You are on page 1of 12

Omega 82 (2019) 193–204

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Optimal production, replenishment, delivery, routing and inventory


management policies for products with perishable inventoryR
Yuzhuo Qiu a,b, Jun Qiao a, Panos M. Pardalos b,∗
a
Jiangsu Key Laboratory of Modern Logistics, School of Marketing and Logistics Management, Nanjing University of Finance and Economics, Nanjing
210023, China
b
Department of Industrial and Systems Engineering, Faculty of Engineering, University of Florida, Gainesville, FL 32611, United States

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a generalized production-inventory-routing model with perishable inventory. We an-
Received 12 May 2017 alyze the optimal integrated decisions of when and how much to deliver and sell products with varying
Accepted 11 January 2018
manufacturing periods. We discuss main inventory management policies to demonstrate the applicability
Available online 12 January 2018
of the model in real-world applications for production routing problems (PRPs) with perishable inventory.
Keywords: Furthermore, an exact branch-and-cut algorithm is developed and discussed. We introduce new families
Routing of logical, strengthened lot-sizing and lifted Miller–Tucker–Zemlin subtour elimination constraints for the
Production planning PRP with perishable inventory. Finally, we test the performance of the algorithm. We also implement and
Branch-and-cut compare 8 suboptimal delivery and selling priority policies with an optimized policy to develop manage-
Strengthened lot-sizing inequalities rial implications.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction inventory. In contrast, for PRPs not considering perishable inven-


tory and backorder, the total amount of production, which equals
Production routing problems (PRPs) that jointly optimize deci- to total demand minus total initial inventory, is actually irrele-
sions of production, inventory, distribution and routing have re- vant and predetermined. The reason is that, the dynamics of to-
cently received a considerable attention [4]. In practice, many com- tal initial inventory, i.e., remains of initial inventory in each pe-
panies such as Kellogg and Frito-Lay have set up similar integrated riod, are unaffected by varying selling policies without perishable
planning systems and obtained multi-million cost savings [4]. Such inventory, but changed by different selling policies with perishable
integrated optimization problems are of practical relevance to suc- inventory.
cess in supply chain competitions, especially in modern logisti- As a reflection, consideration of perishable products has ap-
cal practices of vendor managed inventory (VMI) and just-in-time peared recently in research of inventory routing problems [18].
(JIT). However, in inventory routing problems, the amount of production
In the seminal papers [15,16,20], PRPs were brought forward as is a parameter, although suboptimal selling policies were included.
multi-product multi-vehicle production routing problems. In the When the amount of production is variable, we can considerate
past decade, PRPs were extended in various ways, such as mul- more suboptimal delivery policies thus greatly enrich management
tiple plants and heterogeneous fleets of vehicles [26], incapaci- insights. Our aim is to bridge this gap by investigating the opti-
tated production [7], multiple homogeneous capacitated vehicles mal and suboptimal management policies in PRPs with perishable
[2], and under demand uncertainty [5]. The most recent develop- inventory (PRPPIs).
ments were PRPs with backorder [14], and PRPs with carbon emis- There has been extensive research in PRPs with heuristics:
sions [30]. approximation algorithms [15,16], decoupled heuristics [20], the
Incorporation of perishable inventory is essential both in lot- greedy randomized adaptive search procedure [12], memetic al-
sizing [22] and PRPs. Deciding how much to produce is critical gorithms [13], tabu searches [8,10], adaptive large neighborhood
and affected by the deterioration rate of products with perishable searches [3], iterative mixed integer programming [1], and particle-
swarm optimization [24]. Research on the exact algorithms has
only been addressed in a few publications. For example, branch-
R
This manuscript was processed by Associate Editor Eksioglu.

and-price [9,11,30] and branch-and-cut algorithms for single prod-
Corresponding author.
E-mail addresses: yzqiu@nufe.edu.cn (Y. Qiu), 9119931015@nufe.edu.cn (J. Qiao),
uct PRPs were recently investigated [3,7].
pardalos@ufl.edu (P.M. Pardalos).

https://doi.org/10.1016/j.omega.2018.01.006
0305-0483/© 2018 Elsevier Ltd. All rights reserved.
194 Y. Qiu et al. / Omega 82 (2019) 193–204

The PRP involves combinatorial optimization of both deliv- hi0t unit inventory holding cost at node i at the end of period
ery and routing decisions. Current branch-and-cut algorithms for t for products from initial inventory;
multi-vehicle PRPs can only solve instances of up to 50 customers, hiτ t unit inventory holding cost at node i at the end of period
3 periods and 3 vehicles in 2 h by using parallel computing [2]. t for products manufactured in period τ , 1 ≤ τ ≤ |T|;
A limitation in these algorithms is the lack of lot-sizing-problems- cij transportation cost over arc (i, j);
related (LSP-related) valid inequalities. As indicated in [4], LSPs are dit customer demand at retailer i in period t;
a major component of PRPs. LSP-related valid inequalities are thus Q vehicle capacity;
expected to enhance performance of exact methods to solve PRPs. Li maximum or target inventory level at node i;
In this paper, we aim to strengthen the formulation of PRPPIs with C production capacity;
these valid inequalities.
Decision variables
The contributions of this paper are summarized as follows. First,
di0t customer demand at retailer i (i ∈ N) in period t (t ∈ T) sat-
we introduce a real-world variant of the PRP which addresses per-
isfied by initial inventory;
ishable inventory. Second, we formulate the PRPPI as a mixed inte-
diτ t customer demand at retailer i (i ∈ N) in period t (t ∈ T) sat-
ger linear programming (MILP) problem. Main inventory manage-
isfied by products manufactured in period τ (τ ∈ T);
ment policies are discussed and classified to demonstrate the ap-
Ii0t inventory amount at node i (i ∈ N0 ) by the end of period t
plicability of the model in real-world applications. Third, we intro-
(t ∈ T) remaining from initial inventory, and let Ii00 be the
duce three families of valid inequalities to tighten the MIP problem
initial inventory of node i;
and design a branch-and-cut algorithm to find an optimal solu-
Iiτ t inventory amount at node i (i ∈ N0 ) by the end of period
tion. Finally, we implement and compare eight suboptimal delivery
t (t ∈ T) remaining from products manufactured in period
and selling priority policies with an optimized policy to develop
managerial implications. Our PRPPI model, algorithm and compu-
τ (τ ∈ T);
pi0t binary variable, equal to 1 if delivery amount of products
tational results can serve as a stepping stone for further research
from initial inventory to retailer i (i ∈ N) in period t (t ∈ T)
of PRPs with other extensions [4].
is positive, 0 otherwise;
The remainder of this paper is organized as follows. In
piτ t binary variable, equal to 1 if delivery amount of products
Section 2, we provide a formal description of the problem. Three
manufactured in period τ (τ ∈ T) to retailer i (i ∈ N) in pe-
families of valid inequalities are then introduced in Section 3.
riod t (t ∈ T) is positive, 0 otherwise;
Based on these valid inequalities and MILP formulation, we devise
qt production amount in period t (t ∈ T);
a branch-and-cut algorithm in Section 4. We present the computa-
ri0t delivery amount of products from initial inventory to re-
tional results of the algorithm in Section 5, followed by conclusions
tailer i (i ∈ N) in period t (t ∈ T);
in Section 7.
riτ t delivery amount of products manufactured in period τ
(τ ∈ T) to retailer i (i ∈ N) in period t (t ∈ T);
ritk delivery amount to retailer i in period t by vehicle k;
2. Problem description and model formulation
vitk product amount carried on vehicle k before visiting re-
tailer i in period t if retailer i is visited by vehicle k in
2.1. Problem description and assumptions
period t, 0 otherwise;
wi0t binary variable, equal to 1 if customer demand at retailer
The PRP with perishable inventory (PRPPI) is defined on a com-
i in period t is partially satisfied by initial inventory, 0
plete graph G = (N0 , A ), where the node set N0 = N ∪ {0} consists
otherwise;
of a set N = {1, 2, . . . , n} of retailers and a depot represented by
wiτ t binary variable, equal to 1 if customer demand at retailer
node 0, and the arc set is A = {(i, j ) : i, j ∈ N0 , i = j}. A product
i in period t is partially satisfied by products manufac-
with perishable inventory is produced at the depot and shipped
tured in period τ , 0 otherwise;
to the retailers over a finite set T = {1, 2, . . . , |T |} of planning peri-
xijkt binary variable, equal to 1 if arc (i, j) is traversed in pe-
ods. The known parameters are fixed production setup costs, unit
riod t by vehicle k, 0 otherwise;
production costs, production capacity, product deterioration rates,
yt binary variable, equal to 1 if the product is set up for pro-
unit inventory holding costs, inventory capacities, traveling costs
duction in period t, 0 otherwise;
between nodes, vehicle capacity and customer demands. Given the
zitk binary variable, equal to 1 if node i is visited in period t
initial inventory levels of the product at the retailers and the de-
by vehicle k, 0 otherwise.
pot, the problem is to determine the amount of product to manu-
facture at the depot, the amount of product that has accumulated Before we formulate the problem, we make a few assumptions
in previous periods or was just produced to deliver to each retailer, concerning production, inventory, deterioration and backorder.
the amount of product that has accumulated in previous periods or
was just produced to satisfy the demand at a retailer and the set of Assumption 1. Production occurs instantaneously at the beginning
routes in each period, minimizing production, inventory and rout- of each period.
ing costs. We differentiate inventory stocks by their production pe-
Assumption 2. Inventory of each retailer in period t is defined as
riods because deterioration rates and inventory carrying costs may the total amount of product carried by the end of period t, ex-
be age-dependent. Notations for parameters and decision variables
cluding the amount used to satisfy the demand at each retailer in
of production, inventory and routing are as follows.
period t. Inventory of the depot in period t is thus defined as the
Notation and parameters
total amount of product carried by the end of period t, excluding
T0 set of time periods, indexed by τ ∈ {0, 1, . . . , |T |};
the deliveries to all retailers in period t. The inventories in previ-
K set of identical vehicles, indexed by k ∈ {1, . . . , |K |};
ous periods are assumed to deteriorate through the current period
f fixed production setup cost;
by known deterioration rates.
u unit production cost;
α 0t deterioration rate in period t for products from initial in- Note that in most practical applications, the longer a product
ventory; is held in stock, the faster it may deteriorate and the higher its
ατ t deterioration rate in period t for products manufactured inventory holding cost. Thus, we make the following assumptions
in period τ ; for deterioration rates and inventory costs.
Y. Qiu et al. / Omega 82 (2019) 193–204 195

Assumption 3.

t
ατ1 t ≥ ατ2 t , ∀ t ∈ T , τ1 ∈ T0 , τ2 ∈ T0 , 0 ≤ τ1 ≤ τ2 ≤ t. (1) diτ t = dit , ∀i ∈ N, t ∈ T , (8)
τ =0
Assumption 4.
hiτ1 t ≥ hiτ2 t , ∀ i ∈ N0 , t ∈ T , τ1 ∈ T0 , τ2 ∈ T0 , 0 ≤ τ1 ≤ τ2 ≤ t. (2) 
t 
r iτ t = ritk , ∀i ∈ N, t ∈ T , (9)
Remark 1. The equalities in (1) and (2) usually hold in the con- τ =0 k∈K
texts of products with a short shelf life.

Assumption 5. Backorder is not allowed. qt ≤ Cyt , ∀t ∈ T , (10)

The final assumption we discuss relates to the management of


products of different ages held in inventory at the depot and the ritk ≤ Qzitk , ∀i ∈ N, t ∈ T , k ∈ K, (11)
retailers. It is up to the depot to decide which items with specific
manufacturing periods to deliver to retailers. Three different deliv-

ery priority policies can be conceived. The first consists of applying ritk ≤ Qz0tk , ∀t ∈ T , k ∈ K, (12)
a first produce first deliver (FPFD) policy in which the supplier al- i∈N
ways delivers the items produced earliest first. The second policy
is the reverse. Under a last produce first deliver (LPFD) policy, the 
t
items produced later are delivered first. The third policy is an op- Iiτ t ≤ Li , ∀i ∈ N, t ∈ T , (13)
timized delivery (OD) policy determined by the model. This means τ =0
that the supplier may prefer to deliver a combination of items pro- 
duced in different periods and keep inventories composed of items xi jkt = zitk , ∀i ∈ N0 , t ∈ T , k ∈ K, (14)
produced in different periods to minimize costs, depending on the j∈N0
parameter settings.
 
Similarly, retailers need to decide which items to sell to cus- xi jkt = x jikt , ∀i ∈ N0 , t ∈ T , k ∈ K, (15)
tomers. Three corresponding selling priority policies are as follows. j∈N0 j∈N0
The first is a first produce first sell (FPFS) policy. The second is a
last produce first sell (LPFS) policy. The third policy is an optimized
selling (OS) policy. vitk − v jtk + Qxi jkt ≤ Q − r jtk ,
Therefore, under each set of parameters, we have nine combi- ∀(i, j ) ∈ A, i ∈ N, j ∈ N, t ∈ T , k ∈ K, (16)
nations of inventory management policies, i.e., (1) FPFD–FPFS; (2)
FPFD–LPFS; (3) FPFD–OS; (4) LPFD–FPFS; (5) LPFD–LPFS; (6) LPFD–
OS; (7) OD–FPFS; (8) OD–LPFS; (9) OD–OS. Note that the OD–OS qt ≥ 0, yt ∈ {0, 1}, ∀t ∈ T , (17)
policy is the optimal one whereas the rest of the policies are sub-
optimal.
Iiτ t ≥ 0, ∀i ∈ N0 , τ ∈ T0 , t ∈ T , (18)
Remark 2. The old first (OF) policy in [18] corresponds to (1), (4),
and (7), the fresh first (FF) policy in [18] corresponds to (2), (5),
and (8), and the optimal policy (OP) in [18] corresponds to (9). d iτ t , r iτ t ≥ 0 , ∀i ∈ N, τ ∈ T0 , t ∈ T , (19)
Thus, the discussion of inventory management policies proposed
in this paper generalizes the one in [18].
ritk ≥ 0, ∀i ∈ N, t ∈ T , k ∈ K, (20)
2.2. MIP formulation for the PRP with perishable inventory

Given the notations above, let A(S ) = {(i, j ) : i, j ∈ S, i = j}, zitk ∈ {0, 1}, ∀i ∈ N0 , t ∈ T , k ∈ K, (21)
  
xtk (S ) = (i, j )∈A(S ) xi jkt , rtk (S ) = i∈S ritk , and ztk (S ) = i∈S zitk ,
where S is a node set, S⊆N, |S| ≥ 2, the PRP with perishable inven-
tory can be formulated as follows: xi jkt ∈ {0, 1},
 
 
t   ∀(i, j ) ∈ A, t ∈ T , k ∈ K, (22)
min uqt + f yt + hiτ t Iiτ t + ci j xi jkt (3)
t∈T i∈N0 τ =0 (i, j )∈A k∈K
ritk ≤ vitk ≤ Q, ∀i ∈ N, t ∈ T , k ∈ K. (23)
subject to
 The objective function (3) measures the operational costs of pro-
qt − ritt = I0tt , ∀ t ∈ T, (4) duction, inventory and routing. Constraints (4) and (5) model flow
i∈N
balance of production, inventory and delivery at the depot. Con-
 straints (6) and (7) model flow balance of inventory, delivery and
(1 − ατ ,t−1 )I0τ ,t−1 − riτ t = I0τ t , ∀ t ∈ T , τ ∈ T0 , τ ≤ t − 1, consumption at the retailers. Constraints (8) and (9) reflect the
i∈N
composition of demand and delivery for a retailer in each period,
(5) respectively. Constraints (10) ensure that setup binary variables are
one if production takes place. Constraints (11) permit a positive de-
ritt − ditt = Iitt , ∀ i ∈ N, t ∈ T , (6) livery to node i in period t only if this node is visited in period t.
Constraints (12) and (13) ensure the vehicle and inventory capacity
limits are met, respectively. Constraints (14) serve as degree con-
(1 − ατ ,t−1 )Iiτ ,t−1 + riτ t − diτ t = Iiτ t , straints when the node is visited. Constraints (15) ensure vehicle
flow balance. Constraints (16) are MTZ-type constraints of vehicle
∀ i ∈ N, t ∈ T , τ ∈ T0 , τ ≤ t − 1, (7)
196 Y. Qiu et al. / Omega 82 (2019) 193–204


capacity for the PRPPI. Note that the MTZ formulation introduced Proof. If pi,η+1,t = 0, inequalities (28) which reduce to τ =0 I0τ ,t ≤
by [28] for the TSP is known as a relaxation of subtour elimination L0 is certainly true since we have inequalities (13). If pi,η+1,t = 1,

constraints. This formulation reduces the number of constraints at inequalities (28) which reduce to τ =0 I0τ ,t ≤ 0 is also true be-
the cost of weaker lower bounds. Constraints (17)–(23) are non- cause, when pi,η+1,t = 1, with the FPFD policy, there should be
negativity, binarity and range constraints. no item manufactured before period η + 1 left at the depot, thus

Remark 3. When ατ t = 0, ∀t ∈ T , τ ∈ T0 and hiτ t = hit , ∀i ∈ N0 , t ∈ τ =0 I0τ ,t = 0. 
T , τ ∈ T0 , PRPPI formulation (3)–(23) reduces to a PRP without per- η
Remark 5. In other words, when
 η τ =0 (1 − ατ ,t−1 )I0τ ,t−1 −
ishable inventory. It coincides with the PRP formulation in which r > 0 , there are items manufactured before period
i∈N τ =0 iτ t
the lot-sizing part is a facility location based formulation. Thus η + 1 left at the depot, but with the FPFD policy, there should not
PRPPI formulation (3)–(23) generalizes PRP formulation, e.g., see be delivery of items manufactured after period η. Thus, pi,η+1,t is
[4]. forced to take the value zero, which can be derived from (28).
Remark 4. According to Assumption 5, backorder is not allowed. To model the LPFD policy, we have the following proposition.
This implies that
Proposition 2. A PRPPI with an LPFD policy can be modeled through
diτ t = 0, ∀i ∈ N, τ ∈ T0 , t ∈ T , τ > t, adding the following constraints:

t−1  
t

Iiτ t = 0, ∀i ∈ N0 , τ ∈ T0 , t ∈ T , τ > t, (1 − ατ ,t−1 )I0τ ,t−1 + qt − riτ t ≤ L0 (1 − piηt ),


τ =η+1 i∈N τ =η+1
and
∀i ∈ N, t ∈ T , 0 ≤ η ≤ t − 1, (29)
riτ t = 0, ∀i ∈ N, τ ∈ T0 , t ∈ T , τ > t. into the extended formulation (3)–(23), (25) and (27).
t−1  t
2.3. An extended formulation for the PRP with perishable inventory Proof. When τ =η+1 (1 − ατ ,t−1 )I0τ ,t−1 + qt − i∈N τ =η+1 riτ t >
0, there are items manufactured after period η left at the depot,
To restrict choice of products with different manufacturing pe- but with the LPFD policy, there should not be delivery of items
riods for delivery or sale, we add the following extra variables and manufactured before period η + 1. Thus, piηt is forced to take the
constraints to PRPPI formulation (3)–(23) in Section 2.2. value zero, which can be derived from (29). 

diτ t ≤ dit wiτ t , ∀i ∈ N, τ ∈ T0 , t ∈ T , (24) Remark 6. In contrast with Proposition 1, qt must be included to
derive the correct formulation.

2.5. Modeling retailer inventory management/selling policy


r iτ t ≤ Q p iτ t , ∀i ∈ N, τ ∈ T0 , t ∈ T , (25)
For retailers, the three selling priority policies in Section 2.1,
i.e., FPFS (first produce first sell), LPFS (last produce first sell), and
w iτ t ∈ { 0 , 1 } , ∀i ∈ N, τ ∈ T0 , t ∈ T , (26)
OS (optimized selling) policies, can be modeled as follows. It is
straightforward to model the OS policy with the extended formu-
lation (3)–(23), (24) and (26). For FPFS and LPFS policies we have
p iτ t ∈ { 0 , 1 } , ∀i ∈ N, τ ∈ T0 , t ∈ T . (27) the following propositions.
Constraints (24) ensure that if products with manufacturing pe- Proposition 3. A PRPPI with an FPFS policy can be modeled through
riod τ are used to satisfy the customer demand at retailer i in adding the following constraints:
period t, the corresponding binary variables are one. Constraints
η
 η

(25) ensure that if production in period τ is delivered to retailer
i in period t, the corresponding binary variables are one. Con-
(1 − ατ ,t−1 )Iiτ ,t−1 + riτ t − dit + 1 ≤ Li (1 − wi,η+1,t ),
τ =0 τ =0
straints (26) and (27) are binarity constraints. Adding the above ex-
tra variables and constraints into formulation (3)–(23) constitutes ∀ i ∈ N, t ∈ T , 0 ≤ η ≤ t − 1, (30)
extended formulations which we will discuss in the following sec- into the extended formulation (3)–(23), (24) and (26).
tions. η η
Proof. When τ =0 (1 − ατ ,t−1 )Iiτ ,t−1 + τ =0 riτ t − dit > 0 , items
manufactured before period η + 1 are not enough to satisfy the
2.4. Modeling depot inventory management/delivery policy
customer demand at retailer i in period t, with the FPFS policy,
there should be no selling of items manufactured after period η.
Recall that we proposed three different delivery priority poli-
Thus wi,η+1,t is forced to take the value zero, which can be derived
cies in Section 2.1, i.e., FPFD (first produce first deliver), LPFD (last
from (30). 
produce first deliver), and OD (optimized delivery) policies. It is
straightforward to model the OD policy with the extended formu- Proposition 4. A PRPPI with an LPFS policy can be modeled through
lation (3)–(23), (25) and (27). To model the FPFD policy, we have adding the following constraints:
the following proposition.

t−1 
t

Proposition 1. A PRPPI with an FPFD policy can be modeled through (1 − ατ ,t−1 )Iiτ ,t−1 + riτ t − dit + 1 ≤ Li (1 − wiηt ),
adding the following constraints: τ =η+1 τ =η+1

η η ∀ i ∈ N, t ∈ T , 0 ≤ η ≤ t − 1. (31)
 
(1 − ατ ,t−1 )I0τ ,t−1 − riτ t ≤ L0 (1 − pi,η+1,t ), into the extended formulation (3)–(23), (24) and (26).
τ =0 i∈N τ =0 t−1 t
∀ i ∈ N, t ∈ T , 0 ≤ η ≤ t − 1, (28)
Proof. When τ =η+1 (1 − ατ ,t−1 )Iiτ ,t−1 + τ =η+1 riτ t − dit > 0,
items manufactured after period η are not enough to satisfy the
into the extended formulation (3)–(23), (25) and (27). customer demand at retailer i in period t, with the LPFS policy,
Y. Qiu et al. / Omega 82 (2019) 193–204 197

there should be no selling of items manufactured before period 3.3. Strengthened lot-sizing inequalities
η + 1. Thus wiηt is forced to take the value zero, which can be
derived from (31).  Lot sizing problems (LSPs) are also essential to PRPs. When
there is only one supplier but no retailers, [29] proposed a form of
Remark 7. Recall that in Section 2.1, we propose nine combina- (l, S) inequalities for the LSP when costs are Wagner-Whitin costs,
tions of inventory management policies, i.e., (1) FPFD–FPFS; (2) i.e., ut + ht > ut+1 , as follows:
FPFD–LPFS; (3) FPFD–OS;(4) LPFD–FPFS; (5) LPFD–LPFS; (6) LPFD–
OS;(7) OD–FPFS; (8) OD–LPFS; (9) OD–OS. Using the formulation η

in Sections 2.4 and 2.5, the extended formulations for these in- It−1 ≥ dτ (1 − yt − yt+1 − · · · − yτ ), ∀t ∈ T , η ∈ T , t < η ≤ |T |,
ventory management policies can be easily introduced. For in- τ =t
stance, inventory management policy (1) FPFD–FPFS can be mod- (37)
eled through extended formulation (3)–(27), (28) and (30). Other
where It−1 is the inventory of supplier by the end of period t − 1,
inventory management policies can be modeled similarly.
dτ is the demand for period τ at the supplier, and yτ is the binary
setup variable for period τ . Recall that the LSP has the following
3. Valid inequalities constraints [29]:

It−1 + qt = It + dt , ∀t ∈ T , (38a)
In this section, we review known valid inequalities for the PRP
and introduce several families of new valid inequalities that can be
used to strengthen LP relaxation of the PRPPI formulation (3)–(23) |T |

and its extended formulations. The valid inequalities we propose qt ≤ dτ yt , ∀t ∈ T , (38b)
include logical inequalities, strengthened lot-sizing inequalities and τ =t
lifted MTZ-type inequalities.

qt ≥ 0, yt ∈ {0, 1}, ∀t ∈ T , (38c)


3.1. Known inequalities

The PRPPI formulation can be strengthened by adding a priori It ≥ 0, ∀t ∈ T . (38d)


the following valid inequalities [2,6]:
Lemma 1 [29]. (l, S) inequalities (37) plus the constraints (38a)–
zitk ≤ z0tk , ∀i ∈ N, t ∈ T , k ∈ K. (32) (38d) can give a complete linear description of the convex hull for
the uncapacitated lot-sizing problem with Wagner–Whitin costs.
Constraints (32) are known as logical inequalities.
Based on the information from the initial inventory and de- In our model, ut = ut+1 = u. Therefore, our model features
mand of each retailer, the following valid inequalities focus on the Wagner–Whitin costs. In the PRPPI, we also must address demand
minimum number of visits per retailer in each period [17]: and inventories at the retailers. Thus, inspired by the aggregated
t subtour elimination constraints presented in [2] and Lemma 1, we

t
dil − Ii00 have the following proposition.
zilk ≥ l=1
, ∀i ∈ N, t ∈ T . (33)
Q
k∈K l=1 Proposition 5. The following strengthened lot-sizing inequalities
The PRPPI formulation can also be strengthened by adding a 
t−1 η 
 η  η 

priori the following symmetry-breaking constraints [2,6]: (1 − ατ ,t−1 )Ii,τ ,t−1 + d iθ y τ ≥ diτ (1 − yt ),
i∈N0 τ =0 τ =t+1 θ =τ i∈N τ =t i∈N
z0tk ≤ z0t,k−1 , ∀t ∈ T , k ∈ K \ {1}. (34)
∀t ∈ T , η ∈ T , t < η ≤ |T | (39)

3.2. Logical inequalities are valid for the PRPPI.

Proof. The (l, S) inequalities (37) for the LSP can be rearranged as
In addition to the known inequalities, we also add the following [29]:
valid inequalities
η 
 η η

zitk ≤ ritk , ∀i ∈ N, t ∈ T , k ∈ K. (35) It−1 + dk yτ ≥ dτ (1 − yt ), ∀η : t < η ≤ |T |. (40)
τ =t+1 k=τ τ =t
The inequalities (35) ensure that the visit number to node i equals
For the PRPPI, the inventory at time period t − 1 capable of satis-
zero if there is no delivery to node i, which is not guaranteed in
fying the customer demand at retailer i at time period t is either
the formulation, but implied by the optimal solution if the trian-
at the depot or at the retailer i. Thus, when deriving aggregated
gular inequality cik ≤ ci j + c jk , ∀i, j, k ∈ N0 holds strictly.
LSP inequalities for PRPPIs that look like (40), we should not only
Similarly, we also add the following logical inequalities
aggregate the demand at the retailers, but we should also sum up
d iτ t ≥ w iτ t , ∀i ∈ N, τ ∈ T0 , t ∈ T , (36a) the inventories of the depot and the retailers and consider the de-
terioration rate. 

r iτ t ≥ p iτ t , ∀i ∈ N, τ ∈ T0 , t ∈ T . (36b) 3.4. Lifted MTZ-type inequalities

The inequalities (36a) ensure that when amount diτ t of satisfied In Section 2.2, we adopt the basic MTZ-type constraints to for-
demand with items manufactured in period τ is zero, the corre- mulate the PRPPI. As shown in [19] and [23], MTZ-type Inequalities
sponding binary variable is zero. The inequalities (36b) ensure that can be lifted. However, in PRPPIs, the lifted MTZ-type constraints
when amount riτ t of delivery with items manufactured in period τ are nonlinear. We use an approach proposed by [21] to linearize
is zero, the corresponding binary variable is zero. the lifted MTZ-type constraints.
198 Y. Qiu et al. / Omega 82 (2019) 193–204

Lemma 2. Denote ζi jkt = xi jkt (ritk + r jtk ). For the PRPPI, production-distribution subproblem which could be formulated as
follows:
|T |  
d iτ + d j τ
Ui jkt = , ∀i, j ∈ N, (i, j ) ∈ A, t ∈ T , k ∈ K, (41)  
t 
τ =t
1 − α0τ min uqt + f yt + hiτ t Iiτ t + σi zitk (47)
t∈T i∈N0 τ =0 k∈K t∈T i∈N
is a valid upper bound on ζ ijkt .
subject to (4)–(13), and (17)–(21). Finally, we use the guided
Proof. In the PRPPI formulation, constraints (11) can by strength-
variable neighborhood descent heuristic [25] to quickly obtain
ened as follows:
a solution of the VRP subproblem for each period. The VRP
ritk ≤ min{Q, D}zitk , ∀i ∈ N, t ∈ T , k ∈ K, (42) solutions combined with values of other variables from the
solved production-distribution subproblem give us the initial up-
where
per bound.
|T |
 d iτ
D= , ∀i ∈ N, t ∈ T . (43)
τ =t
1 − α0τ 4.2. Separation procedures and strategy

The inclusion of deterioration rate ensures that delivery to retailer For instances of small to medium sizes, all constraints and vari-
i in period t is enough to satisfy demand from period t to the end ables can be generated and inserted into programs for the model
of the planning horizon. Therefore, Uijkt is a valid upper bound on presented in Section 2.2. These programs can be solved by an
ζ ijkt .  MIP solver. However, for instances of realistic sizes, the number of
Proposition 6. ∀i, j ∈ N, (i, j) ∈ A, t ∈ T, k ∈ K, and Uijkt given as (41), lifted MTZ-type Inequalities (44a)–(44e) is too large and the cor-
the following lifted MTZ-type constraints responding lower bound too weak. In these contexts, we replace
the MTZ-type constraints with the generalized subtour elimination
vitk − v jtk + r jtk − ζi jkt ≤ Q (1 − xi jkt − x jikt ), (44a) constraints

Qxtk (S ) ≤ Qztk (S ) − rtk (S ), ∀t ∈ T , k ∈ K (48)


ζi jkt ≤ r jtk + ritk , (44b)   
where xtk (S ) = (i, j )∈A(S ) xi jkt , rtk (S ) = i∈S ritk , ztk (S ) = i∈S zitk ,
A(S ) = {(i, j ) : i, j ∈ S, i = j} and S is a node set, S⊆N, |S| ≥ 2.
We thus propose a classic branch-and-cut algorithm in which
ζi jkt ≥ r jtk + ritk − Ui jkt (1 − xi jkt ), (44c) the generalized subtour elimination constraints are only generated
and incorporated into the programs whenever they are found to be
violated.
ζi jkt ≤ Ui jkt xi jkt , (44d) A linear programming relaxation of the PRPPI formulation with
a subset of the generalized subtour elimination constraints is
solved at each node in the branch-and-cut tree. If the solution at
ζi jkt ≥ 0, (44e) the current node is fractional, we call connected-component-based
heuristics [27,31] for generalized subtour elimination constraints to
are valid inequalities for the PRPPI formulation (3)–(23) and its ex-
search for violated cuts. If at least one such cut is found, we add
tended formulations.
the violated constraints, and the LP is reoptimized. The algorithm
Proof. The constraints is outlined as follows in Algorithm 1.

vi − v j + Qxi j + (Q − ri − r j )x ji ≤ Q − r j , ∀i, j ∈ N, (i, j ) ∈ A (45)


5. Computational results
are valid inequalities for the CVRP [19,23], where ui is the prod-
uct amount carried on a vehicle before visiting retailer i, and ri is The experiments of this section were run using a PC with Intel
customer demand amount at retailer i. For PRPPI, ∀i, j ∈ N, (i, j) ∈ A, Core i7-6700 CPU of 3.40 GHz in Windows 7 (64 bits) with 16GB
t ∈ T, k ∈ K, the valid inequalities become RAM. We implemented the branch-and-cut algorithm in Microsoft
Visual C++ 2015 based on C++ API of IBM ILOG CPLEX version 12
vitk − v jtk + Qxi jkt + (Q − ritk − r jtk )x jikt ≤ Q − r jtk . (46)
release 7.
Since ritk are variables now, (46) are nonlinear. Using the approach
proposed by [21], and Lemma 2, we can get inequalities (44a)–
5.1. Instances generation
(44e). 

We adapt the data set of [2] combined with the parameters of


4. Branch-and-cut algorithm the deterioration rate and inventory holding cost. The instances are
characterized with n = 10 to 50, to 30, and to 20 retailers for time
4.1. Initial upper bound horizons with |T | = 3, 6, and 9 periods, respectively. The number
of vehicles is set to |K | = 2 or 3 for the instances with n < 25 and
We construct initial solutions for the PRPPI by solving two de- to |K | = 3 or 4 for the instances 25 ≤ n ≤ 50. One hundred instances
composed problems of production-distribution and the VRP [3]. are generated for each inventory management policy.
More specifically, visiting cost is defined as follows [3]: Inventory holding costs hiτ t = hi + (t − τ ) ∗ γ , ∀i ∈ N0 , τ ∈
 
T0 , t ∈ T , where γ is the base rate of extra costs for holding
σi = min 2c0i , min (c ji + cik ) . inventory produced in earlier periods, and hi is given in the data
j,k∈N0 , j=k
set of [2]. Deterioration rate ατ t = β ∗ (t − τ ), where β is the
Then, we fix the shipment-quantity variables ritk and site-visiting base rate between adjacent periods. The details of the instances
indicator variables zitk according to the solution of the solved are provided in the supplement or can be found in the online
supplement of [2].
Y. Qiu et al. / Omega 82 (2019) 193–204 199

Algorithm 1: Branch-and-cut algorithm


1 Initialize the upper bound U ∗ and the incumbent solution.
2 Initialize the node pool N with the root node.
3 Generate and insert all known valid inequalities and the
proposed logical, strengthened lot-sizing into the program at
the root node of the search tree. Also, generate and insert
lifted MTZ-type valid inequalities if the problem size is
medium.
4 repeat
5 Selection: Select the next node in N to evaluate and
remove it from N .
6 Lower bound: Solve the LP relaxation at the current
node,let U l be the obtained lower bound of the current
node:
7 if the current solution is feasible then
8 if U l > U ∗ then
9 go to the termination check.
else
10 U∗ ← Ul.
11 Update the incumbent solution.
12 Prune nodes with lower bound U > U ∗ .
end
end
13 Cut generation:
14 if the current solution of the LP relaxation violates any cuts,
then Fig. 1. Route plan for Monday in the OS–OD policy.
15 determine the violated subtour elimination cuts by
connected component heuristics adapted from
Lysgaard et~al. (2004); Toth and Vigo (2014).
16 Add violated cuts.
end ities added both at the root node and in subsequent branch-and-
cut trees.
17 Branching: If U l > U ∗ , go to the termination check.
For the effects of valid inequalities, we can reach the following
until N = ∅ or time limit is met (termination check)
conclusions. First, as shown from the table, with every valid in-
18 Stop with the optimal solution and the corresponding cost
equality added to the formulation, either the average lower bounds
U ∗.
or the CPU times or both are improved. Second, the effects of the
valid inequalities are stronger in terms of nodes reduction in the
Table 1 branch-and-cut tree when fewer numbers of vehicles are available.
Effects of valid inequalities on average lower bounds at the root node under the Finally, as expected, adding all the cuts together provides the best
OD+OS policy.
results.
n = 50, |T | = 3

|K | = 2 |K | = 3
5.2.2. The effects of suboptimal inventory management policy
RLB (%) CPU (s) Cuts RLB (%) CPU (s) Cuts
Tables 2–4 summarize the results of branch-and-cut algorithms
Without CPLEX cuts 79.9 15.3 94.4 78.5 23.9 100.0 from small to large instances under the optimal (or OD–OS) pol-
With CPLEX cuts 88.0 18.1 77.4 86.2 16.1 72.8 icy and 8 suboptimal ones. Column % increase reflects the percent-
With (35) 88.8 17.4 60.3 88.3 14.2 61.2
With (36a) and (36b) 88.6 17.0 56.3 88.6 11.4 68.7
age of cost increase for suboptimal policy over optimal policy. Col-
With (39) 89.3 12.6 77.6 87.9 20.0 75.4 umn LB (%) computes average lower bounds over the best solu-
With (44a)–(44e) 89.5 17.5 73.8 88.4 12.0 67.0 tions found. Column CPU (s) indicates average CPU times of sec-
All 90.2 12.5 54.9 90.2 12.2 68.3 onds spent at both the root node and subsequence branch-and-cut
searches.
As shown from the tables, instances can be solved with up to
30 retailers for 3 periods and 3 (4) vehicles, 20 retailers for 6 pe-
5.2. Results and discussion riods and 2 (3) vehicles, or 20 retailers for 9 periods and 2 vehi-
cles. The running times and average lower bounds over the best
5.2.1. The effects of valid inequalities solutions are competitive with the exact algorithm for the multi-
The experiments on the effects of valid inequalities are con- vehicle PRP [2]. Furthermore, with the suboptimal inventory man-
ducted on the instances with n = 50 and |T | = 3 under the optimal agement policy, instances can be solved with almost identical ef-
or OD+OS policy. The average results are shown in Table 1. ficiency. On average, the LPFD–OS suboptimal policy results in the
The numbers in each column have the following interpretation. least cost increase over the optimal policy, whereas the OD–LPFS
The column RLB computes average lower bounds over 5 instances suboptimal policy is slightly worse than the LPFD–OS suboptimal
at the root node over the optimal solutions found. The column CPU policy in terms of cost increase. Finally, the FPFD–LPFS suboptimal
indicates average CPU times of seconds spent at both the root node policy results in the largest cost increase over the optimal policy,
and subsequence branch-and-cut searches. Lastly, the column Cuts although this policy is basically at the same level of cost increase
represents average number of cuts and the indicated valid inequal- with three other suboptimal policies. This is possibly because both
200 Y. Qiu et al. / Omega 82 (2019) 193–204

Table 2
Summary of the computational results for the PRPPI under the FPFD-type policy.

Instance size (1) FPFD–FPFS (2) FPFD–LPFS (3) FPFD–OS

n |T| |K| % increase LB (%) CPU (s) % increase LB (%) CPU (s) % increase LB (%) CPU (s)

10 3 2 0 100 0.1 0 100 0.2 0 100 0.2


10 3 3 0 100 0.4 0 100 0.2 0 100 0.2
20 3 2 0 100 2.2 0 100 2.3 0 100 2.4
20 3 3 0 100 8.3 0 100 7.3 0 100 7.2
30 3 3 0 100 1683.9 0 100 1744.6 0 100 1807.4
30 3 4 0 100 1875.9 0 100 1817.4 0 100 921.3
40 3 3 0 98.1 7200 0 98.7 7200 0 98.8 7200
40 3 4 0 96.9 7200 0 97.4 7200 0 97.7 7200
50 3 3 0 96.6 7200 0 96.5 7200 0 97.5 7200
50 3 4 0 93.6 7200 0 94.1 7200 0 94 7200
10 6 2 0.62 100 156.5 0.64 100 169.3 0.4 100 180.1
10 6 3 0.76 100 411.0 1.58 100 442.8 0.12 100 475.9
20 6 2 1.52 100 1001.2 1.39 100 1091.1 0.79 100 1105.9
20 6 3 1.52 100 1735 1.16 100 1649.3 0.64 100 1613.2
30 6 3 1.59 97.9 7200 1.67 97.7 7200 0.41 98.1 7200
30 6 4 0.67 96.8 7200 1.08 96.8 7200 0.8 97.3 7200
10 9 2 1.86 100 2521.1 1.4 100 2696.4 0.5 100 2988.3
10 9 3 0.75 100 2330.0 1.3 100 2514.6 0.97 100 2678.1
20 9 2 0.72 100 5355.1 0.83 100 5435.7 0.81 100 5732.7
20 9 3 1.23 94.9 7200 0.96 95 7200 0.49 96.3 7200
Average 0.56 98.74 3374.04 0.60 98.81 3398.56 0.30 98.99 3395.64

Note. β = 0.1, γ = 1.

Table 3
Summary of the computational results for the PRPPI under the LPFD-type policy.

Instance size (4) LPFD–FPFS (5) LPFD–LPFS (6)LPFD–OS

n |T| |K| % increase LB (%) CPU (s) % increase LB (%) CPU (s) % increase LB (%) CPU (s)

10 3 2 0 100 0.2 0 100 0.2 0 100 0.2


10 3 3 0 100 0.2 0 100 0.2 0 100 0.2
20 3 2 0 100 2.4 0 100 2.2 0 100 2.2
20 3 3 0 100 4.2 0 100 8.1 0 100 6.8
30 3 3 0 100 742.7 0 100 809.6 0 100 828.8
30 3 4 0 100 1838.3 0 100 1929.5 0 100 1780.1
40 3 3 0 98.5 7200 0 98.5 7200 0 98.7 7200
40 3 4 0 97.3 7200 0 97.5 7200 0 97.6 7200
50 3 3 0 96 7200 0 96 7200 0 96.9 7200
50 3 4 0 94 7200 0 93.2 7200 0 94.9 7200
10 6 2 1.41 100 165.1 0.6 100 176.4 0.22 100 195.2
10 6 3 1.27 100 449.3 0.85 100 483.8 0.1 100 485.2
20 6 2 1.41 100 1047.2 1.01 100 1115.2 0.01 100 1227.6
20 6 3 1.87 100 1653.6 1.02 100 1729.2 0.49 100 1613.2
30 6 3 0.99 97.9 7200 1.1 98.1 7200 0.06 98.7 7200
30 6 4 1.59 97.2 7200 0.89 96.3 7200 0.55 97.2 7200
10 9 2 0.96 100 2601.2 1.9 100 2768.2 0.63 100 3064.8
10 9 3 0.71 100 2740.8 0.99 100 2922.7 0.55 100 2966.6
20 9 2 0.64 100 5180.8 1.15 100 5617.0 0.26 100 5781.2
20 9 3 0.51 95.4 7200 0.83 94.9 7200 0.07 95.9 7200
Average 0.57 98.82 3341.31 0.52 98.73 3398.11 0.15 99.00 3417.61

Note. β = 0.1, γ = 1.

the delivery and selling policies are suboptimal, and they are in- riorated inventory. When we do not consider the environmental
consistent. effect of deteriorated inventory and thus ignore disposal cost, the
optimal decision would lead to more deteriorated inventory and
5.2.3. The effects of varying base rate of deterioration and inventory result in lower inventory holding cost. Thus, incorporating the en-
holding vironmental effects of deteriorated inventory into the framework
Table 5 summarizes the results of the branch-and-cut algorithm of the PRPPI would lead to a more reasonable inventory manage-
for instances with 20 retailers, 6 periods, and 2 vehicles under the ment practice.
optimal or OD–OS policy with varying base rates of deterioration
β and inventory holdings γ .
As shown from the table, total cost increased with the base rate 6. Case study
of inventory holdings γ when the base rate of deterioration re-
mained constant at β = 0.1 as expected. However, the total cost This section presents an implementation of the proposed model
increased unexpectedly with the base rate of deterioration β when on the production and distribution operations of a food company
the base rate of inventory holding remained constant at γ = 1. operating in the city of Nanjing, China. We first describe the case
This is because we have neglected the disposal cost of the dete- and data used, then present the results.
Y. Qiu et al. / Omega 82 (2019) 193–204 201

Table 4
Summary of the computational results for the PRPPI under the OD-type policy.

Instance size (7) OD–FPFS (8) OD–LPFS (9) OD–OS

n |T| |K| % increase LB (%) CPU (s) % increase LB (%) CPU (s) % increase LB (%) CPU (s)

10 3 2 0 100 0.2 0 100 0.1 0 100 0.1


10 3 3 0 100 0.2 0 100 0.2 0 100 0.4
20 3 2 0 100 2.4 0 100 2.1 0 100 2.1
20 3 3 0 100 4.8 0 100 6.6 0 100 8.1
30 3 3 0 100 806.2 0 100 855.8 0 100 870.2
30 3 4 0 100 1859.4 0 100 1784.1 0 100 1758.1
40 3 3 0 98.7 7200 0 98.8 7200 0 99 7200
40 3 4 0 97.7 7200 0 98 7200 0 98.4 7200
50 3 3 0 96.7 7200 0 96.6 7200 0 97.5 7200
50 3 4 0 94.1 7200 0 94.9 7200 0 94.9 7200
10 6 2 0.24 100 6.6 0.49 100 193.8 0 100 210.8
10 6 3 0.4 100 16.1 0.65 100 497.0 0 100 535
20 6 2 0.84 100 197 0.45 100 1163.0 0 100 1233.4
20 6 3 0.77 100 1661.6 0.96 100 1588.3 0 100 1565.2
30 6 3 0.99 98.5 7200 0.84 98.5 7200 0 99 7200
30 6 4 0.18 97.3 7200 0.85 97.5 7200 0 97.8 7200
10 9 2 0.99 100 2678.6 0.7 100 2855.8 0 100 3088.7
10 9 3 0.55 100 2792.1 0.47 100 3037.6 0 100 3184.1
20 9 2 0.34 100 5534.3 0.18 100 5931.0 0 100 5976.0
20 9 3 0.69 95.7 7200 0.16 95.6 7200 0 96.6 7200
Average 0.30 98.94 3297.97 0.29 99.00 3415.77 0.00 99.16 3441.61

Note. β = 0.1, γ = 1.

Table 5
Average values of the varying base rate of deterioration and inventory holdings for the PRPPI under
the OD–OS policy.

β γ % Total % Prod.Q % T.Inv % Trans % N.Visits % N.Vehs % N.Setups

0.1 0.5 98.37 96.28 97.02 97.36 97.45 96.79 100


0.1 0.6 98.69 97.03 98.56 97.89 97.03 97.86 100
0.1 0.7 99.02 97.77 98.37 98.28 97.92 98.92 100
0.1 0.8 99.35 98.51 98.64 98.99 99.31 98.94 100
0.1 0.9 99.67 99.26 99.63 99.43 99.38 99.26 100
0.1 1 100 100 100 100 100 100 100
0.09 1 100.15 100.35 100.32 100.28 100.30 100.25 100
0.08 1 100.31 100.7 100.48 100.34 100.51 100.50 100
0.07 1 100.46 101.06 100.67 100.77 100.51 100.82 100
0.06 1 100.62 101.41 101.18 101.29 101.00 100.96 100
0.05 1 100.77 101.76 101.70 100.87 101.75 101.34 100

Notes. % Total—percentage of average total cost over the one with the base rate of deterioration β =
0.1 and that of inventory holding γ = 1; % Prod.Q—percentage of average production quantity over
the one with the base rate of deterioration β = 0.1 and that of inventory holding γ = 1; % T.Inv—
percentage of average total inventory cost over the one with the base rate of deterioration β = 0.1
and that of inventory holding γ = 1; % Trans—percentage of average transportation cost over the one
with the base rate of deterioration β = 0.1 and that of inventory holding γ = 1; % N.Visits—percentage
of average number of visits over the one with the base rate of deterioration β = 0.1 and that of
inventory holding γ = 1; % N.Vehs—percentage of average number of vehicles used over the one with
the base rate of deterioration β = 0.1 and that of inventory holding γ = 1; % N.Setups—percentage of
average number of production setups over the one with base rate of deterioration β = 0.1 and that of
inventory holding γ = 1.

6.1. Description and data ing to the purchasing statistics, the Suguo community stores have
demand estimates for the product for the following week. The food
In the city of Nanjing, China, most consumer products are dis- company will decide how many items to produce, how many items
tributed through the chain stores of power retailers. Suguo com- with different manufacturing periods to deliver, and the routing
pany is among the biggest power retailers which covers the city plans. The Suguo community stores will decide how many items
with hundreds of stores. Among these stores, Suguo community with different manufacturing periods to sell on each day in the
stores are the target store for the food company. The food com- following weeks. The shortest paths between these sites are calcu-
pany, situated on the north side of the Yangtze river, provides a lated using Google Maps and the cost to distribute these products
certain kind of fresh meat product to Suguo community stores. Ex- is 30 RMB dollars per kilometer. Other parameters are included in
cept for one store, other stores situated on the south side of the the Appendix.
Yangtze river.
The vehicle used for deliveries can carry 8 tons of meat prod-
6.2. Results
ucts. In a typical operations scenario, the food company has only
200 kg of fresh meat products in stock at the beginning of the
Before we bring forward the model and algorithm, the company
planning period. The 40 Suguo community stores have only limited
utilized a three-phase heuristic procedure to obtain a feasible pro-
products in stock at the beginning of the planning horizon. Accord-
duction, delivery and routing plan for a planning horizon of one
202 Y. Qiu et al. / Omega 82 (2019) 193–204

Fig. 2. Route plans for Tuesday to Saturday in the OS–OD policy.

week, assuming the retailers’ selling policy is FPFS. The company vide branch-and-cut algorithms based on logical, strengthened lot-
used a FPFD policy and produced seven days a week. The routing sizing, and lifted MTZ-type valid inequalities; and (iv) to present
plans for the week are as shown in Fig. 2. extensive computational analyses that capture the effects of these
Using the proposed model and algorithms, we can compute si- valid inequalities and different combinations of inventory manage-
multaneously the production plans and the routing plans of the ment policies in various scenarios, from which managerial insight
following week for the food company, and the selling plans of the can be drawn.
following week for the retailers within 2 h. The optimal solution Several extensions are possible for PRPPIs. One worth mention-
suggests that the food company will execute continuous produc- ing here is the possibility of incorporating carbon emissions in
tion of the products from Monday to Saturday given the initial vehicle routing. Another extension would be to consider the per-
stock at the plant and the retailers. The routing plan for Monday is ishable inventory in multi-level production and routing problems
shown in Fig. 1, where only 23 retailers are visited, and the rout- with time windows. Finally, the branch-and-cut algorithm could be
ing plans for Tuesday to Saturday are shown in Fig. 2, where forty improved by incorporating column generation techniques to reduce
retailers are all visited. The total cost of this recommended policy CPU times when dealing with a larger number of available vehicles.
resulted in a 12% decrease.

7. Conclusions Acknowledgments

We have introduced, modeled and analyzed PRPs with perish- P.M. Pardalos is partially supported by the Paul and Heidi
able inventory, a generalization of the multi-vehicle production Brown Professorship. Y. Qiu’s work is supported by NSFC under
routing problem. The contributions of this paper are: (i) to de- grant no. 71571092. Y. Qiu’s work is also supported in part by
scribe a modeling approach enriching the lot-sizing part of produc- General Research Project for Humanities and Social Sciences from
tion routing problems with perishable inventory and correspond- Chinese Ministry of Education under grant no. 11YJCZH137 and the
ing inventory management policies; (ii) to offer a mixed integer Priority Academic Program Development of Jiangsu Higher Educa-
linear programming formulation for PRPPIs and extended formu- tion Institutions (PAPD). The programming help from Dr. Jack Gao
lations for different inventory management policies; (iii) to pro- is also acknowledged.
Y. Qiu et al. / Omega 82 (2019) 193–204 203

Appendix Table A.2


Data for the case study.

# II D D D D D D D IC IHC
Table A.1 0 200 0 0 0 0 0 0 0 60 0 0 40
1 188 94 87 84 98 94 96 86 1500 80
ID Location
2 198 80 85 91 99 89 91 99 1500 80
0 Sushi Logistics Peigong Center 3 180 91 88 87 96 89 81 81 1500 80
1 Suguo Community Store, Baoyan South Road 4 166 84 81 92 80 96 85 90 1500 80
2 Suguo Community Store, Mufu East Road 5 184 95 83 94 81 95 93 94 1500 80
3 China Resources Suguo, Beizushi’an 6 200 83 81 80 90 81 85 87 1500 80
4 Suguo Community Store, Zhong Yang Bei Lu 7 174 92 99 88 86 81 90 80 1500 80
5 Suguo Community Store, Jing Wu Li Jiao Qiao 8 196 98 93 83 93 91 85 95 1500 80
6 Suguo Community Store, Sanpailou Street 9 198 98 96 91 87 93 82 99 1500 80
7 Suguo Community Store, Jiangsu Road 10 180 80 100 88 87 94 83 82 1500 80
8 Suguo Community Store, Nanchang Road 11 164 97 91 94 91 84 83 94 1500 80
9 China Resources Suguo, Suojin East Road 12 198 97 94 93 92 87 87 91 1500 80
10 China Resources Suguo, Huayuan Road 13 200 92 100 92 94 86 81 99 1500 80
11 Suguo Community Store, Lijiang Road 14 162 82 89 85 95 94 96 85 1500 80
12 China Resources Suguo, Fenghuang East Street 15 166 97 100 100 98 91 94 89 1500 80
13 Suguo Community Store, Dashamao Alley 16 166 89 96 87 87 85 89 100 1500 80
14 Suguo Community Store, Bei’anmen Street 17 196 93 94 89 90 99 98 80 1500 80
15 Suguo Community Store, Bei’anmen Street 18 170 82 89 81 94 87 81 85 1500 80
16 Suguo Community Store, Hanzhongmen Street 19 160 82 83 89 90 98 90 85 1500 80
17 Suguo Community Store, Jiefang Road 20 198 95 95 94 96 82 85 87 1500 80
18 Suguo Community Store, Zhongshanmen Street 21 182 83 98 98 87 92 92 100 1500 80
19 China Resources Suguo, Wen’an Street 22 160 92 81 98 95 94 97 82 1500 80
20 China Resources Suguo, Daguang Road 23 186 96 99 99 96 97 97 85 1500 80
21 China Resources Suguo, Yudao Street 24 182 88 85 100 92 96 94 91 1500 80
22 Suguo Community Store, Ruijin Road 25 170 99 86 92 81 93 84 97 1500 80
23 Suguo Community Store, Qinhong Road 26 176 99 100 100 84 86 91 82 1500 80
24 Suguo Community Store, Changhong Road 27 184 88 98 84 93 100 83 98 1500 80
25 China Resources Suguo, Yongle Road 28 170 86 95 94 96 85 96 90 1500 80
26 China Resources Suguo, Yuren South Road 29 180 90 88 91 91 98 82 84 1500 80
27 Suguo Community Store, Guanghua Road 30 168 90 85 93 92 100 83 90 1500 80
28 China Resources Suguo, Yujinxiang Road 31 162 92 88 89 90 83 86 81 1500 80
29 China Resources Suguo, Heyan Road 32 198 81 82 91 82 95 82 100 1500 80
30 Suguo Community Store, Jianning Road 33 192 98 84 92 84 90 83 81 1500 80
31 Suguo Community Store, Yutang East Street 34 184 80 95 95 94 82 88 91 1500 80
32 Suguo Community Store, Jiqingmen Street 35 200 85 90 92 85 87 86 82 1500 80
33 Suguo Community Store, Guangzhou Road 36 194 89 84 87 97 91 81 85 1500 80
34 Suguo Community Store, Hankou West Road 37 186 83 89 88 91 96 94 93 1500 80
35 China Resources Suguo, Changle Road 38 190 97 87 82 80 93 83 82 1500 80
36 Suguo Community Store, Fu Zi Miao Shang Mao Cheng 39 180 80 100 96 86 94 84 88 1500 80
37 China Resources Suguo, Bairun West Road 40 200 87 99 82 100 93 86 91 1500 80
38 Suguo Community Store, Shengzhou Road
Notes. #: site number; II: initial inventory of the site (in kg);
39 Suguo Community Store, Jianye Road
D: demand of the following week (in kg);
40 China Resources Suguo, Hongwu Road
IC: inventory capacity (in kg); IHC: inventory holding cost (in RMB).
Number of vehicles: 3; vehicle capacity: 80 0 0 (in kg);
production capacity: 80,0 0 0 (in kg); setup cost: 1500 (in RMB dollars);
unit production cost: 15 (in RMB Dollars).

References [12] Boudia M, Louly M, Prins C. A reactive GRASP and path relinking
for a combined production-distribution problem. Comput Oper Res
[1] Absi N, Archetti C, Dauzere-Peres S, Feillet D. A two-phase iterative heuristic 2007;34(11):3402–19.
approach for the production routing problem. Transp Sci 2015;49(4):784–95. [13] Boudia M, Prins C. A memetic algorithm with dynamic population man-
[2] Adulyasak Y, Cordeau J-F, Jans R. Formulations and branch-and-cut algorithms agement for an integrated production-distribution problem. Eur J Oper Res
for multivehicle production and inventory routing problems. INFORMS J Com- 2009;195(3):703–15.
put 2014;26(1):103–20. [14] Brahimi N, Aouam T. Multi-item production routing problem with backorder-
[3] Adulyasak Y, Cordeau J-F, Jans R. Optimization-based adaptive large neighbor- ing: a milp approach. Int J Prod Res 2015;54(4):1076–93.
hood search for the production routing problem. Transp Sci 2014;48(1):20–45. [15] Chandra P. A dynamic distribution model with warehouse and customer re-
[4] Adulyasak Y, Cordeau J-F, Jans R. The production routing problem: a review of plenishment requirements. J Oper Res Soc 1993;44(7):681–92.
formulations and solution algorithms. Comput Oper Res 2015;55:141–52. [16] Chandra P, Fisher ML. Coordination of production and distribution planning.
[5] Adulyasak Y, Cordeau J-F, Jans R. Benders decomposition for production rout- Eur J Oper Res 1994;72(3):503–17.
ing under demand uncertainty. Oper Res 2015;63(4):851–67. [17] Coelho LC, Laporte G. A branch-and-cut algorithm for the multi-product mul-
[6] Archetti C, Bertazzi L, Laporte G, Speranza MG. A branch-and-cut algorithm for ti-vehicle inventory-routing problem. Int J Prod Res 2013;51(23–24):7156–69.
a vendor-managed inventory-routing problem. Transp Sci 2007;41(3):382–91. [18] Coelho LC, Laporte G. Optimal joint replenishment, delivery and in-
[7] Archetti C, Bertazzi L, Paletta G, Speranza MG. Analysis of the maxi- ventory management policies for perishable products. Comput Oper Res
mum level policy in a production-distribution system. Comput Oper Res 2014;47:42–52.
2011;38(12):1731–46. [19] Desrochers M, Laporte G. Improvements and extensions to the Miller-Tuck-
[8] Armentano V, Shiguemoto A, Lokketangen A. Tabu search with path re- er-Zemlin subtour elimination constraints. Oper Res Lett 1991;10(1):27–36.
linking for an integrated production-distribution problem. Comput Oper Res [20] Fumero F, Vercellis C. Synchronized development of production, inventory, and
2011;38(8):1199–209. distribution schedules. Transp Sci 1999;33(3):330–40.
[9] Bard JF, Nananukul N. Heuristics for a multiperiod inventory routing problem [21] Glover F. Improved linear integer programming formulations of nonlinear in-
with production decisions. Comput Ind Eng 2009;57(3):713–23. teger problems. Manage Sci 1975;22(4):455–60.
[10] Bard JF, Nananukul N. The integrated production-inventory-distribution-rout- [22] Hsu VN. Dynamic economic lot size model with perishable inventory. Manage
ing problem. J Scheduling 2009;12(3):257–80. Sci 20 0 0;46(8):1159–69.
[11] Bard JF, Nananukul N. A branch-and-price algorithm for an integrated produc-
tion and inventory routing problem. Comput Oper Res 2010;37(12):2202–17.
204 Y. Qiu et al. / Omega 82 (2019) 193–204

[23] Kara I, Laporte G, Bektas T. A note on the lifted Miller-Tucker-Zemlin subtour [27] Lysgaard J, Letchford AN, Eglese RW. A new branch-and-cut algorithm for the
elimination constraints for the capacitated vehicle routing problem. Eur J Oper capacitated vehicle routing problem. Math Prog 20 04;10 0(2):423–45.
Res 2004;158(3):793–5. [28] Miller CE, Tucker AW, Zemlin RA. Integer programming formulation of travel-
[24] Kumar RS, Kondapaneni K, Dixit V, Goswami A, Thakur L, Tiwari M. Multi-ob- ing salesman problems. J ACM 1960;7(4):326–9.
jective modeling of production and pollution routing problem with time win- [29] Pochet Y, Wolsey LA. Production planning by mixed integer programming.
dow: a self-learning particle swarm optimization approach. Comput Ind Eng Springer Science & Business Media; 2006.
2015;99:29–40. [30] Qiu Y, Qiao J, Pardalos P. A branch-and-price algorithm for production routing
[25] Kytöjoki J, Nuortio T, Bräysy O, Gendreau M. An efficient variable neighborhood problems with carbon cap-and-trade. Omega 2017;68:49–61.
search heuristic for very large scale vehicle routing problems. Comput Oper [31] Toth P, Vigo D. Vehicle routing: problems, methods, and applications. Society
Res 2007;34(9):2743–57. for Industrial and Applied Mathematics; 2014.
[26] Lei L, Liu S, Ruszczynski A, Park S. On the integrated production, inventory, and
distribution routing problem. IIE Trans 2006;38(11):955–70.

You might also like