the vehicle routing
problem with time windows. European Journal of Operational Research, 2004. To appear.
[33] T. Ibaraki, S. Imahori, T. Masuda, T. Uno, and M. Yagiura. A route hybrid genetic approach for the
vehicle routing problem with time windows. Transportation Science. To appear.
[34] Stefan Irnich and Daniel Villeneuve. The shortest path problem with resource constraints and k-cycle
elimination for k ≥ 3. Technical Report G-2003-55, GERAD, Montreal, Canada, September 2003.
[35] B. Kallehauge, J. Larsen, and O. B. G. Madsen. Lagrangean duality applied on vehicle routing
with time windows - experimental results. Technical Report IMM-REP-2000-8, Informatics and
Mathematical Modelling, Technical University of Denmark, DTU, Richard Petersens Plads, Building
321, DK-2800 Kgs. Lyngby, 2001.
[36] Niklas Kohl, Jacques Desrosiers, Oli B. G. Madsen, Marius M. Solomon, and François Soumis. 2-
path cuts for the vehicle routing problem with time windows. Transportation Science, 33(1):101 –
116, 1999.
[37] D.M. Lambert and M.C. Cooper. Issues in supply chain management. Marketing Management,
29:65–83, 2000.
[38] Gilbert Laporte and Frédéric Semet. Classical heuristics for the capacitated vrp. In Paulo Toth and
Daniele Vigo, editors, The Vehicle Routing Problem, volume 9 of SIAM Monographs on Discrete
Mathematics and Applications, chapter 5, pages 109–128. SIAM, Philidelphia, 2002.
[39] Jesper Larsen. Parallellization of the Vehicle Routing Problem with Time Windows. PhD thesis,
Department of Mathematical Modelling, Technical University of Denmark, 1999. IMM-PHD-1999-
62.
[40] Feiyue Li, Bruce Golden, and Edward Wasil. Very large-scale vehicle routing: new test problems,
algorithms, and results. Computers and Operations Research, 2004. To appear.
[41] Jens Lysgaard, Adam N. Letchford, and Richard W. Eglese. A new branch-and-cut algorithm for the
capacitated vehicle routing problem. Mathematical Programming, Ser. A, 100:423–445, 2004.

[42] David Mester and Olli Bräysy. Active guided evolution strategies for large-scale vehicle routing
problems with time windows. Computers and Operations Research, 2004. to appear.

28
[43] Barindra Nag, Bruce L. Golden, and Arjang Assad. Vehicle routing with site dependencies. In B.L.
Golden and A.A. Assad, editors, Vehicle Routing: Methods and Studies, pages 149–159. Elsevier
Science Publishers B.V, 1988.

[44] David Pisinger. Upper bounds and exact algorithms for p-dispersion problems. Computers and
Operations Research, 2004. to appear.
[45] Jean-Yves Potvin and Jean-Marc Rousseau. A parallel route building algorithm for the vehicle routing
and scheduling problem with time windows. European Journal of Operational Research, 66:331–340,
1993.
[46] Christian Prins. A simple and effective evolutionary algorithm for the vehicle routing problem. Com-
puters and Operations Research, 31:1985–2002, 2004.
[47] M. Reimann, K. Doerner, and R.F. Hartl. D-ants: Savings based ants divide and conquer the vrp.
Computers and Operations Research, 31:563–591, 2004.
[48] Jacques Renaud, Gilbert Laporte, and Fayez F. Boctor. A tabu search heuristic for the multi-depot
vehicle routing problem. Computers and Operations Research, 23(3):229–235, 1996.
[49] Stefan Ropke and David Pisinger. An adaptive large neighborhood search heuristic for the pickup and
delivery problem with time windows, 2004. Submitted to Transportation Science.
[50] Stefan Ropke and David Pisinger. A unified heuristic for a large class of vehicle routing problems
with backhauls. European Journal of Operational Research, 2004. to appear.
[51] D. Sariklis and S. Powell. A heuristic method for the open vehicle routing problem. Journal of the
Operational Research Society, 51:564–573, 2000.
[52] Gerhard Schrimpf, Johannes Schneider, Hermann Stamm-Wilbrandt, and Gunter Dueck. Record
breaking optimization results using the ruin and recreate principle. Journal of Computational Physics,
159(2):139–171, 2000.
[53] Paul Shaw. Using constraint programming and local search methods to solve vehicle routing prob-
lems. In CP-98 (Fourth International Conference on Principles and Practice of Constraint Program-
ming), volume 1520 of Lecture Notes in Computer Science, pages 417–431, 1998.
[54] Mikkel Sigurd, David Pisinger, and Michael Sig. The pickup and delivery problem with time windows
and precedences. Transportation Science, 38:197–209, 2004.
[55] Marius M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35(2):254–265, 1987.
[56] C.D. Tarantilis and C.T. Kiranoudis. Boneroute: an adaptive memory-based method for effective fleet
management. Annals of Operations Research, 115:227–241, 2002.
[57] Paulo Toth and Daniele Vigo. The granular tabu search and its application to the vehicle-routing
problem. INFORMS Journal on Computing, 15(4):333–346, 2003.
[58] Michael A. Trick. A linear relaxation heuristic for the generalized assigment problem. Naval Research
Logistics, 39:137–152, 1992.
[59] Jr. William A. Dees and Patrick G. Karger. Automated rip-up and reroute techniques. In Proceedings
of the 19th conference on Design automation, pages 432–439. IEEE Press, 1982.

29
 Best known ALNS 25K ALNS 50K
n t type cost ref avg. best avg. avg. avg. best avg. avg.
sol. sol. gap time sol. sol. gap time
(%) (s) (%) (s)
P01 50 4 C 576.87 CGW 576.87 576.87 0.00 14 576.87 576.87 0.00 29
P02 50 4 C 473.53 RLB 473.53 473.53 0.00 14 473.53 473.53 0.00 28
P03 75 2 C 641.19 CGW 641.19 641.19 0.00 32 641.19 641.19 0.00 64
P04 100 2 C 1001.59 CGL 1008.49 1001.59 0.74 42 1006.09 1001.04 0.50 88
P05 100 2 C 750.03 CGL 753.04 751.86 0.40 58 752.34 751.26 0.31 120
P06 100 3 C 876.5 RLB 884.36 880.42 0.90 47 883.01 876.70 0.74 93
P07 100 4 C 885.8 CGL 889.14 881.97 0.81 43 889.36 881.97 0.84 88
P08 249 2 CD 4437.68 CGL 4426.86 4387.38 0.90 166 4421.03 4390.80 0.77 333
P09 249 3 CD 3900.22 CGL 3902.18 3874.75 0.74 182 3892.50 3873.64 0.49 361
P10 249 4 CD 3663.02 CGL 3676.93 3655.18 0.74 180 3666.85 3650.04 0.46 363
P11 249 5 CD 3554.18 CGL 3592.82 3552.27 1.32 174 3573.23 3546.06 0.77 357
P12 80 2 C 1318.95 RLB 1319.70 1318.95 0.06 38 1319.13 1318.95 0.01 75
P13 80 2 CD 1318.95 RLB 1321.10 1318.95 0.16 30 1318.95 1318.95 0.00 60
P14 80 2 CD 1360.12 CGL 1360.12 1360.12 0.00 29 1360.12 1360.12 0.00 58
P15 160 4 C 2505.42 CGL 2517.96 2505.42 0.50 125 2519.64 2505.42 0.57 253
P16 160 4 CD 2572.23 RLB 2577.28 2572.23 0.20 92 2573.95 2572.23 0.07 188
P17 160 4 CD 2709.09 CGL 2709.65 2709.09 0.02 90 2709.09 2709.09 0.00 179
P18 240 6 C 3702.85 CGL 3751.85 3727.58 1.32 209 3736.53 3702.85 0.91 419
P19 240 6 CD 3827.06 RLB 3846.35 3839.36 0.50 158 3838.76 3827.06 0.31 315
P20 240 6 CD 4058.07 CGL 4065.32 4058.07 0.18 151 4064.76 4058.07 0.16 300
P21 360 9 C 5474.84 CGL 5576.82 5519.47 1.86 293 5501.58 5474.84 0.49 582
P22 360 9 CD 5702.16 CGL 5731.10 5714.46 0.51 228 5722.19 5702.16 0.35 462
P23 360 9 CD 6095.46 CGL 6107.84 6078.75 0.48 223 6092.66 6078.75 0.23 443
PR01 48 4 CD 861.32 CGL 861.32 861.32 0.00 16 861.32 861.32 0.00 30
PR02 96 4 CD 1307.61 CGL 1311.54 1307.34 0.32 52 1308.17 1307.34 0.06 103
PR03 144 4 CD 1806.6 CGL 1810.90 1806.53 0.24 106 1810.66 1806.60 0.23 214
PR04 192 4 CD 2072.52 CGL 2080.55 2066.64 0.95 146 2073.16 2060.93 0.59 296
PR05 240 4 CD 2385.77 CGL 2352.59 2341.65 0.63 188 2350.31 2337.84 0.53 372
PR06 288 4 CD 2723.27 CGL 2695.15 2685.35 0.36 232 2695.74 2687.60 0.39 465
PR07 72 6 CD 1089.56 CGL 1089.56 1089.56 0.00 29 1089.56 1089.56 0.00 58
PR08 144 6 CD 1666.6 CGL 1677.31 1665.80 0.75 105 1675.74 1664.85 0.65 207
PR09 216 6 CD 2153.1 CGL 2148.85 2136.42 0.58 173 2144.84 2136.42 0.39 350
PR10 288 6 CD 2921.85 CGL 2913.34 2889.49 0.83 228 2905.43 2889.82 0.55 455
Tot. 80394 80651.59 80249.57 3894 80448.26 80133.89 7809
Avg. 0.52 118 0.34 237
< PB 14 14
#B 18 20 27

Table 9: Multi depot vehicle routing problems. The leftmost column shows the problem name, while the
rest of the table is divided into three major columns that display the previously best known results, and the
results obtained by the ALNS-25K and ALNS-50K configurations. The sub columns should be interpreted
like this: n — number of customers, t — number of depots, type — the type of the instance (C indicates that
the instance is capacity constrained, while D indicates that route duration constraints are present), cost —
the cost of the previously best known solution (the cost is calculated as the total distance traveled), ref —
where the solution first was reported. The following abbreviations are used: CGW — Chao et al. [12], RLB
— Renaud et al. [48], CGL — Cordeau et al. [17]. The last 10 instances were introduced by Cordeau et al.
[17] and the two other heuristics have not been applied to these instances. The Columns best sol. and avg.
sol. show the cost of the best solution and the average cost of the solutions obtained during 10 experiments.
avg. gap shows how far the average solution cost is from the best known solution. avg. time shows how
much time the heuristic spends to perform one experiment. The rows Tot. and Avg. sums and averages
key columns. <PB indicates for how many instances an improved solution was found and #B displays the
number of best known solutions obtained. Entries written in bold indicate best known solutions.

30
 Best known ALNS 25K ALNS 50K
n t cost ref avg. best avg. avg. avg. best avg. avg.
sol. sol. gap time sol. sol. gap time
(%) (s) (%) (s)
P01 50 3 642.66 CL 645.04 640.32 0.74 10 642.93 640.32 0.41 20
P02 50 2 598.1 CL 599.40 598.10 0.22 10 598.82 598.10 0.12 19
P03 75 3 959.36 CL 962.36 958.14 0.56 20 963.14 957.04 0.64 40
P04 75 2 854.43 CL 858.05 854.43 0.42 18 856.22 854.43 0.21 36
P05 100 3 1020.22 CL 1012.46 1007.51 0.89 34 1009.08 1003.57 0.55 68
P06 100 2 1036.02 CL 1034.09 1028.70 0.54 35 1032.67 1028.52 0.40 69
P07 27 3 391.3 CGW 391.30 391.30 0.00 4 391.30 391.30 0.00 8
P08 54 3 664.46 CGW 664.46 664.46 0.00 12 664.46 664.46 0.00 24
P09 81 3 948.23 CGW 958.69 948.23 1.10 24 961.36 948.23 1.38 47
P10 108 3 1223.88 CL 1229.42 1218.75 0.88 38 1225.28 1218.75 0.54 76
P11 135 3 1464.98 CL 1488.28 1468.38 1.70 58 1475.85 1463.33 0.86 116
P12 162 3 1695.67 CL 1697.98 1690.56 1.17 78 1689.62 1678.40 0.67 157
P13 54 3 1196.73 CL 1194.40 1194.18 0.02 12 1194.91 1194.18 0.06 24
P14 108 3 1962.66 CL 1961.11 1960.62 0.02 36 1960.83 1960.62 0.01 72
P15 162 3 2751.45 CL 2712.10 2695.22 1.01 77 2701.61 2685.09 0.61 152
P16 216 3 3491.18 CL 3421.74 3402.94 0.75 109 3411.50 3396.36 0.45 213
P17 270 3 4230.96 CL 4109.62 4084.92 0.60 146 4114.26 4085.61 0.72 291
P18 324 3 4929.71 CL 4821.55 4775.35 1.39 177 4795.31 4755.50 0.84 346
P19 100 3 850.39 CL 852.09 846.35 0.71 43 848.54 846.07 0.29 85
P20 150 3 1046.14 CL 1048.75 1042.21 1.74 83 1042.10 1030.78 1.10 168
P21 199 3 1337.83 CL 1281.58 1272.41 0.77 110 1283.03 1271.75 0.89 217
P22 120 3 1012.17 CL 1010.30 1008.78 0.16 65 1008.81 1008.71 0.01 130
P23 100 3 818.75 CL 807.67 803.29 0.55 37 807.00 803.29 0.46 73
PR01 48 4 1384.15 CL 1387.37 1380.77 0.48 10 1393.85 1380.77 0.95 19
PR02 96 4 2320.97 CL 2311.54 2311.54 0.00 32 2330.60 2311.54 0.82 63
PR03 144 4 2623.31 CL 2608.31 2590.01 0.71 71 2607.66 2602.13 0.68 140
PR04 192 4 3500.79 CL 3510.26 3481.44 1.04 98 3489.51 3474.01 0.45 191
PR05 240 4 4479.34 CL 4430.28 4382.65 1.09 123 4431.16 4416.38 1.11 251
PR06 288 4 4546.79 CL 4475.52 4452.93 0.70 159 4465.18 4444.52 0.47 314
PR07 72 6 1955.11 CL 1926.52 1889.82 1.94 19 1916.50 1889.82 1.41 39
PR08 144 6 3082.32 CL 3001.88 2976.76 0.84 66 3007.99 2977.50 1.05 135
PR09 216 6 3664.22 CL 3581.58 3548.22 1.28 113 3567.15 3536.20 0.88 226
PR10 288 6 4739.43 CL 4675.65 4646.96 0.62 162 4673.67 4648.76 0.57 322
PR11 1008 4 13227.96 CL 12987.58 12888.47 2.11 433 12810.71 12719.65 0.72 847
PR12 720 6 9621.99 CL 9510.37 9437.14 1.30 332 9437.56 9388.07 0.53 658
Tot. 90274 89169.30 88541.88 2853 88810.17 88273.77 5658
Avg. 0.80 81 0.60 162
< PB 29 30
#B 5 18 30

Table 10: Site dependent vehicle routing problems. The table should be interpreted like Table 9. Column t
shows the number of vehicle types. CL refers to the the heuristic by Cordeau and Laporte [18] and CGW
refers to the heuristic by Chao et al [11]. The ALNS heuristic was applied 10 times for each problem.

31
 Christofides Golden et al. Li et al.
Heuristic CPU % Minutes % Minutes % Minutes
TV P-200MHz 0.64 3.84 2.88 17.55 - -
LGV Athlon 1Ghz - - 1.05 - 1.20 3.16
CGLM P4-2GHz 0.56 24.62 1.46 56.11 - -
EOS P3-733MHz 0.24 30.95 3.77 137.95 - -
P P3-1GHz 0.24 5.19 0.92 66.90 - -
TK P2-400MHz 0.23 5.22 - - - -
MB Best P4-2GHz 0.03 7.72 0.01 72.94 - -
MB Fast P4-2GHz 0.07 0.27 0.94 0.63 - -
BB P-400MHz 0.49 21.25 - - - -
RDH P-900Mhz - - 0.67 49.33 - -
ALNS 25K Best of 10 P4-3GHz 0.15 9.33 0.67 53.00 0.88 243.17
ALNS 25K Avg. P4-3GHz 0.39 0.93 1.25 5.30 2.40 24.32
ALNS 50K Best of 10 P4-3GHz 0.11 17.50 0.49 107.67 0.50 497.90
ALNS 50K Avg. P4-3GHz 0.31 1.75 1.02 10.77 1.90 49.79
14 instances 20 instances 12 instances
50-200 customers 240-483 customers 560-1200 customers

Table 11: Capacitated vehicle routing problems. The table compares the ALNS heuristic to nine heuristics
proposed in the literature recently. The first column indicates the heuristic considered. TV — granular tabu
search by Toth and Vigo [57], LGV — variable-length neighbor list record-to-record travel heuristic by Li
et al. [40], CGLM — unified tabu search by Cordeau et al. [17, 19], EOS — very large scale neighborhood
search by Ergun et al [22], P — evolutionary algorithm by Prins [46], TK — bone route heuristic by
Tarantilis and Kiranoudis [56], MB — AGES heuristic by Mester and Bräysy [42] (two configurations
of this heuristic is included in the table), BB — hybrid genetic algorithm by Berger and Barkaoui [3],
RDH — ants system algorithm by Reimann et al. [47]. The table contains four rows for the ALNS
heuristic. For each of the configurations ALNS-25K and ALNS-50K we report the best solution quality
in ten experiments and the average solution quality (averaged over the same ten experiments). The CPU
column lists the CPU used, P is used as an abbreviation for Pentium. The rest of the table contains three
major columns, one for each dataset. For each of the datasets we report the gap between the solution
obtained by the heuristic and the best known solution and we report the time spend on average by the
heuristic to solve one instance. When reporting solution times for finding the best solution of ten runs, the
time of all runs has been included. The ALNS heuristic is the only heuristic that has been applied to all
datasets, which explains the missing entries. It should be noted that some of the numbers reported in the
table were obtained from the survey by Cordeau et al. [16].

32
 Best known ALNS 25K ALNS 50K
n veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
P01 50 5 408.5 FEL 416.67 5.0 416.06 5 2.00 12 416.45 5.0 416.06 5 1.95 23
P02 75 10 570.6 FEL 570.81 10.0 567.14 10 0.65 36 568.86 10.0 567.14 10 0.30 53
P03 100 8 617 FEL 642.93 8.0 641.76 8 4.20 85 642.32 8.0 641.76 8 4.10 128
P04 150 12 734.5 FEL 734.34 12.0 733.13 12 0.17 179 733.49 12.0 733.13 12 0.05 279
P05 199 16 953.4 B 912.54 16.0 897.93 16 1.84 124 907.03 16.0 896.08 16 1.22 237
P06 50 6 400.6 FEL 412.96 6.0 412.96 6 3.08 20 412.96 6.0 412.96 6 3.08 31
P07 75 10 634.5 B 592.16 10.0 584.15 10 1.54 18 588.72 10.0 583.19 10 0.95 33
P08 100 9 638.2 FEL 646.23 9.0 645.31 9 1.26 73 646.28 9.0 645.16 9 1.27 114
P09 150 13 785.2 B 766.42 13.1 759.35 13 1.13 108 764.32 13.1 757.84 13 0.85 185
P10 199 17 884.6 B 882.33 17.0 875.67 17 0.76 120 878.42 17.0 875.67 17 0.31 224
P11 120 7 683.4 B 682.68 7.0 682.12 7 0.08 73 682.39 7.0 682.12 7 0.04 141
P12 100 10 534.8 FEL 534.81 10.0 534.24 10 0.11 80 534.44 10.0 534.24 10 0.04 118
P13 120 11 943.7 B 911.98 11.0 909.80 11 0.24 61 911.12 11.0 909.80 11 0.15 116
P14 100 11 597.3 B 591.87 11.0 591.87 11 0.00 40 591.89 11.0 591.87 11 0.00 75
F11 71 4 175 FEL 177.00 4.0 177.00 4 1.14 69 177.00 4.0 177.00 4 1.14 104
F12 134 7 778.5 FEL 770.59 7.0 770.17 7 0.06 237 770.31 7.0 770.17 7 0.02 359
Tot. 156 10340 10246.32 156.10 10198.67 156 1336 10225.99 156.10 10194.19 156 2222
Avg. 1.14 83 0.97 139
< PB 11 11
#B 5 8 11

Table 12: Open vehicle routing problem instances. The table should be interpreted like Table 9. The
abbreviations used in the References column are: B - Brandao’s heuristic [6], FEL - the heuristic by Fu et
al. [25]. The column veh. indicates the number of vehicles used in the previous best solution, avg. #veh.
indicates the number of vehicles used on average by the particular ALNS configuration (averaged over ten
experiments). The column best #veh. indicates the number of vehicles used in the best found solution (out
of 10 experiments).

33
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R101 19 1645.79 H (2000) 1650.86 19.0 1650.80 19 0.31 55 1650.80 19.0 1650.80 19 0.30 85
R102 17 1486.12 RT (1995) 1486.89 17.0 1486.12 17 0.05 62 1486.75 17.0 1486.12 17 0.04 94
R103 13 1292.68 LLH (2001) 1294.89 13.0 1292.68 13 0.17 64 1294.04 13.0 1292.68 13 0.11 97
R104 9 1007.24 M (2002) 987.85 9.8 1013.13 9 -1.93 61 987.85 9.8 1013.13 9 -1.93 96
R105 14 1377.11 RT (1995) 1378.77 14.0 1377.11 14 0.12 56 1378.11 14.0 1377.11 14 0.07 85
R106 12 1251.98 M (2002) 1258.40 12.0 1252.03 12 0.51 61 1255.52 12.0 1252.03 12 0.28 92
R107 10 1104.55 S97 (1997) 1118.18 10.0 1113.70 10 1.23 52 1115.19 10.0 1104.76 10 0.96 85
R108 9 960.88 BBB (2001) 969.37 9.0 963.91 9 0.88 40 965.36 9.0 960.88 9 0.47 75
R109 11 1194.73 HG (1999) 1213.09 11.1 1194.73 11 1.54 47 1211.44 11.1 1194.73 11 1.40 77
R110 10 1118.59 M (2002) 1149.56 10.0 1119.14 10 2.77 41 1148.92 10.0 1119.14 10 2.71 71
R111 10 1096.72 RGP (2001?) 1112.14 10.0 1096.74 10 1.41 46 1105.36 10.0 1096.73 10 0.79 78
R112 9 982.14 GTA (1999) 983.16 9.5 1000.60 9 0.10 58 982.62 9.5 1000.60 9 0.05 91
C101 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 29 828.94 10.0 828.94 10 0.00 57
C102 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 59 828.94 10.0 828.94 10 0.00 91
C103 10 828.06 RT (1995) 828.06 10.0 828.06 10 0.00 65 828.06 10.0 828.06 10 0.00 99
C104 10 824.78 RT (1995) 824.78 10.0 824.78 10 0.00 69 824.78 10.0 824.78 10 0.00 105
C105 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 31 828.94 10.0 828.94 10 0.00 59
C106 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 32 828.94 10.0 828.94 10 0.00 62
C107 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 32 828.94 10.0 828.94 10 0.00 62
C108 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 61 828.94 10.0 828.94 10 0.00 93
C109 10 828.94 RT (1995) 828.94 10.0 828.94 10 0.00 64 828.94 10.0 828.94 10 0.00 99
RC101 14 1696.94 TBGGP (1997) 1688.35 14.2 1697.43 14 -0.51 53 1688.17 14.2 1697.43 14 -0.52 80
RC102 12 1554.75 TBGGP (1997) 1547.04 12.1 1554.75 12 -0.50 56 1555.06 12.1 1554.75 12 0.02 84
RC103 11 1261.67 S98 (1998) 1270.78 11.0 1262.02 11 0.72 58 1268.53 11.0 1262.02 11 0.54 90
RC104 10 1135.48 CLM (2000) 1135.80 10.0 1135.52 10 0.03 60 1135.89 10.0 1135.83 10 0.04 92
RC105 13 1629.44 BBB (2001) 1640.18 13.0 1629.44 13 0.66 54 1640.92 13.0 1633.72 13 0.70 83
RC106 11 1424.73 BBB (2001) 1413.07 11.5 1432.12 11 -0.82 49 1411.92 11.5 1432.12 11 -0.90 76
RC107 11 1230.48 S97 (1997) 1232.48 11.0 1230.95 11 0.16 56 1231.65 11.0 1230.54 11 0.09 86
RC108 10 1139.82 TBGGP (1997) 1167.55 10.0 1140.87 10 2.43 41 1152.30 10.0 1139.82 10 1.10 71
R201 4 1252.37 HG (1999) 1253.23 4.0 1253.23 4 0.07 133 1253.23 4.0 1253.23 4 0.07 193
R202 3 1191.7 RGP (2001?) 1229.81 3.0 1195.30 3 3.20 96 1223.62 3.0 1195.30 3 2.68 181
R203 3 939.54 M (2002) 944.64 3.0 939.58 3 0.54 164 943.57 3.0 941.08 3 0.43 256
R204 2 825.52 BVH (2001) 841.48 2.0 833.09 2 1.93 182 843.39 2.0 833.09 2 2.16 346
R205 3 994.42 RGP (2001?) 1018.90 3.0 994.43 3 2.46 97 1010.43 3.0 994.43 3 1.61 186
R206 3 906.14 SSSD (2000) 923.91 3.0 915.27 3 1.96 192 921.07 3.0 906.14 3 1.65 282
R207 2 893.33 BVH (2001) 928.28 2.0 893.33 2 3.91 180 927.62 2.0 893.33 2 3.84 332
R208 2 726.75 M (2002) 736.12 2.0 726.82 2 1.29 185 735.76 2.0 726.82 2 1.24 369
R209 3 909.16 H (2000) 926.72 3.0 914.45 3 1.93 101 923.48 3.0 914.13 3 1.58 185
R210 3 939.34 M (2002) 955.02 3.0 954.12 3 1.67 112 955.29 3.0 950.52 3 1.70 204
R211 2 892.71 BVH (2001) 889.99 2.3 925.03 2 -0.30 216 887.93 2.3 926.83 2 -0.54 349
C201 3 591.56 RT (1995) 591.56 3.0 591.56 3 0.00 78 591.56 3.0 591.56 3 0.00 147
C202 3 591.56 RT (1995) 591.56 3.0 591.56 3 0.00 88 591.56 3.0 591.56 3 0.00 163
C203 3 591.17 RT (1995) 591.17 3.0 591.17 3 0.00 96 591.17 3.0 591.17 3 0.00 181
C204 3 590.6 RT (1995) 590.60 3.0 590.60 3 0.00 102 590.60 3.0 590.60 3 0.00 189
C205 3 588.88 RT (1995) 588.88 3.0 588.88 3 0.00 81 588.88 3.0 588.88 3 0.00 155
C206 3 588.49 RT (1995) 588.49 3.0 588.49 3 0.00 83 588.49 3.0 588.49 3 0.00 156
C207 3 588.29 RT (1995) 588.29 3.0 588.29 3 0.00 84 588.29 3.0 588.29 3 0.00 167
C208 3 588.32 RT (1995) 588.32 3.0 588.32 3 0.00 85 588.32 3.0 588.32 3 0.00 161
RC201 4 1406.91 M (2002) 1417.80 4.0 1413.52 4 0.77 83 1414.69 4.0 1413.52 4 0.55 140
RC202 3 1367.09 CC (2002) 1405.16 3.0 1368.04 3 2.78 96 1403.60 3.0 1367.09 3 2.67 177
RC203 3 1049.62 CC (2002) 1075.51 3.0 1068.08 3 2.47 100 1072.57 3.0 1068.60 3 2.19 192
RC204 3 798.41 M (2002) 818.00 3.0 799.27 3 2.45 228 806.81 3.0 798.46 3 1.05 320
RC205 4 1297.19 M (2002) 1318.01 4.0 1302.42 4 1.60 134 1312.75 4.0 1302.42 4 1.20 194
RC206 3 1146.32 H (2000) 1155.91 3.0 1146.32 3 0.84 87 1155.16 3.0 1146.32 3 0.77 166
RC207 3 1061.14 BVH (2001) 1095.29 3.0 1070.85 3 3.22 96 1088.15 3.0 1061.84 3 2.55 182
RC208 3 828.14 IKMUY (2001) 834.83 3.0 829.69 3 0.81 109 829.96 3.0 829.69 3 0.22 196
Tot. 405 57192 57641.28 407.50 57360.86 405 4800 57549.75 407.50 57332.03 405 8182
Avg. 0.77 86 0.61 146
< PB 0 0
#B 56 25 28

Table 13: Solomon VRPTW instances. The table should be interpreted as

table 12. The best known solutions were gathered from the web page:
http://www.sintef.no/static/am/opti/projects/top/vrp/benchmarks.html. See this page for complete
references to where the best known solutions first were identified.

34
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R1_2_1 19 5024.65 B 4809.44 20.0 4798.22 20 -4.28 170 4798.77 20.0 4785.96 20 -4.50 257
R1_2_2 18 4054.44 MB 4091.46 18.0 4066.91 18 0.91 165 4079.18 18.0 4059.57 18 0.61 256
R1_2_3 18 3164.41 LC 3414.89 18.0 3387.64 18 7.92 160 3407.48 18.0 3396.47 18 7.68 260
R1_2_4 18 3067.93 MB 3104.90 18.0 3086.11 18 1.20 200 3100.94 18.0 3086.65 18 1.08 308
R1_2_5 18 4112.88 MB 4184.89 18.0 4125.19 18 1.75 147 4157.28 18.0 4125.19 18 1.08 231
R1_2_6 18 3599.84 MB 3643.10 18.0 3616.52 18 1.57 167 3631.74 18.0 3586.80 18 1.25 258
R1_2_7 18 3151.42 MB 3187.56 18.0 3170.98 18 1.15 169 3186.04 18.0 3160.44 18 1.10 271
R1_2_8 18 2963.9 MB 2992.96 18.0 2971.66 18 0.98 200 2989.62 18.0 2975.59 18 0.87 310
R1_2_9 18 3784.33 MB 3853.46 18.0 3802.55 18 1.83 135 3840.07 18.0 3823.15 18 1.47 223
R1_210 18 3307.78 MB 3363.82 18.0 3333.66 18 1.69 148 3336.35 18.0 3312.44 18 0.86 241
C1_2_1 20 2704.57 GH 2704.57 20.0 2704.57 20 0.00 94 2704.57 20.0 2704.57 20 0.00 181
C1_2_2 18 2917.89 BVH 2977.48 18.0 2948.73 18 2.04 152 2969.62 18.0 2943.83 18 1.77 242
C1_2_3 18 2708.08 MB 2744.41 18.0 2719.62 18 1.34 145 2729.39 18.0 2710.21 18 0.79 245
C1_2_4 18 2644.61 MB 2646.94 18.0 2645.60 18 0.09 146 2646.36 18.0 2644.92 18 0.07 253
C1_2_5 20 2702.05 GH 2702.05 20.0 2702.05 20 0.00 96 2702.05 20.0 2702.05 20 0.00 186
C1_2_6 20 2701.04 GH 2701.04 20.0 2701.04 20 0.00 101 2701.04 20.0 2701.04 20 0.00 193
C1_2_7 20 2701.04 GH 2701.04 20.0 2701.04 20 0.00 184 2701.04 20.0 2701.04 20 0.00 281
C1_2_8 18 2769.19 MB 2791.15 19.0 2775.48 19 0.79 157 2789.38 19.0 2775.48 19 0.73 250
C1_2_9 18 2642.82 MB 2705.26 18.0 2687.83 18 2.36 104 2688.82 18.0 2687.83 18 1.74 196
C1_210 18 2649.26 MB 2650.64 18.0 2645.08 18 0.24 117 2651.55 18.0 2644.25 18 0.28 214
RC1_2_1 18 3691.99 MB 3812.41 18.0 3727.17 18 4.52 93 3731.52 18.0 3647.56 18 2.30 175
RC1_2_2 18 3298.68 MB 3342.07 18.0 3269.91 18 2.21 96 3309.57 18.0 3276.88 18 1.21 185
RC1_2_3 18 3025.9 MB 3053.11 18.0 3036.32 18 0.90 104 3051.91 18.0 3034.45 18 0.86 201
RC1_2_4 18 2879.4 MB 2906.27 18.0 2869.74 18 1.27 109 2887.58 18.0 2873.54 18 0.62 215
RC1_2_5 18 3419.81 MB 3509.40 18.0 3463.01 18 2.62 90 3500.46 18.0 3430.03 18 2.36 173
RC1_2_6 18 3393.09 MB 3473.96 18.0 3398.67 18 3.46 91 3431.75 18.0 3357.90 18 2.20 174
RC1_2_7 18 3266.48 MB 3353.23 18.0 3290.65 18 3.71 93 3302.54 18.0 3233.29 18 2.14 179
RC1_2_8 18 3115.82 MB 3163.78 18.0 3147.87 18 1.71 95 3149.37 18.0 3110.46 18 1.25 183
RC1_2_9 18 3083.41 MB 3152.09 18.0 3114.02 18 2.23 94 3150.15 18.0 3116.47 18 2.16 183
RC1_210 18 3038.85 MB 3063.57 18.0 3020.24 18 1.43 98 3056.83 18.0 3042.24 18 1.21 190
R2_2_1 4 4501.8 MB 4340.82 4.5 4563.55 4 -3.58 527 4329.15 4.5 4571.67 4 -3.84 821
R2_2_2 4 3645.38 MB 3683.64 4.0 3666.72 4 1.05 416 3669.25 4.0 3650.54 4 0.65 795
R2_2_3 4 2932.44 MB 2928.17 4.0 2892.07 4 1.25 458 2924.73 4.0 2892.07 4 1.13 890
R2_2_4 4 1981.29 MB 1992.90 4.0 1981.30 4 0.59 482 1989.24 4.0 1981.30 4 0.40 910
R2_2_5 4 3367.55 SAM::OPT 3431.26 4.0 3382.22 4 1.89 385 3417.75 4.0 3377.18 4 1.49 723
R2_2_6 4 2914.56 MB 2957.14 4.0 2929.72 4 1.46 414 2947.20 4.0 2931.14 4 1.12 807
R2_2_7 4 2453.62 MB 2461.82 4.0 2456.71 4 0.33 461 2465.00 4.0 2459.82 4 0.46 883
R2_2_8 4 1849.87 MB 1874.00 4.0 1850.85 4 1.30 495 1866.03 4.0 1849.87 4 0.87 959
R2_2_9 4 3111.41 MB 3134.41 4.0 3113.74 4 0.74 405 3126.66 4.0 3113.74 4 0.49 768
R2_210 4 2657 MB 2696.24 4.0 2666.10 4 1.48 399 2690.93 4.0 2666.35 4 1.28 784
C2_2_1 6 1931.44 GH 1931.44 6.0 1931.44 6 0.00 214 1931.44 6.0 1931.44 6 0.00 391
C2_2_2 6 1863.16 GH 1863.16 6.0 1863.16 6 0.00 233 1863.16 6.0 1863.16 6 0.00 445
C2_2_3 6 1775.11 M 1784.79 6.0 1776.96 6 0.55 263 1783.42 6.0 1776.96 6 0.47 497
C2_2_4 6 1720.09 MB 1719.58 6.0 1713.46 6 0.36 275 1715.66 6.0 1713.46 6 0.13 527
C2_2_5 6 1878.85 BVH 1881.87 6.0 1879.31 6 0.16 224 1879.27 6.0 1878.85 6 0.02 413
C2_2_6 6 1857.35 B 1859.74 6.0 1857.35 6 0.13 224 1857.35 6.0 1857.35 6 0.00 425
C2_2_7 6 1849.46 GH 1851.62 6.0 1849.46 6 0.12 231 1849.46 6.0 1849.46 6 0.00 431
C2_2_8 6 1820.59 MB 1828.56 6.0 1823.88 6 0.44 233 1823.21 6.0 1820.53 6 0.15 442
C2_2_9 6 1830.18 SAM::OPT 1833.78 6.0 1830.05 6 0.20 241 1834.31 6.0 1830.05 6 0.23 449
C2_210 6 1806.6 M 1809.46 6.0 1808.21 6 0.16 249 1809.47 6.0 1808.21 6 0.16 466
RC2_2_1 6 3103.48 MB 3146.36 6.0 3126.03 6 1.38 428 3143.65 6.0 3129.07 6 1.29 635
RC2_2_2 5 2827.45 M 2870.43 5.0 2828.39 5 1.52 629 2856.08 5.0 2835.67 5 1.01 916
RC2_2_3 4 2617.9 MB 2652.00 4.0 2620.87 4 1.49 444 2631.97 4.0 2613.12 4 0.72 849
RC2_2_4 4 2055.97 MB 2080.99 4.0 2056.93 4 1.38 476 2063.32 4.0 2052.74 4 0.52 923
RC2_2_5 4 2912.57 MB 3039.09 4.0 2913.21 4 4.36 433 3041.01 4.0 2912.13 4 4.43 803
RC2_2_6 4 3086.76 LC 2920.37 4.3 2977.41 4 -1.84 525 2900.85 4.3 2975.13 4 -2.50 846
RC2_2_7 4 2550.56 M 2609.93 4.0 2563.90 4 2.76 408 2572.96 4.0 2539.85 4 1.30 789
RC2_2_8 4 2317.8 MB 2341.46 4.0 2322.52 4 1.16 413 2330.66 4.0 2314.61 4 0.69 788
RC2_2_9 4 2175.61 MB 2216.32 4.0 2175.98 4 1.87 417 2214.61 4.0 2180.81 4 1.79 795
RC2_210 4 2015.6 MB 2046.95 4.0 2020.68 4 1.56 424 2030.64 4.0 2015.61 4 0.75 808
Tot. 692 168997 170589.23 694.80 169370.28 694 15341 169941.42 694.80 169042.17 694 27688
Avg. 1.17 256 0.81 461
< PB 8 18
#B 41 14 26

Table 14: Gehring/Homberger VRPTW instances, 200 customers.. The best known solutions were gathered
from the web page: http://www.sintef.no/static/am/opti/projects/top/vrp/benchmarks.html in January 2005.
This list of best known solutions was supplemented by the solutions found by Mester and Bräysy [42] (MB)
and Le Bouthillier and Crainic [?] (LC). See the aforementioned web page for full references. The same
sources were used for the best known solution columns in tables 4 to 7.

35
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R1_4_1 38 11084 B 10557.71 40.0 10502.22 40 -4.75 368 10485.89 40.0 10432.30 40 -5.40 554
R1_4_2 36 9161.26 MB 9277.12 36.0 9239.87 36 1.77 224 9166.43 36.0 9115.68 36 0.56 401
R1_4_3 36 7941.53 MB 8029.69 36.0 7996.33 36 1.11 269 8053.08 36.0 7988.22 36 1.40 464
R1_4_4 36 7332.93 MB 7468.64 36.0 7449.60 36 1.85 227 7441.43 36.0 7415.81 36 1.48 439
R1_4_5 36 9512.25 MB 9738.05 36.0 9588.45 36 2.73 182 9560.46 36.0 9479.10 36 0.86 347
R1_4_6 36 8534.05 MB 8740.69 36.0 8677.13 36 2.42 227 8613.60 36.0 8556.38 36 0.93 407
R1_4_7 36 7710.41 MB 7812.04 36.0 7769.68 36 1.32 245 7763.04 36.0 7725.97 36 0.68 439
R1_4_8 36 7398.68 MB 7468.69 36.0 7425.43 36 1.05 231 7398.80 36.0 7390.76 36 0.11 444
R1_4_9 36 8878.19 MB 9125.58 36.0 9058.30 36 2.79 217 9053.20 36.0 8970.98 36 1.97 386
R1_410 36 8227.49 MB 8417.50 36.0 8386.75 36 2.31 194 8363.10 36.0 8325.16 36 1.65 377
C1_4_1 40 7152.02 M 7152.06 40.0 7152.06 40 0.00 203 7152.06 40.0 7152.06 40 0.00 387
C1_4_2 37 7357.45 MB 7815.71 36.2 7830.99 36 1.06 260 7759.63 36.2 7733.55 36 0.34 437
C1_4_3 36 7151.17 MB 7208.25 36.0 7174.23 36 1.78 212 7104.35 36.0 7082.13 36 0.31 403
C1_4_4 36 6822.18 MB 6909.71 36.0 6833.32 36 1.37 224 6861.13 36.0 6816.17 36 0.66 431
C1_4_5 40 7152.02 M 7152.06 40.0 7152.06 40 0.00 215 7152.06 40.0 7152.06 40 0.00 404
C1_4_6 40 7153.41 M 7153.45 40.0 7153.45 40 0.00 286 7153.45 40.0 7153.45 40 0.00 479
C1_4_7 39 7668.33 LC 7643.60 39.0 7620.09 39 1.28 297 7621.62 39.0 7546.78 39 0.99 485
C1_4_8 38 7113.4 MB 7814.18 37.0 7661.98 37 3.55 284 7794.27 37.0 7546.32 37 3.29 464
C1_4_9 36 7524.32 MB 8042.29 36.0 7673.65 36 6.88 255 7800.59 36.0 7573.18 36 3.67 428
C1_410 36 6907.26 MB 7617.12 36.0 7446.94 36 10.28 214 7325.70 36.0 7145.92 36 6.06 398
RC1_4_1 36 8960.82 MB 9139.22 36.0 9044.65 36 3.70 207 8939.82 36.0 8813.43 36 1.43 371
RC1_4_2 36 8174.27 MB 8287.21 36.0 8181.05 36 2.08 197 8176.96 36.0 8118.43 36 0.72 370
RC1_4_3 36 7737.99 MB 7744.57 36.0 7668.27 36 1.05 214 7729.95 36.0 7663.73 36 0.86 403
RC1_4_4 36 7411.02 MB 7497.41 36.0 7447.70 36 1.75 226 7433.65 36.0 7368.47 36 0.88 436
RC1_4_5 36 8499.15 MB 8634.51 36.0 8503.19 36 2.47 190 8520.69 36.0 8426.57 36 1.12 356
RC1_4_6 36 8304.99 MB 8640.29 36.0 8533.72 36 4.04 185 8445.05 36.0 8390.24 36 1.69 351
RC1_4_7 36 8051.71 MB 8355.82 36.0 8223.65 36 3.78 192 8331.40 36.0 8227.10 36 3.47 360
RC1_4_8 36 7917.68 MB 8174.94 36.0 8135.05 36 3.25 192 8070.47 36.0 7922.67 36 1.93 363
RC1_4_9 36 7890.45 MB 8067.40 36.0 7953.20 36 2.24 194 8016.28 36.0 7987.55 36 1.59 370
RC1_410 36 7716.32 MB 7861.40 36.0 7805.59 36 1.88 199 7823.83 36.0 7774.83 36 1.39 376
R2_4_1 8 9257.92 MB 9513.88 8.0 9375.10 8 2.76 1002 9432.87 8.0 9338.49 8 1.89 1574
R2_4_2 8 7674.9 MB 7762.67 8.0 7728.27 8 1.47 1313 7744.54 8.0 7649.87 8 1.24 1942
R2_4_3 8 5988.02 MB 6078.27 8.0 5998.04 8 1.51 1426 6053.22 8.0 6034.08 8 1.09 2120
R2_4_4 8 4331.07 MB 4356.73 8.0 4326.48 8 0.70 1565 4345.23 8.0 4327.61 8 0.43 2333
R2_4_5 8 7143.55 MB 7305.24 8.0 7255.52 8 2.26 1207 7277.89 8.0 7252.64 8 1.88 1841
R2_4_6 8 6163.81 MB 6284.34 8.0 6222.32 8 1.96 1326 6229.61 8.0 6212.37 8 1.07 1986
R2_4_7 8 5082.1 MB 5182.15 8.0 5138.58 8 1.97 1441 5154.64 8.0 5136.74 8 1.43 2164
R2_4_8 8 4068.97 MB 4090.90 8.0 4055.22 8 0.88 1587 4076.34 8.0 4060.51 8 0.52 2384
R2_4_9 8 6493.13 MB 6565.87 8.0 6526.20 8 1.12 1222 6537.26 8.0 6507.40 8 0.68 1817
R2_410 8 5895.93 MB 5958.31 8.0 5894.40 8 1.08 1283 5919.14 8.0 5897.46 8 0.42 1891
C2_4_1 12 4116.05 M 4125.50 12.0 4116.33 12 0.23 403 4116.93 12.0 4116.33 12 0.02 753
C2_4_2 12 3930.29 MB 3930.22 12.0 3930.05 12 0.00 477 3930.13 12.0 3930.05 12 0.00 858
C2_4_3 12 3739.72 GH 3782.86 12.0 3775.32 12 1.15 525 3780.81 12.0 3775.54 12 1.10 952
C2_4_4 12 3535.99 MB 3549.80 12.0 3546.66 12 0.39 520 3568.37 12.0 3543.60 12 0.92 965
C2_4_5 12 3939.42 MB 3981.35 12.0 3946.94 12 1.06 434 3951.72 12.0 3946.14 12 0.31 783
C2_4_6 12 3875.94 MB 3883.95 12.0 3875.94 12 0.21 457 3921.04 12.0 3875.94 12 1.16 811
C2_4_7 12 3894.13 M 3937.44 12.0 3903.46 12 1.11 453 3960.36 12.0 3894.98 12 1.70 829
C2_4_8 12 3787.08 MB 3863.49 12.0 3804.12 12 2.02 498 3850.01 12.0 3796.00 12 1.66 884
C2_4_9 12 3876.1 MB 4025.46 12.0 3887.00 12 3.85 471 3964.79 12.0 3881.21 12 2.29 851
C2_410 12 3684.89 MB 3764.34 12.0 3706.87 12 2.16 502 3715.36 12.0 3687.13 12 0.83 896
RC2_4_1 11 7019.89 GH 6876.33 11.2 6834.02 11 0.62 786 6857.62 11.2 6840.51 11 0.35 1198
RC2_4_2 10 5924.84 MB 6166.44 9.8 6356.23 9 -2.98 1029 6125.49 9.8 6355.59 9 -3.62 1553
RC2_4_3 8 5114.76 MB 5139.79 8.0 5073.80 8 1.68 820 5109.29 8.0 5055.02 8 1.07 1503
RC2_4_4 8 3648.64 MB 3737.66 8.0 3666.70 8 2.47 959 3692.45 8.0 3647.39 8 1.24 1694
RC2_4_5 9 6063.46 MB 6107.39 9.4 6257.87 9 0.72 901 6019.04 9.4 6119.44 9 -0.73 1416
RC2_4_6 8 6054.21 GH 6093.66 8.0 5997.24 8 1.61 835 6092.17 8.0 6008.41 8 1.58 1425
RC2_4_7 8 5519.25 MB 5664.90 8.0 5529.42 8 3.44 714 5623.09 8.0 5476.57 8 2.68 1324
RC2_4_8 8 4854.16 MB 4949.32 8.0 4877.39 8 1.96 1284 4933.15 8.0 4891.18 8 1.63 1903
RC2_4_9 8 4628.26 MB 4736.64 8.0 4674.88 8 2.94 1168 4662.33 8.0 4601.30 8 1.33 1779
RC2_410 8 4316.36 MB 4415.46 8.0 4400.68 8 2.30 1290 4401.00 8.0 4355.52 8 1.96 1937
Tot. 1386 392070 399377.24 1386.60 395969.66 1385 34732 396157.93 1386.60 393210.00 1385 56699
Avg. 1.80 579 1.05 945
< PB 12 24
#B 35 7 21

Table 15: Gehring/Homberger VRPTW instances, 400 customers..

36
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R1_6_1 59 21131.09 MB 21881.08 59.0 21767.25 59 3.55 514 21743.91 59.0 21677.41 59 2.90 763
R1_6_2 54 19603.7 MB 20892.38 54.0 20719.50 54 6.57 276 20253.42 54.0 20045.49 54 3.31 504
R1_6_3 54 17400.6 MB 18399.70 54.0 18154.60 54 5.74 289 17886.87 54.0 17733.91 54 2.79 535
R1_6_4 54 15993.8 MB 16640.46 54.0 16550.00 54 4.04 300 16459.06 54.0 16374.29 54 2.91 569
R1_6_5 54 20395 MB 22399.39 54.0 22051.85 54 9.83 359 21462.92 54.0 21243.24 54 5.24 577
R1_6_6 54 18620.26 MB 19759.95 54.0 19610.14 54 6.12 263 19206.68 54.0 18948.53 54 3.15 494
R1_6_7 54 17107.91 MB 17915.15 54.0 17773.37 54 4.72 279 17483.82 54.0 17438.28 54 2.20 527
R1_6_8 54 15725.86 MB 16509.06 54.0 16436.50 54 4.98 295 16245.90 54.0 16146.17 54 3.31 560
R1_6_9 54 19372.96 MB 21316.90 54.0 20860.58 54 10.03 276 20548.47 54.0 20375.70 54 6.07 494
R1_610 54 18235.57 MB 19909.33 54.0 19776.64 54 9.18 258 19193.80 54.0 18902.19 54 5.25 485
R2_6_1 11 18325.6 MB 19066.50 11.0 18865.57 11 4.04 872 18937.51 11.0 18837.28 11 3.34 1622
R2_6_2 11 15346.42 MB 15318.18 11.0 15222.07 11 1.65 928 15187.30 11.0 15069.24 11 0.78 1727
R2_6_3 11 11663.06 MB 11422.68 11.0 11395.17 11 1.16 1001 11386.17 11.0 11291.52 11 0.84 1903
R2_6_4 11 8386.64 MB 8331.34 11.0 8264.60 11 2.06 1115 8251.65 11.0 8163.24 11 1.08 2021
R2_6_5 11 15640.6 MB 15637.54 11.0 15430.80 11 1.42 862 15558.66 11.0 15418.00 11 0.91 1621
R2_6_6 11 12937.47 MB 13133.25 11.0 13038.58 11 1.52 920 13026.65 11.0 12936.28 11 0.70 1766
R2_6_7 11 10536.84 MB 10487.56 11.0 10437.39 11 2.12 1000 10352.03 11.0 10269.96 11 0.80 1904
R2_6_8 11 8023.64 MB 7886.24 11.0 7849.32 11 1.72 1095 7805.77 11.0 7752.78 11 0.68 2086
R2_6_9 11 13567.84 MB 14181.45 11.0 14016.38 11 4.52 876 14000.78 11.0 13885.52 11 3.19 1627
R2_610 11 12607.09 MB 12799.15 11.0 12775.18 11 1.83 881 12706.72 11.0 12568.79 11 1.10 1690
C1_6_1 60 14095.64 GH 14095.64 60.0 14095.64 60 0.00 286 14095.64 60.0 14095.64 60 0.00 540
C1_6_2 56 14325.96 MB 14446.21 56.0 14179.06 56 1.92 495 14278.31 56.0 14174.12 56 0.74 737
C1_6_3 56 13898.99 MB 13866.54 56.0 13842.83 56 0.46 509 13842.21 56.0 13803.50 56 0.28 767
C1_6_4 56 13610.66 MB 13626.16 56.0 13615.92 56 0.35 538 13603.40 56.0 13578.66 56 0.18 812
C1_6_5 60 14085.7 BVH 14085.72 60.0 14085.72 60 0.00 306 14085.72 60.0 14085.72 60 0.00 564
C1_6_6 60 14089.7 BVH 14089.66 60.0 14089.66 60 0.00 413 14089.66 60.0 14089.66 60 0.00 674
C1_6_7 59 14659.74 GH 14832.65 58.6 15017.03 58 -1.23 473 14803.08 58.6 15032.51 58 -1.42 726
C1_6_8 57 14976.88 GH 14690.74 57.0 14409.78 57 2.42 433 14510.17 57.0 14343.05 57 1.17 675
C1_6_9 56 13733.56 MB 14265.06 56.0 14017.73 56 3.87 483 13883.26 56.0 13767.45 56 1.09 723
C1_610 56 13758.19 MB 14128.71 56.0 13906.05 56 3.22 492 13788.90 56.0 13688.57 56 0.73 742
C2_6_1 18 7774.1 MB 7789.40 18.0 7780.84 18 0.20 553 7791.82 18.0 7786.86 18 0.23 987
C2_6_2 18 7486.88 MB 7764.29 17.8 8800.94 17 -11.76 727 7763.97 17.8 8799.38 17 -11.77 1224
C2_6_3 17 8371.07 GH 7676.89 17.6 7795.66 17 0.96 762 7613.00 17.6 7604.00 17 0.12 1275
C2_6_4 17 7216.45 MB 7269.90 17.2 7054.65 17 3.95 722 7088.64 17.2 6993.77 17 1.36 1266
C2_6_5 18 7576.35 MB 7694.89 18.0 7592.79 18 1.56 581 7606.34 18.0 7578.12 18 0.40 1007
C2_6_6 18 7478.63 MB 8515.65 18.0 7984.40 18 13.87 635 7910.69 18.0 7554.61 18 5.78 1088
C2_6_7 18 7560.53 MB 8474.41 18.0 7520.34 18 12.69 727 8234.69 18.0 7610.04 18 9.50 1190
C2_6_8 18 7352.42 MB 7771.07 17.8 8696.15 17 -10.64 672 7734.91 17.8 8782.31 17 -11.05 1159
C2_6_9 18 7350.94 MB 7609.44 18.0 7356.19 18 3.52 669 7384.14 18.0 7364.93 18 0.45 1148
C2_610 17 7523.34 MB 7781.30 17.6 8334.99 17 3.43 656 7697.89 17.6 7938.94 17 2.32 1136
RC1_6_1 55 17454.39 MB 18210.19 55.0 17987.59 55 4.33 275 17928.76 55.0 17751.33 55 2.72 494
RC1_6_2 55 16208.24 MB 16883.37 55.0 16718.63 55 4.17 436 16686.63 55.0 16548.43 55 2.95 671
RC1_6_3 55 15524.33 MB 15968.19 55.0 15907.78 55 3.03 333 15642.26 55.0 15499.02 55 0.92 584
RC1_6_4 55 15180.72 MB 15295.11 55.0 15214.81 55 1.47 358 15192.70 55.0 15072.90 55 0.79 621
RC1_6_5 55 17468.57 MB 17981.80 55.0 17879.49 55 3.34 329 17543.75 55.0 17401.34 55 0.82 551
RC1_6_6 55 17248.87 MB 17913.64 55.0 17646.26 55 3.85 329 17466.21 55.0 17355.10 55 1.26 548
RC1_6_7 55 16454.79 MB 17484.20 55.0 17159.31 55 6.26 410 17143.21 55.0 17058.40 55 4.18 636
RC1_6_8 55 16462.49 MB 17043.31 55.0 16955.52 55 3.53 336 16705.09 55.0 16510.65 55 1.47 568
RC1_6_9 55 16153 MB 16806.32 55.0 16609.24 55 4.04 378 16525.18 55.0 16435.71 55 2.30 604
RC1_610 55 16030.86 MB 16483.29 55.0 16388.47 55 2.82 265 16391.34 55.0 16316.51 55 2.25 498
RC2_6_1 15 13275.93 GH 13415.25 15.0 13314.03 15 1.92 1020 13322.77 15.0 13163.03 15 1.21 1573
RC2_6_2 12 12071.4 GH 11652.71 12.8 12039.89 12 -1.70 1250 11539.21 12.8 11853.72 12 -2.65 1970
RC2_6_3 11 9978.25 MB 10220.80 11.0 10032.99 11 3.62 1006 10066.43 11.0 9863.35 11 2.06 1889
RC2_6_4 11 7349.88 MB 7409.35 11.0 7344.31 11 2.46 1069 7274.39 11.0 7231.64 11 0.59 1993
RC2_6_5 13 11919.72 MB 12224.60 12.8 12560.43 12 -2.67 1286 12188.40 12.8 12612.91 12 -2.96 1954
RC2_6_6 12 10700.42 LC 12498.61 11.2 12464.98 11 1.76 1242 12405.28 11.2 12282.52 11 1.00 1963
RC2_6_7 11 11687.04 MB 11510.79 11.0 11347.57 11 4.15 927 11309.89 11.0 11052.49 11 2.33 1706
RC2_6_8 11 10474.95 MB 10744.43 11.0 10627.04 11 2.57 894 10617.44 11.0 10488.75 11 1.36 1658
RC2_6_9 11 10113.82 MB 10094.97 11.0 9982.66 11 2.15 895 10060.58 11.0 9882.71 11 1.80 1661
RC2_610 11 9339.41 MB 9611.45 11.0 9510.51 11 2.91 900 9500.07 11.0 9340.06 11 1.72 1660
Tot. 2076 798645 823814.02 2076.40 818863.38 2071 37731 811014.16 2076.40 807470.21 2071 65718
Avg. 2.83 629 1.28 1095
< PB 22 30
#B 29 6 28

Table 16: Gehring/Homberger VRPTW instances, 600 customers..

37
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R1_8_1 79 39612.2 BVH 37859.00 80.0 37631.40 80 -4.43 684 37756.07 80.0 37492.04 80 -4.69 1010
R1_8_2 72 33548.54 MB 34705.63 72.0 34435.01 72 3.45 613 34273.25 72.0 33816.69 72 2.16 917
R1_8_3 72 30151.9 MB 31065.56 72.0 30746.68 72 3.03 645 30593.64 72.0 30317.49 72 1.47 969
R1_8_4 72 26838.04 MB 29002.34 72.0 28831.80 72 8.06 678 28672.14 72.0 28568.78 72 6.83 1025
R1_8_5 72 34741.53 MB 36198.65 72.0 36038.57 72 4.19 524 35739.41 72.0 35503.63 72 2.87 809
R1_8_6 72 31737.47 MB 32820.77 72.0 32757.13 72 3.41 610 32487.44 72.0 32360.07 72 2.36 913
R1_8_7 72 29538.4 MB 30493.16 72.0 30393.12 72 3.23 644 30089.26 72.0 29979.63 72 1.86 967
R1_8_8 72 28342.64 MB 28803.77 72.0 28622.63 72 1.63 676 28509.27 72.0 28341.21 72 0.59 1020
R1_8_9 72 34231.38 MB 34961.84 72.0 34856.18 72 2.17 576 34437.83 72.0 34218.41 72 0.64 864
R1_810 72 31730.45 MB 33144.45 72.0 32665.95 72 4.46 586 32729.29 72.0 32569.97 72 3.15 879
R2_8_1 15 28440.28 MB 29209.61 15.0 28923.27 15 2.71 951 29086.28 15.0 28822.48 15 2.27 1811
R2_8_2 15 23335.67 MB 23655.16 15.0 23524.65 15 1.64 1070 23492.15 15.0 23274.22 15 0.94 1964
R2_8_3 15 17992.25 MB 18188.06 15.0 18103.52 15 1.09 1127 18137.61 15.0 18078.82 15 0.81 2091
R2_8_4 15 13625.25 MB 13658.48 15.0 13584.57 15 1.82 1213 13525.52 15.0 13413.79 15 0.83 2322
R2_8_5 15 24611.39 MB 25479.70 15.0 25260.54 15 3.53 978 25255.01 15.0 25077.09 15 2.62 1803
R2_8_6 15 20697.06 MB 21104.29 15.0 20969.81 15 1.97 1032 21014.57 15.0 20973.12 15 1.53 1962
R2_8_7 15 17058.3 MB 17114.71 15.0 16977.49 15 0.81 1119 17128.01 15.0 16980.58 15 0.89 2134
R2_8_8 15 13053.31 MB 13187.89 15.0 13054.95 15 1.87 1254 13063.15 15.0 12945.52 15 0.91 2365
R2_8_9 15 22588.02 MB 23303.95 15.0 23138.51 15 3.17 982 23061.61 15.0 22877.21 15 2.10 1849
R2_810 15 21551.26 MB 21372.15 15.0 21240.42 15 1.33 979 21233.28 15.0 21092.27 15 0.67 1841
C1_8_1 80 25030.36 M 25184.38 80.0 25184.38 80 0.62 397 25184.38 80.0 25184.38 80 0.62 741
C1_8_2 75 25518.17 GH 25711.25 74.2 25667.72 74 0.68 664 25634.80 74.2 25536.76 74 0.38 993
C1_8_3 72 25438.6 BVH 25359.87 72.0 24756.97 72 2.96 373 24728.90 72.0 24629.86 72 0.40 682
C1_8_4 72 24040.47 MB 24256.32 72.0 24118.80 72 1.33 378 24005.77 72.0 23938.33 72 0.28 706
C1_8_5 80 25166.3 BVH 25166.28 80.0 25166.28 80 0.00 417 25166.28 80.0 25166.28 80 0.00 762
C1_8_6 80 25160.9 BVH 25162.17 80.0 25160.85 80 0.01 560 25162.21 80.0 25160.85 80 0.01 913
C1_8_7 79 25518.85 GH 25481.02 79.0 25425.92 79 0.22 623 25449.95 79.0 25428.67 79 0.09 972
C1_8_8 76 25379.85 MB 25740.77 75.2 25622.69 75 1.14 608 25538.76 75.2 25450.99 75 0.34 930
C1_8_9 73 24713.38 MB 26318.36 72.2 26169.29 72 2.26 575 25673.55 72.2 25737.46 72 -0.25 868
C1_810 72 29536.81 GH 27097.82 72.0 26382.98 72 5.45 473 26151.75 72.0 25697.68 72 1.77 770
C2_8_1 24 11654.72 MB 11678.08 24.0 11665.21 24 0.20 730 11672.47 24.0 11664.00 24 0.15 1238
C2_8_2 24 11422.34 MB 11456.70 24.0 11428.07 24 0.30 807 11440.98 24.0 11433.46 24 0.16 1397
C2_8_3 23 11554.18 MB 11312.58 24.0 11184.67 24 -2.09 839 11212.69 24.0 11188.30 24 -2.96 1468
C2_8_4 23 10963.49 MB 11511.87 23.2 11440.25 23 5.00 955 11180.00 23.2 10999.42 23 1.97 1627
C2_8_5 24 11432.92 MB 12110.19 24.0 11902.99 24 5.92 896 11565.06 24.0 11451.57 24 1.16 1441
C2_8_6 24 11357.86 MB 12282.80 24.4 12342.70 24 8.14 812 11909.95 24.2 11403.57 24 4.86 1360
C2_8_7 24 11397.54 MB 12058.86 24.6 11540.25 24 5.80 881 11871.66 24.4 11412.08 24 4.16 1443
C2_8_8 24 11206.32 MB 12728.62 23.8 13892.26 23 -8.28 860 12371.35 23.8 13878.40 23 -10.86 1414
C2_8_9 24 11249 MB 13015.41 24.0 12358.05 24 15.70 897 12446.59 24.0 11650.10 24 10.65 1469
C2_810 23 11284.46 MB 11837.70 23.8 12103.56 23 4.90 786 11746.59 23.8 12173.74 23 4.10 1358
RC1_8_1 73 31590.23 MB 31990.65 73.0 31851.54 73 2.29 438 31396.64 73.0 31275.38 73 0.39 720
RC1_8_2 72 39696.2 GH 29762.99 73.0 29537.14 73 -25.02 608 29377.34 73.0 29172.08 73 -25.99 912
RC1_8_3 72 35577.87 GH 28634.08 73.0 28466.83 73 -19.52 646 28301.03 73.0 28164.66 73 -20.45 970
RC1_8_4 72 32654.1 GH 27481.23 73.0 27393.06 73 -15.84 685 27303.22 73.0 27201.39 73 -16.39 1029
RC1_8_5 73 30454.15 MB 31228.63 73.0 31067.35 73 2.54 578 30742.88 73.0 30548.23 73 0.95 865
RC1_8_6 73 29674.68 MB 31019.63 73.0 30863.25 73 4.53 573 30749.36 73.0 30511.07 73 3.62 858
RC1_8_7 72 43829.43 GH 30600.17 73.0 30455.56 73 -30.18 575 30135.52 73.0 30007.82 73 -31.24 868
RC1_8_8 72 43694.6 GH 30006.93 73.0 29820.15 73 -31.33 580 29603.68 73.0 29547.96 73 -32.25 872
RC1_8_9 72 41816.7 GH 29918.07 73.0 29812.35 73 -28.45 581 29493.38 73.0 29360.93 73 -29.47 871
RC1_810 72 41182.44 GH 29518.16 73.0 29373.39 73 -28.32 586 29147.32 73.0 28993.52 73 -29.22 884
RC2_8_1 20 19989.12 MB 20734.14 19.8 21005.11 19 -1.05 1385 20605.53 19.8 20954.95 19 -1.67 2046
RC2_8_2 17 18099.68 MB 18369.12 17.0 18184.31 17 1.86 1728 18208.97 17.0 18032.89 17 0.98 2585
RC2_8_3 15 15116.26 MB 15033.15 15.0 14800.78 15 1.57 1212 14920.27 15.0 14810.81 15 0.81 2172
RC2_8_4 15 11392.25 MB 11592.05 15.0 11402.27 15 1.97 1196 11440.47 15.0 11368.19 15 0.64 2263
RC2_8_5 16 19105.75 MB 19293.34 16.4 19214.57 16 0.98 1529 19181.34 16.4 19180.13 16 0.40 2328
RC2_8_6 15 18882.3 MB 19560.50 15.0 19173.09 15 3.59 1063 19210.08 15.0 19075.89 15 1.74 1918
RC2_8_7 15 17461.44 MB 17798.95 15.0 17519.63 15 2.71 990 17643.28 15.0 17329.32 15 1.81 1858
RC2_8_8 15 16529.24 MB 16756.53 15.0 16485.06 15 3.26 994 16368.61 15.0 16226.78 15 0.87 1859
RC2_8_9 15 15823.5 MB 16071.87 15.0 15979.71 15 2.45 989 15902.97 15.0 15687.20 15 1.38 1825
RC2_810 15 14892.29 MB 15013.68 15.0 14944.14 15 0.82 1001 15048.01 15.0 14953.29 15 1.05 1873
Tot. 2754 1429914 1381184.08 2762.60 1372619.40 2758 48411 1365178.33 2762.20 1358291.43 2758 81648
Avg. -0.86 807 -2.07 1361
< PB 15 25
#B 35 5 22

Table 17: Gehring/Homberger VRPTW instances, 800 customers..

38
 Best known ALNS 25K ALNS 50K
veh. cost References avg. avg. best best avg. avg. avg. avg. best best avg. avg.
sol. #veh. sol. #veh. gap time sol. #veh. sol. #veh. gap time
(%) (s) (%) (s)
R110_1 100 54145.31 MB 55493.78 100.0 55108.89 100 2.49 825 55029.87 100.0 54720.19 100 1.63 1229
R110_2 91 56367.45 GH 54167.93 91.6 57478.64 91 -2.27 698 52844.31 91.6 55428.79 91 -4.66 1066
R110_3 91 46621.19 MB 52196.91 91.0 51840.30 91 11.96 435 50296.23 91.0 49634.84 91 7.88 807
R110_4 91 43461.84 MB 46878.36 91.0 46645.90 91 7.86 435 45626.47 91.0 45303.47 91 4.98 829
R110_5 91 70838.01 GH 54671.65 92.0 54270.08 92 -22.82 705 53259.01 92.0 53089.15 92 -24.82 1061
R110_6 91 49059.8 MB 54109.35 91.4 55826.83 91 10.29 541 52485.80 91.4 54555.32 91 6.98 905
R110_7 91 45847.84 MB 50656.58 91.0 49880.51 91 10.49 429 48869.74 91.0 48141.47 91 6.59 801
R110_8 91 42767.77 MB 46752.56 91.0 46512.13 91 9.32 452 45286.98 91.0 44853.70 91 5.89 847
R110_9 91 51391.8 MB 53216.59 92.0 53163.89 92 3.55 706 52139.44 92.0 52015.72 92 1.45 1067
R11010 91 49348.36 MB 50861.54 92.0 50592.40 92 3.07 674 50007.62 92.0 49769.85 92 1.34 1038
R210_1 19 42922.56 BSJ 44524.99 19.0 44213.65 19 3.73 1192 43904.40 19.0 43264.68 19 2.29 2083
R210_2 19 34918.49 BSJ 34969.52 19.0 34698.44 19 1.60 1800 34564.64 19.0 34417.47 19 0.43 2795
R210_3 19 25689.62 BSJ 26067.93 19.0 25964.09 19 2.63 2088 25807.15 19.0 25400.16 19 1.60 3230
R210_4 19 18858.24 BSJ 18594.33 19.0 18425.77 19 1.43 2289 18477.00 19.0 18332.77 19 0.79 3480
R210_5 19 37265.32 BSJ 38149.38 19.0 37773.72 19 2.37 1328 37833.57 19.0 37746.01 19 1.52 2247
R210_6 19 30725.2 BSJ 31253.33 19.0 30975.00 19 1.72 1453 31007.89 19.0 30778.85 19 0.92 2500
R210_7 19 24363.83 BSJ 24340.48 19.0 24243.39 19 1.45 1923 24228.74 19.0 23991.71 19 0.99 3009
R210_8 19 18185.38 BSJ 18361.52 19.0 18139.74 19 2.90 2313 18037.86 19.0 17844.36 19 1.08 3540
R210_9 19 33777.76 BSJ 35005.84 19.0 34872.05 19 3.64 1345 34496.05 19.0 34349.70 19 2.13 2265
R21010 19 31599.84 BSJ 32006.08 19.0 31782.57 19 1.29 1645 31803.51 19.0 31682.52 19 0.64 2586
C110_1 100 42478.95 GH 42478.95 100.0 42478.95 100 0.00 499 42478.95 100.0 42478.95 100 0.00 915
C110_2 92 42920.7 BVH 42339.69 91.6 42667.84 91 0.21 798 42222.18 91.6 42249.60 91 -0.06 1188
C110_3 90 40934.87 MB 41395.50 90.0 40915.89 90 2.52 506 40904.59 90.0 40376.43 90 1.31 884
C110_4 90 40410.58 MB 40681.78 90.0 40441.12 90 1.76 515 40222.27 90.0 39980.07 90 0.61 902
C110_5 100 42469.2 BVH 42469.50 100.0 42469.18 100 0.00 542 42469.18 100.0 42469.18 100 0.00 968
C110_6 100 42471.3 BVH 42472.69 100.0 42471.29 100 0.00 678 42471.57 100.0 42471.29 100 0.00 1103
C110_7 99 42711.39 GH 42726.27 99.0 42673.51 99 0.12 739 42708.94 99.0 42688.64 99 0.08 1159
C110_8 96 42170.31 MB 42641.48 95.4 42402.12 95 0.67 757 42539.98 95.4 42359.27 95 0.43 1150
C110_9 91 45386.93 GH 42048.67 91.2 41586.54 91 1.37 645 41774.68 91.2 41482.00 91 0.71 1005
C11010 90 40894.38 MB 43409.67 90.0 43132.22 90 6.15 612 42554.17 90.0 42214.60 90 4.06 962
C210_1 30 16879.24 LL 16905.00 30.0 16879.24 30 0.15 888 16893.15 30.0 16879.24 30 0.08 1514
C210_2 29 17228.82 MB 17446.99 29.4 17677.61 29 1.27 1066 17314.77 29.4 17563.06 29 0.50 1719
C210_3 29 16367.59 MB 16938.59 30.0 16253.60 30 3.49 971 16446.58 30.0 16109.71 30 0.48 1690
C210_4 29 17153.19 MB 16845.74 29.0 16712.08 29 5.21 1151 16063.32 29.0 16011.30 29 0.32 1905
C210_5 30 16586.46 GH 17613.87 30.6 16825.34 30 6.19 964 16888.66 30.4 16596.69 30 1.82 1575
C210_6 30 16371.65 MB 17393.97 30.4 17596.06 30 6.26 1070 16696.06 30.2 16369.10 30 2.00 1697
C210_7 31 16578.42 MB 17348.99 31.0 16878.12 31 4.65 978 17057.54 31.0 16590.48 31 2.89 1617
C210_8 29 17219.59 LC 18921.39 29.6 19122.58 29 9.88 1047 17790.97 29.6 18407.27 29 3.32 1700
C210_9 30 16651.96 MB 17626.12 30.0 16679.15 30 8.17 1104 16999.89 30.0 16294.72 30 4.33 1771
C21010 29 16178.26 MB 18856.35 29.0 18447.85 29 16.55 1103 18375.30 29.0 17582.15 29 13.58 1759
RC110_1 90 47143.9 MB 51246.49 90.0 50976.00 90 8.70 517 49693.36 90.0 48933.68 90 5.41 863
RC110_2 90 44906.58 MB 47283.88 90.0 46913.77 90 5.29 539 46647.41 90.0 46165.33 90 3.88 904
RC110_3 90 43782.57 MB 45167.52 90.0 44833.81 90 3.16 562 44408.40 90.0 44014.81 90 1.43 938
RC110_4 90 41917.14 MB 43355.81 90.0 43144.87 90 3.43 668 42844.52 90.0 42607.34 90 2.21 1071
RC110_5 90 47632.31 MB 50533.91 90.0 50226.31 90 6.09 431 49082.31 90.0 48934.53 90 3.04 772
RC110_6 90 46391.6 MB 50436.65 90.0 49703.43 90 8.72 402 49131.04 90.0 48766.98 90 5.91 745
RC110_7 90 46157.71 MB 49716.92 90.0 49238.95 90 7.71 460 48308.95 90.0 48005.94 90 4.66 806
RC110_8 90 45585.08 MB 48391.77 90.0 47670.50 90 6.16 396 47416.90 90.0 47122.61 90 4.02 743
RC110_9 90 45405.54 MB 48343.65 90.0 47930.01 90 6.47 513 46998.60 90.0 46889.79 90 3.51 864
RC11010 90 45041.64 MB 47210.76 90.0 46716.69 90 4.82 466 46284.90 90.0 46080.51 90 2.76 822
RC210_1 22 30320.41 BSJ 30930.47 21.2 30478.44 21 1.76 1429 30618.08 21.2 30396.13 21 0.73 2316
RC210_2 19 26592.4 BSJ 26301.14 19.4 27552.05 18 -4.54 1955 26412.31 19.4 27681.62 18 -4.14 2953
RC210_3 18 20588.38 BSJ 21313.73 18.0 20983.66 18 3.52 1324 21060.93 18.0 20811.18 18 2.30 2443
RC210_4 18 16480.17 BSJ 16617.79 18.0 16254.55 18 3.81 1345 16499.16 18.0 16007.59 18 3.07 2544
RC210_5 18 29352.08 LC 29008.22 18.0 28647.57 18 2.26 1249 28610.45 18.0 28368.48 18 0.85 2198
RC210_6 18 27003.3 MB 29267.17 18.0 28825.98 18 8.38 1136 29005.97 18.0 28746.61 18 7.42 2035
RC210_7 18 26161.91 BSJ 27503.47 18.0 27110.84 18 5.13 1106 26958.52 18.0 26765.43 18 3.04 2067
RC210_8 18 24995 BSJ 25445.17 18.0 25211.63 18 1.94 1103 25128.20 18.0 24961.29 18 0.67 2097
RC210_9 18 23582.89 MB 24729.65 18.0 24420.99 18 4.86 1129 24417.63 18.0 24113.72 18 3.54 2079
RC21010 18 22481.03 BSJ 23544.52 18.0 23193.63 18 4.73 1101 23143.27 18.0 23056.75 18 2.95 2066
Tot. 3438 2099741 2157188.54 3443.80 2146751.97 3438 57741 2119549.93 3443.40 2110924.81 3438 95895
Avg. 3.73 962 1.89 1598
< PB 16 22
#B 38 6 22

Table 18: Gehring/Homberger VRPTW instances, 1000 customers..

39
 Optimal ALNS 25K
cost ref avg. best avg. avg.
sol. sol. gap time
(%) (s)
R101***25 617.1 KDMSS99 617.1 617.1 0.00 3
R102***25 547.1 KDMSS99 547.1 547.1 0.00 3
R103***25 454.6 KDMSS99 454.6 454.6 0.00 4
R104***25 416.9 KDMSS99 416.9 416.9 0.00 4
R105***25 530.5 KDMSS99 530.5 530.5 0.00 3
R106***25 465.4 KDMSS99 465.4 465.4 0.00 3
R107***25 424.3 KDMSS99 424.3 424.3 0.00 4
R108***25 397.3 KDMSS99 397.3 397.3 0.00 4
R109***25 441.3 KDMSS99 441.3 441.3 0.00 3
R110***25 444.1 KDMSS99 444.1 444.1 0.00 4
R111***25 428.8 KDMSS99 428.8 428.8 0.00 4
R112***25 393 KDMSS99 393.0 393.0 0.00 4
C101***25 191.3 KDMSS99 191.3 191.3 0.00 4
C102***25 190.3 KDMSS99 190.3 190.3 0.00 4
C103***25 190.3 KDMSS99 190.3 190.3 0.00 4
C104***25 186.9 KDMSS99 186.9 186.9 0.00 4
C105***25 191.3 KDMSS99 191.3 191.3 0.00 4
C106***25 191.3 KDMSS99 191.3 191.3 0.00 4
C107***25 191.3 KDMSS99 191.3 191.3 0.00 4
C108***25 191.3 KDMSS99 191.3 191.3 0.00 5
C109***25 191.3 KDMSS99 191.3 191.3 0.00 4
RC101***25 461.1 KDMSS99 461.1 461.1 0.00 4
RC102***25 351.8 KDMSS99 351.8 351.8 0.00 4
RC103***25 332.8 KDMSS99 332.8 332.8 0.00 4
RC104***25 306.6 KDMSS99 306.6 306.6 0.00 4
RC105***25 411.3 KDMSS99 411.3 411.3 0.00 4
RC106***25 345.5 KDMSS99 345.5 345.5 0.00 4
RC107***25 298.3 KDMSS99 298.3 298.3 0.00 4
RC108***25 294.5 KDMSS99 294.5 294.5 0.00 4
R201***25 463.3 L99 463.3 463.3 0.00 4
R202***25 410.5 L99 410.5 410.5 0.00 4
R203***25 391.4 L99 391.4 391.4 0.00 4
R204***25 355 C03 355.2 355.0 0.06 6
R205***25 393 L99 393.0 393.0 0.00 5
R206***25 374.4 CR99 374.4 374.4 0.00 5
R207***25 361.6 KLM01 361.6 361.6 0.00 5
R208***25 328.2 FDGG04 328.2 328.2 0.00 11
R209***25 370.7 KLM01 371.5 370.7 0.21 6
R210***25 404.6 CR99 404.6 404.6 0.00 5
R211***25 350.9 KLM01 350.9 350.9 0.00 6
C201***25 214.7 L99 214.7 214.7 0.00 7
C202***25 214.7 L99 214.7 214.7 0.00 8
C203***25 214.7 L99 214.7 214.7 0.00 8
C204***25 213.1 CR99 214.4 213.1 0.59 8
C205***25 214.7 L99 214.7 214.7 0.00 8
C206***25 214.7 L99 214.7 214.7 0.00 7
C207***25 214.5 L99 214.5 214.5 0.00 9
C208***25 214.5 L99 214.5 214.5 0.00 7
RC201***25 360.2 L99 360.2 360.2 0.00 4
RC202***25 338 CR99 338.0 338.0 0.00 4
RC203***25 326.9 FDGG04 326.9 326.9 0.00 4
RC204***25 299.7 FDGG04 299.7 299.7 0.00 5
RC205***25 338 L99 338.0 338.0 0.00 4
RC206***25 324 KLM01 324.0 324.0 0.00 4
RC207***25 298.3 KLM01 298.3 298.3 0.00 5
RC208***25 269.1 C03 269.1 269.1 0.00 6
Tot. 18551.0 18553.2 18551.0 276
Avg. 0.02 5
< PB 0
#B 56 56

Table 19: Solomon VRPTW instances with 25 customers, comparison to exact solutions (distances and
travel times are truncated to one decimal and traveled distance is minimized). The table should be read as
the preceding tables. The abbreviations in the ref column refers to the following papers: C03 – Chabrier
[10], CR99 – Cook and Rich [14], DP03 – Danna and Le Pape [20], FDGG04 – Feillet [24], IV03 – Irnich
and Villeneuve [34], KLM01 – Kallehauge et al. [35], KDMSS99 – Kohl et al. [36] and L99 – Larsen [39].

40
 Optimal ALNS 25K
cost ref avg. best avg. avg.
sol. sol. gap time
(%) (s)
R101***50 1044 KDMSS99 1044.0 1044.0 0.00 9
R102***50 909 KDMSS99 909.0 909.0 0.00 10
R103***50 772.9 KDMSS99 772.9 772.9 0.00 10
R104***50 625.4 KDMSS99 626.1 625.4 0.12 11
R105***50 899.3 KDMSS99 899.8 899.3 0.05 9
R106***50 793 KDMSS99 793.0 793.0 0.00 10
R107***50 711.1 KDMSS99 711.1 711.1 0.00 10
R108***50 617.7 CR99 617.7 617.7 0.00 11
R109***50 786.8 KDMSS99 786.8 786.8 0.00 10
R110***50 697 KDMSS99 697.0 697.0 0.00 10
R111***50 707.2 L99 707.2 707.2 0.00 10
R112***50 630.2 L99 635.1 635.0 0.77 11
C101***50 362.4 KDMSS99 362.4 362.4 0.00 9
C102***50 361.4 KDMSS99 361.4 361.4 0.00 11
C103***50 361.4 KDMSS99 361.4 361.4 0.00 11
C104***50 358 KDMSS99 358.0 358.0 0.00 12
C105***50 362.4 KDMSS99 362.4 362.4 0.00 10
C106***50 362.4 KDMSS99 362.4 362.4 0.00 10
C107***50 362.4 KDMSS99 362.4 362.4 0.00 10
C108***50 362.4 KDMSS99 362.4 362.4 0.00 11
C109***50 362.4 KDMSS99 362.4 362.4 0.00 12
RC101***50 944 KDMSS99 944.0 944.0 0.00 9
RC102***50 822.5 KDMSS99 822.8 822.5 0.04 10
RC103***50 710.9 KDMSS99 710.9 710.9 0.00 10
RC104***50 545.8 KDMSS99 545.8 545.8 0.00 10
RC105***50 855.3 KDMSS99 855.3 855.3 0.00 10
RC106***50 723.2 KDMSS99 723.2 723.2 0.00 9
RC107***50 642.7 KDMSS99 643.7 642.7 0.16 10
RC108***50 598.1 KDMSS99 598.1 598.1 0.00 10
R201***50 791.9 L99 795.8 791.9 0.49 13
R202***50 698.5 L99 698.5 698.5 0.00 14
R203***50 605.3 C03 608.2 605.9 0.49 15
R204***50 506.4 IV03 506.4 506.4 0.00 24
R205***50 690.1 C03 698.2 696.7 1.17 15
R206***50 632.4 C03 634.0 632.4 0.25 16
R207***50 - - 576.1 576.1 0.01 22
R208***50 - - 489.6 487.7 0.39 29
R209***50 600.6 C03 602.5 600.6 0.32 15
R210***50 645.6 C03 648.3 645.6 0.42 16
R211***50 535.5 IV03 549.8 543.3 2.67 25
C201***50 360.2 L99 360.2 360.2 0.00 25
C202***50 360.2 CR99 360.2 360.2 0.00 27
C203***50 359.8 CR99 359.8 359.8 0.00 27
C204***50 350.1 KLM01 350.1 350.1 0.00 29
C205***50 359.8 CR99 359.8 359.8 0.00 30
C206***50 359.8 CR99 359.8 359.8 0.00 26
C207***50 359.6 CR99 359.6 359.6 0.00 27
C208***50 350.5 CR99 350.5 350.5 0.00 28
RC201***50 684.4 L99 684.8 684.8 0.06 12
RC202***50 613.6 FDGG04 613.6 613.6 0.00 12
RC203***50 555.3 C03 555.3 555.3 0.00 15
RC204***50 444.2 DP03 444.2 444.2 0.00 19
RC205***50 630.2 FDGG04 630.2 630.2 0.00 12
RC206***50 610 FDGG04 610.3 610.0 0.05 13
RC207***50 558.6 FDGG04 558.6 558.6 0.00 16
RC208***50 - - 497.9 481.8 3.33 24
Tot. 32560.9 32519.7 841
Avg. 0.19 15

Table 20: Solomon VRPTW instances with 50 customers, comparison to exact solutions .

41
 Optimal ALNS 25K
cost ref avg. best avg. avg.
sol. sol. gap time
(%) (s)
R101 1637.7 KDMSS99 1638.6 1637.7 0.05 30
R102 1466.6 KDMSS99 1467.7 1467.6 0.08 33
R103 1208.7 CR99 1208.9 1208.7 0.01 34
R104 971.5 IV03 977.1 976.0 0.58 34
R105 1355.3 KDMSS99 1355.8 1355.3 0.03 31
R106 1234.6 L99 1234.6 1234.6 0.00 33
R107 1064.6 L99 1068.2 1064.6 0.34 33
R108 - - 943.5 933.7 1.05 36
R109 1146.9 CR99 1150.2 1146.9 0.29 31
R110 1068 CR99 1083.1 1075.6 1.41 33
R111 1048.7 CR99 1049.2 1048.7 0.05 33
R112 - - 952.2 948.6 0.38 35
C101 827.3 KDMSS99 827.3 827.3 0.00 29
C102 827.3 KDMSS99 827.3 827.3 0.00 32
C103 826.3 KDMSS99 826.3 826.3 0.00 34
C104 822.9 KDMSS99 822.9 822.9 0.00 36
C105 827.3 KDMSS99 827.3 827.3 0.00 30
C106 827.3 KDMSS99 827.3 827.3 0.00 31
C107 827.3 KDMSS99 827.3 827.3 0.00 31
C108 827.3 KDMSS99 827.3 827.3 0.00 32
C109 827.3 KDMSS99 827.3 827.3 0.00 34
RC101 1619.8 KDMSS99 1629.8 1619.8 0.61 28
RC102 1457.4 CR99 1475.1 1463.5 1.22 30
RC103 1258 CR99 1272.2 1267.0 1.13 31
RC104 - - 1132.8 1132.6 0.01 33
RC105 1513.7 KDMSS99 1514.2 1513.8 0.04 30
RC106 - - 1376.1 1373.9 0.16 29
RC107 1207.8 IV03 1213.0 1209.3 0.43 30
RC108 1114.2 IV03 1124.6 1114.2 0.94 31
R201 1143.2 KLM01 1153.9 1148.5 0.94 45
R202 - - 1041.0 1036.9 0.40 54
R203 - - 876.5 872.4 0.47 60
R204 - - 731.5 731.3 0.03 67
R205 - - 952.4 949.8 0.27 58
R206 - - 880.6 880.6 0.00 61
R207 - - 796.4 794.0 0.30 72
R208 - - 703.1 701.2 0.27 86
R209 - - 860.2 855.8 0.52 60
R210 - - 914.0 908.4 0.61 59
R211 - - 758.3 752.3 0.80 67
C201 589.1 CR99 589.1 589.1 0.00 69
C202 589.1 CR99 589.1 589.1 0.00 74
C203 588.7 KLM01 588.7 588.7 0.00 80
C204 588.1 IV03 588.1 588.1 0.00 84
C205 586.4 CR99 586.4 586.4 0.00 76
C206 586 CR99 586.0 586.0 0.00 72
C207 585.8 CR99 585.8 585.8 0.00 74
C208 585.8 KLM01 585.8 585.8 0.00 74
RC201 1261.8 KLM01 1272.3 1262.6 0.84 42
RC202 1092.3 C03 1097.4 1095.8 0.47 46
RC203 - - 937.6 923.7 1.50 56
RC204 - - 788.1 785.8 0.29 68
RC205 1154 C03 1154.0 1154.0 0.00 45
RC206 - - 1062.5 1051.1 1.08 52
RC207 - - 976.2 966.6 0.99 55
RC208 - - 790.5 777.3 1.70 65
Tot. 54752.7 54579.5 2649
Avg. 0.36 47

Table 21: Solomon VRPTW instances with 100 customers, comparison to exact solutions .

42
 Best known ALNS 25K ALNS 50K
n type cost avg. best avg. best avg. avg. best avg. best avg.
sol. sol. gap above time sol. sol. gap above time
(%) B.K (s) (%) B.K (s)
P01 50 C 524.61 524.61 524.61 0.00 0.00 12 524.61 524.61 0.00 0.00 21
P02 75 C 835.26 841.81 838.87 0.78 0.43 20 839.62 835.26 0.52 0.00 38
P03 100 C 826.14 828.18 826.14 0.25 0.00 46 826.99 826.14 0.10 0.00 85
P04 150 C 1028.42 1037.43 1031.23 0.88 0.27 96 1034.20 1029.56 0.56 0.11 176
P05 199 C 1291.29 1309.36 1298.92 1.40 0.59 124 1306.63 1297.12 1.19 0.45 233
P06 50 CD 555.43 555.43 555.43 0.00 0.00 12 555.43 555.43 0.00 0.00 21
P07 75 CD 909.68 913.03 909.68 0.37 0.00 19 911.78 909.68 0.23 0.00 36
P08 100 CD 865.94 867.65 865.94 0.20 0.00 42 866.97 865.94 0.12 0.00 78
P09 150 CD 1162.55 1169.06 1164.24 0.56 0.15 86 1167.68 1163.68 0.44 0.10 160
P10 199 CD 1395.85 1408.19 1404.17 0.88 0.60 116 1410.27 1405.88 1.03 0.72 219
P11 120 C 1042.11 1042.37 1042.12 0.03 0.00 73 1042.46 1042.12 0.03 0.00 132
P12 100 C 819.56 819.56 819.56 0.00 0.00 43 819.56 819.56 0.00 0.00 79
P13 120 CD 1541.14 1543.77 1542.86 0.17 0.11 61 1543.54 1542.86 0.16 0.11 113
P14 150 CD 866.37 866.37 866.37 0.00 0.00 40 866.37 866.37 0.00 0.00 73
Tot. 13664 13726.83 13690.13 789 13716.09 13684.21 1464
Avg. 0.39 0.15 56 0.31 0.11 105
< PB 0 0
#B 14 7 8

Table 22: Christofides et al. CVRP problems [13]. The column type indicates if the problem is capacity
constrained (C) or both capacity and duration constrained (CD). The column best above B.K indicates how
much the best solution found differs from the best known solution from the literature (in percent). The best
known solutions where obtained from Cordeau et al. [16].

Best known ALNS 25K ALNS 50K

n type cost ref avg. best avg. best avg. avg. best avg. best avg.
sol. sol. gap above time sol. sol. gap above time
(%) B.K (s) (%) B.K (s)
KELLY01 240 C 5627.54 MB 5667.04 5660.88 0.70 0.59 193 5662.57 5650.91 0.62 0.42 393
KELLY02 320 C 8447.92 MB 8499.27 8478.73 0.61 0.36 321 8487.94 8469.32 0.47 0.25 672
KELLY03 400 C 11036.22 MB 11067.48 11045.81 0.28 0.09 482 11052.72 11047.01 0.15 0.10 1015
KELLY04 480 C 13624.52 MB 13752.13 13635.31 0.94 0.08 654 13748.50 13635.31 0.91 0.08 1328
KELLY05 200 C 6460.98 TK 6479.80 6478.09 0.29 0.26 306 6482.49 6466.68 0.33 0.09 629
KELLY06 280 C 8412.8 MB 8507.29 8415.67 1.12 0.03 418 8543.30 8416.13 1.55 0.04 876
KELLY07 360 C 10195.56 MB 10273.80 10231.34 0.90 0.35 464 10265.15 10181.75 0.82 -0.14 941
KELLY08 440 C 11663.55 MB 11804.08 11721.35 1.20 0.50 499 11766.07 11713.62 0.88 0.43 1011
KELLY09 255 CD 583.39 MB 591.75 584.48 1.43 0.19 225 590.33 585.14 1.19 0.30 437
KELLY10 323 CD 742.03 MB 753.48 749.47 1.54 1.00 305 751.36 748.89 1.26 0.92 616
KELLY11 399 CD 918.45 MB 933.42 926.63 1.63 0.89 371 926.57 922.70 0.88 0.46 761
KELLY12 483 CD 1107.19 MB 1129.53 1125.11 2.02 1.62 458 1125.22 1119.06 1.63 1.07 911
KELLY13 252 CD 859.11 MB 878.22 876.01 2.22 1.97 142 874.24 864.68 1.76 0.65 285
KELLY14 320 CD 1081.31 MB 1107.97 1096.92 2.47 1.44 194 1103.53 1095.40 2.06 1.30 393
KELLY15 396 CD 1345.23 MB 1370.94 1355.91 1.91 0.79 236 1366.23 1359.94 1.56 1.09 468
KELLY16 480 CD 1622.69 MB 1652.00 1639.81 1.81 1.05 279 1645.67 1639.11 1.42 1.01 549
KELLY17 240 CD 707.79 MB 713.76 710.36 0.84 0.36 153 710.59 708.90 0.39 0.16 304
KELLY18 300 CD 998.73 MB 1008.11 1003.20 0.94 0.45 196 1007.84 1002.42 0.91 0.37 387
KELLY19 360 CD 1366.86 MB 1380.49 1377.52 1.00 0.78 227 1377.88 1374.24 0.81 0.54 449
KELLY20 420 CD 1821.15 MB 1843.08 1828.35 1.20 0.40 245 1834.70 1830.80 0.74 0.53 488
Tot. 88623 89413.66 88940.91 6369 89322.91 88832.02 12914
Avg. 1.25 0.66 318 1.02 0.48 646
< PB 0 1
#B 19 0 1

Table 23: Golden et al. CVRP problems ([30]). The best known solutions where obtained from Cordeau et
al. [16], MB refers to the heuristic by Mester and Bräysy [42] (the results are not given in [42], but can be
found in [16]), TK refers to the heuristic by Tarantilis and Kiranoudis [56].

43
 Best known ALNS 25K ALNS 50K
n cost ref avg. best avg. best avg. avg. best avg. best avg.
sol. sol. gap above time sol. sol. gap above time
(%) B.K (s) (%) B.K (s)
CVRP_L_21 560 16212.83 EST 16488.67 16296.21 1.70 0.51 869 16391.23 16224.81 1.10 0.07 1735
CVRP_L_22 600 14641.64 ORTR 14737.97 14638.37 0.73 -0.02 569 14644.06 14631.08 0.09 -0.07 1168
CVRP_L_23 640 18801.13 EST 19155.50 18925.36 1.88 0.66 1097 19112.56 18837.49 1.66 0.19 2268
CVRP_L_24 720 21389.43 EST 22024.22 21652.78 2.97 1.23 1259 21913.83 21522.48 2.45 0.62 2739
CVRP_L_25 760 17053.26 EST 17170.49 17082.81 1.59 0.17 650 17115.78 16902.16 1.26 -0.89 1320
CVRP_L_26 800 23977.74 EST 24577.43 24084.92 2.50 0.45 1425 24405.05 24014.09 1.78 0.15 3081
CVRP_L_27 840 17651.6 ORTR 17833.67 17749.35 1.25 0.55 723 17769.75 17613.22 0.89 -0.22 1504
CVRP_L_28 880 26566.04 EST 27315.94 26651.15 2.82 0.32 1692 27172.63 26791.72 2.28 0.85 3441
CVRP_L_29 960 29154.34 EST 30117.04 29487.26 3.30 1.14 1887 29976.86 29405.60 2.82 0.86 3921
CVRP_L_30 1040 31742.64 EST 32828.86 32133.28 3.42 1.23 2192 32607.06 31968.33 2.72 0.71 4348
CVRP_L_31 1120 34330.94 EST 35617.70 34962.16 3.75 1.84 2395 35472.51 34770.34 3.33 1.28 5003
CVRP_L_32 1200 36919.24 EST 37989.05 37401.49 2.90 1.31 2750 37818.65 37377.35 2.44 1.24 5321
Tot. 288441 295856.55 291065.13 17509 294399.98 290058.65 35849
Avg. 2.40 0.78 1459 1.90 0.40 2987
< PB 1 3
#B 9 0 3

Table 24: Li et al. CVRP problems [40]. EST refers to a solution found by hand by Li et al [40] (the
instances are highly symmetrical which makes it easy to construct good solutions by hand). ORTR refers
to a solution found by a heuristic by Li et al. [40].

44