Computers & Operations Research 55 (2015) 141152

Contents lists available at ScienceDirect

Computers & Operations Research

journal homepage: www.elsevier.com/locate/caor

The production routing problem: A review of formulations

and solution algorithms
Yossiri Adulyasak n, Jean-Franois Cordeau, Raf Jans
HEC Montral and CIRRELT, 3000 chemin de la Cte-Sainte-Catherine, Montral, Canada H3T 2A7

art ic l e i nf o a b s t r a c t

Available online 7 February 2014 The production routing problem (PRP) combines the lot-sizing problem and the vehicle routing problem,
Keywords: two classical problems that have been extensively studied for more than half a century. The PRP is solved
Integrated supply chain planning in an attempt to jointly optimize production, inventory, distribution and routing decisions and is thus a
Production routing generalization of the inventory routing problem (IRP). Although the PRP has a complicated structure,
Inventory routing there has been a growing interest in this problem during the past decade in both academia and industry. This
Exact algorithms article provides a comprehensive review of various solution techniques that have been proposed to solve the
Heuristics PRP. We attempt to provide an in-depth summary and discussion of different formulation schemes and of
Review algorithmic and computational issues. Finally, we point out interesting research directions for further
developments in production routing.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction monitors the inventory at retailers and also decides on the replen-
ishment policy for each retailer. The supplier acts as the central
In a typical supply chain which consists of sequential activities decision maker who solves an integrated supply chain planning
of production, storage and distribution, each individual process is problem. The advantage of a VMI policy with respect to the
often planned and optimized using predetermined decisions from traditional retailer managed inventory system lies in a more overall
its preceding activities. For example, a production planner makes efcient resource utilization. The PRP connects two well-known
production lot-sizing decisions in order to minimize production problems, namely the lot-sizing problem (LSP) and the vehicle
and inventory costs at the production facility. The planned lot- routing problem (VRP), to produce an optimal solution when con-
sizing decisions are then used as inputs in subsequent steps of sidering the total system cost. The PRP is also a generalization of the
distribution planning. Since the decisions are limited by the plan lot-sizing problem with direct shipment and of the inventory routing
of the former process, the benets of coordination in the planning problem (IRP). Solving the PRP becomes challenging as it is a
process have been left behind. An integrated supply chain opera- combined version of the LSP and VRP and it incorporates the
tional planning system is a tool that is used to jointly optimize constraints of these two difcult problems. We aim to provide an
several planning decisions thereby capturing the additional ben- in-depth review of the PRP, particularly with respect to the formula-
ets of coordination between sequential activities in the chain. tions and solution algorithms. Different formulation schemes of the
In recent years, many companies, such as Kellogg [21] and Frito- PRP are examined. The approaches to compute lower bounds, exact
Lay [25], have set up integrated planning systems and achieved algorithms and heuristics are thoroughly reviewed. We further
multi-million cost savings. The key to success is an application that discuss future research directions.
is not only able to produce solutions with minimal costs, but that In the rest of this section, we rst provide a brief overview of
can also be used in an effective and timely manner. the three integrated problems. The network representations of
The production routing problem (PRP) is an integrated opera- these problems in the case of a single supplying facility and
tional planning application that jointly optimizes production, inven- multiple customers in a discrete time nite horizon are presented
tory, distribution and routing decisions. It is of practical relevance in a in Fig. 1. Note that the supplying facility can be a plant with setup
Vendor Managed Inventory (VMI) approach, in which the supplier costs and production decisions or a warehouse with xed ordering
costs and ordering decisions. In each period, a single or multiple
products can be made available at the supplying facility and they
n
are transported to the customers in order to satisfy demands. The
Corresponding author.
E-mail addresses: [email protected] (Y. Adulyasak), products can be stored at the plant or at the customers, thus
[email protected] (J.-F. Cordeau), [email protected] (R. Jans). incurring inventory holding costs.

http://dx.doi.org/10.1016/j.cor.2014.01.011
0305-0548 & 2014 Elsevier Ltd. All rights reserved.
142 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

Fig. 1. Network representations of the integrated problems.

1.1. Integrated lot-sizing with direct shipment two-level productiontransportation problems with capacitated
production and vehicles.
In this problem, the products are directly transported from the There is a link between the lot-sizing problem with truckload
manufacturing plant to the customers. The production, setup, cost structure and the classical lot-sizing problem with batch size
inventory and direct shipment costs are minimized over the where the batch quantity is smaller than the maximum produc-
planning horizon. This problem typically incorporates various tion quantity in one period. The truck capacity can be viewed as
production aspects, e.g., production setup cost and/or setup time, the xed batch quantity limit and the cost of dispatching one truck
and involves distribution decisions where the xed and unit costs can also be considered as the xed production cost of one batch.
of delivery are customer specic. There is also a link between the lot-sizing problem with transship-
The integrated production and direct shipment distribution ments and the lot-sizing problem with production substitution
planning was studied by several researchers. Most of the studies where a product can be used to substitute for the demand of
considered the distribution cost as a xed cost or a complex cost another product [48]. The cost of transshipment between custo-
function. Li et al. [56] focused on the lot-sizing problem with a mer locations can be viewed as the cost of production substitution.
piecewise linear transportation cost function where the supplier
has the option to deliver by direct delivery with truckload (TL) or 1.2. Inventory routing problem (IRP)
less-than-truckload (LTL) transportation. They developed a
dynamic programming approach to solve the one-product and When the routing aspect is included and the production aspect
one-customer problem. Jaruphongsa et al. [50] proposed other is disregarded, the problem is transformed into the inventory
dynamic programming algorithms to solve the problem with TL routing problem (IRP). In the IRP, the starting point is a warehouse
and LTL cost structures. A problem with a special cost structure, where there is no production decision as the production quantities
where the supplier can get a discount from transportation capacity made available in each period are typically given. A vehicle can
reservation, was studied by van Norden and van de Velde [79]. visit more than one customer by travelling along its route. As a
A more general piece-wise linear transportation cost function was generalization of the VRP, which consists of the decisions on
addressed by Rizk et al. [67]. They decomposed the integrated delivery quantities and routes to serve customers, the IRP also
problem into uncapacitated lot-sizing and time-independent sub- includes the timing to serve the customers' demands. This makes
problems and applied a Lagrangian relaxation technique to obtain the problem much more difcult than the classical VRP due to the
lower bounds. In the more general case of multiple customers, complex periodic routing and inventory decisions. The IRP is
Chand et al. [26] developed a dynamic programming algorithm to obviously NP-hard since it contains the VRP as a special case [32].
solve the problem in which backlogging is allowed. Jaruphongsa The IRP rst appeared in a gas delivery study by Bell et al. [15].
and Lee [51] considered the problem with split delivery under The problem was solved using a Lagrangian relaxation method and
time window restrictions and employed dynamic programming was decomposed by time period and by vehicle. Carter et al. [24]
algorithms to solve the problem. A special problem of lot-sizing and Campbell and Savelsbergh [22] proposed efcient heuristic
with truckload shipment where transshipments between the procedures by decomposing the IRP into an allocation problem
customers are allowed was considered by Herer and Tzur [47]. (AP) and a vehicle routing problem (VRP). Since the IRP is a
The multi-item problem with one customer was considered by Lee complicated combinatorial problem, several metaheuristics, e.g.,
et al. [53]. In the case of uncapacitated production and uncapaci- tabu search [70], genetic algorithm [1], greedy randomized adap-
tated vehicles, the problem with direct shipments is also known as tive search procedure (GRASP) [71], hybrid heuristic with com-
the one-warehouse multi-retailer problem (OWMR). Federgruen bined tabu search and MIPs [8], and adaptive large neighborhood
and Tzur [36] considered the OWMR with multiple items and search (ALNS) [31,32], have been proposed. Gaur and Fisher [42]
developed a time-partitioning heuristic to solve the problem. discussed a periodic IRP where the demand pattern is repeated
Solyal and Sral [76] proposed a new strong formulation based and developed a heuristic to solve the problem.
on the combined transportation and shortest path model to solve As mentioned in Andersson et al. [6], few exact algorithms have
the OWMR with a single product. Melo and Wolsey [61] discussed been proposed to solve the IRP due to its complexity. Notable
several formulations and proposed hybrid heuristics for the exceptions include a branch-and-cut procedures to solve the IRP
 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152 143

with a single capacitated vehicle by Archetti et al. [7] and Solyal uncapacitated production and a single capacitated vehicle, and
and Sral [75]. Archetti et al. [7] introduced several valid inequal- introduced several valid inequalities to solve the problem. Adu-
ities to solve the problem under three different inventory replen- lyasak et al. [4] focused on the PRP with multiple vehicles and
ishment policies. In the rst policy, called order-up-to level (OU), proposed two branch-and-cut approaches based on different
a visited customer receives exactly the amount which brings its formulation schemes to solve the problem.
inventory up to a predened target stock level (TSL). The second The paper is organized as follows. Section 2 provides the
policy, called maximum level (ML), allows delivery quantities to be problem description and formulations of the basic version of the
any positive value but the inventory at each customer cannot PRP which consists of a network of a single product, a single plant
exceed its maximum stock level. The third policy is similar to the and multiple customers. Solution approaches for several variants
ML policy but there is no maximum stock level imposed at the of the PRP are discussed next. Section 3 provides details on the
customers. Solyal and Sral [75] strengthened the formulation of approaches developed to compute lower bounds and Section 4
the IRP with the OU policy of Archetti et al. [7] by using a shortest reviews recent developments in exact algorithms. Section 5 pre-
path network reformulation. Adulyasak et al. [4] and Coelho and sents various heuristics and recent computational results. We
Laporte [34] extended the approach of Archetti et al. [7] to the IRP further discuss future research opportunities in Section 6 and this
with multiple vehicles. The IRP with transshipment and the IRP is followed by the conclusion.
with several consistency features were also considered by Coelho
et al. [31].
Variants of the IRP have been proposed as well. Christiansen [29] 2. Notation and formulations for the PRP
introduced an IRP application in a maritime context, called the
inventory pickup and delivery problem, and applied a Dantzig 2.1. Description and notation of the PRP
Wolfe decomposition and column generation approach to solve the
problem. Savelsbergh and Song [72] considered the IRP with A PRP network is dened on a complete directed graph G
continuous moves where a product is distributed from a set of N; A where N represents the set of the plant and the customers
plants to a set of customers by multiple vehicles. The authors indexed by i A f0; ; ng and A fi; j : i; j A N; i a jg is the set of arcs.
developed a branch-and-cut approach to solve the problem. We The plant is represented by node 0 and we further dene the set of
refer to Andersson et al. [6] and Coelho et al. [33] for more customers N c N\f0g. Over a nite set of time periods T f1; ; lg,
comprehensive reviews of the IRP. a single product can be produced at the plant and delivered by a
set of identical vehicles K f1; ; mg to the customers to satisfy
1.3. Production routing problem (PRP) the demands in each period. The parameters and the decision
variables are dened as follows.
The two integrated problems discussed in the previous sections Parameters:
each disregard one important aspect of the supply chain opera-
tional planning process, i.e., the integrated lot-sizing problem with u unit production cost;
direct shipment does not incorporate routing decisions, while the f xed production setup cost;
IRP disregards the production part. In the PRP, the plant must hi unit inventory holding cost at node i;
decide in each period whether or not to make the product and cij transportation cost from node i to node j;
determine the corresponding lot size. If production does take dit demand at customer i in period t;
place, this process incurs a xed setup cost as well as unit C production capacity;
production costs. In addition, the lot size cannot exceed the Q vehicle capacity;
production capacity. Deliveries are made from the plant to the Li maximum or target inventory level at node i;
retailers by a limited number of capacitated vehicles and routing Ii0 initial inventory available at node i.
costs are incurred. If products are stored at the plant or at the
retailers, unit inventory holding costs are also incurred. Table 1 Decision variables:
provides a summary of the PRP literature.
The PRP has received more attention in recent years. The benets pt production quantity in period t;
of coordination in the PRP were rst discussed by Chandra [27] and Iit inventory at node i at the end of period t;
Chandra and Fisher [28]. They showed that 320% cost savings can be yt equal to 1 if there is production at the plant in period t,
achieved by solving the PRP compared to sequentially solving the 0 otherwise;
separate problems. As for the case of the IRP, most of the previous z0t the number of vehicles leaving the plant in period t;
studies employed heuristic procedures to solve the problem. Several zit equal to 1 if customer i is visited in period t, 0 otherwise,
metaheuristics, such as GRASP [17], memetic algorithm [19], tabu 8 i A Nc ;
search [11,9], and ALNS [5] have been employed. Archetti et al. [8] xijt if a vehicle travels directly from node i to node j in period t,
discussed the PRP under the ML and OU policies and developed a 0 otherwise;
mixed integer programming (MIP) heuristic to solve the problem. qit quantity delivered to customer i in period t;
Bard and Nananukul [10,12] introduced a heuristic based on a wit load of a vehicle before making a delivery to customer i in
branch-and-price framework. period t.
Due to the complexity of the problem, few studies have
introduced exact algorithms or methods to compute strong lower ~ it minfLi ; Q ;
We further let M t minfC; lj t i A Nc dij g and M
bounds. Fumero and Vercellis [41] and Solyal and Sral [74] lj t dij g.
developed a Lagrangian relaxation approach to obtain lower
bounds based on the multi-commodity ow formulation. Ruoko- 2.2. Formulations for the PRP
koski et al. [69] and Archetti et al. [8] employed a branch-and-cut
approach similar to that of Archetti et al. [7] to solve the PRP. We rst present a model based on the basic LSP and VRP
Ruokokoski et al. [69] explored different lot-sizing reformulations formulations. This model contains a polynomial number of vari-
for the PRP with uncapacitated production and a single uncapa- ables and constraints. The PRP is formulated with variables that
citated vehicle, while Archetti et al. [8] focused on the PRP with control the amounts delivered by a homogenous eet of vehicles.
144 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

Table 1
Summary of algorithms for the PRP.

Author(s) Problem characteristics

Production Inventory Distribution Solution method

N.Plants Product Cap. Policy Cap. Fleet N.Vehs Cap. Type Approach

Chandra [27] Single Multiple ML Hom. Unlimited H Decomposition

Chandra and Fisher [28] Single Multiple ML Hom. Unlimited H Decomposition
Fumero and Vercellis [41]a Single Multiple ML Hom. Limited H/L Lagrangian relaxation
Lei et al. [54] Multiple Single ML Het. Limited H Decomposition
Boudia et al. [17] Single Single ML Hom. Limited H GRASP
Boudia et al. [18] Single Single ML Hom. Limited H Decomposition
Boudia and Prins [19] Single Single ML Hom. Limited H Memetic
Bard and Nananukul [11] Single Single ML Hom. Limited H Tabu search
Solyal and Sral [74] Single Single OU Hom. Limited H/L Lagrangian relaxation
Bard and Nananukul [10,12] Single Single ML Hom. Limited H/L Branch-and-price
Ruokokoski et al. [69] Single Single ML Hom. Single E Branch-and-cut
Armentano et al. [9] Single Multiple ML Hom. Limited H Tabu search
Archetti et al. [8] Single Single ML/OU Hom. Single E/H Branch-and-cut/MIP Heuristic
Adulyasak et al. [3]b Single Single ML Hom. Multiple E Benders-based branch-and-cut
Adulyasak et al. [5] Single Single ML Hom. Multiple H ALNS
Adulyasak et al. [4] Single Single ML/OU Hom. Multiple E/H Branch-and-cut/ALNS
Absi et al. [2] Single Single ML Hom. Multiple H Iterative MIP heuristic

Note. Hom., homogeneous; Het., heterogeneous; H, heuristic, E, exact; L, approach to compute lower bound.
a
Considered a PRP variant with unit transportation costs.
b
Considered a two-stage stochastic PRP with demand uncertainty.

A basic formulation based on that of Bard and Nananukul [10,12] is 0 0

as follows.
!
PRP1 : min upt fyt hi I it cij xijt 1 1
tAT iAN i;j A A
1
3 3
s:t: I 0;t  1 pt qit I 0t 8t AT 2
i A Nc
2 2
I i;t  1 qit dit I it 8 i A Nc ; 8 t A T 3

pt r M t yt 8t AT 4
Fig. 2. Illustration of vehicle restriction and subtour elimination constraints (11).

I 0t r L0 8t AT 5
Constraints (4) are the setup forcing and production capacity
I i;t  1 qit rLi 8 i A Nc ; 8 t A T 6
constraints. The constraints force the setup variable to be one if
~ it zit production takes place in a given period and limit the production
qit r M 8 i A Nc ; 8 t A T 7
quantity to the minimum value between the production capacity
and the total demand in the remaining periods. Constraints
xijt zit 8 i A Nc ; 8 t A T 8
jAN (5) and (6) limit the maximum inventory at the plant and
customers, respectively. The inventory part of this model is
xjit xijt 2zit 8 i A N; 8 t A T 9 controlled by the so-called maximum level (ML) policy as dened
jAN jAN
by Archetti et al. [7], where the delivery quantity can be any
positive number but the resulting inventory level after delivery
z0t r m 8t AT 10
prior to demand consumption cannot exceed the maximum
~ it 1  xijt
wit  wjt Z qit  M 8 i; j A A; 8 t A T 11 inventory level. The remaining constraints, i.e., (7)(12), are the
vehicle loading and routing restrictions. Constraints (7) allow a
0 r wit r Qzit 8 iA Nc ; 8 t A T 12 positive delivery quantity only if customer i is visited in period t
and each customer can be visited by at most one vehicle (8).
pt ; I it ; qit Z0 8 i A N; 8 t A T 13 Constraints (9) are the vehicle ow conservation. Constraints (10)
limit the number of trucks that can be used to the number of
yt ; xijt A f0; 1g 8 i; j A N; 8 t A T 14 available trucks. Constraints (11) are the vehicle loading restric-
tions and subtour elimination constraints in the form of the
zit A f0; 1g 8 i A Nc ; 8 t A T 15 MillerTuckerZemlin inequalities [63]. These constraints do not
allow taking an arc that generates a subtour as shown in Fig. 2; the
z0t A Z 8 t A T: 16
arc 3; 1; t cannot be taken because w1t  w2t Z q1t is not valid.
The objective function (1) minimizes the total production, Constraints (12) are the vehicle capacity constraints.
setup, inventory and routing costs. Constraints (2)(6) represent In Bard and Nananukul [10,12], the subtour elimination con-
the lot-sizing part of the problem. Constraints (2) and (3) are the straints (11) and vehicle capacity (12) constraints are used. How-
inventory ow balance at the plant and customers, respectively. ever, in the VRP, this subtour elimination constraint set can lead to
 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152 145

I i;t  1 qikt dit I it 8 i A Nc ; 8 t A T 22

kAK

pt rM t yt 8t AT 23

I 0t r L0 8t AT 24

I i;t  1 qkit rLi 8 i A Nc ; 8 t A T 25

kAK

~ it zikt
qikt r M 8 k A K; 8 i A N c ; 8 t A T 26

zikt r 1 8 i A Nc ; 8 t A T 27
kAK

Fig. 3. Illustration of FCCs (left) and GFSECs (right).

xjikt xijkt 2zikt 8 k A K; 8 iA N; 8 t A T 28
jAN jAN

a weak formulation in the routing part [78]. Constraints (11) and (12) xijkt r jSj 1 8 S D Nc : jSj Z 2; 8 k A K; 8 t A T 29
can be replaced with other subtour elimination constraints, i.e., i A Sj A S

 fractional capacity constraints (FCCs) [55] as presented in qikt rQz0kt 8 k A K; 8 t A T 30

i A Nc
Chandra and Fisher [28]:
xijt Z qit =Q 8 S D Nc : jSj Z 1; 8 t A T 17 pt ; I it ; qikt Z0 8 i A N; 8 k A K; 8 t A T 31
i2
= Sj A S iAS
yt ; zikt ; xijkt A f0; 1g 8 i; j A N; 8 k A K; 8 t A T: 32
 generalized fractional subtour elimination constraints (GFSECs): The objective function (20) and constraints (21)(28) are
equivalent to the objective function (1) and constraints (2)(9) in
xijt rjSj qit =Q 8 S D N c : jSj Z2; 8 t A T: 18
i A Sj A S iAS the PRP1, respectively. Constraints (29) are in fact the subtour
elimination constraints (SECs) similar to those of the travelling
salesman problem (TSP) and constraints (30) impose the vehicle
capacity on each vehicle. Archetti et al. [7] and Adulyasak et al. [4]
Fig. 3 depicts how these constraints eliminate subtours. Sup-
found that the following form of the SECs, which was originally
pose that q1t q2t q3t =Q 1. The FCC forces one arc from
developed for the selective TSP [44], is more efcient than (29)
outside the set S to connect to a node inside S, while the GFSEC
when solving the problem with a branch-and-cut algorithm:
does not allow more than two arcs in the set S. Both of the
constraints can eliminate subtours in the set S. xijkt r zikt zekt 8 S DN c : jSj Z 2;
i A Sj A S iAS
Unlike in the VRP where the delivery quantity to each customer
8 e A S; 8 k A K; 8 t A T: 33
is known a priori, the value i A S qit =Q cannot be rounded up in
the PRP because the delivery quantity qit is a decision variable and Another advantage of this formulation is that it can be used to
this would result in a non-linear formulation. Adulyasak et al. [4] solve the problem with a heterogeneous eet of vehicles, e.g.,
used a strengthened version of constraints (18) where the para- various eet size, consumption or costs. In that case, the trans-
meter jSj is replaced with i A S zit . They also observed that the portation costs cij and the vehicle capacity Q in the formulation
form (18) is numerically unstable due to the fractional term qit =Q . PRP2 can be replaced with the cost and capacity associated with
The alternative form of the GFSECs presented in Adulyasak et al. [4] each vehicle size k, i.e., cijk and Qk, respectively.
is as follows:
Q xijt r Qzit  qit 8 S D N c : jSj Z2; 8 t A T: 19 2.3. Remarks on LSP and VRP formulation schemes
i A Sj A S iAS
Because the PRP includes the structure of the LSP and VRP, this
However, using these constraints in the formulation instead of (11),
section provides a brief summary of the different formulation schemes
the problem becomes much larger due to an exponential number of
for these two problems. For the LSP, a review of basic formulation and
subsets. A branch-and-cut (BC) procedure is typically an efcient
reformulation schemes was presented by Pochet and Wolsey [66].
approach to solve the problem. In this procedure, these constraints
In short, the basic LSP formulation is a weak formulation that gives
are initially removed and added iteratively during the branch-and-
poor quality lower bounds. Many reformulation schemes can be used
bound process. Note that in the case where an undirected graph is
to strengthen the formulation. The major LSP reformulation schemes
assumed, the formulations can be converted by using the method
considered for the PRP include the shortest path [35] and facility
presented by Toth and Vigo [78].
location [52] reformulations. In the single level uncapacitated LSP,
To overcome the disadvantage of constraints (19) which are
these two formulations have the integrality property, i.e., feasible
fractional, one can explicitly formulate the problem with a vehicle
mixed-integer solutions are obtained by solving the LP relaxation. The
index and impose routing constraints on each vehicle separately.
efciency of different LSP reformulations in the PRP was studied by
In this formulation, the variables qikt, zikt and xijkt have the same
Ruokokoski et al. [69].
interpretation as qit, zit and xijt but they are associated with vehicle
In the VRP, different formulations are typically used to solve the
k only. The formulation with a vehicle index based on the
problems with different characteristics. For example, the basic
formulation presented by Boudia et al. [17,18] is as follows.
! formulation is a compact formulation that is suitable for a homo-
geneous eet and a limited number of vehicles. To handle a
PRP2 : min upt fyt hi I it cij xijkt 20
t AT iAN i;j A A kAK
heterogeneous eet (e.g., various eet size, consumption, speed, or
costs), one should employ a formulation in which a vehicle index can
s:t: I 0;t  1 pt qikt I 0t 8t AT 21 be incorporated, e.g., a multi-commodity ow formulation. In some
i A Nc k A K VRP applications in which the vehicle routes are predened or have
146 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

few possibilities, e.g., a truck can only serve the customers in the indexed by r, is characterized by the delivery quantity to each
same cluster, or in maritime applications where the vessel schedule customer and routing decisions. The binary variable rt is equal to
option is predened, it is more appropriate to use a set-partitioning one if the delivery plan r for period t is selected. The parameter crt is
formulation. This formulation is also generally used as the master the total cost of using delivery plan r in period t, and irt is the amount
problem in a column generation process. Another issue in VRP delivered to customer i with delivery plan r in period t. The RMP is
formulations is the set of subtour elimination constraints. Applying formulated as follows.
different subtour elimination constraints can result in different lower !
bounds [55]. As in the VRP, some formulations were used in the PRP min upt fyt hi I it crt rt 34
to deal with specic issues, e.g., a formulation with a vehicle index tAT iAN r A Rt

for the PRP with heterogeneous eet by Lei et al. [54] and a path- s.t. (4)(6), (13) and
based formulation used in the column generation approach of Bard
and Nananukul [10,12]. I 0;t  1 pt irt rt I 0t 8t AT 35
i A N c r A Rt

3. Approaches to compute lower bounds I i;t  1 irt rt dit I it 8 iA N c ; 8 t A T 36

r A Rt

As the PRP is a complicated combinatorial optimization pro- rt r 1 8t AT 37

blem containing a large number of binary variables, the quality of r A Rt

the lower bound of the basic PRP formulation obtained by solving

the LP relaxation is generally very poor. Nananukul [65] found that rt A f0; 1g 8 t A T; 8 r A Rt: 38
the LP relaxation of the formulation PRP1 is not practical in The objective function (34) and constraints (35)(36) are
providing relaxed solutions in exact algorithms such as branch- equivalent to (1)(3), respectively. At most one delivery plan can
and-bound or measuring the quality of other solution approaches. be selected in each period (37).
Hence, alternative relaxation methods have been developed to The subproblem is the delivery schedule generator. Denote by
obtain better lower bounds. 1t ; 2it and 3t the dual variables of the RMP associated with the
constraints (35), (36) and (37), respectively. The subproblem is
3.1. Lagrangian relaxation decomposed into a VRP subproblem in each time period. The
subproblem for time period t, referred to as SVRPt, is as follows.
Lagrangian relaxation (see [40]) is an approach to obtain lower
min cij xijt  1t 2it qit 3t 39
bounds by dualizing constraints with Lagrangian multipliers and i;j A A i A Nc
decomposing the problem into subproblems which are more
easily solvable. A Lagrangian relaxation for a variant of the PRP, s.t. (13)(16) and
where unit transportation costs are assumed, was proposed by ~ it zit
qit r M 8 iA N c 40
Fumero and Vercellis [41]. They decomposed the basic LSP and the
multi-commodity ow VRP reformulation into subproblems by xijt zit 8 i A Nc 41
dualizing the plant inventory constraints and the vehicle capacity jAN

constraints. The problem is then transformed into four subpro-

xjit xijt 2zit 8iAN 42
blems, i.e., production (PROD), inventory (INV), distribution (DIS), jAN jAN
and routing (ROU) subproblems. The rst two subproblems can be
solved by inspection and the DIS subproblem can be solved by an z0t r m 43
LP solver. The lower bound of the ROU subproblem is calculated by
the minimum cost network ow problem. Instances with up to ~ it 1  xijt
wit  wjt Zqit  M 8 i; j A A 44
8 periods, 12 customers and 10 products were tested and the
algorithm could obtain lower bounds with an average gap of 5.5% 0 r wit r Qzit 8 i A Nc : 45
compared to an upper bound obtained by a heuristic. Constraints (40)(45) are equivalent to constraints (7)(12).
A similar Lagrangian relaxation approach was used by Solyal A lower bound on the original problem can be obtained when the
and Sral [74] to solve the PRP with the order-up-to level (OU) integrality constraints on the variables rt are relaxed and the
policy. However, the lower bounds obtained by this approach were problem is solved to optimality. However, in the experiments of
weak compared to the case where the unit transportation costs are Nananukul [65] on instances with up to 10 customers, 6 periods and
used as in Fumero and Vercellis [41]. On the instances with 5 vehicles, this relaxation could only give a lower bound which on
8 customers and 5 periods, the lower bound produced by the average was 0.43% higher compared to the original model PRP1.
Lagrangian relaxation has an average deviation of 33.16% from the
optimal value.
4. Exact solution algorithms
3.2. Column generation
Exact solution algorithms for PRP are very scant. To the best of
In a column generation procedure, a basic formulation is our knowledge, three exact algorithms were proposed to solve the
decomposed into a restricted master problem (RMP) and subpro- PRP: the branch-and-cut with strong reformulation of Ruokokoski
blems. The original variables are replaced with a convex combina- et al. [69] for the PRP with a single uncapacitated vehicle, the
tion of extreme points of the subproblems which are generated branch-and-cut of Archetti et al. [8] for the PRP with a single
and added iteratively by solving the subproblems. More details capacitated vehicle, and the branch-and-cut of Adulyasak et al. [4]
about recent general column generation approaches can be found for the PRP with multiple capacitated vehicles. Although the
in Lbbecke and Desrosiers [58]. branch-and-price algorithm of Bard and Nananukul [10] can be
Bard and Nananukul [12] proposed a RMP and subproblem used to obtain an optimal solution, the subproblem SVRPt employs
formulations for the PRP and developed a branch-and-price procedure. the VRP structure and it is very time consuming to solve the
Let R(t) be the sets of delivery plans in period t where a delivery plan, problem to optimality. In their experiments, only instances with
 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152 147

10 customers, 2 periods and 5 vehicles were solved to optimality. algorithm is tested on generated test instances with 14 customers
Therefore, Bard and Nananukul [10,12] used a heuristic to handle and 6 periods. Most of the instances are solved to optimality
this subproblem and developed a heuristic based on a branch-and- within a few seconds.
price framework. We later provide the summary of their heuristic
algorithm in Section 5.1.2. We also discuss the Benders decom- 4.3. Branch-and-cut algorithms of Adulyasak et al. [4]
position algorithm of Adulyasak et al. [3] that was proposed to
handle the PRP with demand uncertainty in Section 4.4. Adulyasak et al. [4] extended the branch-and-cut approach of
Archetti et al. [8] to the PRP with multiple vehicles. They compared
4.1. Branch-and-cut algorithm of Ruokokoski et al. [69] the performance of two formulations. The rst one is the formu-
lation without a vehicle index, i.e., the PRP1 with constraints
The problem with an uncapacitated plant and a single uncapa- (11) and (12) being replaced by the GFSECs (19). The authors also
citated truck was considered by Ruokokoski et al. [69]. They used developed inequalities to strengthen the routing part of the
the stronger LSP reformulations, e.g., facility location and shortest formulation. The second one is the formulation with a vehicle
path reformulations, to solve the PRP. The formulation is similar to index, i.e., the PRP2 with constraints (11) and (12) being replaced
PRP2 but the vehicle index k is dropped and the subtour elimina- by (19). This formulation is enhanced with symmetry breaking
tion constraints (29) are replaced with constraints which disallow alternative solutions that can be
created due to the fact that vehicles are identical.
xijt xijt Z 2zet 8 S D N c : jSj Z2; 8 eA S; 8 t A T: 46
i2
= Sj A S i A Sj 2
=S Both formulations are further enhanced with the inequalities of
Archetti et al. [8]. For the formulation without a vehicle index, the
Ruokokoski et al. [69] investigated the quality of lower bounds authors employed the separation algorithms of Lysgaard et al. [59],
by using different LSP reformulation schemes compared to the which were developed for the VRP, and they also developed a
basic LSP formulation. The computational results show that the LP greedy heuristic algorithm for the new cuts. The formulation with
relaxation values of the shortest path and facility location refor- a vehicle index is solved by a branch-and-cut algorithm similar to
mulations when the subtour elimination constraints are dropped that of Ruokokoski et al. [69]. The experiments showed that the
are much improved compared to the basic formulation. The LP formulation with a vehicle index is superior in nding optimal
bound obtained by the shortest path reformulation is greater than solutions. Instances with up to 35 customers, 3 periods and
or equal to the LP bound of the facility location reformulation 3 vehicles were solved to optimality in 2 h. The formulation
which follows from the proof in Solyal and Sral [76], but the without a vehicle index, however, could generally produce better
difference between the bounds is very small. The facility location lower bounds at the root node and found better lower bounds for
reformulation provides better computational performance when the instances that were not solved to optimality in 2 h. The
adding the valid inequalities for the routing problem (including authors also tested the performance of the formulation with a
the subtour elimination constraints) and solving the integer vehicle index with an eight-core machine using parallel comput-
problem using a branch-and-cut process. ing and instances with up to 50 customers, 3 periods and 3 vehicles
In their branch-and-cut implementation, the authors used a were solved to optimality in 12 h. The approaches of Adulyasak
heuristic and an exact separation procedure based on a minimum et al. [4] were also adapted for the PRP with the OU policy and IRP
s  t cut problem to detect subtours. They also adapted the with the ML and OU policies.
generalized comb and 2-matching inequalities presented by [39]
which were developed for a generalized travelling salesman 4.4. Benders-based branch-and-cut for the stochastic production
problem (GTSP) to the PRP. The results show that when all valid routing problem of Adulyasak et al. [3]
inequalities are used, the facility location LSP reformulation
provides the best LP relaxation. This algorithm can solve problems Adulyasak et al. [3] specically addressed the PRP with demand
with 80 customers and 8 periods within about 30 min. uncertainty and introduced the stochastic PRP (SPRP) under demand
uncertainty in a two-stage decision process. The rst stage consists of
4.2. Branch-and-cut algorithm of Archetti et al. [8] making setup and routing decisions before the realization of demand,
and the second stage involves production and delivery quantity
Archetti et al. [8] studied the PRP with uncapacitated produc- decisions made when the demand for the entire planning horizon
tion and a single capacitated vehicle. They used the formulation becomes known. They developed exact solution approaches based on
PRP2 but with the SECs (29) being replaced with (33) and the Benders decomposition to solve the problem and two different
vehicle index k is dropped. The authors also added the inequalities Benders reformulation schemes were proposed. The rst Benders
of Archetti et al. [7] for the inventory routing problem to reformulation separates the rst- and second-stage decisions in the
strengthen the inventory and routing parts including the following master and subproblems, respectively. In the second reformulation,
inequalities: the routing decisions are also projected out into the subproblems. The
!
s s two Benders reformulations were enhanced with inequalities to
I i;t  s  1 Z di;t  j 1  zi;t  j 8 i A N c ; 8 t A T; s 0; 1; ; t  1: improve the lower bound, aggregate Benders cuts using scenario
j0 j0
groups, and Pareto-optimal cuts [60]. The Benders algorithm was
47
implemented within a branch-and-cut framework in which Benders
The inequalities (47) can be interpreted as follows: if there is no cuts are generated at the nodes of the branching tree for the master
shipment delivered during the time interval t  s; t, the inventory problem. The computational experiments show that this imple-
level in period t  s  1 must be sufcient to satisfy the demand in mentation outperforms the standard implementation of the Bend-
this interval. By adding them to the PRP, they could strengthen the ers algorithm where the master problem is solved from scratch at
lot-sizing part of the customer replenishment and provide better each iteration. This new algorithm also provides superior results to
lower bounds. In Archetti et al. [8], they also proposed inequalities the branch-and-cut approach of Adulyasak et al. [4] when solving a
specically for the PRP with uncapacitated production. large number of scenarios. They further discussed the reoptimiza-
In the branch-and-cut process, the subtour elimination con- tion capabilities of the Benders approach which can be particularly
straints (33) are removed and only the violated cuts are added useful in two stochastic environments, namely, a sample average
iteratively during the branching process. The performance of the approximation scheme (SAA) to handle a large number of scenarios,
148 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

and a rolling horizon framework (RH) for a dynamic and stochastic 5.1.2. Branch-and-price heuristic
variant of the PRP. Bard and Nananukul [10,12] presented a heuristic based on the
branch-and-price framework using the RMP and subproblems as
described in Section 3.2. The branch-and-price scheme is a
decomposition based procedure which involves a branching pro-
5. Heuristics cess. At each branching node, column generation is performed to
add variables to the RMP and this updated RMP is solved again
This section provides a comprehensive review of the heuristics until an optimal LP relaxation value of the node is found. Then, the
for the PRP. We categorize the heuristics into two groups, i.e., branching process continues until an optimal solution to the
general heuristics and metaheuristics. The details on publicly original problem is obtained. The readers are referred to Barnhart
available benchmarks and recent computational results of the et al. [13] for more details on branch-and-price.
heuristics are presented in Section 5.3. The branching process starts with the production setup vari-
ables (yt) until all these variables have integer values. Subse-
quently, the variables rt are considered. Branching on rt directly,
5.1. General heuristics however, results in an unbalanced branching tree. When the
variable rt is set to one, the delivery plan r is used and all
This section presents a review of general heuristics for the PRP, corresponding qit and xijt variables are xed. But when rt 0, it is
i.e., decomposition-based approach, branch-and-price heuristic, very difcult to manage the variables to only exclude the delivery
MIP heuristic and iterative MIP heuristic. plan r. Therefore, it is more appropriate to branch on the xijt
variables. This branching scheme is similar to the branch-on-edge
approach presented in Bramel and Simchi-Levi [20]. The depth-
5.1.1. Decomposition-based heuristics rst-search strategy is used to quickly nd an incumbent solution.
These heuristics solve the PRP by decomposing the problem To improve the branch-and-price procedure, several features are
into production and distribution planning subproblems. The initial included in the process. First, an initial solution is generated by the
solution is obtained by sequentially solving each problem and a tabu search heuristic presented by Bard and Nananukul [11].
heuristic procedure is called to improve the solutions. This Second, during the column generation process, instead of solving
approach was rst introduced by Chandra [27] and Chandra and the subproblems to optimality, they are solved by the separation
Fisher [28] to solve the multi-product PRP. The integrated problem based heuristic algorithms of Bard and Nananukul [10]. Third, the
is decoupled into the capacitated lot-sizing problem and the branching scheme is modied to branch on groups of variables.
distribution scheduling problem. The lot-sizing problem is solved And fourth, a rounding heuristic procedure is used. With these
to optimality and a distribution schedule for each period is modications, the performance of the branch-and-price process is
produced by applying a simple heuristic together with a 3-opt substantially improved. The experiments on instances with up to
procedure [57]. The result is further improved by allowing pro- 50 customers and 8 periods showed that this branch-and-price
duction shifting across periods if the total cost is reduced. This heuristic provides better solutions than those obtained by CPLEX
heuristic algorithm provides approximately 6% cost savings com- (solution costs are improved by 12.2% on average) within 1 h of
pared to the uncoordinated approach with no improvement computing time.
heuristic procedure on the small test instances.
Instead of focusing only on the lot-sizing part in the rst phase
5.1.3. MIP-based heuristic
of the algorithm, Lei et al. [54] incorporated the distribution part
Archetti et al. [8] decomposed the PRP into the uncapacitated
and developed a two-phase heuristic to solve the PRP with
lot-sizing and inventory-routing subproblems and developed a
multiple plants and a heterogeneous eet of vehicles. In the rst
heuristic to solve the decomposed problems. The algorithm starts
phase, they assume that the deliveries are made by direct ship-
by xing production quantities equal to the demand in each period
ments from plants to customers and solve the integrated lot-sizing
and solving the IRP by a heuristic procedure. In this process, each
problem with direct shipment. In the second phase, the decisions
retailer is selected sequentially and a search tree is explored to
in the rst phase except the direct transportation routes are xed
determine the time periods and vehicles used to serve that
and the authors used a VRP heuristic to determine the routes for
retailer. After that the uncapacitated lot-sizing subproblem is
each vehicle at each plant in each period. In the experiments using
solved to further explore whether the production plan can be
instances with one plant and up to 12 customers, 2 vehicles and
improved by shifting some production quantity to reduce the
4 periods, this approach could generally provide better solutions
production and inventory costs. A heuristic procedure is applied to
with much shorter computing times compared to the solutions
the current solution to obtain further improvements by removing
obtained by solving the full model with CPLEX. They also tested
two retailers, and then a problem is formulated to nd the
the algorithm using a real world dataset.
minimum insertion cost of these retailers. If the total cost is
Boudia et al. [18] developed an improved decomposition based
reduced, the uncapacitated LSP subproblem is solved again and
approach by rst determining production lot sizes as large as
the process is repeated until there is no improvement. The authors
possible to cover some future periods. The distribution plan in each
evaluated the performance of this algorithm by comparing it to the
period is constructed by the savings algorithm [30]. The algorithm
best solutions found by the exact branch-and-cut solution proce-
nds opportunities to reduce production costs by adopting the
dure as described in Section 4.2 on instances with 14 customers,
WagnerWhitin algorithm [80] for the LSP. Then, a local search
6 periods and one vehicle. This heuristic provides solutions within
procedure based on 3-opt moves, insertion, and swap heuristics
1% of optimality in a few seconds.
is called to improve the solution. The algorithm is called H2 and
it is tested on the large instances generated by Boudia et al. [16]
with 50200 customers and 20 periods. It provides 1015% cost 5.1.4. Iterative MIP heuristic
savings compared to the two-phase decoupled heuristic, called H1, Absi et al. [2] introduced an iterative MIP heuristic to solve the
which basically provides a solution from the production plan PRP with uncapacitated production. The MIP is formulated by
identied by the WagnerWhitin method, and the delivery plans replacing the routing costs and variables in the original PRP model
generated by a 3-opt procedure. with xed costs SCvit representing the approximate cost of visiting
 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152 149

customer i in period t with vehicle v. At each iteration, this MIP is solutions. The basic idea of a genetic algorithm is to generate new
solved to optimality (or until a maximum CPU time is reached) to solutions from a population of initial solutions which are repre-
obtain production, inventory and customer visit decisions. Then, a sented by chromosomes (or bitstrings) using natural evolution, i.e.,
heuristic is called to determine the routes to serve the visited crossover or mutation, to create new offsprings. In a memetic
customers. Next, the costs SCvit are set to a minimum insertion cost algorithm, a local search procedure is additionally applied to
corresponding to the current solution and the next iteration is improve both the initial population and offsprings of the genetic
performed. A diversication mechanism that modies visit costs is algorithm. This approach was rst introduced by Moscato [64].
also applied when the current solution is not improved for a Boudia and Prins [19] applied this approach to the PRP with a
certain number of iterations. The process stops when a maximum special feature, called memetic algorithm with population man-
number of iterations is reached. The authors proposed two agement (MA-PM) [77].
variants of the iterative approach. In the rst variant, an individual In the study of Boudia and Prins [19], an initial population is
vehicle capacity is taken into account when solving the MIP and created through a simple heuristic procedure that preliminarily
each route is determined by a TSP heuristic, while an aggregate sets a production plan in each period equal to the total demand.
vehicle capacity is considered and the routes are determined by a Then, a savings heuristic is used to generate the delivery plan and
VRP heuristic in the second variant. The results show that the the production plan is adjusted by the WagnerWhitin algorithm.
second variant is superior to the rst and can generally produce Next, the algorithm uses a crossover to generate new offsprings.
better solutions in a shorter computing time. Although the In the following step, the authors adopted the local search of
performance of this algorithm relies heavily on the performance Boudia et al. [17] to improve the generated offsprings and applied
of the MIP solver, the results show that it outperformed other the population management approach to select them, i.e., the new
heuristics on the data set with uncapacitated production of offsprings are accepted only if they improve the current solution
Archetti et al. [8]. by more than a threshold value. The algorithm terminates when
the maximum number of iterations is reached.
5.2. Metaheuristics
5.2.3. Tabu search
We rst describe the different metaheuristics in Sections 5.2.1 The concept of tabu search was introduced by Glover [45].
5.2.4 and then provide a summary of the computational results of In this procedure, the search moves at each iteration from the
these metaheuristics in Section 5.3. current solution to the best neighbor solution. In order to avoid
cycling and to get out of local optima, all visited solutions are
5.2.1. Greedy randomized adaptive search procedure (GRASP) stored in a tabu list and these solutions are forbidden from the
GRASP was rst introduced by Feo and Resende [37]. Basically, search procedure. The tabu search approach is known to be one of
the procedure consists of two main phases, i.e., construction and the most efcient solution methods for the VRP [43].
local search. In the construction phase, an initial solution is A tabu search for the PRP was proposed by Bard and Nananukul
provided via an adaptive randomized greedy algorithm. Then the [11]. They used a reactive tabu list of variable size and this method is
local search phase is applied to improve the solution. called reactive tabu search (RTS) [14]. In their algorithm, an initial
A GRASP for the PRP was developed by Boudia et al. [17]. In the solution is created by solving a modied PRP obtained by dropping
construction phase, an initial solution is generated by sequentially the routing constraints (8)(12), removing variables xijt, and assum-
developing a production and delivery plan. Starting from the rst ing that the delivery cost is equal to the round trip transportation
period onwards, the production plan is preliminarily determined by cost. Then, a subsequent routing decision is made by applying a
producing a sufcient amount to cover the demand in the period capacitated vehicle routing problem (CVRP) subroutine based on a
without excess production which creates inventory at the plant. tabu search proposed by Carlton and Barnes [23]. The generated
Then, delivery routes in the period are constructed by an insertion solution is improved by a local search procedure using swap or
algorithm. Next, the algorithm checks if some customer demands can transfer moves. The swap move examines two customers in two
be moved to prior periods without violating production, inventory periods and exchanges maximum possible delivery quantities
and vehicle capacities and the insertion process is again performed to between these two customers. The transfer move nds the delivery
insert these quantities. Then, the saving heuristic is called to nd a quantity of a customer that can be combined with another delivery
better routing solution. All routing plans are xed and the production in a previous period in order to reduce the transportation cost. The
plan improvement algorithm is applied to shift production quantities moves that lead to an improved solution are stored in the tabu list
to combine with a production plan in earlier periods if the cost is and infeasible solutions are not allowed.
lower, i.e., the incurred storage cost at the plant is less than the setup Armentano et al. [9] developed a tabu search with path relinking
cost. In the local search phase, the routing plan of each period is (TSPR) procedure for the PRP. In their algorithm, an initial solution is
improved by using a 3-opt procedure, inserting, and swapping. created by setting delivery quantities equal to demands and applying
Moves across periods are also considered if the cost can be reduced. the Wagner and Whitin [80] algorithm and the Clarke and Wright [30]
Boudia et al. [17] also developed a path relinking procedure (see [46]) savings algorithm to obtain the production and routing decisions,
as a post-processor. In this process, solutions obtained during the respectively. In the neighborhood search of Armentano et al. [9],
GRASP are ranked according to their total cost and a limited number a move similar to the transfer move of Bard and Nananukul [11] is
of solutions are stored in a pool of elite solutions. Then, any two used but it also allows a delivery quantity to be combined with
solutions in the pool are chosen to create a new solution by another delivery in a future period. At each iteration, the algorithm
transferring some delivery quantities in one of these solutions to also solves an LP to optimize the production and inventory quantities
another period according to the delivery quantities in the other at the plant. The local search terminates when it reaches the maxi-
solution to reduce the differences between these two solutions. This mum number of iterations. Then, a path relinking procedure similar to
process could slightly improve the solutions obtained by GRASP. that of Boudia et al. [17] is used to diversify the search.

5.2.2. Memetic algorithm (MA) 5.2.4. Adaptive large neighborhood search (ALNS)
Informally speaking, a memetic algorithm is a modied genetic The adaptive large neighborhood search (ALNS) framework was
algorithm (GA) that uses some form of local search to improve introduced by Ropke and Pisinger [68] to solve the VRP with
150 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

pickup and delivery. The basic idea of the ALNS is to repeatedly solutions is not comparable to the metaheuristics. The best
destroy and repair a part of a solution to obtain an improved solutions are provided by the ALNS by Adulyasak et al. [5].
solution using search operators. These operators are probabilisti- Although this algorithm spent higher computing times on average
cally selected based on empirical scores. This procedure was compared to other metaheuristics, better solutions were obtained
recently applied to various routing applications. in the early stage of the algorithm as reported in Adulyasak et al. [5],
The ALNS of Adulyasak et al. [5] incorporated three main where the average computing times are comparable to those of the
features. First, an enumeration scheme is used to create several GRASP developed by Boudia et al. [17].
different initial solutions. Each initial solution is generated by Table 4 shows the results obtained by the heuristic H of
solving the two decomposed problems, similarly to the approach Archetti et al. [8], the ALNS of Adulyasak et al. [5] and the best
of Bard and Nananukul [11], and local branching inequalities [38] variant of the iterative MIP heuristic (IM) of Absi et al. [2] on the
are applied to generate another solution with a different produc- Archetti et al. [8] benchmark test set. For the 467 instances of A1
tion conguration. Second, the authors developed two types of solved to optimality, the rst variant of IM provides an average gap
operators, called selection and transformation. At each iteration, of 1.22%, whereas H and ALNS have a gap of 1.71% and 2.66%
one operator of each type is probabilistically selected. The selec- respectively. For the multi-vehicle instances A2 and A3, the best
tion operator is applied rst to create a list of node candidates version of the IM heuristics is the second variant. We note that for
(customer-period combinations) and then the transformation the uncapacitated production problems, the IM heuristic of Absi
operator is applied to remove and reinsert node candidates in et al. [2] generally provides the best solutions.
the list to the current solution. Third, when a new solution is found
during the transformation process, a minimum cost ow (MCF)
problem is solved to optimize the production, inventory and 6. Future research directions
delivery quantities. The algorithm stops when it reaches the
maximum number of iterations. Since the quantity ow part is As an integration of several areas in production and distribu-
optimized by the MCF, this algorithm is referred to as an tion planning, there are many future research directions that have
optimization-based ALNS (Op-ALNS). not been explored. We summarize them here.

Exact algorithms. Although some exact algorithms were proposed,

5.3. Benchmark instances and computational evaluation
the best ones are still limited to medium instance size
of the heuristics
with a relatively small number of time periods. Several
techniques that were used to enhance these algorithms
In this section, we provide details on the benchmark instances
were adapted from the methods developed for the LSP
that are publicly available as well as recent computational results.
and VRP. However, inequalities or formulations that
The rst benchmark was introduced by Boudia et al. [16] and it
strengthen the coordination between these two parts
consists of three problem sets with 50, 100, 200 customers and 20
of the problem can be further explored. Moreover, a
time periods; there are 30 instances per set. This dataset was used
complex decomposition based approach like branch-cut-
in the experiments of Boudia et al. [18] and all the metaheuristics
and-price may be effective to handle the PRP. By applying
described in the previous section. This dataset can be found on the
this approach to the formulations where subtour elimina-
website: https://sites.google.com/site/YossiriAdulyasak/publications.
tion constraints are used, a column generation technique
The second benchmark was introduced by Archetti et al. [8]. It consists
of sets of problems with 14, 50, 100 customers and 6 time periods.
These instances are smaller compared to those of Boudia et al. [16].
However, the set contains instances with four different aspects, i.e.,
Table 3
standard, high production cost, high transportation cost and zero Average computing times in seconds obtained by different heuristics on the Boudia
customer inventory cost. This data set was used in the experiments of et al. [16] benchmark.
Archetti et al. [8], Absi et al. [2] and Adulyasak et al. [5] and it can be
downloaded at: http://www-c.eco.unibs.it/ bertazzi/ml.zip. Set jN c j l H1a H2a GRASPa MA-PMa RTSb TSPRc ALNSd

A comparison of the computational performance of the heur- B1 50 20 0.1 0.2 93.5 172.7 330.6 317.0 481.3
istics of Boudia et al. [18] and the metaheuristics on the Boudia B2 100 20 0.5 1.1 415.9 1108.1 975.6 1147.6 1569.9
et al. [16] dataset is reported in Tables 2 and 3. The computational B3 200 20 2.1 10.4 1893.8 4098.5 2492.3 3926.4 5794.2
experiments were conducted with the test instances from Boudia a
Executed on 2.30 GHz PC.
et al. [16]. All test evaluations were performed on workstations b
Executed on 2.53 GHz PC.
with comparable CPU performances. The decomposition-based c
Executed on 2.80 GHz PC.
heuristics H1 and H2 were very quick but the quality of the d
Executed on 2.10 GHz Duo CPU PC.

Table 2
Average total costs obtained by different heuristics on the Boudia et al. [16] benchmark.

Set jN c j l H1a H2a GRASPb MA-PMc RTSd TSPRe ALNSf

B1 50 20 511,579 453,462 443,264 393,263 369,662 361,704 346,878

B2 100 20 963,649 831,019 791,839 714,627 712,294 685,898 636,962
B3 200 20 1,312,612 1,112,392 1,070,026 1,001,634 1,034,923 951,638 876,761

a
Boudia et al. [16].
b
Boudia et al. [17].
c
Boudia and Prins [19].
d
Bard and Nananukul [11].
e
Armentano et al. [9].
f
Adulyasak et al. [5].
 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152 151

Table 4 problem, the production routing problem has an interesting, but

Average total costs and computing times in seconds obtained by different heuristics complex, structure. Therefore, the majority of the studies proposed
on the Archetti et al. [8] benchmark.
decomposition-based heuristics and metaheuristics to handle the
Set jN c j l Average total cost Average CPU time problem, while few exact algorithms were developed. We dis-
cussed these techniques that have been developed for several
Ha ALNSb IMc Ha ALNSb IMc variants of the PRP and summarized the computational results.
Although various contributions have been made to this problem,
A1 14 6 180,830 181,803 n/a 9.0 11.3
A2 50 6 592,608 590,210 587,973 11.0 46.6 24.4
there are still a number of interesting issues that have not been
A3 100 6 1,092,509 1,089,589 108,4320 188.4 212.5 80.8 addressed. We encourage researchers in this eld to pursue
further developments in this promising research area.
, computational times are negligible.
a
Executed on 2.40 GHz PC.
b
Executed on 2.10 GHz CPU PC.
c
Executed on 2.67 GHz CPU PC.
Acknowledgments

This work was supported by the Natural Sciences and Engi-

can be used to control the ow part of the problem where neering Research Council of Canada under grants 227837-09 and
the subtour elimination constraints are handled by a 342182-09. This support is gratefully acknowledged.
cutting-plane technique.
Other variants of the LSP. As a major component of the PRP, one
may consider other interesting variants of the LSP. For References
example, the multi-product problem where the produc-
tion setup for each product must be done separately. [1] Abdelmaguid TF, Dessouky MM. A genetic algorithm approach to the inte-
grated inventory-distribution problem. Int J Prod Res 2006;44(21):444564.
Although some heuristics have been proposed for this [2] Absi N, Archetti C, Dauzre-Prs S, Feillet D. A two-phase iterative heuristic
variant [41,9], no studies have discussed exact methods approach for the production routing problem, Transp Sci, Forthcoming.
to solve the problem. One can also consider the multi- [3] Adulyasak Y, Cordeau J-F, Jans R. Benders decomposition for production
routing under demand uncertainty. GERAD Technical Report G-2012-57, HEC
product variant where production startups must also be
Montral, Canada; 2012.
taken into account. We refer to Jans and Degraeve [49] [4] Adulyasak Y, Cordeau JF, Jans R. Formulations and branch and cut algorithms
for the details of other variants of the LSP. for multi-vehicle production and inventory routing problems. INF J Comput
2014;26(1):103-20.
PRP with customer visit and delivery time windows. Time windows
[5] Adulyasak Y, Cordeau J-F, Jans R. Optimization-based adaptive large neighbor-
are one of the most common issues in transportation hood search for the production routing problem. Transp Sci 2014;48(1):20-45.
operations. There are two different types of time win- [6] Andersson H, Hoff A, Christiansen M, Hasle G, Lkketangen A. Industrial
dows that can be considered in the PRP. The rst type, aspects and literature survey: combined inventory management and routing.
Comput Oper Res 2010;37(9):151536.
called customer visit time windows, is imposed on [7] Archetti C, Bertazzi L, Laporte G, Speranza MG. A branch-and-cut algorithm for
periods where some customers should be visited to a vendor-managed inventory-routing problem. Transp Sci 2007;41(3):38291.
satisfy operational requirements. The second type, called [8] Archetti C, Bertazzi L, Paletta G, Speranza MG. Analysis of the maximum level
policy in a production-distribution system. Comput Oper Res 2011;38
delivery time window, is imposed during the day of (12):173146.
delivery and is a well-known variant of the VRP. [9] Armentano VA, Shiguemoto AL, Lkketangen A. Tabu search with path
Robust PRP. In stochastic environments, it is possible that a relinking for an integrated production-distribution problem. Comput Oper
Res 2011;38(8):1199209.
probabilistic description of the uncertainty is not avail- [10] Bard JF, Nananukul N. Heuristics for a multiperiod inventory routing problem
able and one cannot use the SPRP to solve the problem. with production decisions. Comput Ind Eng 2009;57(3):71323.
In this case, instead of minimizing the expected total [11] Bard JF, Nananukul N. The integrated production-inventory-distribution-
routing problem. J Sched 2009;12(3):25780.
cost, one is interested in obtaining a solution that is [12] Bard JF, Nananukul N. A branch-and-price algorithm for an integrated
immune to any realization of the uncertainty in a given production and inventory routing problem. Comput Oper Res 2010;37
set and therefore the product availability must be guar- (12):220217.
[13] Barnhart C, Johnson EL, Nemhauser GL, Savelsbergh MWP, Vance PH. Branch-
anteed. A study of the robust approach for the IRP can be
and-price: column generation for solving huge integer programs. Oper Res
found in Solyal et al. [73]. 1998;46(3):31629.
Tactical PRP. In many applications, customers must be clustered [14] Battiti R, Tecchiolli G. The reactive tabu search. ORSA J Comput 1994;6
and a driver is assigned to serve a specic cluster. The (2):12640.
[15] Bell WJ, Dalberto LM, Fisher ML, Greeneld AJ, Jaikumar R, Kedia P, et al.
cluster decisions are xed for a long term horizon. Improving the distribution of industrial gases with an on-line computerized
Examples of this variant can be found in Michel and routing and scheduling optimizer. Interfaces 1983;13(6):423.
Vanderbeck [62] and Coelho et al. [31]. One possible [16] Boudia M, Louly MAO, Prins C. Combined optimization of production and
distribution. In: CD-ROM proceedings of the international conference on
approach is to decompose the problem. The rst problem industrial engineering and systems management (IESM'05). Marrakech, Mor-
would be used to identify clusters, while the other occo; 1619 May 2005.
problem would be a modied PRP to determine the [17] Boudia M, Louly MAO, Prins C. A reactive GRASP and path relinking for a
combined production-distribution problem. Comput Oper Res 2007;34
remaining decisions. (11):340219.
[18] Boudia M, Louly MAO, Prins C. Fast heuristics for a combined production
planning and vehicle routing problem. Prod Plan Control 2008;19(2):8596.
[19] Boudia M, Prins C. A memetic algorithm with dynamic population manage-
ment for an integrated production-distribution problem. Eur J Oper Res
7. Conclusion 2009;195(3):70315.
[20] Bramel J, Simchi-Levi D. Set-covering-based algorithms for the capacitated
VRP. In: Toth P, Vigo D, editors. The vehicle routing problems. Philadelphia,
This paper provides an in-depth review of the production
USA: Society for Industrial and Applied Mathematics; 2001. p. 85108.
routing problem to support the growing interest in this research [21] Brown G, Keegan J, Vigus B, Wood K. The Kellogg company optimizes
area. We rst described the relevance of this problem and of the production, inventory, and distribution. Interfaces 2001;31(6):115.
other two integrated supply chain planning problems, namely, the [22] Campbell AM, Savelsbergh MWP. A decomposition approach for the
inventory-routing problem. Transp Sci 2004;38(4):488502.
lot-sizing problem with direct shipment and the inventory routing [23] Carlton WB, Barnes JW. Solving the traveling-salesman problem with time
problem. As a combination of the lot-sizing and vehicle routing windows using tabu search. IIE Trans 1996;28(8):61729.
152 Y. Adulyasak et al. / Computers & Operations Research 55 (2015) 141152

[24] Carter M, Farvolden J, Laporte G, Xu J. Solving an integrated logistics problem [53] Lee W-S, Han J-H, Cho S-J. A heuristic algorithm for a multi-product dynamic
arising in grocery distribution. INFOR J 1996;34:290306. lot-sizing and shipping problem. Int J Prod Econ 2005;98(2):20414.
[25] etinkaya S, Uster H, Easwaran G, Keskin BB. An integrated outbound logistics [54] Lei L, Liu S, Ruszczynski A, Park S. On the integrated production, inventory, and
model for Frito-Lay: coordinating aggregate-level production and distribution distribution routing problem. IIE Trans 2006;38(11):95570.
decisions. Interfaces 2009;39(September 1 (5)):46075. [55] Letchford AN, Salazar-Gonzlez JJ. Projection results for vehicle routing. Math
[26] Chand S, Hsu VN, Sethi S, Deshpande V. A dynamic lot sizing problem with Program 2006;105(2):25174.
multiple customers: customer-specic shipping and backlogging costs. IIE [56] Li C-L, Hsu VN, Xiao W-Q. Dynamic lot sizing with batch ordering and
Trans 2007;39(11):105969. truckload discounts. Oper Res 2004;52(4):63954.
[27] Chandra P. A dynamic distribution model with warehouse and customer [57] Lin S. Computer solutions of the traveling salesman problem. Bell Syst Technol
replenishment requirements. J Oper Res Soc 1993;44(7):68192. J 1965;44(10):224569.
[28] Chandra P, Fisher M. Coordination of production and distribution planning. [58] Lbbecke ME, Desrosiers J. Selected topics in column generation. Oper Res
Eur J Oper Res 1994;72(3):50317. 2005;53(6):100723.
[29] Christiansen M. Decomposition of a combined inventory and time constrained [59] Lysgaard J, Letchford AN, Eglese RW. A new branch-and-cut algorithm for the
ship routing problem. Transp Sci 1999;33(1):316. capacitated vehicle routing problem. Math Program 2004;100(2):42345.
[30] Clarke G, Wright JW. Scheduling of vehicles from a central depot to a number [60] Magnanti TL, Wong RT. Accelerating Benders decomposition: algorithmic
of delivery points. Oper Res 1964;12(4):56881.
enhancement and model selection criteria. Oper Res 1981;29(3):46484.
[31] Coelho LC, Cordeau J-F, Laporte G. Consistency in multi-vehicle inventory-
[61] Melo RA, Wolsey LA. MIP formulations and heuristics for two-level produc-
routing. Transport Res CEmerg Technol 2012;24:27087.
tion-transportation problems. Comput Oper Res 2012;39(11):277686.
[32] Coelho LC, Cordeau J-F, Laporte G. The inventory-routing problem with
[62] Michel S, Vanderbeck F. A column-generation based tactical planning method
transshipment. Comput Oper Res 2012;39(11):253748.
for inventory routing. Oper Res 2012;60(2):38297.
[33] Coelho LC, Cordeau J-F, Laporte G. Thirty years of inventory-routing. Transp Sci
[63] Miller CE, Tucker AW, Zemlin RA. Integer programming formulation of
2014;48(1):1-19.
traveling salesman problems. J Assoc Comput Mach 1960;7(4):3269.
[34] Coelho LC, Laporte G. The exact solution of several classes of inventory-routing
[64] Moscato P. Memetic algorithms: a short introduction. In: New ideas in
problems. Comput Oper Res 2013;40(2):55865.
[35] Eppen GD, Martin RK. Solving multi-item capacitated lot-sizing problems optimization. UK: McGraw-Hill Ltd.; 1999. p. 21934.
using variable redenition. Oper Res 1987;35(6):83248. [65] Nananukul N. Lot-sizing and inventory routing for production, inventory and
[36] Federgruen A, Tzur M. Time-partitioning heuristics: application to one ware- distribution systems. PhD thesis, Graduate Program in Operations Research
house, multi-item, multi-retailer lot-sizing problems. Nav Res Logist 1999;46 and Industrial Engineering. The University of Texas, Austin; 2008.
(5):46386. [66] Pochet Y, Wolsey LA. Single-item uncapacitated lot-sizing. In: Production
[37] Feo TA, Resende MGC. Greedy randomized adaptive search procedures. planning by mixed integer programming. Springer Series in Operations
J Global Optim 1995;6(2):10933. Research and Financial Engineering. Springer, New York; 2006. p. 20734.
[38] Fischetti M, Lodi A. Local branching. Math Program 2003;98(1):2347. [67] Rizk N, Martel A, Ramudhin A. A Lagrangean relaxation algorithm for multi-
[39] Fischetti M, Salazar-Gonzlez JJ, Toth P. A branch-and-cut algorithm for the item lot-sizing problems with joint piecewise linear resource costs. Int J Prod
symmetric generalized traveling salesman problem. Oper Res 1997;45(3): Econ 2006;102(2):34457.
378394. [68] Ropke S, Pisinger D. An adaptive large neighborhood search heuristic for the
[40] Fisher ML. The Lagrangian relaxation method for solving integer programming pickup and delivery problem with time windows. Transp Sci 2006;40
problems. Manag Sci 1981;27(1):118. (4):45572.
[41] Fumero F, Vercellis C. Synchronized development of production, inventory, [69] Ruokokoski M, Solyal O, Cordeau J-F, Jans R, Sral H. Efcient formulations
and distribution schedules. Transp Sci 1999;33(3):33040. and a branch-and-cut algorithm for a production-routing problem. GERAD
[42] Gaur V, Fisher ML. A periodic inventory routing problem at a supermarket Technical Report G-2010-66, HEC Montral, Canada; 2010.
chain. Oper Res 2004;52(6):81322. [70] Rusdiansyah A, Tsao D. An integrated model of the periodic delivery problems
[43] Gendreau M, Laporte G, Potvin J-Y. Metaheuristics for the capacitated VRP. In: for vending-machine supply chains. J Food Eng 2005;70(3):42134.
Toth P, Vigo D, editors. The vehicle routing problems. Philadelphia, USA: [71] Savelsbergh MWP, Song J-H. Inventory routing with continuous moves.
Society for Industrial and Applied Mathematics; 2001. p. 12954. Comput Oper Res 2007;34(6):174463.
[44] Gendreau M, Laporte G, Semet F. A branch-and-cut algorithm for the [72] Savelsbergh MWP, Song J-H. An optimization algorithm for the inventory
undirected selective traveling salesman problem. Networks 1998;32 routing problem with continuous moves. Comput Oper Res 2008;35
(4):26373. (7):226682.
[45] Glover F. Tabu searchPart I. ORSA J Comput 1989;1(3):190206. [73] Solyal O, Cordeau J-F, Laporte G. Robust inventory routing under demand
[46] Glover F. Tabu search and adaptive memory programmingadvances, appli-
uncertainty. Transp Sci 2012;46(3):32740.
cations and challenges. In: Barr R, Helgason R, Kennington J, editors. Interfaces
[74] Solyal O, Sral H. A relaxation based solution approach for the inventory
in computer science and operations research. Boston: Kluwer Academic
control and vehicle routing problem in vendor managed systems. In: Neogy
Publishers; 1996. p. 175.
SK, Das AK, Bapat RB, editors. Modeling, computation and Optimization.
[47] Herer YT, Tzur M. The dynamic transshipment problem. Nav Res Logist
Singapore, Sinagpore: World Scientic; 2009. p. 17189.
2001;48(5):386408.
[75] Solyal O, Sral H. A branch-and-cut algorithm using a strong formulation and
[48] Hsu VN, Li C-L, Xiao W-Q. Dynamic lot size problems with one-way product
substitution. IIE Trans 2005;37(3):20115. an a priori tour-based heuristic for an inventory-routing problem. Transp Sci
[49] Jans R, Degraeve Z. Modeling industrial lot sizing problems: a review. Int 2011;45(3):33545.
J Prod Res 2008;46(6):161943. [76] Solyal O, Sral H. The one-warehouse multi-retailer problem: reformulation,
[50] Jaruphongsa W, etinkaya S, Lee C-Y. Outbound shipment mode considera- classication, and computational results. Ann Oper Res 2012;196(1):51741.
tions for integrated inventory and delivery lot-sizing decisions. Oper Res Lett [77] Srensen K, Sevaux M. MA-PM: memetic algorithms with population manage-
2007;35(6):81322. ment. Comput Oper Res 2006;33(5):121425.
[51] Jaruphongsa W, Lee C-Y. Dynamic lot-sizing problem with demand time [78] Toth P, Vigo D. An overview of vehicle routing problems. In: Toth P, Vigo D,
windows and container-based transportation cost. Optim Lett 2008;2 editors. The vehicle routing problem. PA: SIAM Monographs on Discrete
(1):3951. Mathematics and Applications; 2001. p. 126.
[52] Krarup J, Bilde O. Plant location, set covering and economic lot size: an O(mn)- [79] van Norden L, van de Velde S. Multi-product lot-sizing with a transportation
algorithm for structured problems. In: Collatz L, Wetterling W, editors. capacity reservation contract. Eur J Oper Res 2005;165(1):12738.
International series of numerical mathematics, vol. 36. Basel: Birkhaeuser; [80] Wagner HM, Whitin TM. Dynamic version of the economic lot size model.
1977. p. 15580. Manag Sci 1958;5(1):8996.