Vehicle routing problems with multiple trips

Diego Cattaruzza, Nabil Absi, Dominique Feillet

To cite this version:

Diego Cattaruzza, Nabil Absi, Dominique Feillet. Vehicle routing problems with multiple trips.
4OR: A Quarterly Journal of Operations Research, Springer Verlag, 2016, 14 (3), pp.223-259.
�10.1007/s10288-016-0306-2�. �emse-01250603�

HAL Id: emse-01250603

https://hal-emse.ccsd.cnrs.fr/emse-01250603
Submitted on 3 Feb 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Vehicle routing problems with multiple trips
Diego Cattaruzza1 , Nabil Absi2 , and Dominique Feillet3
1
Univ. Lille, CNRS, Centrale Lille, UMR 9189 - CRIStAL - Centre de Recherche
en Informatique Signal et Automatique de Lille, F-59000 Lille, France , Tel.:
+33-3-20335494
2,3
Ecole des Mines de Saint-Etienne and LIMOS UMR CNRS 6158, CMP Georges
Charpak, F-13541 Gardanne, France

Abstract
This paper presents a survey on the Multi-Trip Vehicle Routing Problem (MTVRP) and on
related routing problems where vehicles are allowed to perform multiple trips. The first part of
the paper focuses on the MTVRP. It gives an unified view on mathematical formulations and
surveys exact and heuristic approaches. The paper continues with variants of the MTVRP and
other families of routing problems where multiple trips are sometimes allowed. For the latter, it
specially insists on the motivations for having multiple trips and the algorithmic consequences.
The expected contribution of the survey is to give a comprehensive overview on a structural
property of routing problems that has seen a strongly growing interest in the last few years and
that has been investigated in very different areas of the routing literature.

1 Introduction
Since the introduction of the Vehicle Routing Problem (VRP) by Dantzig and Ramser in 1959 [40], the
vast majority of papers interested in vehicle route optimization share the same assumption (largely
realistic in practice) that each vehicle is restricted to perform at most one trip. The first visible
attempt to address a vehicle routing problem with multiple trips dates back from Fleischmann [44]
under the name Vehicle Routing Problem with Multiple Use of Vehicles, in the context of the solution
of a series of distribution problems involving a heterogeneous fleet of vehicles and time windows (see
Section 3).
Despite the clear improvements that could be obtained, in some situations, by allowing vehicles
to perform multiple trips, this feature remained very punctually investigated for years. However,
a significant increase of the number of publications dealing with this subject can be observed in
the recent literature, as shown in Figure 1. One important reason for this gain of interest is the
development of new distribution schemes in cities. Nuisances related to congestion and pollution
pushed scholars, communities and enterprises to study new delivery policies to increase city livability.
Common approaches envisage the use of electrical vehicles and/or forbid heavy trucks to enter city
centers. Moreover, physical city structures often force final deliveries to be accomplished by small-
sized vans that can go through narrow streets.
The usage of electrical vehicles or small-sized vans ends up in delivery trips shorter than the
working day due to, respectively, limited autonomy and capacity. Unless multiple trips are allowed
for vehicles, a consequence is the bad exploitation of the time horizon and the need of an oversized
fleet to satisfy all the customers (Cattaruzza et al. [22]). Operations, then, assume the possibility to

1
 10

number of papers
6

0
1990 1995 2000 2005 2010 2015
publication year

Figure 1 – Number of publications per year cited in this survey and dealing with multi-trip trans-
portation (PhD theses excluded)

re-load vans when they are back at the depot and route them for another trip. As it will be seen in
the subsequent sections, many other sources of motivation can however be found, in domains as far
to city logistics as maritime transportation.
The aim of this survey is to give a clear overview of the work devoted to vehicle routing problems
with multiple trips, and to highlight how the possibility to link up several trips for a vehicle impacts
solution methods.
An important difficulty to meet our goal was the naming of these problems. The Multi-Trip
Vehicle Routing Problem (MTVRP) appears in the literature under several names. In addition to
the already mentioned VRP with multiple use of vehicles used by Fleischmann [44], it has been
addressed as multitrip VRP (Prins [72]), VRP with multiple routes (Azi et al. [8]), VRP with multiple
trips (Olivera and Viera [66]), VRP with multiple depot returns (Tsirimpas et al. [89]) and multiple
trip VRP (Battarra et al. [12]). Taniguchi and Van Der Heijden [87] allow vehicles to make multiple
traverses, while the multiple usage of vehicles has been called recycling of trucks in Van Buer et
al. [92]. The differences in the problem naming, made the research of papers difficult. Moreover,
several works dealing with multi-trip routing problems do not mention this aspect in the title, nor
in the abstract, making the hunt even harder. For this reason there could probably be some fishes
that escaped our net.
A previous literature review was proposed by Şen and Bülbül [39] but with a very limited scope:
only the pure MTVRP was surveyed. In this paper we address a broader class of problems, whose
limits are precisely defined in the next paragraphs. However, before making clear the class of problems
we consider, we need to clarify some terminology.
All along the paper we refer to a trip as a sequence of customer services preceded and followed
by a visit to a depot and without intermediate stop at a depot. Note that the empty trip is included
in this definition. A sequence of trips performed by the same vehicle is called journey. Note that
the definitions of trip and journey are not limited to the single-depot case, i.e., the starting and the
ending depots can differ if the problem involves several depots. In the literature trip and journey
can, for example, be respectively referred to trip and tour, tour and multi-tour (Aghezzaf et al. [2]),
route and schedule (Mingozzi et al. [62]) or to voyage and route (in maritime context, Section 3.3).
To further clarify, we say that a problem involves the multi-trip aspect, when vehicles have the
possibility to accomplish, during the same working day, journeys made by more than one trip.
Stochastic routing problems where returns to the depot can be implied by recourse strategies
are not considered in this survey. In these problems, anticipated returns to the depot result from
the realization of the uncertainty, for example because of underestimated demands. However, since
these returns are not part of the initial plan, we decided to exclude this category of problems from

2
the survey. Similarly, several papers deal with special routing problems, where the weight of routing
decisions is very limited. An example is given by Ronen [78] that considers multiple trips in Full
Truckload distribution. Another example is found in Tsirimpas et al. [89], where the optimization of
an automated guided vehicle system along fixed paths is addressed. These papers are also excluded
from the survey.
On the contrary, we decided to include multi-trip routing problems where the fleet is composed
of a single vehicle. In the VRP literature the single vehicle case (Traveling Salesman Problem, Gutin
and Punnen [50]) and the multi-vehicle case (VRP, Toth and Vigo [88]) are usually distinguished
because of the strong differences between these problems (one can just look at the largest instances
solved exactly for these two problems to be convinced). When multiple trips are allowed, a single-
vehicle solution has the same characteristics as a VRP solution and we believe that this case thus
deserves to be part of this survey.
The paper is structured as follows. In Section 2 we introduce the MTVRP and survey the related
literature. Section 3 deals with important variants of the MTVRP. The section is decomposed in five
subsections. In Subsection 3.1, we present direct extensions of the MTVRP for which benchmarks
exist and an on-going “horse race” is run. Subsection 3.2 is devoted to multiple trips in the context of
combined routing, production and/or inventory optimization. In Subsection 3.3, we focus on multi-
trip routing in the context of maritime transportation. Subsection 3.4 investigates multi-trip vehicle
routing in the context of multi-level distribution. This section is finally concluded with Subsection 3.5,
where other types of constraints met in multi-trip vehicle routing problems are reviewed. Section 4
concludes the paper.

2 The Multi-Trip Vehicle Routing Problem

As explained above, the Multi-Trip Vehicle Routing Problem was introduced in different papers,
under different names and sometimes with some slight differences in the definition. We propose here
to retain the following definition.
Let G = (N , A) be a directed graph, where N = {0, 1, . . . , N } is the set of nodes and A =
{(i, j)|i, j ∈ N } is the set of arcs. Arcs (i, j) ∈ A are characterized by their travel time Tij . Node 0
represents the depot where a fleet V of identical vehicles with limited capacity Q is available at time
0 and has to be returned at time TH . Nodes 1, . . . , N represent the customers to be served, each one
requiring a certain non-negative quantity Qi of a product. The MTVRP calls for the determination
of a set of trips and an assignment of each trip to a vehicle, such that the traveled time is minimized
and the following conditions are satisfied:

(i) each trip starts and ends at the depot,

(ii) each customer is visited exactly once,

(iii) the sum of the demands of the customers in any trip does not exceed Q,

(iv) the sum of the durations of the trips assigned to the same vehicle (journey) does not exceed
TH (a trip duration being the sum of the travel times on arcs used in the trip).

Note that vehicles are not necessarily all used. In papers on the pure MTVRP, loading times are
not introduced. Service times are sometimes included (e.g., Taillard et al. [85]), sometimes not (e.g.,
Olivera and Viera [66]). In the benchmark instances issued from Taillard et al. [85] (see Subsection
2.2), they are set to zero. For this reason we did not include service times nor loading times in this
definition. Note that one can however consider them being included in the travel time matrix.
In most papers (and in the aforementioned instances), the travel time matrix is symmetric. The
problem can then indifferently be defined on a directed or an undirected graph. Depending on the

3
papers, one of this two options is chosen. To be as general as possible, and following usual trends
when time is part of the problem definition (and induces an implicit orientation of routes), we decided
to define the MTVRP on a directed graph. As far as we can see in the surveyed literature, the issue
of the travel time matrix being symmetric or not is not a major issue.
Olivera and Viera [66] shortly prove that the MTVRP is N P-hard as being a generalization of
the VRP. Quoting Olivera and Viera [66], “any P VRP instance can be transformed to an equivalent
MTVRP instance, setting |V| = N and TH = (i,j)∈A tij ”. Apart from this short proof, we did
not find any discussion on the complexity of the MTVRP in the literature. Implicitly, the proof by
Olivera and Viera [66] assumes that an unlimited fleet of vehicles is available in the VRP. We propose
below a more formal proof, where this assumption is removed.

Proposition 1. The MTVRP is an N P-hard problem

Proof. We prove it by reduction to the VRP. Given a VRP instance, wePduplicate it to construct
a MTVRP instance where: (i) parameter TH is added and is set to 2 × (i,j)∈A Tij and, (ii) travel
times on arcs returning to the depot are increased by T2H . Note that no vehicle can perform more
than one trip in this MTVRP instance. The solution sets of the VRP and MTVRP instances are
identical, with identical cost evaluation functions. Seeing that the VRP is N P-hard (Lenstra and
Rinnooy Kan [58]), it proves that the MTVRP is N P-hard.

2.1 Mathematical formulations for the MTVRP

As far as we know, the only compact formulation proposed for the MTVRP is by Koc and Karaoglan
[55]. However, several very different formulations exist for extensions of the MTVRP, from which
formulations can be deduced for the MTVRP. We review these formulations below. We exclude
extended formulations that are discussed in Section 2.3.
For the sake of consistency, we sometimes slightly modify the models found in the papers (in
addition to removing all the variables and constraints that do not relate to the MTVRP). In these
models, M represents a large enough constant. The objective in this section is not to compare these
formulations (information is lacking) but rather to present an unified view of the literature.

4-index formulations
The most common formulation is an extension of the 3-index vehicle flow formulation (3VFF-VRP)
initially introduced for the VRP, where a fourth index is added for trips. In this formulation a pair
composed of a trip and the vehicle that performs this trip exactly plays the role played by a vehicle
in 3VFF-VRP. All constraints from the 3VFF-VRP can thus just be translated. An additional
constraint is needed in order to limit the total duration of the journey for each vehicle.
A limit of this formulation is that the number of trips that can be envisaged for each vehicle is
unknown. An easy upper bound is N , but it then conducts to a very weak formulation with a large
number of variables. In some papers where the maximal number of trips per vehicle is bounded, this
limit is withdrawn. We present this 4-index formulation below, adapted from the 3VFF-VRP given
in Chapter 1 of Toth and Vigo [88]. It relies on the following decision variables:
(
1 if trip r ∈ R of vehicle v ∈ V travels through arc (i, j) ∈ A,
xvr
ij =
0 otherwise,
(
1 if trip r ∈ R of vehicle v ∈ V visits vertex i ∈ N ,
yivr =
0 otherwise,

4
 where R = {0, . . . , N − 1}. The model is then as follows:
X XX
min Tij xvr
ij (1)
(i,j)∈A v∈V r∈R

s.t.
XX
yivr = 1 ∀i ∈ N \ {0}, (2)
v∈V r∈R
X X
xvr
ij = xvr vr
ji = yi ∀i ∈ N , v ∈ V, r ∈ R, (3)
j∈N j∈N
X
Qi yivr ≤ Q ∀v ∈ V, r ∈ R, (4)
i∈N \{0}
XX
xvr
ij ≤ |S| − 1 ∀S ⊆ N \ {0}, |S| ≥ 2, v ∈ V, r ∈ R, (5)
i∈S j∈S
X X
Tij xvr
ij ≤ TH ∀v ∈ V, (6)
r∈R (i,j)∈A

xrv
ij ∈ {0, 1} ∀v ∈ V, r ∈ R, (i, j) ∈ A, (7)
yirv ∈ {0, 1} ∀v ∈ V, r ∈ R, i ∈ N . (8)

Constraints (1)-(5) are taken from Toth and Vigo [88]. Constraints (6) limit to TH each vehicle’s
journey. Trip index on xvrij variables is needed because of capacity constraints (4). Vehicle index is
needed because of constraint set (6) and because the fleet size is limited (one cannot simply bound
the outgoing flow from the depot to limit the number of vehicles used, because of the multi-trip
feature). To the best of our knowledge, the first appearance of this formulation dates back from
2006, in Gribkovskaia et al. [49], for a complex livestock collection problem. Since then, several
authors proposed the same type of formulation (e.g., Azi et al. [8], Alonso et al. [3]).
Lei et al. [57] propose a different 4-index formulation extended from the single-commodity flow
formulation of the VRP (Toth and Vigo [88]). In this formulation, flow variables qijvr are added and
represent the load in vehicle v when traversing arc (i, j) during its trip r. Constraints (4)-(5) are
then replaced with:
X X
vr
qji − qijvr = Qi ∀i ∈ N \ {0}, v ∈ V, r ∈ R, (9)
j∈N j∈N
qijvr ≤ Qxvr
ij ∀(i, j) ∈ A, v ∈ V, r ∈ R, (10)
vr
qij ≥ 0 ∀(i, j) ∈ A, v ∈ V, r ∈ R. (11)

Note that these constraints manage both vehicle capacity and subtour elimination.

3-index formulations with vehicle index (without trip index)

Surprisingly, Lei et al. [57] conserved the trip index in the above formulation while it is apparently
not necessary anymore. Actually, Aghezzaf et al. [2] propose an equivalent formulation where the
trip index is removed:
X X
min Tij xvij (12)
(i,j)∈A v∈V

5
s.t.
X
yiv = 1 ∀i ∈ N \ {0}, (13)
v∈V
X X
xvij = xvji = yiv ∀i ∈ N , v ∈ V, (14)
j∈N j∈N
X X
v
qji − qijv = Qi ∀i ∈ N \ {0}, v ∈ V, (15)
j∈N j∈N
qijv ≤ Qxvij ∀(i, j) ∈ A, v ∈ V, (16)
X
Tij xvij ≤ TH ∀v ∈ V, (17)
(i,j)∈A

xvij ∈ {0, 1} ∀v ∈ V, (i, j) ∈ A, (18)

yiv ∈ {0, 1} ∀v ∈ V, i ∈ N , (19)
qijv ≥ 0 ∀(i, j) ∈ A, v ∈ V. (20)

Constraints (12)-(14) and (17) come from (1)-(3) and (6), respectively. Constraints (15)-(16) come
from (9)-(10).
Coming back to vehicle flow formulations, Buhrkal et al. [18] propose a 3-index formulation
extended from the 2-index vehicle flow formulation of the VRP. To get rid of the trip index, capacity
constraints are expressed using Miller-Tucker-Zemlin-like constraints [61]. The resulting formulation
is similar to (12)-(20), replacing (15)-(16),(20) with:

qiv + Qi ≤ qjv + Q(1 − xvij ) ∀i ∈ N \ {0}, j ∈ N , v ∈ V, (21)

0 ≤ qiv ≤ Q ∀i ∈ N , v ∈ V, (22)

where qiv represents the accumulative load in vehicle v before serving customer i. Note that these
constraints also eliminate subtours.
In both formulations, the vehicle index is kept because of constraints (17) and the limited fleet
size.

3-index formulations with trip index (without vehicle index)

Two 3-index formulations have been proposed where the vehicle index is removed instead of the trip
index (Azi et al. [9], Hernandez et al. [54]). As stated above, the difficulty is then to handle the
time-horizon constraint and the size of the fleet.
In Azi et al. [9], the problem addressed involves time windows and time variables tri are present:
tri indicates the time at which trip r visits customer i, if i is visited in r, 0 otherwise; tr0 indicates the
starting time of trip r; trN +1 indicates its ending time. These variables are needed when translating
the model to the MTVRP, even if the time windows are removed. New binary variables zrs (with
r < s) are also introduced to detect consecutive trips: zrs = 1 if and only if trip r is followed by trip
s on the same vehicle. The resulting model is:
X X
min Tij xrij (23)
(i,j)∈A r∈R

6
s.t.
X
yir = 1 ∀i ∈ N \ {0}, (24)
r∈R
X X
xrij = xrji = yir ∀i ∈ N , r ∈ R, (25)
j∈N j∈N
X
Qi yir ≤ Q ∀r ∈ R, (26)
i∈N \{0}

tri + Tij ≤ trj + M (1 − xrij ) ∀i ∈ N , j ∈ N \ {0}, r ∈ R, (27)

tri + Ti0 ≤ trN +1 + M (1 − xri0 ) ∀i ∈ N \ {0}, r ∈ R, (28)
trN +1 ≤ ts0 + M (1 − zrs ) ∀r, s ∈ R, r < s, (29)
trN +1 ≤ TH ∀r ∈ R, (30)
X X
|R| − zrs ≤ |V| , (31)
r∈R s∈R,r<s
X
zrs ≤ 1 ∀s ∈ R, (32)
r∈R,r<s
X
zsr ≤ 1 ∀s ∈ R, (33)
r∈R,s<r
xrij ∈ {0, 1} ∀r ∈ R, (i, j) ∈ A, (34)
yir ∈ {0, 1} ∀r ∈ R, i ∈ N , (35)
tri ≥ 0 ∀r ∈ R, i ∈ N ∪ {N + 1}, (36)
zrs ∈ {0, 1} ∀r, s ∈ R, r < s. (37)

Constraints (23)-(26) correspond to (1)-(4). Time variables are computed with constraints (27)-(29)
The time-horizon constraint is expressed with constraints (30). Constraints (31) are added to limit
the number of vehicles. The left-hand side of the constraint indicates the number of vehicles used,
based on the observation that the number of consecutive trips in a journey is equal to the number of
trips in the journey minus one. Constraints (32) and (33) prevent variables zrs from being artificially
set to one. They are needed to make sure that the left-hand side of Constraint (31) indeed represents
the number of vehicle used. Note that these two constraints are forgotten in Azi et al. [9].
Hernandez et al. [54] propose an alternative for constraints (31)-(33). They replace binary
v v
variables zrs with variables zrs such that zrs = 1 when r is scheduled before s on vehicle v (note
that these variables are defined for all r, s ∈ R and that they are equal to 1 even if r and s are not
consecutive). They introduce binary variables σrv where σrv = 1 indicates that a trip r is assigned to
a vehicle v. Constraints (29), (31)-(33), (37) are replaced with:

trN +1 ≤ ts0 + M (1 − zrs

v
) ∀v ∈ V, r, s ∈ R, r =
6 s (38)
v v
1 − zrs − zsr ≥ 0 ∀v ∈ V, r, s ∈ R, r =6 s, (39)
X
σrv = y0r ∀r ∈ R, (40)
v∈V
σrv + σsv ≤ 1 + zrs
v v
+ zsr ∀v ∈ V, r, s ∈ R, r 6= s, (41)
v
zrs ∈ {0, 1} v ∈ V, r, s ∈ R, r 6= s, (42)
v
σr ∈ {0, 1} ∀v ∈ V, r ∈ R. (43)

Constraints (38) come from (29). Constraints (39) impose an order between trips assigned to the
same vehicle. The assignment of trips to vehicles (variables σrv ) is managed with constraints (40).
Constraints (41) are coupling constraints.

7
2-index formulations (without vehicle index nor trip index)
Finally, two two-index formulations were also proposed by Koc and Karaoglan [55] and Rivera et al.
[76].
In Koc and Karaoglan [55], binary variables x0ij are introduced to detect when a trip finishing
with customer i is followed (on the same vehicle and after a stop at the depot) by a trip visiting
j as first customer (x0ij = 1). The role played by these variables is very similar to the role played
by variables zrs in the 3-index formulation proposed by Azi et al. [9], and so are the constraints
involved.

X
min Tij xij (44)
(i,j)∈A

s.t.
X
xij = 1 ∀i ∈ N \ {0}, (45)
j∈N
X X
xij = xji ∀i ∈ N , (46)
j∈N j∈N
qi + Qi ≤ qj + Q(1 − xij ) ∀i ∈ N \ {0}, j ∈ N , (47)
ti + Tij ≤ tj + M (1 − xij ) ∀i ∈ N , j ∈ N \ {0}, (48)
ti + (Ti0 + T0j ) ≤ tj + M (1 − x0ij ) ∀i, j ∈ N \ {0}, (49)
T0i ≤ ti ≤ TH − Ti0 ∀i ∈ N \ {0}, (50)
X
x0ij ≤ xi0 ∀i ∈ N \ {0}, (51)
j∈N \{0}
X
x0ij ≤ x0j ∀j ∈ N \ {0}, (52)
i∈N \{0}
X X X
x0j − x0ij ≤ |V| , (53)
j∈N \{0} i∈N \{0} j∈N \{0}

xij ∈ {0, 1} ∀(i, j) ∈ A, (54)

x0ij ∈ {0, 1} ∀i, j ∈ N \ {0}, (55)
0 ≤ qi ≤ Q ∀i ∈ N , (56)
ti ≥ 0 ∀i ∈ N \ {0}, (57)

Note that in this formulation yi variables are not introduced anymore, because they would directly
be constrained to 1. Note also that variable t0 is not introduced. The objective function (44)
corresponds to (1), while Constraints (45)-(46) correspond to (2)-(3). Capacity constraints (47),(56)
are equivalent to (21)-(22). Constraints (48) participate to the computation of visiting times and
are equivalent to (27). Constraints (49)-(50) complete the computation of time variables and impose
finishing journeys before the end of the time horizon. Constraints (53) impose the fixation of variables
x0ij to satisfy the fleet size. Constraints (51)-(52) then connect these variables with returns to the
depot.
Rivera et al. [76] propose a formulation based on the same type of variables, but instead of
coming back to the depot between two consecutive trips, vehicles use replenishment arcs that have
been added to the graph. Variables x0ij then represent the flow on these arcs. Constraints (51)-(52)

8
are removed, constraints (45)-(46) are replaced by
X X
xij + x0ij = 1 ∀i ∈ N \ {0}, (58)
j∈N j∈N \{0}
X X
xji + x0ji = 1 ∀i ∈ N \ {0}, (59)
j∈N j∈N \{0}

and constraints (53) are replaced by

X
x0i ≤ |V| . (60)
i∈N \{0}

Note that in addition capacity constraints are managed with a single-commodity flow formulation
(see Constraints (9)-(11)). Constraints (47),(56) are then changed to:
X X
qji − qij = Qi ∀i ∈ N \ {0}, (61)
j∈N j∈N
qij ≤ Qxij ∀i ∈ N \ {0}, j ∈ N , (62)
X
q0j ≤ Q(x0j + x0ij ) ∀j ∈ N \ {0}, (63)
i∈N \{0}

qij ≥ 0 ∀(i, j) ∈ A. (64)

2.2 Instances for the MTVRP

The benchmark of instances for the MTVRP was introduced in Taillard et al. [85] and is constructed
from the instances 1–5 and 11–12 proposed by Christofides et al. [29] (usually denoted CMT1–
CMT5 and CMT11–CMT12) and instances 11–12 proposed by Fisher [43] (F11-F12) for the VRP.
All these instances are Euclidean: customer locations are described with coordinates and travel times
computed accordingly. As a consequence the triangle inequality is satisfied.
From each VRP instance, several instances are constructed with different values for the number h i
∗
of available vehicles |V| and two different values for the time horizons TH , given by TH1 = 1.05z |V|
h ∗i
2 1.1z ∗
and TH = |V| where z is the solution cost of the original VRP instance found by Rochat and
Taillard [77] and [x] represents the closest integer to x. It results in a benchmark set of 104 instances.
Surprisingly, Taillard et al. [85] adopted the following convention when reporting their experi-
ments. First, they changed the problem definition by allowing overtime, penalized by a factor 2.
Then, they only reported solution costs when they were not able to find a feasible solution. When
a feasible solution was found, they just indicated it. Note that in this case, unfeasible solutions can
cost less than feasible solution. Aside from a few exceptions, next researchers followed the same
scheme.
Optimal solutions are known for 42 out of the 104 instances. Feasible solutions have been found
for all remaining instances except for 5. Optimal and best known solution values on this benchmark
set are reported in Appendix A. For the 5 instances for which no feasible solution has been found,
the cost including the penalty for overtime is given.
One can observe from the results given in Appendix A that for all instances but 7 for which the
optimal solution is known (that is 35 out of 42), the value obtained is equal to z ∗ . When |V| = 1 it is
not surprising as it is a direct consequence of the way the instances are constructed: the optimal VRP
solution can be interpreted as a set of trips assigned to the single vehicle involved in the instance.
When, |V| > 1 it tends to indicate that the routes present in the optimal VRP solution can in many
cases be combined to form feasible journeys. It also means that a solution method able to first

9
find the optimal VRP solution and then to find an optimal assignment of the trips obtained to the
vehicles, would certainly obtain many optimal solutions. Actually, next sections will show that it is
a scheme that has been followed by many researchers. A second observation that can be made from
these results, is that when the number of vehicles increases (and the horizon decreases), solution costs
tend to increase and it seems more difficult to find optimal or even feasible solutions. It probably
means that the packing of trips into vehicles is more challenging and that the two components of the
problem (construction of trips and packing of these trips) are more deeply interlinked.
We review in the following subsections the solution algorithms proposed for the MTVRP. We
restrict ourselves to algorithms that present numerical results on the above instances, which is the
case for all papers addressing the pure MTVRP plus a few papers that tackled more general problems.

2.3 Exact methods

The first exact method we are aware is by Koc and Karaoglan [55]. They propose a branch-and-cut
algorithm where cuts are valid inequalities taken from the VRP literature that remain valid for the
MTVRP. They propose preliminary results limited to the 8 instances constructed from CMT1 in the
benchmark set (50 customers) and are able to find 3 optimal solutions.
Mingozzi et al. [62] propose a much more sophisticated exact method based on branch-and-
price. This method extends previous works from the same authors applied to the VRP and variants.
It combines two set partitioning formulations. The first formulation defines columns as trips. The
master problem is then in charge of selecting trips and combining them to form journeys. The second
formulation defines columns as journeys. The master problem then only has to select journeys. These
two formulations are successively used to compute the best possible lower bound. It is interesting to
note that in the first formulation the pricing problem is very similar to what can usually be found
when applying branch-and-price to vehicle routing problems, while it is the master problem for the
second formulation. This algorithm is able to find an optimal solution for 42 instances, the ones
indicated above for which the optimal solution is known.
Note there are some inconsistencies between the branch-and-cut method by Koc and Karaoglan [55]
and the branch-and-price method by Mingozzi et al. [62]. Indeed, Koc and Karaoglan [55] report an
upper bound on instance CMT1, |V| = 4, TH = 144 lower than the optimal solution value reported
later by Mingozzi et al. [62].

2.4 Heuristics
The first heuristic approach for the MTVRP can be found in Taillard et al. [85]. They propose a
two-phase algorithm. In the first phase, several VRP solutions are created and the trips forming these
solution are inserted in a list. In the second phase, MTVRP solutions are constructed with a Bin
Packing Problem (BPP) heuristic where the trips are the objects to be packed into bins with capacity
equal to TH . Once a trip is selected from the list, trips serving the same customers are discarded.
The VRP solutions of the first phase are built using a tabu search algorithm with adaptive memory
(Taillard [84]). The BPP heuristic is a simple greedy algorithm completed with swaps of trips.
Petch and Salhi [70] reproduce this two-phase scheme but completes it with a third phase improv-
ing the MTVRP solutions found. In the first phase, VRP solutions are obtained from two sources:
a parametric savings algorithm (Yellow [94]) and a sweep approach. The second phase is carried out
with a simple BPP heuristic similar to the one used in Taillard et al. [85]. Compared to Taillard et
al. [85], the effort made in these two phases is very limited, which explains (and is explained by) the
third phase.
The authors propose several components to improve the MTVRP solutions in the third phase.
First, they introduce an original local search operator defined as follows:
• a return to the depot is inserted within a trip,

10
 • one of the two resulting trips is moved to another vehicle.

Note that this operator cannot improve the length of the trips (as long as the triangle inequality is
satisfied) but can contribute to a better packing (decreasing overtime). In addition, a standard VRP
local search is applied to each vehicle separately, by addressing the trips forming the journey of a
vehicle as a VRP solution. Finally, this local search is completed with attempts of reallocation of
customers between vehicles.
Other authors abandon this principle of a two-phase algorithm focusing first on VRP solutions.
Instead, they apply and adapt existing metaheuristic schemes to the MTVRP.
Several authors propose tabu search (TS) algorithms. Brandão and Mercer [17] adapt to the
MTVRP a TS algorithm initially designed for a practical problem involving multi-trip vehicle routes
(Brandão and Mercer [17]). Alonso et al. [3] propose a TS approach for a periodic MTVRP. In both
cases, the algorithm uses standard VRP local search operators. Apart from the construction of the
initial solutions, the fact that multiple trips are assigned to vehicles only influences the evaluation of
the solutions with regard to overtime.
Olivera and Viera [66] follow Taillard et al. [85] by using a TS algorithm with adaptive memory,
but applying the approach to MTVRP solutions. As usual in TS with adaptive memory, the memory
is composed of trips. When a set of trips is selected to form a new VRP solution, this solution is
changed to a MTVRP solution with a BPP heuristic. Then, after each move of the TS algorithm,
the BPP heuristic is repeated to search for better assignment of the trips to the vehicles. The
local search operators used in the TS are standard VRP local search operators that move customers
between trips: no operator can change the assignment of trips to vehicles.
Salhi and Petch [80] address the problem with a genetic algorithm, completed with local search.
In this method a chromosome is a sequence of strictly increasing angles, measured with respect to
the depot, and dividing the plane into sectors. Chromosomes are decoded as follows: customers
are assigned to the sector they occupy; in each cluster, the Clarke and Wright [31] savings heuristic
is applied; the resulting VRP solution is finally transformed to a MTVRP solution with a BPP
heuristic. Local search is managed as in Petch and Salhi [70].
Cattaruzza et al. [21] also propose a population-based algorithm. They apply the giant tour
coding popularized by Prins [73]. To decode chromosomes, they first execute the split procedure
initially proposed by Prins [73] for the VRP. They then compute by dynamic programming the best
MTVRP solution that can be obtained with the trips of this solution. Local search is composed of
standard VRP local search operators improved with what the authors call combined local search.
They observe that completing a move with a swap of trips can make beneficial a non-beneficial move.
They propose several rules to apply these combined moves with a limited impact on computing times.
Note that the local search operator introduced by Petch and Salhi [70] can be seen as an example of
combined move.
Figure 2 gives algorithm comparison. The reported CPU times are the original ones: no scale
factor is considered. The machines used to run the algorithms are reported in Table 1. State-of-
the-art algorithms are provided by Olivera and Viera [66], Cattaruzza et al. [21]. Detailed algorithm
comparison can be found in Cattaruzza et al. [21]. The algorithm proposed by Alonso et al. [3] is
proposed to solve a richer problem that considers a multi-period horizon and accessibility constraints.

3 Variants of the MTVRP

This section surveys several important categories of routing problems with multiple trips. The first
category (Section 3.1) consists of direct (academic) extensions of the MTVRP for which benchmarks
exist and are subject to competition between authors. This section contains two other papers (Anaya-
Arenas et al. [4] and Chbichib et al. [25]) that are not involved in this competition but address the
same problems motivated by a real-life application. The second category consists of complex routing,

11
 100
CAFV OV
95
Feasible solution found

AAB Symbol Meaning

90 BM
TLG Taillard et al. [85]
85
TLG
BM Brandão and Mercer [17]
80 PS Petch and Salhi [70]
75 PS SP Salhi and Petch [80]
70
OV Olivera and Viera [66]
AAB Alonso et al. [3]
65
SP CAFV Cattaruzza et al. [21]
60
10 100 1000 10000
CPU times (seconds)

Figure 2 – Algorithm comparison on benchmark instances

Paper Machine RAM

TLG 100 Mhz Silicon Graphics Indingo -
BM HP Vectra XU Pentium Pro 200 Mhz -
PS Ultra Enterprise 450 dual processor 300 Mhz -
SP Ultra Enterprise 450 dual processor 300 Mhz -
OV 1.8 Ghz AMD Athlon XP 2200+ 480 Mb
AAB DELL Dimensio 8200 1.6 Ghz 256 Mb
CAFV Intel Xeon 2.8 Ghz -

Table 1 – Machines used to run the algorithms

12
production and/or inventory optimization problems where multiple trips are allowed for distribution.
The third category is composed of multi-trip routing problems motivated by maritime transportation.
The fourth category comprises multi-trip routing problems appearing in the context of multi-level
distribution.
These four categories are reviewed in the next four subsections. A final subsection reviews other
types of constraints that have punctually been considered in multi-trip routing problems.

3.1 Main academic extensions of the MTVRP

In this section, we are interested in a relatively limited number of papers and problems summarized in
Table 2. We first describe the different extensions investigated and then discuss about mathematical
formulations, available benchmark instances and solution methods. Detailed numerical results on all
the instances tackled by the methods surveyed in this section are provided in Appendices B and C.
Reference Time windows Service-dependent loading time Limited trip duration Profits
Hernandez et al. [53] ? ?
Cattaruzza et al. [20] ? ?
Hernandez et al. [54] ? ? ?
Azi et al. [9] ? ? ? (?)
Macedo et al. [60] ? ? ? (?)
Wang et al. [93] ? ? ? ?
Azi et al. [11] ? ? ? ?
Anaya-Arenas et al. [4] ? ?
Chbichib et al. [25] ?

Table 2 – Main academic extensions of the MTVRP

Time windows (TW) indicates that with each customer i is associated a time interval [Ei , Li ]
during which service should take place. Arriving at customer location earlier than Ei is allowed,
but makes the driver wait until the opening of the time window. On the opposite, late arrival is
forbidden. The time horizon can then equivalently be represented by a time window [E0 , L0 ] = [0, TH ]
associated with the depot. When introducing time windows, authors generally also introduce service
times Si at customers and a loading time S0 at the depot. Also, authors often introduce the notions
of travel cost or travel distance on arcs, both equal to Tij , in order to distinguish clearly between the
objective function and journey duration. The introduction of time windows has several important
consequences. First, the duration of a trip is a function of the time at which it starts. Condition (iv)
introduced when defining the MTVRP thus does not make sense anymore. Each trip is time-stamped
in a solution and journeys are composed of trips that do not overlap. Furthermore local search
becomes more complicated as modifications in a trip impact the rest of the journey. Secondly,
solution schemes cannot exploit anymore a decomposition of the problem in two successive phase: a
routing phase where trips are constructed and a packing phase where they are packed together to
obtain journey. With time windows, trips cannot be properly evaluated as long as their starting time
is not known and the construction of journeys is not anymore a question of packing but of scheduling.
A simple illustration is given by the case where a single vehicle is available and the time horizon is
large: while the MTVRP is equivalent to the VRP (with an unlimited fleet of vehicles), it is unlikely
that a VRPTW solution can be transformed to its multi-trip counterpart.
Service-dependent loading times indicates that the loading time of a vehicle at the depot depends
on the customers visited during the trip. All authors that consider service-dependent loading times
consider a loading time proportional to the total service time of the trip. We note β the coefficient
applied. A difficulty met with service-dependent loading times is the fact that any modification of
a trip can influence the departure time of the vehicle from the depot (this is however not specific to
the multi-trip situation).

13
 Limited trip duration indicates that trip duration is limited by a value Tmax . For most authors,
the duration of a trip is the elapsed time between the departure time of the vehicle from the depot,
after it has been loaded, and the moment the service starts at the last customer in the trip. These
constraints are motivated by the fact that perishable goods must be delivered before a certain amount
of time has passed from the moment they have been loaded. The only exception is Anaya-Arenas et
al. [4] that address biomedical sample collection. The trip length is then the elapsed time between
the arrival at the first customer of the trip and the arrival at the laboratory.
Profits indicates that serving all customers is not mandatory and a profit Pi is associated with
each customer i and represents the profit of serving customer i. In practice, the proposed instances
consider equal profit customers and the objective function consists in minimizing the number of
customers not served, breaking ties in favor of the minimum traveled distance. Parenthesis are
added in Table 2 for Azi et al. [9] and Macedo et al. [60] because in their experiments all customers
are visited most of the times; comparisons with Hernandez et al. [54] (that enforce the visit of all
customers) are thus possible.
All these extensions are trivially N P-hard because they admit the MTVRP as a special case.

Mathematical formulations
Though no compact formulations have been proposed for the variants cited above, one can easily
update the 4-index model introduced in Section 2 from other papers in the literature.
For time windows, we propose to add variables tvr i defined as the time variables added in the
three-index formulation of the MTVRP proposed by Azi et al. [9]: tvr i indicates the time at which
trip r of vehicle v visits customer i, if i is visited in the trip, 0 otherwise; tvr
0 indicates the starting
vr
time of trip r of vehicle v; tN +1 indicates its ending time. Time windows are then added by replacing
Constraints (5)-(6) with the following constraints:
tvr vr vr
i + Si + Tij ≤ tj + M (1 − xij ) ∀i ∈ N , j ∈ N \ {0}, v ∈ V, r ∈ R, (65)
tvr vr vr
i + Si + Ti0 ≤ tN +1 + M (1 − xi0 ) ∀i ∈ N \ {0}, v ∈ V, r ∈ R, (66)
v,r+1
tvr
N +1 ≤ t0 ∀v ∈ V, r ∈ R \ {N − 1}, (67)
Ei yivr ≤ tvr
i ≤ Li y i
vr
∀i ∈ N \ {0}, v ∈ V, r ∈ R, (68)
tvr
N +1 ≤ TH ∀v ∈ V, r ∈ R, (69)
tvr
i ≥ 0 ∀v ∈ V, r ∈ R, i ∈ N ∪ {N + 1}. (70)
Constraints (65)-(67) compute time variables. Constraints (68) guarantee customer time windows.
Constraints (69) impose that trips are finished before the end of the time horizon.
To take the service-dependent loading time into account, variables S0vr are added, where S0vr is
the loading time for trip r of vehicle v. Constraints (65) are then changed to:
tvr vr vr
i + Si + Tij ≤ tj + M (1 − xij )∀i ∈ N \ {0}, j ∈ N \ {0}, v ∈ V, r ∈ R, (71)
tvr
0 + S0vr + T0j ≤ tvr
j + M (1 − xvr
0j )∀j ∈ N \ {0}, v ∈ V, r ∈ R, (72)
and constraints
X
S0vr = β Si yivr ∀v ∈ V, r ∈ R, (73)
i∈N \{0}

S0vr ≥0 ∀v ∈ V, r ∈ R, (74)
are added. Constraints (71)-(72) decompose constraints (65) to deal with the fact that the loading
time at the depot becomes a variable. Constraints (73) compute this variable.
Limited trip duration is obtained adding:
tvr vr vr
i ≤ t0 + S0 + Tmax ∀i ∈ N \ {0}, v ∈ V, r ∈ R, (75)

14
except for Anaya-Arenas et al. [4] where it is instead

tvr vr
N +1 ≤ ti + Tmax ∀i ∈ N \ {0}, v ∈ V, r ∈ R. (76)

Finally, profits can be introduced by changing the objective function (1) to

X XX X XX
max MP Pi yivr − Tij xvr
ij (77)
i∈N \{0} v∈V r∈R (i,j)∈A v∈V r∈R

and changing constraints (2) to

XX
yivr ≤ 1 ∀i ∈ N \ {0}. (78)
v∈V r∈R

where MP is a large constant that always makes beneficial increasing the profit, whatever the impact
on the traveled distance. Constraints (78) relax the obligation to visit customers.

Instances
Experiments in the above papers are all carried out on benchmark instances obtained from Solomon’s
instances [82] and their extension proposed by Gehring and Homberger [45]. Solomon’s instances
with small TW (groups R1, C1 and RC1) have however been excluded as, according to Azi et al. [9],
they are not adapted to journeys with multiple trips. Following current practices for the VRPTW,
smaller instances are generated by considering only the first customers.
Several parameters as the number of vehicles, the capacity of these vehicles or service times have
been modified to increase the chance that vehicles perform multiple trips without obliterating the set
of feasible solutions. For the case of service-dependent loading times, parameter β is always set to
0.2. For the case of limited trip duration, different values have been proposed for Tmax . For the case
of profits, profits are set to 1 for all customers. Customer locations in Gehring and Homberger [45]
instances have been normalized in order to fit the 100×100 square as in Solomon instances.

Exact methods
Hernandez et al. [53] propose a branch-and-price algorithm for the MTVRP with time windows
(MTVRPTW) and service-dependent loading times. They explain that the method developed by
Mingozzi et al. [62] for the MTVRP cannot easily be extended because time windows strongly modify
the structure of the problem. They investigate two formulations. In the first formulation, columns are
journeys; in the second formulation column are trips (with a fixed schedule). The first formulation
does not specially bring difficulties up, except for the size of the pricing problem solution space.
The second formulation necessitates to solve a complex pricing problem, where column reduced
cost depends on their timing. Note that in both cases the authors tackle the difficulty implied by
service-dependent loading times by applying backward strategies when solving the pricing problems.
Experiments show that the second formulation is more efficient. It is able to solve most instances
with 25 customers (25 out of 27) and a few instances with 50 customers (5 out of 27). Values are
reported in Appendix B, Table 4.
Several authors address the special case where trip duration is limited. This feature plus the
presence of TW and the limited vehicle capacity that allows only few customers to be served in each
trip, make the enumeration of all feasible trips achievable for small and medium size instances. The
proposed exact methods are based on this observation. They all proceed in two phases. The first
phase consists in listing all the candidate trips. The second phase aims at constructing the vehicle
routes.

15
 In the first phase, all authors apply the same method, based on labeling. Trips are constructed
by extending labels from the depot in all feasible directions. Dominance rules are applied to delete
trips that do not visit customers following the optimal sequence.
Three different methods are proposed for the second phase. In Azi et al. [9], an auxiliary graph
is constructed, with a vertex and a TW for each trip. Arcs represent possible successions of trips.
Phase 2 is then addressed as a VRPTW with profits and solved with a branch-and-price algorithm.
In Macedo et al. [60], the second phase relies on a time-indexed graph. Nodes correspond to discrete
time instants and arcs either correspond to waiting times at the depot or elapsed time during a given
(time-stamped) trip. Phase 2 can then be addressed as a min-cost flow problem. In Hernandez et
al. [54], a set covering formulation is proposed and solved with branch-and-price. Columns represent
time-stamped trips. The pricing problem simply consists of finding a convenient starting time for
trips.
In addition, Azi et al. [8] address the single vehicle case. They introduce the same auxiliary
graph as in Azi et al. [9]. Because a single vehicle route is searched for, the problem they face is
very similar to the pricing problem of Azi et al. [9]. They apply an equivalent dynamic programming
procedure.
Experiments show that Macedo et al. [60] and Hernandez et al. [54] obtain comparable results
and significantly improve upon Azi et al. [9]. The methods developed in Macedo et al. [60] and Azi et
al. [9] are however more general as they solve the problem with the Profits feature described above.
Optimal solutions are reported in Appendix C, Table 5 when all customers can be visited and Table
6 when the optimal solution only visits a subset of customers.
It can be observed from the results that, as expected, instances with short time limits and small
feasible trip sets are easier to solve. Actually, one could note that none of the above approaches is
specifically designed for problems with limited trip duration: they can be applied as soon as, and
only when, the set of feasible trips is small.

Heuristics
We are not aware of any algorithm developed for the heuristic solution of the MTVRPTW. Cattaruzza
et al. [20] however propose a population-based procedure for the more general MTVRPTW with
Release Dates (see Section 3.5) and evaluate this procedure on the instances used by Hernandez et
al. [53]. The principle of the algorithm is the same as in Cattaruzza et al. [21] (see Section 2), the
main difference being the split procedure used for decoding chromosomes. In order to deal with
time constraints, the assignment of trips to vehicles is managed within the split procedure, which
turns out to search for a resource constrained shortest path. An adaptive heuristic dominance rule
is introduced to limit computing times. All optimal solutions computed in Hernandez et al. [53] are
obtained except one. Best known solutions are reported in Appendix B, Table 4.
When trip duration limits are added, we are aware of three papers. Anaya-Arenas et al. [4]
propose two constructive heuristics completed by local search. The first heuristic creates trips and
packs them into journeys. The second heuristic directly create journeys. Local search considers
insertion of visits in later positions. The procedure is evaluated on a set of dedicated instances
developed by the authors and based on real-world biomedical sample transportation. In both other
papers, profits are considered. The first is from Azi et al. [11]. They propose an adaptive large
neighborhood search. Destruction and insertion operators are developed for customers, trips and
journeys. The second is by Wang et al. [93]. They propose an algorithm based on the Adaptive
Memory Procedure paradigm. Each solution that is constructed is inserted in a memory M. When
M reaches a certain size, a solution is constructed selecting trips randomly with a distribution based
on the quality of the solution they belong to. Results obtained by Wang et al. [93] outperform those
provided by Azi et al. [11] with respect to quality, but the procedure is on average three time more
expensive in terms of CPU time.

16
 Chbichib et al. [25] consider the MTVRP with profits but without TW nor limited trip duration.
A two-phase algorithm is designed. Solution construction is guided by the net profit: next inserted
customers maximize the difference between the profit and the extra routing cost needed to serve the
customer. Improvement is carried out with a variable neighborhood descent algorithm and a hill
climbing procedure. Both are based on swap, insertion and cross-exchange moves.

3.2 Production, inventory, routing problems

Many times, vehicle routing optimization is only a step in a more complex decision-making process
involving also production and inventory management. Traditionally, these problems are decoupled
and optimized separately. There are however several reasons that can motivate to address them
simultaneously. Firstly, because of strong temporal constraints, it might be necessary to synchronize
the different types of operations (production and transport). Secondly, a coordinated optimization
can often allow reaching a better global efficiency. Recently, these issues have received a lot of
attention in the literature. It resulted in the publication of several survey papers (Chen [26], Coelho
et al. [32], Bertazzi et al. [14], Adulyasak et al. [1]). In this large amount of literature, a few papers
allow multiple routes in the routing phase. We focus on these papers below. We organize the section
by splitting them in three subgroups: production-routing, inventory-routing, production-inventory-
routing.
Synchronization between production and routing is particularly important in make-to-order poli-
cies where customized products are produced and delivered to clients within short time. No (or almost
not) inventory is considered and then addressing simultaneously production and transportation is cru-
cial. In the context of newspaper production and distribution, Van Buer at al. [92] explain that (daily
overnight) distribution starts long before the production is finished, vehicle routes are relatively short
and it is important to organize journeys with successive trips to satisfy drop-off time windows with a
limited fleet of vehicles. Ullrich [91] considers an academic but practice-oriented production-routing
problem with similar characteristics. Lee et al. [56] address the integrated production-distribution of
radioisotopes for medical treatments. Because of the very short lifespan of this medicine, distribution
with multiple trips is required. Besides, due to the natural decrease of radiation levels, the quantities
produced depend on the delivery times. Geismar et al. [46] finally consider the distribution of an
adhesive of limited lifespan with a single vehicle. In their case, production is performed at a fixed
production rate and the successive trips of the vehicle have to be scheduled in accordance with the
product availability. A few other papers also consider multiple trips, but with very small routes
(one or two customers, e.g., Chang and Lee [24]), which falls beyond the scope of this survey. In all
the aforementioned papers, solution methods concentrate on the complex synchronization between
production and routing and are not specifically innovative in the way multiple trips are handled.
The only exception can be found in Lee et al. [56] that give a polynomial time procedure to assign
a set of time-stamped trips to a minimum number of vehicles.
Aghezzaf et al. [2] introduce the possibility of multiple trips in inventory routing. In this class of
problems, a set of customers consumes a product at a certain rate; the supplier needs to determine,
over a time horizon of several periods, when to visit customers, which quantity to deliver at each
visit and how to organize routes. By allowing multiple trips in routes, Aghezzaf et al. [2] permit
to replenish with a vehicle a set of customers whose demand is larger than the vehicle capacity.
Higher delivery quantities are then possible at the cost of traveling additional distance to reload
the vehicle at the depot. The authors show how it impacts the economic order quantities and
eventually decreases total inventory holding costs and ordering costs. Raa and Aghezzaf [75] and
Raa and Aghezzaf [74] pursue this line of research, but limiting the driving time per period and
giving the possibility of assigning different frequencies to the trips of a vehicle, respectively. Solution
approaches in the three papers are based on an equivalent heuristic column generation procedure
where new columns are obtained with a modified savings algorithm. In this algorithm, the savings

17
obtained when joining two vehicle journeys is computed by evaluating the savings (and feasibility)
for all possible combinations of ending trip for the first journey and starting trip for the second.
Cornillier, et al. [33] address inventory routing in the context of petrol station replenishment.
Vehicles are allowed to perform several trips during each period. Several side constraints are present:
the fleet of vehicles is heterogeneous, vehicles have different compartments, trip duration is limited,
among others. A multi-phase heuristic is proposed. Trips are first constructed for each period and
then packed into vehicles. Cornillier et al. [36] extend this work to the case of time windows.
Oppen and Løkketangen [68] and Oppen et al. [69] study a livestock collection problem that
deals with transportation of live animals from farmers to slaughterhouses. Inventory constraints are
imposed at the slaughterhouse: on one side, non-working periods should be avoided, on the other side
animals cannot spend more than one overnight at the slaughterhouse. Rules to secure animal welfare
are considered: mixing animals of different types in the same vehicle compartment is not allowed;
animals with and without horns cannot be mixed; herds that are infected should be collected as
the last of the trip. The fleet of vehicles is heterogeneous. A tabu search and a column generation
approach are proposed.
Finally, Lei et al. [57] seek to optimize conjointly production, inventory and distribution in the
chemical industry. The problem involves several production sites, each one with its own fleet, and
vehicles are allowed to perform several delivery trips at each period. The authors propose a two-phase
approach. In the first phase, a mixed-integer program is solved where vehicles are allowed to serve
customers only via direct shipment. In a second phase shipments are heuristically consolidated.

3.3 Multi-trip in maritime transportation

Vehicle routing problems have mainly been developed in relationship with the transportation of
goods or passengers on road networks. Models used for vessel route optimization however share
obvious similarities with the VRP: optimized sequences of port visits where cargoes are picked-up
and/or delivered are searched for. There are however important differences that we list below, before
moving to MTVRP-related papers. Readers interested in maritime transportation are referred to
Christiansen et al. [28] and Christiansen et al. [27].
First of all, the usual goal in maritime transportation problems is to minimize sailing costs and
time charter costs, i.e., costs due to the rent of vessels. Classical vehicle routing problems usually
focus on minimizing travel costs; when vehicle costs are considered, because the fleet needs to be
sized, it is simply modeled with fixed costs. Conversely, the integration of charter costs can be more
tricky. Second, in maritime transportation the size of vessels is generally critical: different sizes of
vessels are usually available, with different renting costs and characteristics (capacity, sailing speed,
accessibility to ports, etc.). Classical vehicle routing problems usually consider a homogeneous fleet.
Other differences are the length of the time horizon TH that goes from a working day in standard
vehicle routing problems to weeks in the maritime case; truck routing times are shorter than vessel
sailing times; problem sizes are generally bigger in land transportation than in maritime. Another
remarkable difference is that usually the notion of depot is absent in maritime problems.
There are however a few practical situations that fall into the multi-trip definition we gave in
Section 1. We identified two: the supply of offshore platforms from an onshore depot and short-haul
hub-and-spoke feedering.
Fagerholt [41] investigates a liner shipping problem (LSP) in order to find a set of vessels and an
order of visits to offshore platforms to minimize sailing and charter costs respecting capacity and
time constraints. Vessels are allowed to be re-sailed once they are back at the onshore supply depot.
The author proposes a three-phase method to solve the problem. First, for each type of vessel, all
the feasible trips are generated. Second they are packed together into feasible journeys. The largest
size vessel associated with trips forming the journey is associated to the journey itself. Third, a set
partitioning problem is formed on the set of trips and journeys determined in phase one and two.

18
A real instance formed by 15 nodes and 5 different vessels available is solved. Moreover, three test
instances with 20, 30 and 40 customers are evaluated.
Suprayogi et al. [83] adapt this algorithm for the case when port’s water depth is taken into
account and ships cannot have access to all ports. The study is motivated by oily liquid waste
collection from ports in the Indonesian territory.
Fagerholt and Lindstad [42] extend the work initiated in [41] by introducing temporal aspects.
Some offshore platforms close overnight, generating a time window for service. The algorithm pro-
posed follows the same scheme: first, all the feasible trips are generated, then a solution is constructed
solving an integer programming model.
Halvorsen-Weare et al. [52] pursue this line of research with a much more complicated problem.
Offshore platforms require several visits during the time horizon. Departures from the onshore supply
depot to the same offshore platform need to be evenly spread. Onshore and offshore platforms are
constrained by opening hours. Finally trips need to be neither too short (for better capacity exploita-
tion) nor too long (sailing times uncertainty increases in long time periods). Again, Halvorsen-Weare
et al. [52] adapt the scheme initially proposed in Fagerholt [41]. Instances with up to 14 offshore
platforms are considered. Visits to platforms vary between 1 and 6 during the time horizon. Shyshou
et al. [81] address the same problem with a large neighborhood search. An interesting issue deserves
to be underlined here. The problem can somewhat be classified in between the classes of multi-trip
and periodic problems. It is not a standard periodic problem (which would fall out of the scope
of this survey) because journeys can cover several periods. However, contrary to usual multi-trip
problems, time is discretized in the sense that a vessel is not allowed to finish and start a trip on
the same day. Improving the duration of a trip by a few hours can thus sometimes save a whole day.
This is all the more important in this context given that the main objective of the problem (and the
main strategy of the local search procedure) is to limit the number of vessels used.
Bendall and Stent [13] are concerned with short-haul hub-and-spoke feedering. They investigate
the profitability of fast cargo services, where a great number of round voyages can be achieved quickly.
In their model, ships start from the container terminal and can visit any number of times each port.
The objective is to maximize the revenue earned from transported containers minus shipping costs.
The solution method starts by computing a set of possible trips. A mathematical model is then
solved in order to determine how many times each trip is performed (possibly zero). Trips are finally
heuristically packed into journeys. A simple case study based on South East Asia is given as an
example.

3.4 Multi-level distribution with multi-trips

Another category of problems involving multiple trips can be found in the context of multi-level
distribution. The main peculiarity in these problems concerning the multi-trip aspect is the presence
of multiple depots. Contrary to the common practice for the multi-depot VRP, the fact that vehicles
can perform multiple trips leaves the possibility that a trip starting at a depot ends at another depot.
The literature on the multi-depot VRP has largely been influenced by the possibility of a two-phase
strategy: first assign customers to a depot, then optimize routes for this depot. With the new option
offered when multiple trips are allowed, algorithms based on this strategy are dismissed and the
movements of the vehicles planned from the different depots are much more deeply interlinked.
Examples of papers dealing with this issue are Crainic et al. [37], Nguyen et al. [64], Nguyen et
al. [65] and Grangier et al. [48]. All these papers address multi-level distribution for city logistics.
Delivery flows are organized as follows. Direct delivery to customer is forbidden. Goods are first
delivered to satellites by first-tier vehicles, from where they are loaded in small-capacity vehicles
(second-tier vehicles) for final delivery to customers. No storage is allowed at the satellites. A journey
of the second-tier vehicles typically starts from (and ends at) a central depot and is constituted of
several successive trips, each possibly starting and ending at different satellites.

19
 In Crainic et al. [37], Nguyen et al. [64] or Nguyen et al. [65], second tier-vehicles have to reach
satellites during very restrictive time windows, derived from the arrival times of first-tier vehicles.
Thanks to these time windows, the optimization of the trips can almost be decoupled. This is
the approach proposed in Crainic et al. [37]. They first compute vehicle trips for the different
satellites and then constitute journeys by solving a minimum cost network flow problem. As shown
in Nguyen et al. [64], the weakness of this approach is to badly intertwine the routing and the satellite
assignments. Nguyen et al. [64] propose instead a tabu search approach. In addition to standard
routing moves, relocation and exchange of trips between vehicles are introduced. Nguyen et al. [65]
extend the algorithm to a more complex problem involving both inbound and outbound flows.
Grangier et al. [48] optimize simultaneously the first-tier and two-tier fleets of vehicles. They
thus have to manage the synchronization of the two types of vehicles at the satellites. They tackle
the problem with an adaptive large neighborhood search heuristic. The focus is then more on the
synchronization between the two fleet of vehicles than on the synchronization between the successive
trips of second-tier vehicle.
Plenty of more distant problems where a relatively similar situation is found are treated in the
literature. Illustrative examples are given by the multi-depot VRP with inter depot routes (Crevier
et al. [38]) or other papers dealing with intermediate facilities for recharge and refill (see Gianessi
[47] for a recent survey). Reviewing in details this paper is out of the scope of this survey.

3.5 Other variants of the MTVRP

Many other variants of the MTVRP have been studied in the literature. We propose in this subsection
to elaborate on the characteristics of the problems that interfere the most with the multiple trip
component. We then explain how the challenges raised by these characteristics have been dealt with.

Original temporal constraints

Several original temporal constraints have emerged in the context of VRP with multiple trips. A
first noteworthy constraint is the so-called double time window for the depot considered in Hajri-
Gabouj and Darmoul [51]. In general, the time window associated with the depot is defined by two
parameters: an opening time, before which vehicles are not allowed to start, and a closing time,
before which vehicles must have finished their journey at the depot. Hajri-Gabouj and Darmoul [51]
introduce a third parameter: the latest time at which a vehicle is allowed to start one of its trips
from the depot. This new constraint is motivated by the fact that facilities used for loading the
vehicles can be closed earlier than the depot. Though the constraint sounds important practically,
its impact on solution methods seems rather limited (a genetic algorithm in the case of Hajri-Gabouj
and Darmoul [51]).
Battarra et al. [12] impose the duration of the journey (difference between the instant the vehicle
is back at the depot after performing the last trip and the instant at which loading operations of
the first trip start) not to be longer than the spread time ST . The spread time differs from the time
horizon TH and it is usually strictly shorter (the spread time becomes meaningless when it coincides
with the time horizon). The spread time models the driver working day length, when this is shorter
than the depot opening hours. A vehicle starting its journey before TH − ST (assuming the depot
opens at time 0) must not take longer than the spread time; while on the other case must be back
at the depot before the time horizon is over. Battarra et al. [12] solve the problem with an adaptive
guidance approach that penalises the creation of trips that heavily overlap on certain time intervals.
This is to favour trip packing into journeys.
Cattaruzza et al. [20] introduce so-called release dates. A release date represents the time at
which goods are available at the depot for the final delivery to a customer. Note that commodities
are not transferable between customers. It is motivated by distribution from mutualized platforms:

20
given a customer request, goods are first supplied by trucks in the platform at the time of the release
date, and are then available for final distribution (with eco-friendly vans). These constraints raise
the difficulty of finding a balance between starting trips early to optimize the use of the vehicles
and waiting at the depot to reach a more effective consolidation of the merchandises. The authors
address the MTVRPTW with Release Dates with the genetic algorithm described in Section 3.1.
They explain how the split procedure and local search moves can be adapted when release dates are
introduced.
Mu and Eglese [63] study a relatively similar situation except that a single commodity is involved.
Part of the commodity that has to be delivered is supposed to arrive late at the depot (the time
and the quantity are known). This simulates the disruption of one of the vehicles in charge of
delivering the merchandise to the depot. The objective is then to minimize a weighted sum of service
delay, drivers overtime, and the deviation from the original plan that is calculated supposing all the
merchandise available at the beginning of the day. Vehicles are allowed to wait at the depot until
the late merchandise arrives or can leave the depot for a first delivering trip and possibly perform
a second trip when they are back. Note that journeys are made by not more than two trips: one
that starts before the late arrival of merchandise and one that starts after. The problem is called
Disrupted Capacitated Vehicle Routing Problem with Order Release Delay and is solved by means
of a tabu search algorithm.
Another situation investigated in Azi et al. [10] is the dynamic setting where new requests arrive
during the day. New requests can then either be rejected or be integrated in future trips of the
vehicles.

Distribution of incompatible products

Another interesting situation is found when multiple incompatible commodities are transported (Bat-
tarra et al. [12], Cattaruzza et al. [22]). The context is the distribution of different types of products
to supermarkets located in a regional territory. Commodities are said incompatible because they
cannot be transported together in a vehicle. If multiple trips were not allowed, each vehicle could be
assigned to a commodity and the distribution problem would naturally decomposes in one routing
problem per commodity. On the contrary, in Battarra et al. [12] and Cattaruzza et al. [22], vehicles
can perform several trips; hence, the different trips of the journey of a vehicle can be devoted to
the distribution of different commodities. This possibility strengthens the connection between the
different commodities and complicates the problem. The definition of the problem considers a fixed
maneuvering time between each pair of successive trips in addition to loading times that depend
on the quantity transported in the trip. The maneuvering time can be seen as the time needed to
prepare the vehicle for the next trip and can include cleaning operations when service switches to a
different commodity.

Limited number of trips

An easier type of constraint explored in the literature is the limitation of the number of trips allowed
in the journeys (Povlovitsch and Bergsten Mendes [71], Alonso et al. [3], Gribkovskaia et al. [49]).
As seen in Section 2.1, a positive consequence of this constraint can be a tighter integer pro-
gramming formulation for the problem. Indeed, considering for example the 4-index formulation,
less values are needed for the index on the number of trips and less variables are involved in the
formulation.

Fleet size minimization an other objectives

In several references (Prins [72], Battarra et al. [12], Cattaruzza et al. [22]) the objective of minimizing
the fleet size is introduced. Considering a limited fleet of vehicles or trying to minimize this fleet

21
actually appears very natural when multiple trips are considered: the willingness of operating multiple
trips is necessarily related to a purpose of limiting the fleet size.
If the incentive to find solutions with few vehicles is very strong (either because finding feasible
solutions is difficult with the available fleet of vehicles or because the minimization of the fleet is
integrated as an objective), the computation of trips that can be effectively packed together becomes
more stringent.
Cattaruzza et al. [22] propose detailed computational experiments to evaluate how much allowing
multiple trips enables decreasing the fleet size. These experiments are conducted for the VRPTW
and on Solomon’s instances [82] and their extension by Gehring and Homberger [45]. They show
that the fleet size can be strongly reduced when multiple trips are allowed (up to almost 60% in
their experiments). They also highlight the relationship between this reduction and the so-called
time horizon usage, that is, the average percentage of traveling times for vehicles in the VRPTW
solution.
Apart from the above objective, the usual objective of minimizing routing costs and the already
mentioned objective of profit maximization (see Section 3.1), a few other objectives have been con-
sidered in the literature.
A few papers consider green objectives. Ayadi et al. [6] minimizes CO2 emissions modeled as a
function of the distance and the load transported on each trip, while Cinar et al. [30] minimize the
total fuel consumption as a function of the distance, the load, and the type of the vehicles.
Rivera et al. [76] study the Multi-Trip Cumulative Capacitated VRP where the objective function
is the minimization of the sum of arrival times at required nodes. Applications of the problem can
be found in disaster logistics.

Less influential characteristics

Apart from the above characteristics, several variants of the MTVRP have been investigated with
additional constraints for which the multiple trips feature is relatively neutral.
Many papers consider heterogeneous fleet of vehicles, with or without accessibility restrictions
(e.g., Fleischmann [44], Brandão and Mercer [16], Brandão and Mercer [17], Cornillier et al. [34],
Caceres-Cruz et al. [19]). The fact that multiple trips are considered for distribution problems
with heterogeneous fleets of vehicles is not really surprising. Indeed, a carrier operating a fleet of
heterogeneous vehicles is prone to face under-usage of its smallest vehicles unless it allows multiple
trips for these vehicles. Note that the seminal paper by Fleischmann [44] belongs to this category. It
explains how the well-known savings algorithm can be extended to deal with an heterogeneous fleet
of vehicles and multiple trips.
Legal drivers hours rules have been considered in Brandão and Mercer [16]. Lin and Kwok [59]
study a complex location-routing problem, where multiple trips are allowed for the routing. Avella
et al. [5] tackle a problem where vehicles have multiple compartments. Boschetti and Miniezzo [15]
consider pickups and deliveries and drivers shifts. Ceselli et al. [23] face multiple depots, multiple
capacities, incompatibility constraints, compulsory drivers’ rest periods among others. Tung and
Pinnoi [90] study a real waste collection problem with very specific constraints. Taniguchi and Shi-
mamoto [86] and Taniguchi and Van Der Heijden [87] develop complex models so as to evaluate
the benefits of traffic information systems, cooperation among companies and other city logistics
initiatives. Cornillier et al. [34] and Cornillier et al. [35] investigate single-period variants of the com-
plex petrol station replenishment problems presented in Section 3.2, without or with time windows
respectively. Ong and Suprayogi [67] consider linehaul and backhaul customers (backhaul customers
can be visited after all the linehaul customers in the same trip).

22
4 Conclusions and perspectives
In this paper we presented the first complete survey on the multi-trip vehicle routing problem more
than 25 years after that the concept was introduced by Salhi [79] and 25 years after the problem was
formalized by Fleischmann [44].
With this literature review, we can observe a clear increase in the number of papers interested in
multiple trip routing in the recent years. This growing interest can be explained at least partly by
the recent investigations on new city logistics distribution systems and the development of the usage
of small eco-friendly vehicles. Many recent papers are concerned by these issues.
We can also conclude that the literature on the MTVRP and its academic variants only gives
a very partial vision on the amount of work devoted to multiple trip routing. Multiple trips can
be encountered in many other contexts as maritime transportation, inventory routing or multi-level
distribution. Considering all these contexts is important to achieve complete insight on the subject.
This actually confirms the interest of the survey. An illustrative example is given by the section on
mathematical formulations for the MTVRP (Section 2.1): while, to the best of our knowledge, a
single compact formulation has been proposed for this problem (Koc and Karaoglan [55]), we were
able to extract formulations from various different problems and present an unified view on the
possible formulations for the MTVRP.
Even if the numerous papers cited in this survey demonstrate that the issue of multiple trip
routing already raises the interest of many researchers, a lot of work is still needed. Regarding
urban distribution for example, intrinsic characteristics of cities include dynamic aspects and time-
dependent travel times. A the present time, only one paper deals with multi-trip distribution in
a dynamic context, while none considers time-dependent travel times. Regarding solution meth-
ods, most of the problems cited in this paper have only been tackled with heuristics. Furthermore
more theoretical studies (polyhedral study, lower bounding schemes, algorithms with performance
guarantee, complexity results on special graphs, etc.) are almost non-existent.
All these considerations should motivate researchers in pursuing further research in this area.

References
[1] Y. Adulyasak, J.-F. Cordeau, and R. Jans. The production routing problem: A review of
formulations and solution algorithms. Computers & Operations Research, 155:141–152, 2014.

[2] E.-H. Aghezzaf, B. Raa, and H. Van Landeghem. Modeling inventory routing problems in supply
chains of high consumption products. European Journal of Operational Research, 169(3):1048–
1063, 2006.

[3] F. Alonso, M.J. Alvarez, and J.E. Beasley. A tabu search algorithm for the periodic vehicle
routing problem with multiple vehicle trips and accessibility restrictions. Journal of Operational
Research Society, 59(7):963–976, 2008.

[4] A.M. Anaya-Arenas, T. Chabot, J. Renaud, and A. Ruiz. Biomedical sample transportation: A
case study based on quebec’s healthcare supply chain. Technical report, Faculté des sciences de
l’administration, Universit Laval Québec (Québec) Canada, 2014.

[5] P. Avella, M. Boccia, and A. Sforza. Solving a fuel delivery problem by heuristic and exact
approaches. European Journal of Operational Research, 152(1):170–179, 2004.

[6] R. Ayadi, A.E Elldrissi, Y. Benadada, and A. El Hilali Alaoui. Evolutionary algorithm for a
green vehicle routing problem with multiple trips. In International Conference on Logistics and
Operations Management, pages 148–154. 2014.

23
 [7] N. Azi. Méthodes Exactes et Heuristiques pour le Problème de Tournées avec Fenêtres de Temps
et Réutilisation de Véhicules. PhD thesis, Université de Montréal, 2010.

[8] N. Azi, M. Gendreau, and J.-Y. Potvin. An exact algorithm for a single-vehicle routing problem
with time windows and multiple routes. European Journal of Operational Research, 178(3):755–
766, 2007.

[9] N. Azi, M. Gendreau, and J.-Y. Potvin. An exact algorithm for a vehicle routing problem with
time windows and multiple routes. European Journal of Operational Research, 202(3):756–763,
2010.

[10] N. Azi, M. Gendreau, and J.-Y. Potvin. A dynamic vehicle routing problem with multiple
delivery routes. Annals of Operations Research, 199(1):103–112, 2012.

[11] N. Azi, M. Gendreau, and J.-Y. Potvin. An adaptive large neighborhood search for a vehicle
routing problem with multiple routes. Computers & Operations Research, 41:167–173, 2014.

[12] M. Battarra, M. Monaci, and D. Vigo. An adaptive guidance approach for the heuristic solu-
tion of a minimum multiple trip vehicle routing problem. Computers & Operations Research,
36(11):3041–3050, 2009.

[13] H.B. Bendall and A.F. Stent. A scheduling model for a high speed containership service: A
hub and spoke short-sea application. International Journal of Maritime Economics, 3:262–277,
2001.

[14] L. Bertazzi, M. Savelsbergh, and M.G. Speranza. In The Vehicle Routing Problem - Last Ad-
vances and New Challenges, Operations Research Computer Science Interfaces, chapter Inven-
tory Routing, pages 49–72. Springer, 2008.

[15] M. Boschetti and V. Maniezzo. A set covering based matheuristic for a real-world city logistics
problem. International Transaction in Operational Research, 22:169–195, 2015.

[16] J. Brandão and A. Mercer. A tabu search algorithm for the multi-trip vehicle routing problem.
European Journal of Operational Research, 100(1):180–191, 1997.

[17] J. Brandão and A. Mercer. The multi-trip vehicle routing problem. Journal of the Operational
Research Society, 49:799–805, 1998.

[18] K. Buhrkal, A. Larsen, and S. Ropke. The waste collection vehicle routing problem with time
windows in a city logistics context. Procedia-Social and Behavioral Sciences, 39:241–254, 2012.

[19] J. Caceres-Cruz, A. Grasas, H. Ramalhinho, and A.A. Juan. A savings-based randomized

heuristic for the heterogeneous fixed fleet vehicle routing problem with multi-trips. Journal of
Applied Operational Research, 6(2):69–81, 2014.

[20] D. Cattaruzza, N. Absi, and D. Feillet. The multi-trip vehicle routing problem with time windows
and release dates. Transportation Science, to appear.

[21] D. Cattaruzza, N. Absi, D. Feillet, and T. Vidal. A memetic algorithm for the multi trip vehicle
routing problem. European Journal of Operational Research, 236(6):833–848, 2014.

[22] D. Cattaruzza, N. Absi, D. Feillet, and D. Vigo. An iterated local search for the multi-commodity
multi-trip vehicle routing problem with time windows. Computer & Operations Research, 51:257–
267, 2014.

24
[23] A. Ceselli, G. Righini, and M. Salani. A column generation algorithm for a rich vehicle-routing
problem. Transportation Science, 43(1):56–69, 2009.

[24] T.-C. Chang and C.-Y. Lee. Machine scheduling with job delivery coordination. European
Journal of Operational Research, 158(2):470–487, 2004.

[25] A. Chbichib, R. Mellouli, and H. Chabchoub. Profitable vehicle routing problem with multiple
trips: Modeling and variable neighborhood descent algorithm. American Journal of Operational
Research, 2(6):104–119, 2012.

[26] Z.-L. Chen. Integrated production and outbound distribution scheduling: Review and exten-
sions. Operations Research, 58(1):130–148, 2010.

[27] M. Christiansen, K. Fagerholt, B. Nygreen, and D. Ronen. In Handbooks in Operations Research

& Management Science, chapter Maritime Transportation, pages 189–284. North Holland, 2007.

[28] M. Christiansen, K. Fagerholt, and D. Ronen. Ship routing and scheduling: Status and per-
spectives. Transportation Science, 38(1):1–18, 2004.

[29] N. Christofides, A Mingozzi, and P. Toth. The vehicle routing problem, pages 315–338. Wiley,
Chichester, 1979.

[30] D. Cinar, K. Gakis, and P.M. Pardalos. Reduction of co2 emissions in cumulative multi-trip
vehicle routing problems with limited duration. Environmental Modeling & Assessment, 2014.
doi:10.1007/s10666-014-9434-2.

[31] G. Clarke, J.W. Wright. Scheduling of Vehicles from a Central Depot to a Number of Delivery
Points. Operations Research, 12(4):568–581, 1964.

[32] L.C. Coelho, J.-F. Cordeau, and G. Laporte. Thirty years of inventory routing. Transportation
Science, 48(1):1–19, 2014.

[33] F. Cornillier, F.F. Boctor, G. Laporte, and J. Renaud. A heuristic for the multi-period petrol
station replenishment problem. European Journal of Operational Research, 191:295–305, 2008.

[34] F. Cornillier, F.F. Boctor, G. Laporte, and J. Renaud. An exact algorithm for the petrol station
replenishment problem. Journal of Operational Research Society, 59:607–615, 2009.

[35] F. Cornillier, G. Laporte, F.F. Boctor, and J. Renaud. The petrol station replenishment problem
with time windows. Computers & Operations Research, 36(3):919–935, 2009.

[36] G. Cornillier, F.F. Boctor, and J. Renaud. Heuristic for the multi-depot petrol station replenish-
ment problem with time windows. European Journal of Operational Research, 220(2):361–369,
2012.

[37] T.G. Crainic, Y. Gajapl, and M. Gendreau. Multi-zone multi-trip vehicle routing problem with
time windows. Technical Report 2012 - 36, Centre interuniversitaire de recherche sur les réseaux
d’enterprise, la logistique et le transport, Université de Montréal, Montréal, QC, Canada , 2012.

[38] B. Crevier, J.-F. Cordeau, and G. Laporte. The multi-depot vehicle routing problem with
inter-depot routes. European Journal of Operational Research, 176(2):756–773, 2007.

[39] A. Şen and K. Bülbül. A survey on multi trip vehicle routing problem. In VI International
Logistics & Supply Chain Congress. Instanbul, 2008.

25
[40] G.B. Dantzig and J.H Ramser. The truck dispatching problem. Management Science, 6:80–91,
1959.

[41] K. Fagerholt. Optimal fleet design in a ship routing problem. International Transaction in
Operational Research, 9(5):453–464, 1999.

[42] K. Fagerholt and H. Lindstad. Optimal policies for maintaining a supply service in the norwegian
sea. Omega, 28(3):269–275, 2000.

[43] M.L. Fisher. Optimal solution for vehicle routing problem using minimum k-trees. Operations
Research, 42(4):626–642, 1994.

[44] B. Fleischmann. The vehicle routing problem with multiple use of vehicles. Technical report,
Fachbereich Wirtschaftswissenschaften, Universität Hamburg, 1990.

[45] H. Gehring and J. Homberger. A parallel hybrid evolutionary metaheuristic for the vehicle
routing problem with time windows. In EUROGEN99. 1999.

[46] H.N. Geismar, G. Laporte, L. Lei, and C. Sriskandarajah. The integrated production and
transportation scheduling problem for a product with a short lifespan. INFORMS Journal of
Computing, 20(1):21–33, 2008.

[47] P. Gianessi. Solving strategic and tactical optimization problems in city logistics. PhD thesis,
Université Paris 13, 2014.

[48] P. Grangier, M. Gendreau, F. Lehuédé, and L.-M. Rousseau. An adaptive large neighborhood
search for the two-echelon multiple-trip vehicle routing problem with satellite synchronization.
Technical Report 2014-33, Centre interuniversitaire de recherche sur les réseaux d’enterprise, la
logistique et le transport, Université de Montréal, Montréal, QC, Canada , 2014.

[49] I. Gribkovskaia, B.O. Gullberg, K.J. Hovden, and S.W. Wallace. Optimization model for a
livestock collection problem. International Journal of Physical Distribution & Logistics Man-
agement, 36(2):136–152, 2006.

[50] G. Gutin, A.P. Punnen. The Traveling Salesman Problem and Its Variations. Combinatorial
Optimization, vol. 12, Kluwer Academic Publisher, 2002.

[51] S. Hajri-Gabouj and S. Darmoul. A hybrid evolutionary approach for a vehicle routing problem
with double time windows for the depot and multiple use of vehicles. Studies in Informatics
and Control, 12(4):253–268, 2003.

[52] E.E. Halvorsen-Weare, K. Fagerholt, L.M. Nonås, and B.E. Asbjørnslett. Optimal fleet compo-
sition and periodic routing of offshore supply vessels. European Journal of Operational Research,
223(2):508–517, 2012.

[53] F. Hernandez, D. Feillet, R. Giroudeau, and O. Naud. An exact algorithm to solve the multi-trip
vehicle routing problem with time windows. Technical report, 2013.

[54] F. Hernandez, D. Feillet, R. Giroudeau, and O. Naud. A new exact algorithm to solve the
multi-trip vehicle routing problem with time windows and limited duration. 4OR, 12:235–259,
2014.

[55] C. Koc and I. Karaoglan. A branch and cut algorithm for the vehicle routing problem with
multiple use of vehicles. Proceedings of the 41st International Conference on Computers &
Industrial Engineering, pages 554–559. Los Angeles, USA, 2011.

26
[56] J. Lee, B.-I. Kim, A. L. Johnson, and K. Lee. The nuclear medicine production and delivery
problem. European Journal of Operational Research, 236(2):461–472, 2014.

[57] L. Lei, S. Liu, A. Ruszczynski, and S. Park. On the integrated production, inventory, and
distribution routing problem. IIIE Transactions, 38(11):955–970, 2006.

[58] J.K. Lenstra and H.G. Rinnooy Kan. Complexity of vehicle routing and scheduling problems.
Networks, 11(2):221–227, 1981.

[59] C.K.Y. Lin and R.C.W. Kwok. Multi-objective metaheuristic for a location routing problem
with multiple use of vehicles on real data and simulated data. European Journal of Operational
Research, 175(3):1833–1849, 2006.

[60] R. Macedo, C. Alves, J.M. Valério de Carvalho, F. Clautiaux, and S. Hanafi. Solving the vehicle
routing problem with time windows and multiple routes exactly using a pseudo-polynomial
model. European Journal of Operational Research, 214(3):536–545, 2011.

[61] C.E. Miller, A.W. Tucker, R.A. Zemlin. Integer programming formulation of traveling salesman
problems. Journal of the ACM, 7(4):326–329, 1960.

[62] A. Mingozzi, R. Roberti, and P. Toth. An exact algorithm for the multi-trip vehicle routing
problem. INFORMS Journal on Computing, 25(2):193–207, 2013.

[63] Q. Mu and R.W. Eglese. Disrupted capacitated vehicle routing problem with order release delay.
Annals of Operations Research, 207(1):201–216, 2013.

[64] P.K. Nguyen, T.G. Crainic, and M. Toulouse. A tabu search for time-dependent multi-zone
multi-trip vehicle routing problem with time windows. European Journal of Operational Re-
search, 231(1):43–56, 2013.

[65] P.K. Nguyen, T.G. Crainic, and M. Toulouse. Multi-zone multi-trip pickup and delivery problem
with time windows and synchronization. Technical Report 2014–18, Centre interuniversitaire
de recherche sur les réseaux d’enterprise, la logistique et le transport, Université de Montréal,
Montréal, QC, Canada , 2014.

[66] A. Olivera and O. Viera. Adaptive memory programming for the vehicle routing problem with
multiple trips. Computers & Operations Research, 34(1):28–47, 2007.

[67] J.O. Ong and Suprayogi. Vehicle routing problem with backhaul, multiple trips and time win-
dows. Journal Teknik Industri, 13(1):1–10, 2011.

[68] J. Oppen and A. Lø kketangen. A tabu search approach for the livestock collection problem.
Computers & Operations Research, 35(10):3213–3229, 2008.

[69] J. Oppen, A. Lø kketangen, and J. Desrosiers. Solving a rich vehicle routing and inventory
problem using column generation. Computers & Operations Research, 37(7):1308–1317, 2010.

[70] R.J. Petch and S. Salhi. A multi-phase constructive heuristic for the vehicle routing problem
with multi trips. Discrete Applied Mathematics, 133(1–3):69–92, 2004.

[71] S. Povlovitsch and A. Bergsten Mendes. Column generation for a multitrip vehicle routing
problem with time windows, driver work hours, and heterogeneous fleet. Mathematical Problems
in Engineering, 2013, 2013. Article ID 824961, 13 pages.

[72] C. Prins. Efficient heuristic for the heterogeneous fleet multitrip vrp with application to a
large-scale real case. Journal of Mathematical Modelling and Algorithms, 1:135–150, 2002.

27
[73] C. Prins. A simple and effective evolutionary algorithm for the vehicle routing problem. Com-
puters & Operations Research, 31(12):1985–2002, 2004.

[74] B. Raa and E.-H. Aghezzaf. Designing distribution patterns for long-term inventory routing
with constant demand rates. International Journal of Production Economics, 112(1):255–263,
2008.

[75] B. Raa and E.-H. Aghezzaf. A practical solution approach for the cyclic inventory routing
problem. European Journal of Operational Research, 192(2):429–441, 2009.

[76] J.C. Rivera, H.M. Afsar, and C. Prins. Multistart evolutionary local search for a disaster relief
problem. In Artificial Evolution, pages 129–141. Springer, 2014.

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

[78] D. Ronen. Allocation of trips to trucks operating from a single terminal. Computers & Operations
Research, 19(5):445–451, 1992.

[79] S. Salhi. The integration of routing into location-allocation and vehicle fleet composition prob-
lems. PhD thesis, university of Lancaster, 1987.

[80] S. Salhi and R.J. Petch. A GA based heuristic for the vehicle routing problem with multiple
trips. Journal of Mathematical Modeling and Algorithms, 6(4):591–613, 2007.

[81] A. Shyshou, K. Fagerholt, I. Gribkovskaia, and G. Laporte. A large neighbourhood search

heuristic for a periodic supply vessel planning problem arising on offshore oil and gas operations.
INFOR, 50(4):195–204, 2012.

[82] M.M. Solomon. Algorithms for the vehicle routing and scheduling problem with time windows
constraints. Operations Research, 35:254–265, 1987.

[83] Suprayogi, H. Yamato, and Iskendar. Ship routing design for the oily liquid waste collection.
Journal of the Society of Naval Architects of Japan, 190:325–335, 2001.

[84] É.D. Taillard. Parallel iterative search methods for vehicle routing problems. Networks,
23(8):661–673, 1993.

[85] É.D. Taillard, G. Laporte, and M. Gendreau. Vehicle routeing with multiple use of vehicles.
Journal of Operational Research Society, 47(8):1065–1070, 1996.

[86] E. Taniguchi and H. Shimamoto. Intelligent transportation sustem based dynamic vehicle rout-
ing and scheduling with variable travel times. Transportation Research Part C, 12(3–4):235–250,
2004.

[87] E. Taniguchi and R.E.C.M Van Der Heijden. An evaluation methodology for the city logistics.
Transport Reviews, 20(1):65–90, 2000.

[88] P. Toth and D. Vigo. The Vehicle Routing Problem. Monographs on Discrete Mathematics and
Applications. Siam, 2002.

[89] P. Tsirimpas, A. Tatarakis, I. Minis, and E.G. Kyriakidis. Single vehicle routing with a prede-
fined customer sequence and multiple depot returns. European Journal of Operational Research,
187(2):483–495, 2008.

28
[90] D.V. Tung and A. Pinnoi. Vehicle routing-scheduling for waste collection in hanoi. European
Journal of Operational Research, 125(3):449–468, 2000.

[91] A.C. Ullrich. Integrated machine scheduling and vehicle routing with time windows. European
Journal of Operational Research, 227(1):152–165, 2013.

[92] M.G. Van Buer, D.L. Woodruff, and R.T. Olson. Solving the medium newspaper produc-
tion/distribution problem. European Journal of Operational Research, 115(2):237–253, 1999.

[93] Z. Wang, W. Liang, and X. Hu. A metaheuristic based on a pool of routes for the vehicle routing
problem with multiple trips and time windows. Journal of the Operational Research Society,
65:37–48, 2014.

[94] P.C. Yellow. A computational modification to the savings method of vehicle scheduling. Oper-
ational Research Quarterly, 21(2):281–283, 1970.

29
 APPENDICES

A Benchmark instances for the MTVRP

Table 3 presents the state-of-the-art results on benchmark instances proposed by Taillard et al. [85]
for the MTVRP. The first column indicates the instance name, the number of customers N and
the value z ∗ that served as a basis for the construction of the instance. Columns |V|, TH1 and TH2
give the number of vehicles and the length of the time horizon. Columns Opt report optimal values
when available (all obtained by Mingozzi et al. [62]). Columns Best Known report best known values
(all obtained by the heuristic method proposed in Cattaruzza et al. [21]), when the corresponding
optimal values are not known but feasible solutions have been found. Column Best Unfeas reports
the value (including penalty) of the best unfeasible solution, when no feasible solutions are known.
This column is omitted for instances corresponding to time horizon TH2 since a feasible solution has
been found for all instances. All these values and more detailed algorithm comparisons can be found
in Cattaruzza et al. [21]. State-of-the-art algorithms are Olivera and Viera [66], Cattaruzza et al. [21].
|V| TH1 Opt Best Known Best Unfeas. TH2 Opt Best Known
CMT1 1 551 524.61 577 524.61
N = 50 2 275 533.00 289 529.85
z ∗ = 524.61 3 184 569.54 192 552.68
4 138 564.07 144 546.29
CMT2 1 877 835.26 919 835.26
N = 75 2 439 835.26 459 835.26
z ∗ = 835.26 3 292 835.26 306 835.26
4 219 835.26 230 835.26
5 175 835.80 184 835.26
6 146 858.58 153 839.22
7 125 866.58 131 844.70
CMT3 1 867 826.14 909 826.14
N = 100 2 434 826.14 454 826.14
z ∗ = 826.14 3 289 826.14 303 826.14
4 217 829.45 227 826.14
5 173 832.89 182 832.34
6 145 836.22 151 834.35
CMT4 1 1080 1031.00 1131 1031.07
N = 150 2 540 1031.07 566 1030.45
z ∗ = 1028.42 3 360 1028.42 377 1031.59
4 270 1031.10 283 1031.07
5 216 1031.07 226 1030.86
6 180 1034.61 189 1030.45
7 154 1068.59 162 1036.08
8 135 1056.54 141 1044.32
CMT5 1 1356 1302.43 1421 1299.86
N = 199 2 678 1302.15 710 1305.35
z ∗ = 1291.44 3 452 1301.29 474 1301.03
4 339 1304.78 355 1303.65
5 271 1300.02 284 1300.62
6 226 1303.37 237 1306.17
7 194 1309.40 203 1031.54
8 170 1303.91 178 1308.78
9 151 1307.93 158 1307.25
10 136 1323.01 142 1308.81
CMT11 1 1094 1042.11 1146 1042.11
N = 120 2 547 1042.11 573 1042.11
z ∗ = 1042.11 3 365 1042.11 382 1042.11
4 274 1078.64 287 1042.11
5 219 1042.11 229 1042.11
CMT12 1 861 819.56 902 819.56
N = 100 2 430 819.56 451 819.56
z ∗ = 819.56 3 287 819.56 301 819.56
4 215 819.56 225 819.56
5 172 845.56 180 824.78
6 143 845.48 150 823.14
F11 1 254 241.97 266 241.97
N = 71 2 127 250.85 133 241.97
z ∗ = 241.97 3 85 256.93 89 254.07
F12 1 1221 1162.96 1279 1162.96
N = 134 2 611 1162.96 640 1162.96
z ∗ = 1162.96 3 407 1162.96 426 1162.96

30
 |V| 1
TH Opt Best Known Best Unfeas. 2
TH Opt Best Known

Table 3 – Results on MTVRP instances

B Benchmark instances for the MTVRP with time-windows

and service-dependent loading times
Table 4 presents the state-of-the-art results on benchmark instances proposed by Hernandez et al. [53],
and adapted from Solomon’s, for the MTVRP with time-windows and service-dependent loading
times. In these instances Q is set to 100 and M is set to 2 (instances with 25 customers) or 4
(instances with 50 customers). Other data are kept as in original instances. Distances are Euclidean
and truncated to one decimal place. Column Opt reports optimal values when they are known.
Column Best Known reports best known solution values for other instances.
Optimal values are obtained by Hernandez et al. [53]. All the best known values are obtained by
the heuristic method proposed in Cattaruzza et al. [20]. Two of these best know values are upper
bounds determined by Hernandez et al. [53] (but no optimality certificate is provided).

N = 25 N = 50
Opt Best Known Opt Best Known
RC 201 660.0 1096.6
202 596.8 1001.6
203 530.1 - 941.2
204 - 518.0 - 915.9
205 605.3 - 1058.7
206 575.1 - 1027.4
207 528.2 - 941.7
208 - 506.4 - 916.8
R 201 554.6 909.8
202 485.0 816.0
203 444.2 - 742.4
204 407.5 - 702.3
205 448.4 807.3
206 413.9 - 758.2
207 400.1 - 715.7
208 394.3 - 699.6
209 418.3 - 746.0
210 448.3 - 777.2
211 400.1 - 717.4
C 201 380.8 - 714.2
202 368.6 - 700.1
203 361.7 - 688.0
204 358.8 - 685.1
205 377.2 - 700.0
206 367.2 - 694.6
207 359.1 - 689.7
208 360.9 - 688.6

Table 4 – Results on MTVRP with time-windows and service-dependent loading times instances

C Benchmark instances for the MTVRP with time-windows,

service-dependent loading times, limited trip duration
and profits
Tables 5 and 6 report state-of-the-art results on benchmark instances proposed by Azi et al. [7], based
on Solomon’s, for the MTVRP with time-windows, service-dependent loading times, limited trip
duration and profits (MTVRPTW-LDP). Table 5 reports optimal solution values when all customers
are visited. Table 6 gives optimal solutions for which some customers are left non-visited. The

31
problem was investigated by Azi et al. [9], Macedo et al. [60] and Hernandez et al. [54], with the
exception that the latter does not allow non-visited vertices.
Each instance is solved with two values for N , respectively equal to 25 and 40, and two values for
Tmax . For instances of groups RC and R, Tmax is either set to 75 (indicated as Short Tmax in Table
5) or 100 (Large Tmax ). For group C it is set to 220 (Short Tmax ) or 250 (Large Tmax ). Service times
in C2 instances are equal to 90 for each customer, while in the other cases they are equal to 10. The
number of vehicles is set to 2. β = 0.2.
Additionally, two different conventions have been adopted for computing the distance matrix. In
Azi et al. [9], instances are used with distances rounded at the second decimal place. Macedo et al. [60]
performed their experiments on the same instances but without truncating distances. Hernandez et
al. [54] evaluate their approach on both types of distances. Columns Opt(T) and Opt(NT) of Tables 5
report the optimal value when the distances are truncated (T) or not truncated (NT), respectively.
In Table 6 the objective function is hierarchical and the optimal solution is described by the number
of non-visited customers (Column Nb) and the traveled distance (column Dist). Whether distances
are truncated or not, we follow the literature by reporting travel distances truncated at the second
decimal place.
In both tables a “-” indicates that no optimal solution has been found. In Tables 5 NoSol is
written instead if it is proved that no solution visiting all customers exist. No heuristic algorithms
having been applied to these instances, we do not report upper bounds.
Detailed comparisons of the three approaches when all customers are visited can be found in
Hernandez et al. [54]. They permit to conclude that Macedo et al. [60] and Hernandez et al. [54]
obtain comparable results and significantly improve upon Azi et al. [9], not forgetting that Hernandez
et al. [54] is more specialized (not allowing non-visited vertices).
Short Tmax Large Tmax
Opt (T) Opt(NT) Opt (T) Opt(NT)
RC 201 988.05 988.20 849.33 849.45
N = 25 202 881.49 881.60 679.86 679.95
203 749.15 749.26 593.56 593.63
204 - 744.83 - -
205 840.35 840.47 702.49 702.61
206 761.03 761.14 604.12 604.23
207 - - 514.81 514.90
208 - - - -
R 201 762.43 762.53 698.18 698.26
N = 25 202 645.78 645.86 617.53 617.60
203 621.97 622.04 577.74 577.80
204 579.68 579.75 483.30 483.37
205 634.09 634.17 559.14 559.21
206 596.74 596.81 523.64 523.70
207 585.74 585.81 512.00 512.04
208 579.68 579.75 483.30 483.37
209 602.39 602.47 517.69 517.74
210 636.15 636.24 547.23 547.29
211 575.91 575.97 474.49 474.54
C 201 659.02 659.15 540.90 541.02
N = 25 202 653.37 653.50 533.43 533.55
203 646.40 646.51 532.77 532.88
204 602.46 602.58 525.46 525.57
205 636.39 636.52 529.94 530.05
206 636.39 636.52 527.84 527.95
207 603.22 603.34 525.46 525.57
208 613.20 613.34 525.46 525.57
RC 201 NoSol NoSol NoSol NoSol
N = 40 202 NoSol NoSol NoSol NoSol
203 NoSol NoSol - -

32
 Short Tmax Large Tmax
Opt (T) Opt(NT) Opt (T) Opt(NT)
204 - - - -
205 NoSol NoSol NoSol NoSol
206 NoSol NoSol - -
207 NoSol NoSol - -
208 - - - -
R 201 NoSol NoSol NoSol NoSol
N = 40 202 - - - -
203 - 962.42 816.51 816.65
204 858.22 858.35 - -
205 1017.84 1019.89 872 873.36
206 927.22 931.94 812.31 812.42
207 886.22 890.93 - 764.52
208 858.22 858.35 - -
209 935.81 935.95 768.84 768.99
210 952.92 963.45 - -
211 869.75 869.88 - -
C 201 1168.83 1169.04 966.7 966.89
N = 40 202 1111.15 1111.34 919.85 920.05
203 1088.55 1089.24 - -
204 1039.16 1039.35 - -
205 1083.81 1084.02 921.19 921.37
206 1081.37 1081.57 919.05 919.24
207 1055.04 1055.24 - -
208 1071.99 1072.22 915.41 915.61

Table 5 – Benchmarks for the MTVRPTW-LDP: exact solutions with all customers visited

Truncated Not truncated

Nb Dist Nb Dist
RC (N = 40, Tmax = 75) 201 9 1292.16 9 1292.35
202 3 1457.89 3 1458.09
205 6 1272.12 6 1290.75
R (N = 40, Tmax = 75) 201 2 1130.59 2 1130.73
RC (N = 40, Tmax = 100) 201 6 1157.48 6 1157.65
202 - - 1 1322.08
205 2 1195.32 2 1195.51
R (N = 40, Tmax = 100) 201 - - 1 1075.05

Table 6 – Benchmarks for the MTVRPTW-LDP: exact solutions with non visited customers

33