TRSC 2019 0964

This article was downloaded by: [203.110.242.
13] On: 06 August 2022, At: 20:55

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA
Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
A Benders Decomposition Approach for the Multivehicle

Production Routing Problem with Order-up-to-Level Policy
Zhenzhen Zhang, Zhixing Luo, Roberto Baldacci, Andrew Lim
To cite this article:

Zhenzhen Zhang, Zhixing Luo, Roberto Baldacci, Andrew Lim (2021) A Benders Decomposition Approach for the Multivehicle
Production Routing Problem with Order-up-to-Level Policy. Transportation Science 55(1):160-178. https://doi.org/10.1287/
trsc.2019.0964
Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
Copyright © 2020, INFORMS
Please scroll down for article—it is on subsequent pages
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
Published in Transportation Science on August 19, 2020 as DOI: 10.1287/trsc.2019.0964.
This article has not been copyedited or formatted. The final version may differ from this version.
TRANSPORTATION SCIENCE
Vol. 55, No. 1, January–February 2021, pp. 160–178
http://pubsonline.informs.org/journal/trsc ISSN 0041-1655 (print), ISSN 1526-5447 (online)
A Benders Decomposition Approach for the Multivehicle

Production Routing Problem with Order-up-to-Level Policy
Downloaded from informs.org by [203.110.242.13] on 06 August 2022, at 20:55 . For personal use only, all rights reserved.
Zhenzhen Zhang,a Zhixing Luo,b,* Roberto Baldacci,c Andrew Limd

a
School of Economics and Management, Tongji University, Shanghai 200092, China; b School of Management and Engineering,
Nanjing University, Nanjing 210093, People’s Republic of China; c Department of Electrical, Electronic, and Information Engineering
“Guglielmo Marconi,” University of Bologna, Cesena 47521, Italy; d Department of Industrial Systems Engineering and Management,
National University of Singapore, Singapore 119077
*Corresponding author
Contact: zhenzhenzhang222@gmail.com, https://orcid.org/0000-0001-9776-8355 (ZZ); luozx.hkphd@gmail.com,
https://orcid.org/0000-0002-0491-1501 (ZL); r.baldacci@unibo.it, https://orcid.org/0000-0003-0938-5798 (RB); isealim@nus.edu.sg (AL)
Received: June 24, 2018 Abstract. The production routing problem (PRP) arises in the applications of integrated
Revised: May 2, 2019 supply chain which jointly optimize the production, inventory, distribution, and routing
Accepted: July 19, 2019 decisions. The literature on this problem is quite rare due to its complexity. In this paper,
Published Online in Articles in Advance: we consider the multivehicle PRP (MVPRP) with order-up-to-level inventory replenish-
August 19, 2020 ment policy, where every time a customer is visited, the quantity delivered is such that the
https://doi.org/10.1287/trsc.2019.0964 maximum inventory level is reached. We propose an exact Benders’ decomposition ap-
proach to solve the MVPRP, which decomposes the problem as a master problem and a
Copyright: © 2020 INFORMS slave problem. The master problem decides whether to produce the product, the quantity
to be produced, and the customers to be replenished for every period of the planning
horizon. The resulting slave problem decomposes into a capacitated vehicle routing
problem for each period of the planning horizon where each problem is solved using an
exact algorithm based on the set partitioning model, and the identified feasibility and
optimality cuts are added to the master problem to guide the solution process. Valid in-
equalities and initial optimality cuts are used to strengthen the linear programming re-
laxation of the master formulation. The exact method is tested on MVPRP instances and on
instances of the multivehicle vendor-managed inventory routing problem, a special case of
the MVPRP, and the good performance of the proposed approach is demonstrated.
Funding: This work was supported by the National Natural Science Foundation of China [Grants 71501091,
71732003, and 71571094], the National Research Foundation Singapore [Grant NRFRSS2016-004],
and the Ministry of Education Singapore [Grants R-266-000-096-133, R-266-000-096-731, and R-266-000-
100-646].
Supplemental Material: The e-companion is available at https://doi.org/10.1287/trsc.2019.0964.
Keywords: production routing problem • logic Benders’ decomposition • set partitioning model
1. Introduction simultaneously, including the production, inventory

Inventory routing problems (IRPs) are among the management, and routing decisions. Generally speak-
most important and challenging extensions of vehi- ing, over a multiperiod horizon, these integrated supply
cle routing problems (VRPs; Toth and Vigo 2014). In chain problems require one to determine the period
its basic version, the IRP is concerned with the dis- of production and visit of each customer, the corre-
tribution of a single product from a single facility to sponding quantities of production to be produced and
a set of customers over a given planning horizon. Cus- delivered to the customers, and the detailed routing
tomers consume the product at a given rate and can plan of the vehicles. The aim is to minimize the total
maintain an inventory of the product up to a specific level. production, inventory, and distribution costs.
A fleet of identical vehicles is available for the distribu- The PRP is very difficult to solve in practice and
tion of the product. The objective is to minimize the total finds a very large number of applications, for ex-
distribution cost, computed as the sum of the route costs ample, in so-called vendor-managed inventory (VMI)
and the inventory holding costs, without causing stock- systems, in which a supplier manages the inventory
outs at any of the customers. In the IRP, inventory control replenishment of its customers (retailers). Therefore, the
and routing decisions have to be made simultaneously. supplier has to decide when to replenish the customers
The production routing problem (PRP) generalizes over a given planning horizon, the routes to perform, and
the IRP by taking the production decisions into account. the quantities to deliver at each visit to avoid stockouts
The PRP considers various decisions in supply chain at the customers. The application of vendor-managed
160
Zhang et al.: Benders Decomposition for MVPRP

Transportation Science, 2021, vol. 55, no. 1, pp. 160–178, © 2020 INFORMS 161
resupply principles creates advantages for both supplier replenishment (RP) policy, further relaxes the ML
and customers. The vendor saves on distribution costs by policy by relaxing the constraint on the maximum
being able to better coordinate deliveries to different inventory level and by allowing the delivered quan-
customers. Customers may receive incentives and all save tity to be any positive value.
time and effort on inventory management. The single-vehicle VMIRP with the OU policy was
VMI systems were first introduced in the literature introduced by Bertazzi, Paletta, and Speranza (2002),
through applications in the distribution of liquid air who proposed a heuristic algorithm for its solution.
products, but many different industries are now im- Archetti et al. (2007) developed a branch-and-cut
plementing such systems or are exploring their use. approach for the VMIRP with a single vehicle and
These include the automotive, electronics assembly, the three different replenishment policies described
and chemicals industries, vending machines for juice above. They could solve instances with up to 45 cus-
or foods, chain stores, and maritime logistics, among tomers and three periods and up to 30 customers and
others (Andersson et al. 2010; Coelho, Cordeau, and six periods to optimality within two hours for the
Laporte 2014). The number of applications is in- VMIRP with the OU and ML policies, respectively.
creasing, along with the need for approaches to the Another branch-and-cut algorithm, based on a stronger
PRP that can handle the additional constraints and mathematical formulation, was introduced by Solyalı
complexities found in practical contexts. and Süral (2011), who were able to solve to optimal-
ity instances with up to 60 customers and three periods or
1.1. Literature Review 15 customers and 12 periods. Their formulation relies on
The IRP was introduced by Campbell, Clarke, and the shortest path representation of the lot-sizing prob-
Savelsbergh (2002) in the context of the distribution lem, where decision variables indicate time inter-
of liquid air products. The authors described a two- vals between successive deliveries. Avella, Boccia, and
phase heuristic approach based on decomposing the Wolsey (2015) took advantage of the special structure
set of decisions into the creation of a delivery sched- of test instances and developed tighter reformula-
ule, followed by the construction of a set of delivery tions for the single-vehicle IRP with both OU and ML
routes. This solution methodology was designed for policies. The authors reported computational results on
the solution of large-scale, real-life instances. benchmark instances with 50 customers and six periods.
There is a growing number of references to inventory Heuristic algorithms for the single-vehicle VMIRP have
routing and related problems in the literature. Surveys been investigated by Archetti et al. (2012).
on IRPs can be found in Bertazzi, Savelsbergh, and Coelho and Laporte (2013) introduced a branch-
Speranza (2008) and Andersson et al. (2010). The and-cut algorithm for the multivehicle variant of the
more recent survey papers of Coelho, Cordeau, and VMIRP. Their algorithm uses an extension of the
Laporte (2014) and Adulyasak, Cordeau, and Jans Archetti et al. (2007) formulation, and it could solve
(2015) provide a comprehensive review of the IRP some instances with up to 45 customers, three pe-
literature, based on a new classification of the problems. riods, and three vehicles. The results of Coelho and
IRPs are categorized with respect to their structural Laporte (2013) were further improved by Coelho and
variants and with respect to the availability of in- Laporte (2014) by introducing new valid inequalities
formation on customer demand. based on the relation between demand and available
An important variant of the IRP arises when both capacities. Exact algorithms for the multivehicle VMIRP
inventory control at the depot and inventory costs are have also been proposed by Avella, Boccia, and Wolsey
considered. This problem is often called the vendor- (2018) and Desaulniers, Rakke, and Coelho (2016).
managed inventory routing problem (VMIRP). The Avella, Boccia, and Wolsey (2018) described IRP
PRP further generalizes the VMIRP by considering, in reformulations under the ML replenishment policy,
addition to inventory control, production lot-sizing derived from a single-period substructure, and defined
decisions at the depot. a generic family of valid inequalities. The authors de-
Three main replenishment policies for the cus- scribed a branch-and-cut algorithm and reported com-
tomers have been considered in the inventory rout- putational results for the benchmark instances with 50
ing literature (Archetti et al. 2007). In the first policy, customers and three periods and 30 customers and six
called order-up-to-level (OU) policy, every time a cus- periods. Desaulniers, Rakke, and Coelho (2016) de-
tomer is visited, the quantity delivered is such that the scribed a branch-price-and-cut algorithm based on an
maximum inventory level of the customer is reached. innovative mathematical formulation for the prob-
The second policy, called maximum-level (ML) policy, lem under the ML policy, tightened with the inclu-
relaxes the OU policy by allowing the quantity de- sion of known and new families of valid inequalities.
livered to a customer to be any value such that the The authors reported computational experiments
resulting inventory level is between the current level on a set of 640 benchmark instances involving be-
and the maximum level. The third policy, called tween two and five vehicles, showing that their

162 Transportation Science, 2021, vol. 55, no. 1, pp. 160–178, © 2020 INFORMS
branch-price-and-cut algorithm outperforms the branch- 1.2. Contributions

and-cut algorithm of Coelho and Laporte (2014) on the The main contribution of this paper is to propose a
instances with four and five vehicles. new exact algorithm for the MVPRP specifically tai-
The literature on PRP problem is rather limited. lored for the OU policy which is accounted to be the
Archetti et al. (2011) adapted the branch-and-cut most difficult MVPRP policy (Adulyasak, Cordeau,
method of Archetti et al. (2007) to solve the single- and Jans 2014a). It is a Benders decomposition al-
vehicle PRP under ML policy and reported computa- gorithm based on the path-based formulation pro-
tional experiments on small-size instances involving 14 posed by Solyalı and Süral (2011) and also used by
customers. Exact algorithms for the multivehicle PRP Adulyasak, Cordeau, and Jans (2014a). The problem is
(MVPRP) have been proposed by Bard and Nananukul decomposed into a master problem, which decides
(2010) and by Adulyasak, Cordeau, and Jans (2014a). whether to produce the product, the quantity to be
Bard and Nananukul (2010) introduced a branch-and- produced, and the customers to be replenished for every
price procedure for the PRP with the ML policy and period of the planning horizon, and a slave problem,
with multiple vehicles. Instances with up to 10 cus- which is further decomposed into capacitated VRPs, one
tomers were solved to optimality within 30 minutes. problem for each period of the planning horizon. The
Adulyasak, Cordeau, and Jans (2014a) considered both algorithm relies on procedures used to compute a lower
the multivehicle VMIRP and the MVPRP. They intro- bound on the total routing cost of any optimal MVPRP
duced different mathematical formulations, with and solution and to generate a priori optimality cuts. In order
without a vehicle index, to solve the problems under to evaluate and assess the efficiency of our algorithm,
both the ML and OU inventory replenishment poli- extensive computational experiments were performed
cies. By using parallel computing, the algorithms of on both MVPRP and multivehicle VMIRP instances.
Adulyasak, Cordeau, and Jans (2014a) could solve The remainder of this paper is organized as follows.
instances with up to 45 and 50 customers (with three Section 2 formally defines the problem and presents
periods and three vehicles) for the VMIRP and PRP the formulation proposed by Adulyasak, Cordeau,
with the ML policy, respectively. For the OU policy, and Jans (2014a). The Benders reformulation, the logic-
the algorithms could handle instances with up to 45 based Benders decomposition algorithm, and its algo-
customers (with three periods and three vehicles) and rithmic features are then presented in Section 3. Sec-
35 customers (with six periods and three vehicles) for tion 4 introduces features that improve the efficiency of
the VMIRP and the PRP, respectively. Heuristic al- the algorithm. Section 5 presents the results of extensive
gorithms for both the multivehicle VMIRP and the computational experiments performed on both MVPRP
MVPRP can be found in Adulyasak, Cordeau, and and multivehicle VMIRP instances. Conclusions and
Jans (2014b). future research directions are given in Section 6.
According to the classification introduced by Coelho,
Cordeau, and Laporte (2014), all the IRPs mentioned 2. Problem Description and
above belong to the class of problems with finite time Mathematical Formulation
horizons. The problems are also deterministic and In this section, we formally describe the MVPRP, and
static due to the assumption that consumption rates we give the mathematical formulation of Adulyasak,
are known upfront. If the information on demand is Cordeau, and Jans (2014a) that is in turn based on the
not fully available to the decision maker at the be- formulation of Solyalı and Süral (2011).
ginning of the planning horizon, the problem is not Let G (N, E) be a complete and undirected graph
deterministic. In this case, if the probability distri- where N {0, . . . , n} is the node set and E is the edge
bution of the demand is known, then the problem is set. The set Nc {1, . . . , n} corresponds to n customers,
called the stochastic IRP. Dynamic IRPs arise when and node 0 corresponds to the plant or depot. Each
demand is not fully known in advance, but is grad- edge {i, j} ∈ E is associated to a routing cost cij . We
ually revealed over time, as opposed to what happens consider the problem of supplying the customers with a
in a static context. Solyalı, Cordeau, and Laporte single product from the plant over a discrete planning
(2012) introduced a branch-and-cut algorithm for a horizon of length l. At the beginning of each period, the
robust IRP in which demand is uncertain and its production plant can produce at most C units of prod-
probability distribution is unknown. This algorithm uct with a fixed setup cost equal to f and a unit pro-
could solve instances with up to 30 customers and duction cost equal to u. Node i, i ∈ N, has an inventory
seven periods. Their approach was also adapted to capacity of Li units and an inventory unit cost equal
solve the deterministic IRP with demand backlogging to hi . The quantity of the product held by node i at the
and could solve instances of the same size as for the beginning of the planning horizon is Ii0 . Customer i ∈ Nc
robust IRP. For these last variants of the IRP, the requires dit units of product during period t and can
reader is referred to the survey of Coelho, Cordeau, be visited at most once per period. A set of m identi-
and Laporte (2014). cal vehicles of capacity Q are available at the depot.

The MVPRP consists of simultaneously deciding • zit : binary variable equal to 1 if and only if node i,
(i) when and how much to produce at the plant, (ii) when i ∈ Nc , is visited in period t;
and how much to deliver to each customer, and (iii) what • st : nonnegative integer variable denoting the
routes to use in every period of the planning horizon. number of vehicles used in period t;
The objective is to minimize the sum of the production, • xijt : integer variable that might take value in {0, 1},
inventory, and routing costs during the planning hori- {i, j} ∈ E, i 0, t ∈ T, and value in {0, 1, 2}, {0, j} ∈ E, t ∈ T;
zon without causing stockouts at any of the customers. • λivt : binary variable equal to 1 if and only if node i is
In this paper, we consider the OU policy as replen- visited in period t and the previous visit is in period v.
ishment policy. Moreover, as commonly assumed in the The following additional notation is used in the
literature, for each period the production at the plant formulation. We denote with E(S) the set of edges
takes place before delivery, and the deliveries at the cus- with both nodes in S, that is, {{i, j} ∈ E : i, j ∈ S, S ⊆ N},
tomers are executed at the beginning of the time period. and with δ(S) the set of edges incident to a node in S,
that is, δ(S) {{i, j} ∈ E : i ∈ S, j ∈/ S or i ∈/ S, j ∈ S}; we
2.1. Mathematical Formulation use δ(i) to denote the set of edges incident to node i.
The following notation is adopted, as defined by The mathematical formulation without a vehicle in-
Adulyasak, Cordeau, and Jans (2014a): dex of the MVPRP described by Adulyasak, Cordeau,
• T: set of time periods {1, . . . , l}; to simplify the and Jans (2014a) for⎛ the OU policy is as follows: ⎞
formulation, we introduce a fictitious period l + 1, and ∑⎜⎜ ∑ ⎟
⎟
(F) z(F) min ⎜⎜⎜upt + fyt + h0 I0t + cij xijt ⎟
⎟
⎟
we define T T ∪ {l + 1}; ⎝ ⎠
• givt : total delivery quantity when customer i is t∈T {i,j}∈E
visited in period t and the previous visit is in period v, ∑∑ ∑ t−1
computed as + eivt λivt (1)
t∈T i∈Nc vπ(i,t)
⎪
⎧
⎪ ∑
t−1
∑ ∑
t−1
⎪
⎪ dij + (Li − Ii0 ), v 0,
⎪
⎪ if s.t. I0 t−1 + pt givt λivt + I0t ,
⎪
⎨ j1 i∈Nc vπ(i,t)
givt ⎪ t−1
⎪
⎪ ∑
⎪
⎪ dij , 0 < v < t ≤ l + 1; ∀t ∈ T, (2)
⎪
⎪
⎩
if
jv pt ≤ Cyt , ∀t ∈ T, (3)
• eivt : total inventory holding cost when customer i I0t ≤ L0 , ∀t ∈ T, (4)
∑
t−1
is visited in period t and the previous visit is in period λivt zit , ∀i ∈ Nc , t ∈ T,
v, computed as vπ(i,t)
( ( )) (5)
⎪
⎧
⎪ ∑
t−1 ∑ j
⎪
⎪ μ∑
(i,0)
⎪
⎪ hi Ii0 − dir , if v 0,
⎪
⎨ j1 r1
λi0t 1, ∀i ∈ Nc , (6)
eivt ⎪ ( ( )) t1
⎪
⎪
⎪ ∑
t−1 ∑j
⎪
⎪ Li − dir , if 0 < v < t ≤ l + 1; ∑
t−1 μ∑
(i,t)
⎪
⎪ hi
⎩ λivt − λitv 0, ∀i ∈ Nc , t ∈ T,
jv rv
vπ(i,t) vt+1
• μ(i, t): the latest period after period t when cus- (7)
tomer i can be replenished next without having a ∑
l
λit l+1 1, ∀i ∈ Nc , (8)
stockout, computed as μ(i, t) arg maxt<v≤l+1 {gitv ≤ Li }; tπ(i,l+1)
• π(i, t): the earliest period before period t when ∑
xjj t 2st , t ∈ T, (9)
customer i can be replenished without having a stock-
{ j,j }∈δ(0)
out, computed as π(i, t) arg min0≤v<t {givt ≤ Li }. ∑
Note that, in the definitions listed above, giv l+1 is a xjj t 2zit , ∀i ∈ Nc , t ∈ T,
fictitious delivery, and it is only used to compute { }∈δ(i)
j,j
μ(i, t) and π(i, t). The following decision variables are (10)
used by the formulation: st ≤ m, ∀t ∈ T, (11)
• pt : nonnegative continuous variable denoting the ( )
production quantity in period t; ∑ ∑ ∑
t−1
Q xijt ≤ Qzit − givt λivt ,
• Iit : nonnegative continuous variable denoting
{i,j}∈E(S) i∈S vπ(i,t)
inventory level at node i at the end of period t;
• yt : binary variable equal to 1 if and only if there is ∀S ⊆ Nc ,|S| ≥ 2,
production at the plant in period t; t ∈ T, (12)

pt , Iit ≥ 0, ∀i ∈ N, t ∈ T, (13) models. Basically, it decomposes the original prob-

yt , zit ∈ {0, 1}, ∀i ∈ Nc , t ∈ T, (14) lem into two simpler ones, that is, an integer master
λivt ∈ {0, 1}, ∀i ∈ Nc , v, t ∈ T , (15) problem (BMP) and a linear slave problem or sub-
problem (BSP), which are solved in an iterative fashion
st ∈ {0, . . . , m}, ∀t ∈ T, (16)
{ } by utilizing the solution of one in the other. At each
xijt ∈ {0, 1}, ∀ i, j ∈ E : i 0, t ∈ T, (17) iteration, the master problem actually behaves as a
x0jt ∈ {0, 1, 2}, ∀j ∈ Nc , t ∈ T. (18) relaxation of the original problem and provides fixed
integer variable values for the slave problem to obtain

The objective function (1) minimizes the total cost a feasible solution. A Benders cut is constructed and
consisting of production, setup, inventory, and routing added to the master problem in the next iteration to
costs. Constraints (2) guarantee the inventory flow exclude the solution just obtained in the last master
balance at the plant. Constraints (3) ensure that the problem. Therefore, each solution of the master prob-
production capacity is not violated if the production lem must satisfy all the Benders cuts generated so far to
takes place; otherwise, no product is produced. The avoid repetition. The master problem and the slave
inventory capacity constraint at the plant for each problem are solved in this iterative fashion until an
period is imposed by constraints (4). Constraints (5) optimal solution to the original problem is obtained.
link variables z with variables λ. Constraints (6), (7), The generation of Benders cuts (i.e., optimality and
and (8) are flow conservation constraints on the shortest feasibility cuts) is the core of the Benders decompo-
path network for each customer. Constraints (9), (10), sition algorithm, and they guarantee the convergence
and (12) are the routing-related constraints. Con- of the iterations to the optimal solution of the original
straints (11) limit the number of available vehicles. problem. Furthermore, these cuts also determine how
Constraints (12) impose both connectivity and vehicle fast the algorithm converges. The classic Benders
capacity constraints and are based on the generalized decomposition algorithm was proposed for linear
fractional subtour elimination constraints of the capaci- programming problems (Benders 1962), the cut gen-
tated VRP (Toth and Vigo 2014). Given a feasible F eration of which is based on the strong duality property
solution and a pair (S, t), the term on the right-hand of linear programming. Geoffrion (1972) has extended
side is equal to Q multiplied by the number of cus- it to a larger class of mathematical programming
tomers in S visited on day t (say α) minus the total problems. For more general integer programming,
demand delivered to the visited customers. If the logic-based Benders decomposition was proposed to
inequality is divided by Q, the term at the left-hand generate valid integer Benders cuts (Hooker 2000,
side, that is, the total number of edges in solution with Hooker and Ottosson 2003). The key is to generalize
both nodes in set S, must be less than or equal to α the linear programming dual used in the classical
minus a lower bound on the number of vehicles necessary method to an inference dual, and the solution of the
to visit the customers in S (say, β, computed as the ratio inference dual takes the form of a logical deduction
between the total demand in S and Q), thus imposing that yields valid Benders cuts. In the following, we
that at least β edges leave the customer set S. describe in detail our reformulation of formulation F
based on logic-based Benders decomposition.
3. Benders Decomposition The reformulation is based on the observation that,
In this section, we describe an exact approach based on once variables p, I, y, λ, and z have been fixed, for-
logic-based Benders decomposition, which was formally mulation F decomposes into l subproblems, where each
developed by Hooker (2000) and was applied with subproblem is a capacitated vehicle routing problem
success by Hooker and Ottosson (2003) to 0-1 program- (CVRP) defined on the customer delivery quantities
ming and by Hooker (2007) to planning and scheduling determined by the values of variables z and λ.
problems. The approach was later specialized to mixed Introducing nonnegative continuous extra variables
integer programming by Codato and Fischetti (2006), ωt , ∀t ∈ T, for the routing costs associated with the dif-
who introduced the so-called combinatorial Benders cuts. ferent periods, we can reformulate the MVPRP as follows:
Logic-based Benders decomposition is a generalization ∑( )
of classical Benders decomposition that can be applied (F) z(F) min upt + fyt + h0 I0t + ωt
t∈T
to a much wider variety of combinatorial optimization
problems, since the subproblem may be any combina- ∑∑ ∑
t−1
+ eivt λivt (19)
torial problem, not necessarily a linear or nonlinear t∈T i∈Nc vπ(i,t)
programming problem (Hooker and Ottosson 2003). ∑
s.t. ωt ≥ φ(z, λ),
t∈T
3.1. Logic-Based Benders Reformulation
Benders decomposition was first proposed by Benders (2) − (8), (13), (14), (15),
(1962) to efficiently solve mixed integer programming ωt ≥ 0, ∀t ∈ T, (20)

where i. BSP is infeasible. Add the infeasibility cut (23)

∑ ∑ to BMP.
φ(z, λ) min cij xijt ii. BSP admits a feasible and integer solution such
∑
{i,j}∈E (21) that φ(z∗ , λ∗ ) > t∈T ω∗t . Add the optimality cut (24)
t∈T
s.t. (9) − (12), (16), (17) and (18), to BMP.

Let (p∗ ,I ∗ ,y∗ ,z∗ , λ∗ ,s∗ ,x∗ ) be the corresponding MVPRP
∑
is the separation subproblem, where we assume φ(z, feasible solution of cost ẑ t∈T (up∗t + fy∗t + h0 I0t ∗) +
∑ ∑ ∑t−1
λ) ∞ if the problem is infeasible. ∗ ∗ ∗

t∈T i∈Nc vπ(i,t) eivt λivt + φ(z , λ ). Set UB min
The above reformulation can be handled by solving {UB, ẑ}.
a Benders master problem (BMP) to integer opti- At the different iterations of the above algorithm,
mality before calling a Benders subproblem (BSP) LB represents a valid lower bound on z(F), whereas at
corresponding to the separation subproblem. We have the end of the algorithm UB z(F). (We assume z(F)
three possible outcomes about a solution (p∗ , I ∗ , y∗ , ∞ if problem F does not admit a feasible solution.) In
z∗ , λ∗ , ω∗ ) of BMP: Section 3.2, we describe in detail how problem BSP is
i. Problem BSP is infeasible for (z∗ , λ∗ ), and one can solved at Step 2.b and how Step 2.b.i is implemented
add the following infeasibility cut to BMP: in practice.
fλ∗ (λ) ≥ 1, (22) The following lemma holds about optimality cut (24).
where Lemma 1. The optimality cut (24) satisfies the following

two properties:
∑∑ ∑
t−1
i. g(z∗ ,λ∗ ) (λ∗ ) φ(z∗ , λ∗ ).
f λ ∗ ( λ) (1 − λivt )
t∈T i∈Nc vπ(i,t):λ∗ 1
ii. Any feasible solution (p,I, y, z,λ, ω, s, x) of problem
∑ ∑
ivt
F such that λ λ∗ satisfies t∈T {i,j}∈E cij xijt ≥ g(z∗ ,λ∗ ) (λ).
∑∑ ∑
t−1
+ λivt , (23) Proof. Property (i) follows directly from the definition
t∈T i∈Nc vπ(i,t):λ∗ 0
ivt of the optimality cut. Regarding property (ii), since λ
λ∗ we have fλ∗ (λ) ≥ 1,
which cuts off solution λ∗ . ( ) ( ( ) ) ( )
ii. Problem BSP is feasible for (z∗ , λ∗ ), but φ(z∗ , φ z∗ , λ∗ − φ z∗ , λ∗ − LBR fλ∗ λ
∑
∗
λ ) > t∈T ω∗t ; then one can add the following opti- ( ) ( ( ) )
≤ φ z∗ , λ∗ − φ z∗ , λ∗ − LBR ≤ LBR ,
mality cut to BMP:
∑ ∑ ∑
and the inequality t∈T {i,j}∈E cij xijt ≥ g(z∗ ,λ∗ ) (z) holds
ωt ≥ g(z∗ ,λ∗ ) (λ), (24) ∑ ∑
since we have t∈T {i,j}∈E cij xijt ≥ LBR due to the def-
t∈T
inition of LBR . □
where g(z∗ ,λ∗ ) (λ) φ(z∗ , λ∗ ) − (φ(z∗ , λ∗ ) − LBR )fλ∗ (λ) and The following theorem then shows the correctness
LBR is a lower bound on φ(z, λ), that is, a lower bound of the algorithm.
on the routing cost of any optimal MVPRP solution.
iii. Problem BSP is feasible for (z∗ , λ∗ ), and φ(z∗ , Theorem 1. The logic-based Benders algorithm terminates
∑ ∑ ∑
∗
λ ) t∈T ω∗t t∈T {i,j}∈E cij xijt ∗ , where (s∗ , x∗ ) is an after finitely many steps.
optimal solution of BSP. The solution (p∗ , I ∗ , y∗ ,
z∗ , λ∗ , s∗ , x∗ ) is an optimal MVPRP solution with Proof. Suppose first that the algorithm terminates with
∑ ∗ )+ ∑ ∑ ∑t−1 a finite solution UB. Clearly, UB is an upper bound on
value t∈T (up∗t + fy∗t + h0 I0t t∈T
∗
i∈Nc vπ(i,t) eivt λivt +
∗
φ(z , λ ).∗ the solution cost z(F). Because the algorithm termi-
The exact algorithm based on the above logic-based nated, we have LB UB, and due to Lemma 1, prop-
Benders reformulation is as follows. The initial master erty (ii), LB is a valid lower bound on the optimal
problem BMP is defined by the objective function (19) solution cost. Because the solution corresponding to UB
subject to constraints (2)–(8), (13), (14), (15), and (20). is feasible, it is also optimal. Second, since the do-
The algorithm performs the following steps: main of BMP variables z and λ is finite, only finitely
1. Initialization. Compute the lower bound LBR (see many subproblems can be defined (and corresponding
Section 3.3), and set LB 0 and UB ∞. Benders infeasibility and optimality cuts), and the
2. While LB < UB optimal value is reached after finitely many steps. If no
a. Solution of the master problem. Solve problem feasible MVPRP solution exists, then the algorithm will
BMP. If BMP is infeasible, no feasible MVPRP solution terminate at Step 2.a after finitely many steps; this is
exists, stop. Otherwise, let (p∗ , I ∗ , y∗ , z∗ , λ∗ , ω∗ ) be the due, again, to the fact that the domain of BMP variables
corresponding optimal solution of cost z. Set LB z. z and λ is finite. □
b. Solution of the subproblem. Solve problem BSP; It is worth mentioning that the decomposition
there are two possible outcomes: approach used to solve the MVPRP strongly relies on

the structural properties of the OU policy, that is, on Poggi and Uchoa (2014) and Zhang et al. (2019) for a
the definition of variables λivt that enables us to for- review of exact methods based on the set partitioning
mulate and solve the corresponding subproblem in an formulation.
effective and efficient way. More precisely, under the For the sake of simplicity, we omit the index t in the
OU policy, the set Qi of all possible demand values for description of the formulation. Let 5 be the index set
a customer i ∈ Nc is polynomially sized, being Qi of all feasible routes, and let air be a binary coefficient
equal to {givt ≤ Q : ∀v ∈ {0} ∪ T, ∀t ∈ T , t > v}. How- that is equal to 1 if vertex i ∈ Nct belongs to route r ∈ 5
ever, under the ML policy, the set Qi is pseudo-pol- and takes the value 0 otherwise. (Note that a0r
ynomially sized, since general nonnegative decision 1, ∀r ∈ 5.) Each route r ∈ 5 has an associated cost
variables representing the quantities delivered to the br that is equal to the optimal solution cost of the
customers over the planning horizon are necessary to traveling salesman problem (TSP) instance defined
model the problem. In the case of the ML policy, if we by route r. Let ξr be a binary variable that is equal to 1
are willing to sacrifice the property of having only a if and only if route r ∈ 5 belongs to the optimal so-
polynomial number of variables in the master prob- lution. The formulation for the CVRP(t) is as follows:
lem by using variables λiwt instead of variables λivt , ∑
where λiwt is equal to 1 if a quantity w is delivered to CVRP(t) ωt min br ξr
∑
customer i on period t, and variables qit Q w0 wλiwt
r∈5
∑
are used to denote the quantity delivered in period t to s.t. air ξr λivi t , ∀i ∈ Nct , (25)
customer i, then variables λiwt can still be used sim- r∈5
∑
ilarly as variables λivt to model the subproblem and ξr ≤ m, (26)
derive the corresponding infeasibility and optimality r∈5
cuts. Nevertheless, the complexity of solving the ξr ∈ {0, 1}, ∀r ∈ 5.
master is greatly increased compared with the OU
policy. Constraints (25) specify that each customer i ∈ Nct
must be covered by one route, and constraint (26)
3.2. Solving the Subproblem requires that at most m routes are selected.
In this section, we describe the procedure used in Problem CVRP(t) can be infeasible due to the def-
Step 2.b of the exact algorithm to solve problem BSP inition of the set Nct and the limited number of vehicles
defined by the objective function (21) subject to m; therefore, a corresponding infeasibility cut must be
constraints (9)–(12), (16), (17), and (18). added to BMP. To check if CVRP(t) admits a feasible
Given a BMP solution (p, I, y, z, λ, ω), problem BSP solution, we solve the bin packing problem (BPP)
decomposes into l subproblems, where each sub- instance defined by |Nct | items and weights {qit } with at
problem corresponds to a CVRP. More precisely, the most m bins of capacity equal to Q. If the resulting BPP
CVRP associated with period t ∈ T, denoted CVRP(t), instance is infeasible, then the following infeasibility
is defined on a complete and undirected graph Gt cut can be defined:
(V t , Et ), where V t {0} ∪ Nct is the vertex set and Et is
the edge set defined as Et {{i, j} : i, j ∈ V t , i < j}. The ∑ ( ) ∑ ∑
t−1
1 − λivi t + λivt ≥ 1,
set Nct is defined as {i ∈ Nc : zit 1}, that is, it is the set i∈Nc :zit 1 i∈Nc :zit 1 vπ(i,t):λivt 0
of customers serviced in period t, and vertex 0 cor-
responds to the depot of graph G. A nonnegative cost, which cuts off the solution {λivt }.
ctij cij , is associated with each edge {i, j}. Each cus- The above cut can be strengthened by observ-
tomer i ∈ Nct is associated with a known nonnegative ing that if the solution {λivt } is infeasible, then also
demand, qit givi t , to be delivered, where vi is such that a solution {λivt } is infeasible whenever λivt ≥ λivt ,
λivi t 1. A set of m identical vehicles, each with ca- ∀i ∈ Nc , v π(i, t), . . . , t − 1, that is, delivery to addi-
pacity Q, is available at the depot. The problem tional customers on the same day will also result in an
consists of finding a collection of at most m simple infeasible solution. Hence, the following infeasibility
cycles or routes with minimum cost, defined as the sum cut can be added to BMP:
of the costs of the edges belonging to the routes and
such that (i) each route visits the depot vertex, (ii) each ∑ ( )
1 − λivi t ≥ 1. (27)
customer vertex is visited by exactly one route, and i∈Nc :zit 1
(iii) the sum of the demands of the vertices visited by a
route does not exceed the vehicle capacity Q. The above cut can be lifted by observing that for a
given i ∈ Nc we have giv1 t > giv2 t if v1 < v2 < t, that is,
Solving Problem CVRP(t). To solve each problem if the previous visit to customer i is in between
CVRP(t) we use an exact algorithm based on the set [π(i, t), vi − 1], then the demand to be delivered to i at
partitioning formulation. The reader is referred to period t will be greater than the demand delivered if

the previous visit is at period vi . Hence, the following respect to the dual solution u∗ . Using procedure
strengthened infeasibility cut can be derived: GENROUTE, generate the largest subset R of the
route set 5 such that
∑ ∑
vi
{ }
(1 − λivt ) ≥ 1. (28) |R| ≤ Δ, max br ≤ zUB − zLP ,
i∈Nc :zit 1 vπ(i,t) r∈R
If the BPP instance defined above admits a feasible so- where Δ is a user-defined parameter that is set equal
lution, then problem CVRP(t) admits a finite optimal to 60,000 in the computational experiments reported
solution. In the rest of this section, we briefly describe in Section 6.
the method used to solve to optimality problem 5. Solve problem CVRP(t). We have the following
CVRP(t). The method is based on the route enumeration two cases:
procedure described by Baldacci, Christofides, and i. If |R| < Δ, solve the reduced problem obtained
Mingozzi (2008) and on the hybrid strategy used in from problem CVRP(t) by substituting the route set 5
Pessoa, Uchoa, and Poggi de Aragão (2009) and with set R by the generic branch-and-cut algorithm of
Pessoa, de Aragão, and Uchoa (2008). Since solving the IBM CPLEX solver (IBM CPLEX 2016).
problem CVRP(t) can be time consuming, we also ii. If |R| Δ, solve problem CVRP(t) using a
consider the generation of valid optimality cuts based branch-and-price algorithm, where the procedure
on the linear programming (LP) relaxation of for- GENROUTE is again used to generate feasible routes
mulation CVRP(t). in a column generation fashion and where the branch-
Let u (u0 , ui1 , . . . , ui|Nt | ) be a vector of dual vari- ing on sets strategy is used (Lysgaard, Letchford, and
c
ables, where uih , ih ∈ Nct , and u0 ≤ 0 are associated with Eglese 2004).
constraints (25) and (26), respectively. The dual of the Optimally cut (29) can be lifted by using the same
LP-relaxation of problem CVRP(t) is as follows: observation used to lift infeasibility cut (28) as follows:
∑
DCVRP(t) max λivi t ui + mu0 ∑ ∑
vi
i∈Nct ωt ≥ ui∗ λivt + mu∗0 . (30)
∑ i∈Nct vπ(i,t)
s.t. air ui + u0 ≤ br , ∀r ∈ 5,
i∈Nct
Updating Problem BMP. Problems CVRP(t), ∀t ∈ T,
ui ∈ R, ∀i ∈ Nct , are first checked for feasibility, and any violated in-
u0 ≤ 0. feasibility cut (28) is added to BMP. If at least one in-
feasibility cut has been detected, the procedure termi-
The exact algorithm used to solve problem CVRP(t) nates and the master problem BMP is solved again at the
to optimality is as follows. next main iteration. Otherwise, if all problems CVRP(t),
1. Compute a primal bound. Compute a primal bound ∀t ∈ T, are feasible, then the problems are solved in
zUB on the optimal solution cost ωt by means of a tabu sequence (for t 1, . . . , l) up to Step 3 of the exact
search heuristic based on the algorithm proposed by algorithm used to solve problems CVRP(t), and any
Gendreau, Hertz, and Laporte (1994). violated optimality cut (30) is added to BMP. Also in
2. Solve the LP-relaxation of problem CVRP(t). Solve this case, if any violated cut is found, the procedure
the LP-relaxation of problem CVRP(t) by means of the terminates, and the master problem BMP is solved
standard column generation procedure. Let zLP be the again at the next main iteration.
optimal solution cost, and let ξ∗ and u∗ be the cor- If no infeasibility (28) and optimality (30) cuts have
responding primal and dual solutions, respectively. been detected, then all problems CVRP(t) are solved
The restricted master problem is initialized with the to optimality by executing Steps 4 and 5 of the exact
set of routes forming the primal solution computed at algorithm, and Step 2.b.ii of the exact algorithm for
Step 1, and the procedure GENROUTE described in the MVPRP is then executed to check if a new opti-
Baldacci, Christofides, and Mingozzi (2008) is used to mality cut (24) must be added to BMP.
generate feasible CVRP routes.
3. Add an optimality cut. If zLP > ωt , the following
optimality cut is added to BMP: 3.3. Computing Lower Bound LBR on the
∑ ∗ Routing Cost
ω ≥ u λ + mu∗ ,
t i (29)
i vi t 0 In this section, we describe the relaxation and the
i∈Nct
bounding procedure used to compute the lower bound
and the procedure terminates; otherwise, the next LBR introduced to define the optimality cut (24).
step is executed. Let fit , i ∈ Nc , t ∈ T, be a lower bound on the number
∑
4. Route enumeration. Let br br − i∈Nct air ui∗ − u∗0 be of visits that customer i must receive up to period t.
the reduced cost associated with route r ∈ 5 with For each i ∈ Nc and t ∈ T, let qit be the cumulative

demand of customer i up to period t computed as qit Proof. Consider the dual constraint corresponding to
∑
max{0,−Ii0 + tv 1 div } for t 1, . . . , l. Values fit , ∀i ∈ Nc , route r ∈ R̃. Since for each i visited by route r we have
t ∈ T, can be computed as fit qit / min{Q, Li }. In r ∈ R̃i , we have
addition, let mL and mU be lower and upper bounds on {( ) }
the total number of vehicles needed in the planning φi qi min br − λ(r ) − wL − wU /q(r )
∑ r ∈ R̃i
horizon, respectively. Let q̂i t∈T dit − Ii0 be the total
quantity of product required by customer i ∈ Nc over ≤ qi (br − λ(r) − wL − wU )/q(r).
the planning horizon. Values mL and mU can be de-

∑ Hence,
fined as mL Q1 i∈Nc q̂i and mU m · l. The relaxation
also requires the definition of value qi , ∀i ∈ Nc , defined ∑ ∑
as a lower bound on the quantity delivered to cus- air ui + vL + vU ≤ air qi (br − λ(r) − wL − wU )/
i∈Nc i∈Nc
tomer i during any visit in the planning horizon that
can be computed as qi mint∈T {dit }, i ∈ Nc . q(r) + λ(r) + wL + wU
We define a route as a least cost simple cycle of q(r)(br − λ(r) − wL − wU )/q(r)
graph G passing through depot 0 such that the total + λ(r) + wL + wU br . □
demand of the customers visited computed using
demands {qi } does not exceed the vehicle capacity Q. The optimal solution cost of the following problem
Let R̃ be the index set of all routes, and let air be a
binary coefficient that is equal to 1 if vertex i ∈ Nc max {z(λ, wL , wU )} (35)
belongs to route r ∈ R̃ and takes the value 0 otherwise. λ,wL ,wU
Each route r ∈ R̃ has an associated routing cost br .

A valid lower bound on the routing cost of any provides the best possible lower bound which can be
optimal MVPRP solution is given by the optimal computed by means of Theorem 2. Problem (35)
solution cost of the following integer problem: cannot be solved directly as the computation of so-
∑ lution u, for given penalties λ, wL , and wU , requires the
(RF) z(RF) min br ξr a priori generation of the set R̃. In practice, we use an
r∈R̃ iterative algorithm which computes a lower bound as
∑
s.t. air ξr ≥ fil ∀i ∈ Nc , (31) the cost of a suboptimal solution of problem (35) by
r∈R̃
∑ using a limited subset of set R̃ and by changing the
ξ r ≥ mL , (32) values of vector λ, wL , and wU . At each iteration, the
r∈R̃ procedure uses expressions (34) to find a solution
∑
ξr ≤ mU , (33) (u, vL , vU ) of the reduced problem defined on a route
r∈R̃ subset of R̃. In addition, subgradient vectors are
computed and used to change vectors λ, wL , wU to
ξr ≥ 0 integer , ∀r ∈ R̃.
maximize the value of the lower bound. In the pro-
The lower bound LBR is computed by a bounding cedure, we further relax the requirement that a route
procedure as a near-optimal dual solution of the LP- is a simple cycle of graph G, and we extend the route
relaxation of problem RF. The procedure differs from set R̃ to contain ng-routes (see Baldacci, Mingozzi,
standard column generation methods based on the and Roberti 2011). This relaxation allows us to execute
simplex algorithm as it uses a dual ascent heuristic to in pseudo-polynomial time the pricing algorithm
solve the master problem (see Baldacci et al. 2010). used to identify the route subset whose dual con-
The bounding procedure is based on the following straints are violated by the current solution (u, vL , vU ).
theorem. The above procedure is executed for a fixed number of
iterations, and the lower bound LBR is set equal to the
Theorem 2. Let ui ≥ 0, i ∈ Nc , vL ≥ 0, and vU ≤ 0 be the
maximum of the lower bounds computed at the
dual variables associated with constraints (31), (32), and (33),
different iterations.
respectively. Associate penalties λi ≥ 0, i ∈ Nc , with con-
straints (31) and wL ≥ 0 and wU ≤ 0 with constraints (32)
and (33), respectively. Let φi , i ∈ Nc , be computed as φi 4. Improving the Benders Master Problem
qi minr∈R̃i {(br − λ(r) − wL − wU )/q(r)}, where R̃i ⊆ R̃ is In this section, we describe valid inequalities to
the index set of routes passing through customer i ∈ Nc , strengthen the BMP. In particular, Section 4.1 describes
∑ ∑ two optimality cuts for the master problem used to
λ(r) i∈Nc air λi , and q(r) i∈Nc air qi . A feasible dual
solution (u, vL , uU ) of cost z(λ, wL , wU ) can be computed by accelerate the convergence of the exact algorithm.
means of the following expressions: Adulyasak, Cordeau, and Jans (2014a) described a
number of valid inequalities to strengthen the LP-
ui φi + λi , i ∈ Nc , vL wL , and vU wU . (34) relaxation of formulation F. Among the different valid

inequalities described by Adulyasak, Cordeau, and where constraints (39) state that a route servicing
Jans, we added to the BMP the following inequality: customer i with demand givt must be in the solution in
∑
t period t if and only if λivt is equal to 1, that is, the
yt ≥ 1, (36) previous visit of customer i occurred in period v < t.
t1 The procedure used to derive the optimality cuts is
where t is the earliest period when the plant must based on the following proposition.
produce to prevent a stockout, computed as
Proposition 1. Let problem P be defined as (P) min{cx :
{ { } }

∑ ∑ t Ax b, x ∈ Rn+ }, with A ∈ Rm×n , c ∈ Rn , b ∈ Rm , and let D
t arg min max 0, div − Ii0 − I00 > 0 . be its dual, that is, (D) max{wb : wA ≤ c, w ∈ Rm }, where
1≤t≤l i∈N v1
c
w is the dual vector associated with constraints Ax b. Then
We also added to the master problem the following any feasible dual solution of the following problem, (P )
valid inequalities: min{cx : ai x bi , ∀i ∈ I1 , ai x ≥ bi , ∀i ∈ I2 , x ∈ Rn+ }, where
∑ ∑
t−1 I1 and I2 form a partition of the index set {1, . . . , m} of the
givt λivt ≤ mQ, ∀t ∈ T, (37) rows of matrix A, ai denotes row i of matrix A, b ∈ Rm , is also
i∈Nc vπ(i,t)
a feasible solution of problem D.
and
∑ ∑
t−1 Proof. Let u be a feasible solution of the dual of prob-
λivt ≤ m, ∀t ∈ T, (38) lem P , that is, (D ) max{ub : uA ≤ c, ui ∈ R, ∀i ∈ I1 ,
i∈Nc vπ(i,t):givt >Q/2 ui ≥ 0, ∀i ∈ I2 }. It is easy to see that the solution w of
both based on the fact that a limited number of m vehicles D obtained by setting wi ui , i 1, . . . , n, is a feasible
are available in each period to serve the customers. D solution. □
It is worth mentioning that, due to constraints (5), Let wiv ∈ R, ∀i ∈ Nc , v π(i, t), . . . , t − 1, and u0 ≤ 0
imposing integrality on variables λ ensures the in- be the dual variables associated with constraints (39)
tegrality of variables z. Solyalı and Süral (2011) ob- and (40), respectively. The dual of the LP-relaxation of
served that also the vice versa holds. We conducted CVRP(t) is as follows:
preliminary experiments to define the integrality re-
quirements of variables z and λ, and as a result of our ∑ ∑
t−1
experiments, in the computational experiments re- DCVRP(t) max λivt wiv + mu0
i∈Nc vπ(i,t)
ported in Section 6 we decided to impose integrality
requirements on variables λ only. ∑ ∑
t−1
s.t. aivr wiv + u0 ≤ br ,
4.1. Initial Set of Optimality Cuts i∈Nc vπ(i,t)
In this section, we describe two ways of generating ∀r ∈ R̂,

initial optimality cuts for the BMP. The cuts are based
wiv ∈ R, ∀i ∈ Nc ,
on the following alternative formulation of prob-
lem CVRP(t). v π(i, t), . . . , t − 1,
Let R̂ be the index set of all feasible routes for period t, u0 ≤ 0.
where the demand qit associated with customer i,
i ∈ Nc , belongs to the discrete set of demands {givt : v
Type I Cut. A type I cut is based on the following LP-
π(i, t), . . . , t − 1}, that is, the set of all possible demands
problem derived from problem CVRP(t):
of customer i visited in period t according to the set
{π(i, t), . . . , t − 1} of its potential previous visiting pe- ∑
(T1) min br ξr
riods. Let aivr be a binary coefficient that is equal to 1 if
r∈R̂
and only if customer i belongs to route r and has ∑
associated a demand equal to givt . Problem CVRP(t) s.t. aivr ξr ≥ 1,
r∈R̂
can be formulated as follows:
∑ ∀i ∈ Nc , v π(i, t), . . . , t − 1,
CVRP(t) ωt min br ξ r
(41)
r∈R̂ ∑
∑ ξr ≤ m, (42)
s.t. aivr ξr λivt ,
r∈R̂
r∈R̂
∀i ∈ Nc , v π(i,t),. . .,t − 1, (39) ξr ≥ 0, ∀r ∈ R̂.
∑
ξr ≤ m, (40) Let wiv ≥ 0, ∀i ∈ Nc , v π(i, t), . . . , t − 1, and u0 ≤ 0
r∈R̂ be the dual variables associated with constraints (41)
ξr ∈ {0, 1}, ∀r ∈ R̂, and (42), respectively, and let (w∗ , u0∗ ) be a feasible

dual solution of problem (T1). Based on Proposition 1, To speed up the computation, the route sets R̂ and
the following inequality provides a valid optimality cut: R̂t , ∀t ∈ T, of formulations (T1) and (T2), respectively,
are extended with a relaxation of feasible routes easier
∑ ∑
t−1
ωt ≥ w∗iv λivt + mu0∗ . (43) to compute and based on the nonelementary route
i∈Nc vπ(i,t) relaxation ng-route proposed by Baldacci, Mingozzi,
and Roberti (2011), specifically tailored to consider
Type II Cut. A type II cut is based on the LP prob- variable demands associated with the customers. Let
lem derived from formulation F as follows. Let R̂t be the Oi ⊆ Nc be a set of selected customers for vertex i such
index set of all feasible routes for period t, where that |Oi | 8, ∀i ∈ Nc , and Oi contains i and the seven
the demand qit associated with customer i, i ∈ Nc , nearest customers to i according to routing costs {cij },
belongs to the discrete set of demands {givt : v and let t be a given time period. We define an ng-path
π(i, t), . . . , t − 1}. Let aivrt be a binary coefficient that is (NG, q, i) as a nonnecessarily elementary path P
equal to 1 if customer i belongs to route r of period t (0, i1 , . . . , ik−1 , ik i) starting from the plant at time pe-
and is associated to a demand equal to givt . The for- riod t, visiting a subset of customers of total demand
mulation uses binary variable ξrt that is equal to 1 if equal to q such that NG Π(P), and ending at customer
route r for period t is in the solution and is equal to 0 i such that i ∈/ Π(P ), where P (0, i1 , . . . , ik−1 ). We
otherwise. The formulation is as follows: denote by ψ(NG, q, i) the cost of the least cost ng-path
⎛ ⎞ (NG, q, i), and we define an (NG, q, i)-route to be an
∑⎜ ⎜ ∑ ⎟
⎟ (NG, q, 0)-path where i is the last customer visited
(T2) min ⎜
⎜
⎝upt + fyt + h0 I0t + br ξrt ⎟
⎟
⎠
t∈T r∈R̂t
before arriving at the plant. Functions ψ(NG, q, i)
can be computed using dynamic programming
∑∑ ∑
t−1
+ eivt λivt recursions similar to the recursions described by
t∈T i∈Nc vπ(i,t) Baldacci, Mingozzi, and Roberti (2011) (details are
s.t. (2) − (4), (6) − (8), (13), (36), (37), (38), omitted for the sake of brevity) on a state space graph
∑ H (E , Ψ), where E {(NG,q,i) : git−1t ≤ q ≤ Q, ∀NG ⊆
aivrt ξrt λivt , ∀i ∈ Nc , t ∈ T, ∑
Oi s.t. NG i and j∈NG gjt−1t ≤ q, ∀i ∈ N}, Ψ−1 (NG, q, i)
r∈R̂t
{(NG , q − qit , j) : ∀NG ⊆ Oj s.t. NG j and NG ∩ Oi
v π(i, t), . . . , t − 1, (44) NG\{i}, ∀qit ∈ {givt : v π(i,t),. . .,t − 1}, ∀j ∈ N,j i}, and
∑
ξrt ≤ m, t ∈ T, (45) Ψ {(Ψ−1 (NG, q, i), (NG, q, i)) : ∀(NG, q, i) ∈ E }.
r∈R̂t
0 ≤ yt ≤ 1, ∀t ∈ T, 5. Computational Results
0 ≤ λivt ≤ 1, ∀i ∈ Nc , v, t ∈ T , This section reports on the computational results of the
ξrt ≥ 0, ∀r ∈ R̂t , t ∈ T. exact method described in Section 3, hereafter called
“EXM.” Experiments were conducted on both MVPRP
Let wivt ∈ R, ∀i ∈ Nc , t ∈ T, v π(i, t), . . . , t − 1, and u0 ≤ 0 and multivehicle VMIRP test instances generated by
be the dual variables associated with constraints (44) Adulyasak, Cordeau, and Jans (2014a) using the set of
and (45), respectively. The dual constraints associated instances proposed by Archetti et al. (2007, 2011).
with variables ξrt are Method EXM was coded in Java and executed on a
workstation equipped with an Intel(R) Xeon(R) CPU
∑ ∑
t−1
aivrt wivt + u0 ≤ br , ∀r ∈ R̂t , t ∈ T, E5-2623 clocked at 3.00 GHz and 32 GB RAM, running
i∈Nc vπ(i,t) under a Linux operating system in a single-thread
mode. ILOG CPLEX 12.6.3 (IBM CPLEX 2016) was
and let (w∗ , u∗0 ) be the variables associated with a used as the LP solver and the IP solver in EXM.
feasible dual solution of the above formulation. Based We compare EXM with the branch-and-cut methods
on Proposition 1, the following inequality provides a of Adulyasak, Cordeau, and Jans (2014a) that, to the
valid optimality cut: best of our knowledge, were the only to consider both
∑ ∑∑ ∑
t−1 the MVPRP and the multivehicle VMIRP under the
ωt ≥ w∗ivt λivt + mu0∗ . (46) OU policy. The experiments of Adulyasak, Cordeau,
t∈T t∈T i∈Nc vπ(i,t) and Jans were performed on a workstation equipped
with an Intel(R) Xeon(R) CPU clocked at 2.67 GHz and
24 GB RAM. According to the SuperPi (1M) bench-
4.2. Computing Type I and II Cuts mark (http://www.superpi.net/), an estimate of the
Optimality cuts (43) and (46) are derived by column single-thread speed of a CPU, our machine is about
generation-based procedures to compute optimal dual 10% faster than that used by Adulyasak, Cordeau,
solutions associated with formulations (T1) and (T2). and Jans.

The MVPRP test set consists of instances involving (“t”), and the average percentage of the final lower
up to 50 customers with time horizons equal to three, bound relative to the best upper bound of the in-
six, and nine time periods. The multivehicle VMIRP stances not solved to optimality (“%lb”). For method
test set considers a number of customers ranging from 8c-Veh-Ind, the computing time refers to the wall
5 to 50 with time horizons equal to three and six clock time. For methods EXM and 8c-Veh-Ind the
periods. A total of 168 MVPRP instances and 320 table also reports the number of customers of the
multivehicle VMIRP instances have been considered instance solved to optimality (“n̂”) having the largest
in our experiments. For additional details about the number of customers. The last row of the table reports
instances considered in this section, the reader is re- the total number of instances solved to optimality by
ferred to the online supplement of Adulyasak, Cordeau, the different methods.
and Jans. The complete set of instances and the de- The results obtained can be summarized as follows:
tailed results of Adulyasak, Cordeau, and Jans can be • EXM solved to optimality 21 more instances than
downloaded at the website https://sites.google.com/ 8c-Veh-Ind, which represents the best method among
site/YossiriAdulyasak/publications. the three versions proposed by Adulyasak, Cordeau,
We compare method EXM with the following three and Jans (2014a).
versions of the branch-and-cut algorithm proposed • Taking into account the speed ratio between the
by Adulyasak, Cordeau, and Jans (2014a): machines used by the EXM and 8c-Veh-Ind and the
• Veh-Ind: the vehicle index formulation running fact that 8c-Veh-Ind also uses eight cores, EXM is on
on a single core; average significantly faster than 8c-Veh-Ind.
• Non-Veh-Ind: the nonvehicle index formulation • Instances with 40 customers were solved to opti-
running on a single core; mality by EXM involving three periods, 15 customers
• 8c-Veh-Ind: the vehicle index formulation run- more than the size of the instances that can be solved
ning on eight cores. by 8c-Veh-Ind.
For methods Veh-Ind and Non-Veh-Ind a time limit • The final lower bounds obtained by EXM for the
of two hours was imposed on the execution of the instances not solved to optimality are on average
branch-and-cut algorithm, whereas for 8c-Veh-Ind quite tight.
the maximum computing time was set to 12 hours Tables 2–4 show a detailed comparison of Veh-Ind,
of wall clock time. For method EXM, we imposed a Non-Veh-Ind, and EXM methods that were all run on
time limit of two hours. a single core machine. The first four columns of the
In the following, the results obtained on MVPRP in- tables give details about the instances where col-
stances are given in Section 5.1, whereas Section 5.2 umn c reports the instance class. For the methods of
reports the results obtained on multivehicle VMIRP in- Adulyasak, Cordeau, and Jans (2014a), the tables give
stances. The results obtained by EXM can be down- the value of the best upper bound computed (“Best ub”),
loaded at the website http://www.computational including method 8c-Veh-Ind, the percentage ratio of the
-logistics.org/orlib/MVPRP. lower bound relative to the best upper bound (“%lb”),
and the computing time in seconds (“t”). For method
5.1. Computational Results on the MVPRP EXM, the tables give the cost of the best solution
Table 1 summarizes the results obtained on MVPRP found (“ub”), the value of the final lower bound (“lb”)
instances. In the table, the instances are grouped and the corresponding percentage ratio relative to
according to the number of periods, and columns “n” value ub (“%lb”), the total computing time in sec-
and “ #ist” report the ranges of the number of cus- onds (“t”), the computing time in seconds spent in
tomers and the corresponding number of instances, solving the subproblem (“tS ”), and the percentage ratio
respectively. For each method, Table 1 gives the between the total routing cost and value ub (“%rc”).
number of instances solved to optimality within the Boldface numbers are used to indicate the best results
imposed time limit (“ #opt”), the average computing among the upper bounds. In addition, a symbol “n/a”
time in seconds of the instances solved to optimality indicates that the corresponding results are not
Table 1. Summary Results on the MVPRP
Veh-Ind Non-Veh-Ind 8c-Veh-Ind EXM
l n #ist #opt t %lb #opt t %lb n̂ #opt t %lb n̂ #opt t %lb
3 10-50 72 28 520.4 96.2 24 262.4 96.9 25 31 2,493.2 96.5 40 44 1,141.1 99.3

6 10-40 56 32 292.7 99.1 30 362.1 99.3 40 47 5,531.5 98.3 40 53 489.0 98.5
9 10-30 40 19 970.8 98.2 13 739.1 99.1 30 28 4,473.9 98.1 30 30 456.9 99.1
168 79 67 106 127

Table 2. Detailed Results on the MVPRP Instances with Three Periods
Adulyasak, Cordeau, and Jans (2014a)
Veh-Ind Non-Veh-Ind EXM
n l m c Best ub %lb t %lb t ub lb %lb t tS %rc
10 3 2 1 36,636 100.0 0.3 100.0 0.2 36,636 36,636.0 100.0 3.5 0.4 6.7
10 3 2 2 2,54,526 100.0 0.1 100.0 0.0 2,54,526 2,54,526.0 100.0 3.2 0.0 1.0
10 3 2 3 46,422 100.0 0.9 100.0 0.3 46,422 46,422.0 100.0 2.8 0.1 26.4
10 3 2 4 26,687 100.0 0.3 100.0 0.1 26,687 26,687.0 100.0 3.0 0.0 9.2
10 3 3 1 37,226 100.0 0.5 100.0 0.2 37,226 37,226.0 100.0 2.4 0.2 8.2
10 3 3 2 2,55,116 100.0 0.2 100.0 0.1 2,55,116 2,55,116.0 100.0 2.3 0.0 1.2
10 3 3 3 49,371 100.0 1.1 100.0 0.3 49,371 49,371.0 100.0 2.2 0.1 30.8
10 3 3 4 27,247 100.0 0.5 100.0 0.1 27,247 27,247.0 100.0 2.3 0.0 11.1
15 3 2 1 56,309 100.0 14.0 100.0 38.8 56,309 56,309.0 100.0 12.0 0.6 7.1
15 3 2 2 4,06,122 100.0 14.7 100.0 20.5 4,06,122 4,06,122.0 100.0 12.6 0.8 1.2
15 3 2 3 71,239 100.0 18.0 100.0 41.1 71,239 71,239.0 100.0 11.8 0.4 25.8
15 3 2 4 42,978 100.0 23.7 100.0 116.4 42,978 42,978.0 100.0 15.2 2.9 9.2
15 3 3 1 57,339 100.0 10.4 100.0 46.8 57,339 57,339.0 100.0 6.8 0.3 8.4
15 3 3 2 4,09,891 100.0 11.0 100.0 187.0 4,09,891 4,09,891.0 100.0 8.3 0.8 1.4
15 3 3 3 76,406 100.0 25.0 100.0 152.8 76,406 76,406.0 100.0 6.8 0.3 31.1
15 3 3 4 44,293 100.0 24.0 100.0 551.0 44,293 44,293.0 100.0 7.8 0.7 10.7
20 3 2 1 57,205 100.0 31.8 100.0 132.5 57,205 57,205.0 100.0 42.9 2.2 5.9
20 3 2 2 3,94,852 100.0 107.3 100.0 63.0 3,94,852 3,94,852.0 100.0 46.0 0.9 0.9
20 3 2 3 69,745 100.0 14.0 100.0 79.0 69,745 69,745.0 100.0 38.6 1.8 22.4
20 3 2 4 40,893 100.0 85.3 100.0 91.6 40,893 40,893.0 100.0 46.4 3.2 7.6
20 3 3 1 57,863 100.0 204.0 100.0 160.3 57,863 57,863.0 100.0 21.2 1.6 7.0
20 3 3 2 3,95,363 100.0 185.9 100.0 90.7 3,95,363 3,95,363.0 100.0 20.4 1.1 1.0
20 3 3 3 74,065 100.0 1,700.4 100.0 4,052.5 74,065 74,065.0 100.0 26.7 3.5 27.3
20 3 3 4 41,550 100.0 679.7 100.0 471.5 41,550 41,550.0 100.0 23.9 3.1 9.7
25 3 2 1 78,180 100.0 966.2 99.8 7,200.0 78,180 78,180.0 100.0 182.5 21.5 5.4
25 3 2 2 5,64,868 100.0 4,639.4 100.0 7,200.0 5,64,868 5,64,868.0 100.0 104.9 27.6 0.8
25 3 2 3 94,420 100.0 1,031.6 99.1 7,200.0 94,420 94,420.0 100.0 549.4 30.0 21.3
25 3 2 4 58,528 100.0 4,782.1 99.2 7,200.0 58,528 58,528.0 100.0 1,696.3 114.5 7.5
25 3 3 1 79,151 99.3 7,200.0 99.6 7,200.0 79,151 79,151.0 100.0 288.8 15.3 6.6
25 3 3 2 5,65,800 99.7 7,200.0 99.8 7,200.0 5,65,800 5,65,800.0 100.0 48.9 7.4 1.1
25 3 3 3 99,139 96.2 7,200.0 97.6 7,200.0 99,139 99,139.0 100.0 1,067.2 25.1 25.1
25 3 3 4 59,535 98.3 7,200.0 98.4 7,200.0 59,426 59,426.0 100.0 2,737.2 69.5 8.9
30 3 3 1 82,570 99.0 7,200.0 99.0 7,200.0 82,361 82,361.0 100.0 5,088.8 435.8 6.0
30 3 3 2 5,86,405 99.6 7,200.0 99.7 7,200.0 5,85,391 5,85,391.0 100.0 188.8 38.2 1.0
30 3 3 3 1,02,234 95.5 7,200.0 95.9 7,200.0 1,01,707 1,01,197.9 99.5 7,200.3 590.5 23.4
30 3 3 4 61,945 97.1 7,200.0 97.4 7,200.0 61,010 60,924.7 99.9 7,200.8 1,203.1 8.1
30 3 4 1 83,738 98.1 7,200.0 98.5 7,200.0 83,316 83,284.1 100.0 7,200.3 292.5 7.1
30 3 4 2 5,87,571 99.6 7,200.0 99.7 7,200.0 5,86,933 5,86,933.0 100.0 205.3 102.1 1.1
30 3 4 3 1,09,845 91.4 7,200.0 93.3 7,200.0 1,06,835 1,05,952.8 99.2 7,200.3 1,043.5 27.1
30 3 4 4 63,156 96.4 7,200.0 96.8 7,200.0 62,145 61,961.8 99.7 7,200.3 1,651.6 9.5
35 3 3 1 96,528 96.6 7,200.0 96.8 7,200.0 94,349 94,349.0 100.0 6,840.5 756.6 5.9
35 3 3 2 6,61,386 99.4 7,200.0 99.5 7,200.0 6,58,885 6,58,885.0 100.0 339.3 101.2 0.9
35 3 3 3 1,22,256 90.4 7,200.0 91.1 7,200.0 1,16,478 1,15,231.3 98.9 7,200.2 756.0 23.5
35 3 3 4 70,948 96.3 7,200.0 96.5 7,200.0 69,440 69,440.0 100.0 6,163.5 929.4 7.9
35 3 4 1 98,239 95.4 7,200.0 96.0 7,200.0 95,296 95,296.0 100.0 6,049.2 785.9 6.9
35 3 4 2 6,61,447 99.5 7,200.0 99.6 7,200.0 6,60,203 6,60,203.0 100.0 317.5 171.5 1.2
35 3 4 3 1,28,029 88.8 7,200.0 90.6 7,200.0 1,21,137 1,21,137.0 100.0 6,716.9 833.4 26.6
35 3 4 4 72,571 94.9 7,200.0 95.5 7,200.0 70,467 70,307.6 99.8 7,200.2 1,090.8 9.1
40 3 3 1 1,27,280 98.8 7,200.0 98.8 7,200.0 1,26,821 1,26,821.0 100.0 5,148.8 340.4 5.0
40 3 3 2 8,98,014 99.6 7,200.0 99.6 7,200.0 8,96,020 8,96,020.0 100.0 3,899.1 1,160.1 0.8
40 3 3 3 1,53,029 94.8 7,200.0 95.1 7,200.0 1,51,526 1,49,860.7 98.9 7,200.2 1,676.2 19.8
40 3 3 4 92,922 97.3 7,200.0 97.5 7,200.0 91,929 91,428.7 99.5 7,200.3 2,215.0 6.9
40 3 4 1 1,29,067 97.8 7,200.0 98.1 7,200.0 1,28,157 1,28,034.2 99.9 7,200.3 544.1 6.0
40 3 4 2 9,01,594 99.2 7,200.0 99.3 7,200.0 8,97,145 8,97,145.0 100.0 2,207.4 900.3 0.9
40 3 4 3 1,64,191 89.5 7,200.0 90.6 7,200.0 1,58,049 1,56,334.9 98.9 7,200.3 2,107.6 23.2
40 3 4 4 95,927 94.7 7,200.0 95.2 7,200.0 93,097 92,761.8 99.6 7,200.3 1,895.6 8.0
45 3 3 1 1,41,856 98.3 7,200.0 98.6 7,200.0 1,41,265 1,41,075.0 99.9 7,200.3 571.8 5.0
45 3 3 2 10,25,062 99.3 7,200.0 99.4 7,200.0 10,20,302 10,19,985.0 100.0 7,200.3 2,611.2 0.8

Table 2. (Continued)
45 3 3 3 1,72,851 92.9 7,200.0 93.2 7,200.0 1,69,147 1,67,377.4 99.0 7,200.3 2,627.6 19.7
45 3 3 4 1,06,750 96.5 7,200.0 96.8 7,200.0 1,05,017 1,04,445.3 99.5 7,200.2 2,332.9 6.9
45 3 4 1 1,44,963 96.8 7,200.0 97.1 7,200.0 1,42,958 1,42,507.3 99.7 7,200.3 1,433.0 6.2
45 3 4 2 10,27,296 99.1 7,200.0 99.2 7,200.0 10,21,791 10,21,528.5 100.0 7,200.3 3,337.9 0.9
45 3 4 3 1,83,037 88.7 7,200.0 90.4 7,200.0 1,76,584 1,74,930.9 99.1 7,200.3 1,607.8 23.2
45 3 4 4 1,08,419 95.5 7,200.0 96.1 7,200.0 1,06,684 1,06,047.1 99.4 7,200.3 2,428.5 8.1
50 3 3 1 1,39,164 98.0 7,200.0 98.1 7,200.0 1,38,235 1,37,491.7 99.5 7,200.3 5,122.7 5.8
50 3 3 2 9,80,022 99.4 7,200.0 99.4 7,200.0 9,76,479 9,75,639.3 99.9 7,200.3 6,509.8 0.8
50 3 3 3 1,73,118 91.4 7,200.0 92.1 7,200.0 1,72,256 1,64,668.0 95.6 7,203.7 6,839.8 23.4
50 3 3 4 1,03,031 96.1 7,200.0 96.3 7,200.0 1,00,972 1,00,239.7 99.3 7,200.2 3,985.2 7.5
50 3 4 1 1,411,78 97.2 7,200.0 97.5 7,200.0 1,39,950 1,39,143.0 99.4 7,200.2 2,340.3 6.8
50 3 4 2 9,83,352 99.2 7,200.0 99.2 7,200.0 9,78,174 9,77,274.8 99.9 7,200.3 2,319.0 1.0
50 3 4 3 1,89,474 85.9 7,200.0 87.2 7,200.0 1,78,483 1,73,194.1 97.0 7,201.6 6,875.9 26.4
50 3 4 4 1,04,222 95.8 7,200.0 96.3 7,200.0 1,02,748 1,02,007.8 99.3 7,200.2 2,467.8 9.1
available. For the detailed results about method 8c-Veh- C. the lifted cuts (28) and (30) are not used, but their
Ind the reader is referred to https://sites.google.com/ nonlifted versions (27) and (29) are used instead.
site/YossiriAdulyasak/publications. For the different versions, we used the same time
The detailed results on the whole set of instances limit and settings used for EXM.
can be summarized as follows: Table 5 gives an overview of the results obtained. In
• In terms of number of instances solved to opti- the table, the instances are grouped according to the
mality, EXM outperforms all the three versions pro- number of periods, and columns “opt” and “%opt”
posed by Adulyasak, Cordeau, and Jans on instances report the number of instances solved to optimality
with three and six periods. For instances with nine by the different versions and the corresponding per-
periods, EXM still outperforms method Non-Veh-Ind centage ratios computed over the total number of in-
and cannot solve to optimality one and three instances considered for each group of instances, respec-
stances solved by the Veh-Ind and 8c-Veh-Ind methods, tively. The table then shows the number of optimality
respectively. cuts (24) (“#cuts(24)”), the number of lifted infeasi-
• For 36 out of the 41 instances not solved to opti- bility cuts (28) (“#cuts(28)”), the number of LP-based
mality by EXM, improved upper bounds were com- lifted optimality cuts (30) (“#cuts(30)”), and the
puted by EXM with respect to the upper bounds number of executions of Step 5 of the algorithm used
computed by 8c-Veh-Ind. to solve problems CVRP(t), that is, the number of times
• Regarding the number of instances solved to CPLEX is invoked to solve the reduced problems
optimality for each instance class, the tables show that (“#CVRP(t)”). The next three columns report com-
36, 38, 24, and 29 instances were solved to optimality puting times about the solution of the BPPs used to
for classes c 1, 2, 3, 4, respectively. Therefore, EXM generate infeasibility cuts (“tBPP ”), the heuristic used
performs particularly well on instances of class 2, to compute the upper bounds zUB in solving problems
characterized by optimal solutions with a low ratio CVRP(t) (“theu ”), and the time spent in solving the
between the routing cost and the total cost. subproblem (“tsub ”). The last column reports the total
computing time (“ttot ”). Regarding the number of cuts
5.1.1. Analysis of the Different Components of EXM. In and the computing times, the table reports average
this section, we analyze the impact of the different values, and the last line of each table section dedicated
cuts embedded in method EXM. For the sake of the to a version reports the total number of instances
comparison, we considered the 127 MVPRP instances solved to optimality and average values over the
solved to optimality by EXM, and we compare EXM different columns. Moreover, for version (C) of EXM,
with the following three versions of the method: the table reports the average numbers of nonlifted
A. type I (43) and type II (46) cuts described in cuts added instead of the average number of lifted
Section 4.1 are not used; cuts. The results obtained can be analyzed as follows.
B. none of the two versions of the LP-based cuts • The initialization of EXM using type I and type II
(29) and (30) are added during the solution of the cuts and the use of the lifted versions of infeasibility and
subproblem; LP-based cuts are quite effective, since versions (A)

Table 3. Detailed Results on the MVPRP Instances with Six Periods
10 6 2 1 38,669 100.0 0.4 100.0 0.2 38,669 38,669.0 100.0 5.7 0.1 7.3
10 6 2 2 2,22,269 100.0 0.4 100.0 0.2 2,22,269 2,22,269.0 100.0 5.8 0.1 1.3
10 6 2 3 50,025 100.0 0.5 100.0 0.2 50,025 50,025.0 100.0 6.6 0.2 28.4
10 6 2 4 23,453 100.0 0.7 100.0 0.4 23,453 23,453.0 100.0 6.6 0.9 12.1
10 6 3 1 38,856 100.0 0.6 100.0 0.2 38,856 38,856.0 100.0 4.4 0.1 7.8
10 6 3 2 2,22,456 100.0 1.2 100.0 0.1 2,22,456 2,22,456.0 100.0 4.5 0.1 1.4
10 6 3 3 50,963 100.0 1.4 100.0 0.2 50,963 50,963.0 100.0 4.4 0.1 29.7
10 6 3 4 23,640 100.0 2.1 100.0 0.3 23,640 23,640.0 100.0 4.7 0.1 12.8
15 6 2 1 54,845 100.0 5.7 100.0 2.3 54,845 54,845.0 100.0 25.7 1.6 7.8
15 6 2 2 3,07,565 100.0 5.3 100.0 2.3 3,07,565 3,07,565.0 100.0 26.0 1.6 1.4
15 6 2 3 71,661 100.0 9.9 100.0 44.9 71,661 71,661.0 100.0 28.0 1.0 27.1
15 6 2 4 32,475 100.0 13.4 100.0 91.8 32,475 32,475.0 100.0 33.1 6.1 12.7
15 6 3 1 55,726 100.0 27.9 100.0 10.9 55,726 55,726.0 100.0 12.4 1.2 9.2
15 6 3 2 3,08,446 100.0 28.1 100.0 13.7 3,08,446 3,08,446.0 100.0 12.7 1.7 1.7
15 6 3 3 75,004 100.0 150.3 100.0 176.3 75,004 75,004.0 100.0 16.2 3.6 31.0
15 6 3 4 33,178 100.0 89.0 100.0 386.4 33,178 33,178.0 100.0 17.2 5.1 14.7
20 6 2 1 64,447 100.0 25.7 100.0 120.1 64,447 64,447.0 100.0 146.2 2.9 6.6
20 6 2 2 3,61,987 100.0 16.4 100.0 26.7 3,61,987 3,61,987.0 100.0 152.8 3.1 1.2
20 6 2 3 80,568 100.0 38.3 100.0 483.4 80,568 80,568.0 100.0 174.4 5.3 24.8
20 6 2 4 37,798 100.0 14.2 100.0 496.2 37,798 37,798.0 100.0 151.0 2.0 10.6
20 6 3 1 65,111 100.0 304.7 100.0 656.6 65,111 65,111.0 100.0 72.0 8.0 7.6
20 6 3 2 3,62,651 100.0 151.4 100.0 23.0 3,62,651 3,62,651.0 100.0 69.7 6.0 1.4
20 6 3 3 83,347 100.0 408.5 100.0 417.8 83,347 83,347.0 100.0 62.0 3.2 27.3
20 6 3 4 38,355 100.0 214.1 100.0 91.2 38,355 38,355.0 100.0 64.5 1.8 11.8
25 6 2 1 80,401 100.0 425.9 100.0 2,160.3 80,401 80,401.0 100.0 208.2 10.3 6.1
25 6 2 2 4,30,861 100.0 244.4 100.0 880.3 4,30,861 4,30,861.0 100.0 199.8 8.6 1.1
25 6 2 3 99,385 100.0 445.0 100.0 988.0 99,385 99,385.0 100.0 288.7 31.3 23.7
25 6 2 4 45,070 100.0 199.8 100.0 2,180.9 45,070 45,070.0 100.0 220.2 14.3 10.4
25 6 3 1 81,155 99.7 7,200.0 99.6 7,200.0 81,155 81,155.0 100.0 106.6 21.9 7.1
25 6 3 2 4,31,615 100.0 7,200.0 99.9 7,200.0 4,31,615 4,31,615.0 100.0 100.3 18.4 1.3
25 6 3 3 1,02,924 98.1 7,200.0 98.8 7,200.0 1,02,924 1,02,924.0 100.0 194.0 28.8 25.1
25 6 3 4 45,743 99.4 7,200.0 99.3 7,200.0 45,743 45,743.0 100.0 105.5 10.3 12.0
30 6 3 1 81,067 100.0 651.9 100.0 621.2 81,067 81,067.0 100.0 579.1 160.9 6.9
30 6 3 2 4,58,257 100.0 775.6 100.0 987.7 4,58,257 4,58,257.0 100.0 462.1 48.2 1.2
30 6 3 3 1,02,824 100.0 3,683.1 99.1 7,200.0 1,02,824 1,02,824.0 100.0 1,500.3 964.9 26.0
30 6 3 4 47,649 100.0 1,429.1 99.7 7,200.0 47,649 47,649.0 100.0 611.8 123.2 11.2
30 6 4 1 81,697 99.8 7,200.0 99.8 7,200.0 81,697 81,697.0 100.0 219.0 18.6 7.6
30 6 4 2 4,58,887 100.0 7,200.0 100.0 7,200.0 4,58,887 4,58,887.0 100.0 208.3 15.3 1.3
30 6 4 3 1,06,086 97.0 7,200.0 97.9 7,200.0 1,06,086 1,06,086.0 100.0 456.7 120.5 28.4
30 6 4 4 48,296 99.0 7,200.0 99.2 7,200.0 48,296 48,296.0 100.0 310.7 54.7 12.6
35 6 3 1 99,205 n/a n/a n/a n/a 99,205 99,205.0 100.0 653.4 34.2 6.2
35 6 3 2 5,70,355 n/a n/a n/a n/a 5,70,355 5,70,355.0 100.0 695.8 42.0 1.1
35 6 3 3 1,23,688 n/a n/a n/a n/a 1,23,688 1,23,688.0 100.0 1,324.3 561.3 24.2
35 6 3 4 59,046 n/a n/a n/a n/a 59,046 59,046.0 100.0 931.3 219.2 10.1
35 6 4 1 1,00,225 n/a n/a n/a n/a 1,00,225 1,00,225.0 100.0 513.3 242.8 7.1
35 6 4 2 5,71,385 n/a n/a n/a n/a 5,71,375 5,71,375.0 100.0 607.6 335.1 1.3
35 6 4 3 1,29,846 n/a n/a n/a n/a 1,29,922 1,26,773.6 97.6 7,201.1 6,925.8 27.9
35 6 4 4 59,913 n/a n/a n/a n/a 59,878 59,878.0 100.0 3,084.2 1,144.0 11.6
40 6 3 1 1,33,248 n/a n/a n/a n/a 1,33,248 1,33,248.0 100.0 1,695.5 815.4 5.0
40 6 3 2 7,34,268 n/a n/a n/a n/a 7,34,268 7,34,268.0 100.0 1,295.8 455.0 0.9
40 6 3 3 1,60,896 n/a n/a n/a n/a 1,62,069 1,59,057.1 98.1 7,206.3 6,483.8 21.7
40 6 3 4 74,693 n/a n/a n/a n/a 74,621 74,621.0 100.0 6,284.7 4,519.1 9.4
40 6 4 1 1,35,077 n/a n/a n/a n/a 1,34,281 1,34,281.0 100.0 431.5 79.6 5.7
40 6 4 2 7,36,464 n/a n/a n/a n/a 7,35,301 7,35,301.0 100.0 485.6 112.4 1.0
40 6 4 3 1,68,262 n/a n/a n/a n/a 1,64,762 1,64,626.6 99.9 7,200.3 435.5 23.0
40 6 4 4 76,062 n/a n/a n/a n/a 75,505 75,505.0 100.0 1,037.9 459.2 10.6

Table 4. Detailed Results on the MVPRP Instances with Nine Periods
10 9 2 1 63,064 100.0 4.3 100.0 1.9 63,064 63,064.0 100.0 11.7 1.6 7.9
10 9 2 2 3,81,394 100.0 5.3 100.0 2.1 3,81,394 3,81,394.0 100.0 11.9 1.8 1.3
10 9 2 3 82,683 100.0 31.7 100.0 8.6 82,683 82,683.0 100.0 16.1 1.9 29.3
10 9 2 4 40,774 100.0 19.9 100.0 5.6 40,774 40,774.0 100.0 15.9 2.3 12.9
10 9 3 1 63,822 100.0 18.6 100.0 20.2 63,822 63,822.0 100.0 6.9 0.4 9.2
10 9 3 2 3,82,152 100.0 31.9 100.0 21.6 3,82,152 3,82,152.0 100.0 6.6 0.5 1.5
10 9 3 3 86,095 100.0 124.3 100.0 58.4 86,095 86,095.0 100.0 11.6 1.3 32.2
10 9 3 4 41,379 100.0 126.3 100.0 41.9 41,379 41,379.0 100.0 10.2 1.7 13.3
15 9 2 1 91,148 100.0 118.2 100.0 3,440.1 91,148 91,148.0 100.0 49.7 1.6 7.6
15 9 2 2 5,40,698 100.0 295.9 100.0 2,734.2 5,40,698 5,40,698.0 100.0 46.8 1.6 1.3
15 9 2 3 1,18,746 100.0 2,160.8 98.4 7,200.0 1,18,746 1,18,746.0 100.0 470.6 6.9 28.0
15 9 2 4 57,753 100.0 973.7 99.3 7,200.0 57,753 57,753.0 100.0 126.1 4.7 12.0
15 9 3 1 92,632 100.0 3,112.6 99.7 7,200.0 92,632 92,632.0 100.0 23.2 2.5 9.1
15 9 3 2 5,42,182 100.0 4,935.5 100.0 7,200.0 5,42,182 5,42,182.0 100.0 22.5 1.9 1.5
15 9 3 3 1,25,383 95.7 7,200.0 97.6 7,200.0 1,25,383 1,25,383.0 100.0 327.2 6.1 32.1
15 9 3 4 59,386 98.5 7,200.0 98.5 7,200.0 59,386 59,386.0 100.0 1,935.6 21.4 13.9
20 9 2 1 1,03,809 100.0 833.3 100.0 1,854.9 1,03,809 1,03,809.0 100.0 306.1 8.2 7.0
20 9 2 2 6,17,889 100.0 153.9 100.0 70.2 6,17,889 6,17,889.0 100.0 396.3 3.1 1.2
20 9 2 3 1,31,101 100.0 792.5 99.2 7,200.0 1,31,101 1,30,587.8 99.6 7,200.3 347.0 25.3
20 9 2 4 65,859 100.0 583.7 99.7 7,200.0 65,859 65,859.0 100.0 502.2 95.0 12.2
20 9 3 1 1,04,704 99.7 7,200.0 99.9 7,200.0 1,04,704 1,04,704.0 100.0 138.2 4.6 8.4
20 9 3 2 6,18,902 100.0 4,121.9 100.0 1,348.8 6,18,902 6,18,902.0 100.0 135.4 13.0 1.4
20 9 3 3 1,36,443 97.3 7,200.0 97.9 7,200.0 1,36,286 1,35,522.4 99.4 7,200.2 42.1 28.5
20 9 3 4 66,830 99.6 7,200.0 99.4 7,200.0 66,830 66,830.0 100.0 575.5 59.7 13.5
25 9 2 1 1,29,172 n/a n/a n/a n/a 1,29,172 1,29,172.0 100.0 388.6 18.1 5.8
25 9 2 2 7,49,311 n/a n/a n/a n/a 7,49,311 7,49,311.0 100.0 364.0 10.5 1.1
25 9 2 3 1,58,573 n/a n/a n/a n/a 1,58,684 1,57,712.2 99.4 7,200.3 768.4 22.9
25 9 2 4 79,476 n/a n/a n/a n/a 79,496 79,300.9 99.8 7,200.3 216.0 10.6
25 9 3 1 1,30,594 n/a n/a n/a n/a 1,30,550 1,30,550.0 100.0 515.3 14.6 7.0
25 9 3 2 7,50,481 n/a n/a n/a n/a 7,50,481 7,50,481.0 100.0 166.6 14.3 1.2
25 9 3 3 1,67,271 n/a n/a n/a n/a 1,65,430 1,63,824.1 99.0 7,200.3 329.6 25.4
25 9 3 4 81,012 n/a n/a n/a n/a 80,761 80,272.2 99.4 7,200.3 128.1 11.8
30 9 3 1 1,37,463 n/a n/a n/a n/a 1,37,463 1,37,463.0 100.0 931.8 29.3 6.3
30 9 3 2 8,28,543 n/a n/a n/a n/a 8,28,529 8,28,529.0 100.0 2,958.5 331.8 1.2
30 9 3 3 1,74,270 n/a n/a n/a n/a 1,73,868 1,71,090.1 98.4 7,200.6 722.7 25.7
30 9 3 4 87,607 n/a n/a n/a n/a 87,456 86,875.7 99.3 7,200.2 567.2 11.1
30 9 4 1 1,39,909 n/a n/a n/a n/a 1,38,887 1,38,887.0 100.0 1,906.6 143.5 7.4
30 9 4 2 8,30,800 n/a n/a n/a n/a 8,29,697 8,29,697.0 100.0 1,328.2 174.1 1.3
30 9 4 3 1,82,839 n/a n/a n/a n/a 1,81,650 1,77,116.8 97.5 7,200.5 348.8 27.9
30 9 4 4 89,684 n/a n/a n/a n/a 88,541 88,043.4 99.4 7,200.3 447.1 12.3
and (C) cannot solve to optimality 10 and 11 instances Type II cuts. As a result, EXM initialized with only
solved by EXM, respectively. type I cuts solved to optimality 116 instances, whereas
• EXM particularly takes advantage in using LP- EXM with only type II cuts solved 117 instances.
based cuts (30). Indeed, version (B) solved to opti- Finally, the percentage ratio of the lower bound LBR
mality only 64 instances over the 127 instances con- on the routing cost used in the definition of the op-
sidered. As shown by Table 5, if the LP-based cuts are timality cuts (24), computed using the routing costs of
not used, the number of optimality cuts (24) and the the 127 instances considered, is equal to 55.0%, hence
number of times the IP solver of CPLEX is invoked particularly weak. An explanation for its quality is
(and the corresponding average computing times) due to the definition of values {qi }, that is, the lower
increase considerately. bounds on the quantity to be delivered. Nevertheless,
In summary, all the different cuts embedded in EXM its computing time is negligible, being on average
are particularly effective in solving MVPRP instances. equal to 9.2 seconds, and our aim in computing LBR
Concerning version (A) of EXM, we also executed was to quickly compute an initial lower bound to
EXM by selectively disabling the use of type I and properly define the optimality cuts (24).

Table 5. Analysis of the Different Components of EXM on the MVPRP
Version l #inst opt %opt #cuts(24) #cuts(28) #cuts(30) #CVRP(t) tBPP theu tsub ttot
EXM 3 44 44 100.0 13.5 234.8 1,009.4 4.8 55.6 2.4 156.6 1,141.1
6 53 53 100.0 253.6 2.9 425.0 21.2 4.2 1.6 196.8 489.0
9 30 30 100.0 245.4 27.1 730.6 20.3 5.7 1.0 32.6 456.9
127 168.5 89.0 699.7 15.3 22.4 1.7 144.1 702.9
A (no type I and II cuts) 3 44 39 88.6 15.1 353.5 1,729.2 4.3 106.8 3.8 191.5 1,450.5
6 53 51 96.2 183.9 3.3 418.0 23.5 3.9 1.2 491.7 543.0
9 30 27 90.0 328.0 35.0 895.0 26.0 9.0 1.3 37.8 1,034.8
117 159.5 132.1 985.0 17.4 40.8 2.2 280.5 973.6
B (no optimality cuts (29), (30)) 3 44 26 59.1 6,664.7 569.0 — 2,096.1 182.4 10.9 1,059.1 3,183.4
6 53 32 60.4 2,592.4 6.1 — 915.7 19.0 3.3 2,401.9 3,315.7
9 30 6 20.0 9,975.1 91.6 — 4,510.4 96.8 8.9 968.7 5,814.1
64 5,747.2 221.3 2,173.8 94.0 7.2 1,598.1 3,860.0
#cuts(27) #cuts(29)
C (no lifted cuts (28), (30)) 3 44 40 90.9 13.2 232.3 994.6 4.1 52.6 2.4 105.1 1,168.7
6 53 51 96.2 287.7 3.1 677.8 26.9 5.8 2.3 242.6 728.7
9 30 25 83.3 401.3 35.6 1,280.6 31.2 10.3 1.7 53.6 1,752.7
116 219.4 90.2 929.9 20.1 23.1 2.2 150.3 1,123.0
5.2. Computational Results on the and Jans. In particular, we also considered 20 in-
Multivehicle VMIRP stances involving 30 customers and six periods (with
In this section, we report on the results obtained by both low and high inventory costs) that were not run
EXM on multivehicle VMIRP, a special case of the by Adulyasak, Cordeau, and Jans with any of their
MVPRP defined as follows: branch-and-cut versions.
• The fixed production setup cost f and the unit The results about the multivehicle VMIRP can be
production cost u are set equal to 0. analyzed as follows.
• All variables y are set equal to 1, that is, yt 1, • In terms of the number of instances solved to
∀t ∈ T. optimality, EXM outperforms method Non-Veh-Ind
• The production quantity in period t is fixed to Bt , but it is outperformed by the methods based on the
that is, pt Bt , ∀t ∈ T, where Bt is the production vehicle index formulation (Veh-Ind and 8c-Veh-Ind).
quantity made available in each period. The addi- • The detailed results show that EXM is not domi-
tional constraints Iot ≥ Bt , ∀t ∈ T, are added to for- nated by method Veh-Ind as it can solve four instances
∑
mulation F and the term i∈Nc hi Ii0 is added to the to optimality that were not solved by Veh-Ind within the
objective function of F. imposed time limit. In addition, EXM computed 20 new
As assumed by Adulyasak, Cordeau, and Jans improved upper bounds with respect to the best upper
(2014a), the production at the plant takes place be- bounds computed by method 8c-Veh-Ind.
fore delivery, and the deliveries at the customers are • EXM scales particularly well with the number of
executed at the beginning of the time period. periods. Indeed, new upper and lower bounds for the
Table 6 summarizes the results obtained. The no- instances with 30 customers and six periods were
tation used is the same as described in the previous computed by EXM.
section about the MVPRP, and the instances are The detailed results about the MVPRP and the
grouped according to the type of inventory costs (Low multivehicle VMIRP show that the average ratios
or High). We run EXM on the whole set of multivehicle between the routing cost and the total cost of the best
VMIRP instances generated by Adulyasak, Cordeau, solutions obtained (column “%rc”) are equal to 11.3%
Table 6. Summary Results on the Multivehicle VMIRP
Veh-Ind Non-Veh-Ind 8c-Veh-Ind EXM
l class n #ist n̂ #opt t %lb n̂ #opt t %lb n̂ #opt t %lb n̂ #opt t %lb
3 Low 5–50 100 40 59 670.4 89.5 30 45 530.2 94.6 50 78 4,279.1 88.9 30 50 892.5 86.5
3 High 5–50 100 40 60 745.0 95.8 40 47 808.1 97.6 50 77 3,599.7 95.4 30 47 668.0 95.1
6 Low 5–30 60 15 24 671.3 93.8 10 16 622.8 94.8 25 37 4,610.9 87.2 10 20 358.7 91.4
6 High 5–30 60 15 24 658.1 96.4 10 16 617.2 96.9 25 37 4,651.1 93.1 10 20 326.8 95.0
320 167 124 229 137

and 66.7%, respectively. This is due to the cost structure Archetti C, Bertazzi L, Laporte G, Speranza MG (2007) A branch-and-
of the instances and, in particular, to the fact that, in the cut algorithm for a vendor-managed inventory-routing problem.
Transportation Sci. 41(3):382–391.
multivehicle VMIRP, setup and production costs are not Archetti C, Bertazzi L, Paletta G, Speranza MG (2011) Analysis of the
considered. The results obtained show that EXM is maximum level policy in a production-distribution system.
particularly efficient on MVPRP instances. A possible Comput. Oper. Res. 38(12):1731–1746.
explanation for this behavior is due to the fact that, since Avella P, Boccia M, Wolsey LA (2015) Single-item reformulations for
setup, production, and inventory costs dominate the a vendor managed inventory routing problem: Computational
experience with benchmark instances. Networks 65(2):129–138.

routing cost, the master problem provides a tight lower Avella P, Boccia M, Wolsey LA (2018) Single-period cutting planes for
bound on the optimal solution cost, thus speeding up inventory routing problems. Transportation Sci. 52(3):497–508.
the convergence of EXM. Baldacci R, Christofides N, Mingozzi A (2008) An exact algorithm for
the vehicle routing problem based on the set partitioning for-
6. Conclusions and Future Research mulation with additional cuts. Math. Programming 115(2):351–385.
We presented an exact Benders decomposition algo- Baldacci R, Mingozzi A, Roberti R (2011) New route relaxation and
rithm for the multivehicle production routing problem pricing strategies for the vehicle routing problem. Oper. Res.
59(5):1269–1283.
(MVPRP) with order-up-to level inventory replenish- Baldacci R, Bartolini E, Mingozzi A, Roberti R (2010) An exact so-
ment policy and for its special case, the multivehicle lution framework for a broad class of vehicle routing problems.
vendor-managed inventory routing problem (VMIRP). Comput. Management Sci. 7(3):229–268.
We demonstrated through extensive computational Bard JF, Nananukul N (2010) A branch-and-price algorithm for an
experiments that our approach outperforms the state- integrated production and inventory routing problem. Comput.
Oper. Res. 37(12):2202–2217.
of-the-art method for the MVPRP. In particular, the Benders JF (1962) Partitioning procedures for solving mixed-variables
method could solve to optimality MVPRP instances programming problems. Numerical Math. 4(1):238–252.
with up to 40 customers, three periods, and three ve- Bertazzi L, Paletta G, Speranza MG (2002) Deterministic order-up-to
hicles, which were not solved to optimality by the state- level policies in an inventory routing problem. Transportation Sci.
of-the-art method. Moreover, our approach is also 36(1):119–132.
Bertazzi L, Savelsbergh M, Speranza MG (2008) Inventory routing.
competitive on multivehicle VMIRP instances and Golden B, Raghavan S, Wasil E, eds. Operations Research/
could compute new lower and upper bounds for dif- Computer Science Interfaces (Springer, New York), 49–72.
ficult multivehicle VMIRP instances with up to 30 Campbell AM, Clarke LW, Savelsbergh MWP (2002) Inventory
customers, six periods, and four vehicles. routing in practice. Toth P, Vigo D, eds. The Vehicle Routing
The proposed method, due to its decomposition Problem (SIAM, Philadelphia), 309–330.
Codato G, Fischetti M (2006) Combinatorial benders’ cuts for mixed-
structure, can be easily adapted to deal with other integer linear programming. Oper. Res. 54(4):756–766.
intraroute constraints, simply by taking into account Coelho LC, Laporte G (2013) The exact solution of several classes of
such constraints in the route-generation phase. Fu- inventory-routing problems. Comput. Oper. Res. 40(2):558–565.
ture research will therefore investigate the general- Coelho LC, Laporte G (2014) Improved solutions for inventory-
ization of the model described in this paper to deal routing problems through valid inequalities and input order-
ing. Internat. J. Production Econom. 155:391–397.
with the complexity of real-world production routing Coelho LC, Cordeau J-F, Laporte G (2014) Thirty years of inventory
applications, such as time windows and distance routing. Transportation Sci. 48(1):1–19.
constraints. Desaulniers G, Rakke JG, Coelho LC (2016) A branch-price-and-cut
algorithm for the inventory-routing problem. Transportation Sci.
Acknowledgments 50(3):1060–1076.
The authors thank the anonymous reviewers and associate Gendreau M, Hertz A, Laporte G (1994) A tabu search heuristic for
editor for their helpful suggestions and very thorough review the vehicle routing problem. Management Sci. 40(10):1276–1290.
of the paper. Geoffrion AM (1972) Generalized benders decomposition. J. Optim.
Theory Appl. 10(4):237–260.
Hooker JN (2000) Logic-Based Methods for Optimization (John Wiley &
References Sons, New York).
Adulyasak Y, Cordeau J-F, Jans R (2014a) Formulations and branch- Hooker JN (2007) Planning and scheduling by logic-based benders
and-cut algorithms for multivehicle production and inventory decomposition. Oper. Res. 55(3):588–602.
routing problems. INFORMS J. Comput. 26(1):103–120. Hooker JN, Ottosson G (2003) Logic-based benders decomposition.
Adulyasak Y, Cordeau J-F, Jans R (2014b) Optimization-based Math. Programming 96(1):33–60.
adaptive large neighborhood search for the production rout- IBM CPLEX (2016) IBM ILOG CPLEX 12.6.3 callable library.
ing problem. Transportation Sci. 48(1):20–45. Lysgaard J, Letchford AN, Eglese RW (2004) A new branch-and-cut
Adulyasak Y, Cordeau J-F, Jans R (2015) The production routing algorithm for the capacitated vehicle routing problem. Math.
problem: A review of formulations and solution algorithms. Programming 100(2):423–445.
Comput. Oper. Res. 55:141–152. Pessoa A, de Aragão MP, Uchoa E (2008) Robust branch-cut-and-
Andersson H, Hoff A, Christiansen M, Hasle G, Løkketangen A (2010) price algorithms for vehicle routing problems. Golden BL,
Industrial aspects and literature survey: Combined inventory Raghavan S, Wasil E, eds. The Vehicle Routing Problem: Latest
management and routing. Comput. Oper. Res. 37(9):1515–1536. Advances and New Challenges (Springer, New York), 297–325.
Archetti C, Bertazzi L, Hertz A, Speranza MG (2012) A hybrid Pessoa A, Uchoa E, Poggi de Aragão M (2009) A robust branch-cut-
heuristic for an inventory routing problem. INFORMS J. Comput. and-price algorithm for the heterogeneous fleet vehicle routing
24(1):101–116. problem. Networks 54(4):167–177.

Poggi M, Uchoa E (2014) New exact algorithms for the capacitated Solyalı O, Cordeau J-F, Laporte G (2012) Robust inventory routing
vehicle routing problem. Toth P, Vigo D, eds. Vehicle Routing: under demand uncertainty. Transportation Sci. 46(3):327–340.
Problems, Methods, and Applications, 2nd ed., MOS-SIAM Series Toth P, Vigo D (2014) Vehicle Routing: Problems, Methods, and Ap-
on Optimization (SIAM, Philadelphia), 59–86. plications, 2nd ed., MOS-SIAM Series on Optimization (SIAM,
Solyalı O, Süral H (2011) A branch-and-cut algorithm using a Philadelphia).
strong formulation and an a priori tour-based heuristic Zhang Z, Luo Z, Qin H, Lim A (2019) Exact algorithms for the vehicle
for an inventory-routing problem. Transportation Sci. 45(3): routing problem with time windows and combinatorial auction.
335–345. Transportation Sci. 53(2):427–441.

TRSC 2019 0964

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

TRSC 2019 0964

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [203.110.242.

13] On: 06 August 2022, At: 20:55

A Benders Decomposition Approach for the Multivehicle

To cite this article:

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

Copyright © 2020, INFORMS

Please scroll down for article—it is on subsequent pages

A Benders Decomposition Approach for the Multivehicle

Zhenzhen Zhang,a Zhixing Luo,b,* Roberto Baldacci,c Andrew Limd

1. Introduction simultaneously, including the production, inventory

Zhang et al.: Benders Decomposition for MVPRP

Zhang et al.: Benders Decomposition for MVPRP

branch-price-and-cut algorithm outperforms the branch- 1.2. Contributions

Zhang et al.: Benders Decomposition for MVPRP

Zhang et al.: Benders Decomposition for MVPRP

pt , Iit ≥ 0, ∀i ∈ N, t ∈ T, (13) models. Basically, it decomposes the original prob-

integer variable values for the slave problem to obtain

Zhang et al.: Benders Decomposition for MVPRP

where i. BSP is infeasible. Add the infeasibility cut (23)

s.t. (9) − (12), (16), (17) and (18), to BMP.

λ) ∞ if the problem is infeasible. ∗ ∗ ∗

where Lemma 1. The optimality cut (24) satisﬁes the following

Zhang et al.: Benders Decomposition for MVPRP

Zhang et al.: Benders Decomposition for MVPRP

Zhang et al.: Benders Decomposition for MVPRP

the planning horizon. Values mL and mU can be de-

Each route r ∈ R̃ has an associated routing cost br .

Zhang et al.: Benders Decomposition for MVPRP

In this section, we describe two ways of generating ∀r ∈ R̂,

Zhang et al.: Benders Decomposition for MVPRP

Zhang et al.: Benders Decomposition for MVPRP

Table 1. Summary Results on the MVPRP

Veh-Ind Non-Veh-Ind 8c-Veh-Ind EXM

l n #ist #opt t %lb #opt t %lb n̂ #opt t %lb n̂ #opt t %lb

3 10-50 72 28 520.4 96.2 24 262.4 96.9 25 31 2,493.2 96.5 40 44 1,141.1 99.3

Zhang et al.: Benders Decomposition for MVPRP

Table 2. Detailed Results on the MVPRP Instances with Three Periods

Adulyasak, Cordeau, and Jans (2014a)

Veh-Ind Non-Veh-Ind EXM

n l m c Best ub %lb t %lb t ub lb %lb t tS %rc

Zhang et al.: Benders Decomposition for MVPRP

Adulyasak, Cordeau, and Jans (2014a)

Veh-Ind Non-Veh-Ind EXM

n l m c Best ub %lb t %lb t ub lb %lb t tS %rc

Zhang et al.: Benders Decomposition for MVPRP

Table 3. Detailed Results on the MVPRP Instances with Six Periods

Adulyasak, Cordeau, and Jans (2014a)

Veh-Ind Non-Veh-Ind EXM

n l m c Best ub %lb t %lb t ub lb %lb t tS %rc

Zhang et al.: Benders Decomposition for MVPRP

Table 4. Detailed Results on the MVPRP Instances with Nine Periods

Adulyasak, Cordeau, and Jans (2014a)

Veh-Ind Non-Veh-Ind EXM

n l m c Best ub %lb t %lb t ub lb %lb t tS %rc

Zhang et al.: Benders Decomposition for MVPRP

Table 5. Analysis of the Different Components of EXM on the MVPRP

Table 6. Summary Results on the Multivehicle VMIRP

Veh-Ind Non-Veh-Ind 8c-Veh-Ind EXM

Zhang et al.: Benders Decomposition for MVPRP

experience with benchmark instances. Networks 65(2):129–138.

Zhang et al.: Benders Decomposition for MVPRP

You might also like