You are on page 1of 14

Computers & Industrial Engineering 133 (2019) 253–266

Contents lists available at ScienceDirect

Computers & Industrial Engineering


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

The order scheduling problem of product-service system with time windows T


a b c,⁎ d
Yang Zhang , Yiran Dan , Bin Dan , Huali Gao
a
School of Management, Chongqing Technology and Business University, Chongqing, China
b
School of Construction Management and Real Estate, Chongqing University, Chongqing, China
c
School of Economics and Business Administration, Chongqing University, Chongqing, China
d
Business School, Southwest University of Political Science & Law, Chongqing, China

ARTICLE INFO ABSTRACT

Keywords: A product-service system (PSS) is a system offering a mix of tangible products and intangible services to meet the
Product-service system personalized needs of customers. This paper studies a PSS order scheduling problem with time windows (PSS-
Order scheduling OSPTW) for a PSS provider composed of a manufacturing plant and a service centre. In this problem, the product
Time window of each PSS order is first produced by the manufacturing plant, and then the installation service is provided by
Metaheuristics
the service centre. A mathematical model for the PSSOSPTW is presented to minimize the total storage and
tardiness cost. To solve this NP-hard problem, an idle time insertion algorithm is proposed to obtain the optimal
start and completion times of production and service of PSS orders for a given order sequence. Three meta-
heuristics based on iterated local search, simulated annealing and tabu search are developed to search for the
best sequence through incorporating the idle time insertion algorithm. All metaheuristics utilize an adapted
version of the NEH heuristic to generate initial solutions, and apply the idea of variable neighbourhood search to
find local optimum. The performance of the proposed algorithms is evaluated by comparison, and the effects of
some instance factors are also examined. The results show that our algorithms perform very similarly in general
but quite differently in various situations. Besides, the problem size, tardiness factor, due date range factor and
time window width range factor have noticeable impact on their performance.

1. Introduction In equipment manufacturing industry (such as elevator and wind


power, etc.), equipment procurement and installation contracts are a
In order to cope with the increasingly fierce market competition and common type of PSS orders. In such type of PSS orders, a customer-spe-
meet customers’ personalized and diversified service needs, many tra- cified time window is usually given to require the PSS provider to provide
ditional manufacturing firms have shifted from simply selling products the installation service based on this time window. If the product is fin-
to providing comprehensive solutions of products and their related ished earlier than the time window, the PSS provider cannot provide the
services, namely, product service systems (PSSs) (Baines et al., 2007; service immediately due to various reasons, for instance, installation
Boehm & Thomas, 2013). For instance, IBM and GE have transformed conditions are not available. If idle service resources are unavailable when
themselves from pure product manufacturers to PSS providers by pro- the production is completed, it is also impossible for the PSS provider to
viding products and related services such as technology support, start providing services immediately. In both cases, the PSS provider must
maintenance and training (Sawhney, Balasubramanian, & Krishnan, deal with the storage problem and bear the corresponding storage cost,
2004). Research findings have revealed that PSS can not only improve which depends on the storage time and insurance (Behnamian, Ghomi, &
the competitiveness and profitability of firms (Geng, Chu, Xue, & Zandieh, 2010). If the PSS provider fails to deliver an order within its time
Zhang, 2010), but also improve customer satisfaction and loyalty window, it is deemed as a breach of contract and will incur the tardiness
(Baines, Lightfoot, Benedettini, & Kay, 2009; Beuren, Ferreira, & cost correlated with the tardiness time and the penalty stipulated in the
Miguel, 2013). Additionally, PSS is beneficial to saving resources, re- contract. Besides, tardiness may lead to a decrease in customer satisfaction
ducing pollution and improving manufacturing sustainability (Lindahl, and a potential loss of goodwill (Chung, Chan, & Ip, 2011). To reduce
Sundin, & Sakao, 2014; Tukker, 2015). Therefore, PSS is conducive to operational cost and enhance customer satisfaction, how to schedule PSS
achieve a high degree of unity of social, environmental and economic orders reasonably according to the time window limits of the customers is
benefits. a practical problem faced by the PSS providers.


Corresponding author.
E-mail address: danbin@cqu.edu.cn (B. Dan).

https://doi.org/10.1016/j.cie.2019.04.055
Received 10 January 2019; Received in revised form 5 April 2019; Accepted 30 April 2019
Available online 11 May 2019
0360-8352/ © 2019 Elsevier Ltd. All rights reserved.
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

The complexity and difficulty of the PSS order scheduling mainly logistics tasks dynamic optimization model. To improve the profit-
stem from the interaction between the production and service stages. ability of a use-oriented PSS, Liu, Yang, Pei, Liao, and Pohl (2019)
Since the service can start after the completion of the production, the formulate a discrete-time Markov Decision Process to maximize the
hybrid system operated by a PSS provider can be seen as a serial system long-term revenue per period. The above literature provides valuable
composed of a manufacturing subsystem and a service subsystem experiences and references for the marketing and operations manage-
(Wang, Jiang, Li, & Geng, 2013). On the one hand, early completion of ment of PSS, but to our best knowledge, the PSS order scheduling
production will cause the product to wait in the warehouse of the problem remains to be addressed in the field of operations manage-
manufacturing subsystem, thus incurring the storage cost. On the other ment.
hand, if the production cannot complete on time, the provision of ser- The existing literature on integrated production and service sche-
vice by the service subsystem will be delayed, which increases the risk duling mainly focuses on the integrated production and distribution
of tardiness. Therefore, PSS providers need to make a tradeoff between scheduling area. Li, Ganesan, and Sivakumar (2005) address the pro-
the storage and tardiness cost as well as coordinate the production and blem of synchronized scheduling of assembly with air-transportation to
service schedule. In reality, the interaction between the production and minimize cost in a consumer electronics supply chain. Ma, Chan, and
service stages often makes it difficult for PSS providers to effectively Chung (2013) develop an integrated model of production scheduling
manage their systems and businesses (Lockett, Johnson, Evans, & Bastl, with shipping information, and propose a two-level genetic algorithm
2011), causing high operational cost. How to achieve coordinated to deal with the model. Massengale, Larson, Dai, and Jr (2014) study
scheduling of the production and service is another critical problem for the problem of synchronizing production scheduling and rail trans-
PSS providers. portation allocation of orders to optimize customer service at minimum
To deal with the above issues, we consider a PSS provider composed total cost. Karimi and Davoudpour (2015) study coordination between
of a manufacturing plant and a service centre. The manufacturing plant production scheduling and distribution in a supply chain network, and
first produces the products of PSS orders, and then the service centre is present a branch and bound method for solving this problem. Wang,
responsible for providing the corresponding installation services based Ma, Luo, and Qin (2016) develop a hybrid genetic algorithm based on
on time windows specified by customers. We study how to schedule PSS the opposition-based learning to solve the coordinated production and
orders so as to minimize the total storage and tardiness cost, and pro- transportation scheduling in a two-stage assembly flowshop. Through
pose some effective algorithms. The contribution of this research is to combining mixed integer programming, discrete event simulation and a
provide PSS providers with effective modelling and solving tools for genetic algorithm, Frazzon et al. (2018) propose a hybrid approach for
scheduling PSS orders, so that the operational cost can be reduced and the integrated scheduling of production and transport processes in the
the operational efficiency can be improved. framework of supply chain management. Still in supply chain context,
The remainder of this paper is organized as follows. Section 2 re- Marandi and Fatemi Ghomi (2019) investigate an integrated multi-
views relevant literature. Section 3 describes the problem and for- factory production and transportation scheduling problem with the
mulates a mathematical model. Section 4 proposes an idle time inser- objective of minimizing the total tardiness and transportation cost.
tion algorithm. Section 5 develops three metaheuristics. Section 6 More early research on issues of integrated production and distribution
conducts some computational experiments. Section 7 draws conclusions scheduling can be referred to Chen (2010). The above literature con-
and gives suggestions on future work. siders that a transportation vehicle can provide delivery service for
multiple orders at a time, or the products of an order can be split into
2. Literature review multiple batches to be delivered by different transportation vehicles. In
the PSS order scheduling, it is considered that the service centre can
Three streams of literature are closely related to this paper. They are only perform one installation service task at a time, taking into account
PSS, integrated production and service scheduling and order sche- that the installation service cannot be split; additionally, if the in-
duling. stallation cannot start immediately after the product is finished, the
At present, PSS has been widely studied in academia. The majority storage cost will be incurred. Therefore, previous solutions to the in-
of relevant literature is qualitative research on the concept, design and tegrated production and distribution scheduling cannot be applied to
case study of PSS (Mahut, Daaboul, Bricogne, & Eynard, 2017). In re- the PSS order scheduling problem, and new scheduling strategies need
cent years, the quantitative research of PSS has gradually attracted to be developed for this problem.
scholars’ attention. In the field of marketing, Xie, Jiang, Zhao, and Shao Customer order scheduling has become an important and hot issue
(2014) study three common contracts under information asymmetry in in manufacturing and service operations management. Wang and
a supply chain providing PSS, and find that the retail price maintenance Cheng (2007) consider that each customer order contains several dif-
contract benefits the manufacturer while the franchise fee contract does ferent jobs to be processed on multiple dedicated facilities, and study
good to the retailer. Lee, Yoo, and Kim (2016) investigate the compe- the corresponding order scheduling problem to minimize the total
tition between a conventional channel providing separated goods and weighted order completion time. Chen and Pundoor (2009) study an
services and a servitized channel providing PSSs, and reveal that the integrated order scheduling and packing problem under the constraints
service dependency and channel substitutability are critical for the of the capacity of the delivery batch and the customer service level. Xu
manufacturer to decide whether to adopt servitization. Pascual, Siña, et al. (2016) propose a multiple-machine order scheduling problem
Santelices, Román, and Droguett (2017) develop a quantitative tool for with position-based learning effect to minimize the total tardiness.
the design of use-oriented PSS contracts based on non-repairable Based on particle swarm optimization, Lin et al. (2017) develop three
component reliability and risk sharing. In the field of operations man- metaheuristics for a two-agent multi-facility order scheduling with
agement, various optimization models for PSS are developed. For in- ready times. With the objective of minimizing the number of tardy jobs,
stance, Kuo and Wang (2012) use multi-attribute utility analysis to Wu, Liu, Zhao, Wang, and Lin (2017) study a multi-machine order
develop different maintenance service models, in which the operation scheduling problem with learning. Wu et al. (2018) introduce an order
cost, CO2 emission and service satisfaction are decision variables. Li and scheduling problem with multiple objectives, and propose both heur-
Jiang (2013) construct a block structure Markov chain optimization istics and metaheuristics to solve this problem. Framinan and Perez-
model for a PSS with additional service capacity and impatient custo- Gonzalez (2018) develop some constructive heuristics for the order
mers. Shokohyar, Mansour, and Karimi (2014) integrate services and scheduling problem with total tardiness minimization. Wu et al. (2019)
product end-of-life management in sustainable PSS to develop an op- address a customer order scheduling problem with sum-of-processing-
timization model. Zhang, Liu, Liu, and Li (2016) investigate the PSS- time-based learning effect on multiple machines, so that the total tar-
based smart box-enabled sustainable logistics services, and present a diness of the orders is minimized. Unlike the above studies, which only

254
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

consider production orders, this paper deals with the more complicated Order sequence, = ( (1), (2), …, (n)) , where (j ) denotes the j th
scheduling problem of PSS orders, each of which is composed of both order, j Q

product and service that have to be processed by a manufacturing plant


and a service centre, respectively. 3.2. Mathematical formulation

3. Problem description and formulation Based on the above assumptions and notations, a mathematical
model for the PSSOSPTW is formulated as follows:
Consider a PSS provider composed of a manufacturing plant and a
service centre. At time zero, a set N = {1, 2, …, n} of n PSS orders si- min Z = (hi · INVi + bi ·TARi )
(1)
multaneously arrive at the PSS provider. Order i , i N , consists of one i N

unit of product and its corresponding installation service. Associated subject to:
with order i is a time window [ei , di], where ei is the earliest allowable
CTPi = STPi + pti, i N (2)
service time, and di is the due date. Each PSS order must go through the
production and service stages in turn to complete delivery, and all PSS STPj CTPi + M ·(x ij 1), i, j N, i j (3)
orders are processed in the same sequence. In the production stage,
products are produced by the manufacturing plant according to the STSi CTPi, i N (4)
MTO (make-to-order) mode. In the service phase, installation services STSi ei, i N (5)
are provided by the service centre for customers. If the production of
order i completes earlier than its earliest allowable service time, or CTSi = STSi + sti, i N (6)
there is no idle resource to be assigned providing the service, the pro-
STSj CTSi + M·(xij 1), i, j N, i j (7)
duct will be stored in the warehouse at the cost of hi per unit time. If the
service delivery of order i is tardy, the tardiness penalty will be incurred x ij + x ji 1, i, j N (8)
at the cost of bi per unit tardiness time. Generally, the revenue, pro-
duction cost and service cost of each order are fixed, so the objective of x ij 1, i N
the PSS provider is to determine the optimal sequence as well as the j N (9)
optimal start and completion times of production and service of PSS
xij 1, j N
orders to minimize the total storage and tardiness cost. For con- (10)
i N
venience, we simply denote the above described problem by PSSOSPTW
(the PSS order scheduling problem with time windows). (x ij + xji) 1, i N
j N (11)
3.1. Assumptions and notations
x ij = n 1
To formulate the PSSOSPTW, the following assumptions are made: i N j N (12)

INVi STSi CTPi, i N (13)


(1) Preemption is not allowed once the production or service starts.
(2) The manufacturing plant and service centre can only perform one TARi CTSi di , i N (14)
task at a time. STPi 0, CTPi 0, STSi 0, CTSi 0, INVi 0, TARi 0, i N
(3) No interruption occurs during the process of production and ser-
(15)
vice.
x ij {0, 1}, i, j N (16)
The notations for the PSSOSPTW are described as follows:
In Eq. (1), the objective function is to minimize the total storage and
tardiness cost, where the storage cost of each order is the unit storage
Input parameters cost multiplied by the storage time, and the tardiness cost of each order
n Total number of orders is the unit tardiness cost multiplied by the tardiness time. Constraint (2)
Q Collection of positions in an order sequence with n orders, Q = {1, 2, …, n} indicates that the production process is not allowed to be interrupted.
pti Production time of order i
Constraint (3) ensures that the production start time of an order cannot
sti Service time of order i
ei Earliest allowable service time of order i ; To make ei play a function of be earlier than the production completion time of its immediate pre-
constrain, it is assumed that the earliest allowable service time of order i is not decessor. Constraint (4) imposes that the service start time of an order
earlier than its earliest production completion time, namely, ei pti cannot be earlier than its production completion time. Constraint (5)
di Due date of order i assures that the service start time of an order cannot be earlier than its
ceni Centre of the time window [ei, di], ceni = (ei + di) 2
earliest allowable service time. Constraint (6) represents that the ser-
twi Time window width of order i , twi = di ei ; To make twi reasonable, it is
assumed that the time window width of order i is not less than its service time,
vice process is not allowed to be interrupted. Constraint (7) ensures that
namely, sti twi the service start time of an order cannot be earlier than the service
hi Storage cost per unit time of order i completion time of its immediate predecessor. Constraints (8)–(12)
bi Tardiness cost per unit time of order i represent the order processing sequence. Constraints (13) and (14)
A very large positive number
M
determine the storage and tardiness time, respectively. Constraints (15)
Variables and (16) define the domain of the decision variables.
CTPi Production completion time of order i
CTSi Service completion time of order i Theorem 1.. The PSSOSPTW problem is NP-hard.
INVi Storage time of order i
TARi Tardiness time of order i Proof.. See the Appendix.
Decision variables Based on the mathematical model of the PSSOSPTW, we present a
STPi Production start time of order i timetabling algorithm to obtain the earliest start and completion times
STSi Service start time of order i
of production and service of PSS orders for a given sequence. The idea is
xij If order j is immediately after order i , then xij = 1; otherwise, xij = 0
simple, that is, let the manufacturing plant produce products and the
service centre provide services as soon as possible. Pseudocode 1 shows

255
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

the proposed timetabling algorithm, where (0) represents the dummy Proof.. See the Appendix.
order. This algorithm will be used for designing an idle time insertion
It is a very difficult problem to determine whether there should be
algorithm in the next section.
idle time between adjacent PSS orders. Although Theorem 2 partially
solves this problem, it is only a sufficient unnecessary condition, since
Pseudocode 1 some adjacent orders that should be produced successively may not
satisfy the conditions in this theorem. At the end of this section, such a
procedure Timetabling( ) counterexample will be given. Because this problem has not yet been
CTP (0): =0;
solved completely, further research is required in the future. Based on
CTS (0): =0;
Theorem 2, the following corollary can be obtained.
for j: =1 to n do
STP (j ): =CTP (j 1); Corollary 1.. Consider an order sequence , where orders (j) and
CTP (j ): =STP (j ) + pt (j );
(j + 1) satisfy pt (j + 1) st (j) and d (j + 1) e (j) pt (j + 1) + st (j + 1) .
STS (j ): =max{CTS (j 1), max{e (j) , CTP (j )}}; If the service completion time of order (j) is also the service start time of
CTS (j ): =STS (j) + st (j );
order (j + 1) , then INV (j) INV (j + 1) . If both orders (j) and (j + 1)
INV (j ): =STS (j ) CTP (j);
are tardy, and the production completion time of order (j) is also its service
TAR (j ) : =max{CTS (j ) d (j) , 0};
start time, then TAR (j) TAR (j + 1) .
end for
return STPi, CTPi, STSi, CTSi, INVi , TARi Proof.. See the Appendix.
end
Inspired by the net benefit of movement (NBM) algorithm proposed
Proposition 1.. Given an order sequence , a lower bound of the service by Tseng and Liao (2008) for a lot-streaming flowshop scheduling
start time of the last order in this sequence is given as problem, and the sub-block move presented by Kedad-Sidhoum and
Sourd (2010) for a single machine scheduling problem, the idea of our
idle time insertion algorithm for the PSSOSPTW is stated as follows.
¯ = max min ei +
STS sti max sti, max min ei, pti
We first apply the timetabling algorithm to obtain an initial sche-
i N
i N
i N i N
i N (17)
dule. Since orders have been scheduled as soon as possible, only by
A lower bound of the service completion time of the last order is given right shifting move can we reduce the total cost, i.e., by the insertion of
as idle time before orders. However, it is a problem to determine how
many units of idle time should be inserted before each order. Consider
the blocks of successive orders separated by idle times in the production
¯ = max min ei +
CTS sti, max min ei , pti + min sti phase. It is enough to calculate the optimal amount of movement for
i N
i N
i N
i N
i N
(18)
each block, because all the orders in a block move the same amount.
Denoting any of these blocks of orders by B , we divide the orders in B
Proof.. See the Appendix. into four different subsets based on the following two judgements. One
Proposition 1 indicates that the time to complete the delivery of all judgement is whether the service starts later than the production
orders is not earlier than the time defined by Eq. (18). Hence, if the unit completion time or not, and the other is whether the service completes
tardiness cost of an order is very expensive and the due date of this before the due date or not. According to the affirmative (A) or negative
order is earlier than the time defined by Eq. (18), it is suggested that the (N) answer to these two judgements, the four subsets divided are
decision maker should avoid placing this order in the final position. BAA = {i B| CTPi < STSi, CTSi < di} ,
In order to reduce the complexity of the PSSOSPTW problem, we BAN = {i B| CTPi < STSi, CTSi di } ,
decompose it into two sub-problems, namely, determining the optimal BNA = {i B| CTPi = STSi, CTSi < di}
start and completion times of production and service of PSS orders for and BNN = {i B| CTPi = STSi, CTSi di} , respectively. Assume that
an order sequence and finding the optimal sequence. The solutions to the minimum idle time insertion is one unit. If we insert one unit of idle
these two sub-problems will be described in Sections 4 and 5, respec- time before block B , then the storage cost will be reduced by i BA hi ,
tively. where BA = BAA BAN , meanwhile the tardiness cost will be increased
by i B NN bi . Therefore, the condition for determining whether idle
time should be inserted before B is whether the reduced storage cost is
4. An idle time insertion algorithm for the PSSOSPTW
greater than the increased tardiness cost. If i BA hi > i BNN bi , then
we can keep inserting idle time until either set BA or BNN changes.
This section discusses how to deal with the first subproblem. If all
According to this, the maximum idle time insertion for B is calculated
PSS orders are scheduled as soon as possible, then many orders are
as 1 = min{mini BA {INVi }, mini BNA {di CTSi}} . Insertion of more than
likely to complete production before their services start, resulting in
1 units of idle time will cause an order to either move from BA to B NA
high storage cost. Therefore, an idle time insertion algorithm is needed
or move from BNA to BNN . In addition, there might be another block of
to schedule the production tasks as close to their service start times as
orders after B , thus we denote the idle time between both blocks by 2 .
possible. In fact, idle time insertion has been widely applied in the
Considering all the above, the maximum idle time insertion before B
earliness-tardiness production scheduling (e.g., Hendel & Sourd, 2007;
actually is = min{ 1, 2} . The above procedure is applied to all blocks
Kedad-Sidhoum & Sourd, 2010; Pan, Ruiz, & Alfaro-Fernández, 2017),
of orders, starting from the last order to the first order and considering
but its application in the PSS scheduling area is still a gap. Note that
one order at a time.
services in the PSSOSPTW problem are non-storable and do not typi-
cally incur additional cost once completed, so we only need to consider
idle time insertion in the production stage, and we can assume that all Table 1
services are provided as soon as possible. Before putting forward our The parameters of the example.
idle time insertion algorithm for the PSSOSPTW, some conclusions are
Order i pti sti ei di hi bi
drawn.
1 2 4 4 10 2 3
Theorem 2.. Given an order sequence , if orders (j) and (j + 1) satisfy
2 3 2 6 9 1 2
pt (j + 1) st (j) and d (j + 1) e (j) pt (j + 1) + st (j + 1) , then idle time 3 6 3 10 15 3 2
should not exist between these two orders in the production stage.

256
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Fig. 1. Gantt charts for the example with the application of the idle time insertion algorithm, where ‘MP’ and ‘SC’ represent the manufacturing plant and service
centre, respectively.

Next, we present an example to illustrate the idle time insertion


Pseudocode 2
algorithm. The parameters of the example is given in Table 1, and the
procedure IdleTimeInsertion( ) given sequence is = (1, 2, 3) . The initial schedule is given in Fig. 1(a).
j: =n ; Orders 1 and 2 finish production earlier than their service start times
Timetabling( ); (INV1 = 2 , INV2 = 3), respectively. Order 3 completes production within
while { j 1} do its time window. Orders 1 and 3 deliver service on time, whereas order
Construct a block of orders B starting from (j);
2 is tardy (TAR2 = 1). Therefore, the initial total cost Z = 9 . The pro-
Generate BAA , BAN, BNA and BNN from B, respectively;
cedure starts with order 3 to construct a block B = {3} . However,
BA : =BAA BAN ;
BA = does not satisfy the conditions of idle time insertion, so we
BN : =BNA BNN ;
continue to consider orders 2 and 3 to form a new block B = {2, 3} .
if ((BA )&(BNN = )) ( (B , B
A NN )& ( i BA hi > i BNN bi ) ) then Since BAN = {2} , BNA = {3} and BNN = , idle time should be inserted
if BNA then before B . There is no order after B , thus 2 = + . Considering that
1: =min{mini BA {INVi }, mini B NA {di CTSi}};
1 = min{min{1}, min{15 14}} = 1, = 1 time units should be in-
Else
serted before B , resulting in that the total cost becomes Z = 8. This
1: =mini BA {INVi };
schedule is shown in Fig. 1(b). Then, BNN has changed to BNN = {3} .
end if
if there is an order to the right of B then Since i BA hi = 1 is less than i B NN bi = 2 , inserting idle time before
Calculate idle time 2 between B and this order; B will not improve the total cost. Next, order 1 is considered to form a
else block B = {1} . Since BAA = {1} and BNN = , we can improve the total
2 : =+ ; cost by inserting idle time before this block. Notice that order 2 is after
end if
B , the units of idle time between order 2 and B is 2 = 1. Considering
: =min{ 1, 2};
Insert time units before B;
that 1 = min{2} = 2 , = 1 is what we insert before B , generating a
STPi: =STPi + , i B; schedule with Z = 6 as shown in Fig. 1(c). The final block constructed is
CTPi: =CTPi + , i B; B = {1, 2, 3} , where BAA = {1} , BAN = {2} , BNA = and BNN = {3} . Be-
STSi: =STSi + , i BN; cause i BA hi = 3 is greater than i B NN bi = 2 , inserting idle time is
CTSi: =CTSi + , i BN ;
profitable. Similar to the above procedure, it is easy to know that
INVi: =INVi , i BA ;
TARi : =TARi + , i BNN ; 2 = + and 1 = min{1, 2} = 1. Therefore, = 1 time units should be
else inserted before B . Now BNA has changed to BNA = {1} , and further in-
j: =j 1; serting idle time before B will not improve the total cost. As a result, the
end if idle time insertion procedure ends, and the final schedule with the
end while
optimal total cost Z = 5 is shown in Fig. 1(d).
return {STPi, CTPi, STSi, CTSi, INVi and TARi }
end Fig. 1(d) indicates that for the sequence = (1, 2, 3) , all orders
should produce successively to minimize the total cost. However,
pt2 = 3 < 4 = st1 does not satisfy the conditions in Theorem 2. This
Pseudocode 2 shows the idle time insertion algorithm. For a sequence verifies that Theorem 2 is only a sufficient unnecessary condition for
containing n orders, the calculation of 1 may require traversing all deciding whether idle time should exist between adjacent PSS orders.
orders, thus the computational complexity of the idle time insertion
algorithm is O (n2) in the worst case. 5. Metaheuristics for the PSSOSPTW
Theorem 3.. The idle time insertion algorithm produces an optimal schedule
Since the PSSOSPTW is NP-hard, no polynomial time algorithm can
for a given sequence.
possibly solve it. Therefore, we propose simple metaheuristics based on
Proof.. This theorem can be easily proved in a similar way as Theorem iterated local search (ILS), simulated annealing (SA) and tabu search
2 in (Tseng & Liao, 2008), hence the proof is omitted here. (TS) for solving this problem. The reason for using simple

257
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

metaheuristics instead of complex approaches such as genetic algo- Farias, 2010) and the flowshop scheduling problem (e.g., Pan & Ruiz,
rithm, ant colony optimization and etc. has three aspects. First, these 2012). Starting from an initial solution, ILS first performs a local search
single-solution-based metaheuristics are easier to understand, imple- to find a local optimum, which is used to update the current solution
ment and extend to other problem variants. Second, they have fewer according to an acceptance criterion, and then carries out a perturba-
parameters, hence the experimental results are easy to replicate. Third, tion to escape from it. This process iterates until the termination cri-
recent studies on many scheduling problems (e.g., Naderi, Ruiz, & terion is met. Since ILS is simple and effective, we apply it to solving the
Zandieh, 2010; Cheng, Chung, Liao, & Lee, 2013; Gao, Chen, & Deng, PSSOSPTW.
2013) indicate that simple metaheuristics can give state-of-the-art re- The proposed ILS applies a variable neighbourhood search (VNS)
sults when compared to complex approaches. (Mladenović & Hansen, 1997) method to find promising solutions in the
In Section 5.1, we first present an adapted version of the NEH solution space. The commonly used insert and swap neighbourhoods
heuristic as an initial solution generator for our metaheuristics. In form the neighbourhood structures, which are dynamically changed
Sections 5.2–5.4, we introduce the proposed metaheuristics. Note that during the search process. Both neighbourhoods are defined based on
the encoding scheme for the PSSOSPTW is a simple permutation of PSS the moves of orders. The insert move is defined by removing a ran-
orders, and the idle time insertion algorithm is incorporated into these domly selected order from its original position and inserting it into
metaheuristics to evaluate the solutions. another random position, while the swap move is defined by swapping
the positions of two randomly selected orders. In the insert (swap)
neighbourhood search, the insert (swap) move is performed n times on
5.1. An adapted NEH heuristic for the PSSOSPTW the current solution to generate n neighbourhood solutions, and the
best one among them is selected. Pseudocode 4 shows both neigh-
Based on the NEH algorithm proposed by Nawaz and Ham (1983), bourhood searches. The VNS method always prefers the insert neigh-
we develop an adapted NEH algorithm called NEHPSS for the PSSOS- bourhood search to the swap one, unless the former fails to obtain a
PTW. The idea of NEHPSS is as follows. better neighbourhood solution than the current one. The VNS method is
Firstly, three initial sequences are generated by the following three shown in Pseudocode 5.
dispatching rules, respectively:

• EDT (earliest delivery time): sort the PSS orders in the increasing Pseudocode 4

order of their earliest delivery times e + st , i N.


• EDD (earliest due date): sort the PSS orders in the increasing order
i i procedure NeighbourhoodSearch( , i)
:= ;
of their due dates d , i N. for j: =1 to n do

• MSS (minimum service slack): jointly consider the time windows


i
if i = 1 then
: =insert( );
and service times, and sort the PSS orders in the increasing order of
else if{i = 2 } then
(di ei sti ) , i N. : =swap( );
end if
Secondly, the NEH insertion procedure is performed for each initial Add to ;
sequence. That is, the first two orders in a given initial sequence are sorted end for
to form a partial solution, and then the remaining orders are inserted into : =best permutation in ;
the best position of all possible positions in the partial solution one by one. return
Finally, three new complete solutions are obtained and the best one is end

selected as the final schedule. Intuitively, combing the above three dis-
patching rules can yield better results than just using a single dispatching
rule. Pseudocode 3 shows the details of the NEHPSS heuristic, which is used
to generate initial solutions for our proposed metaheuristics.
Pseudocode 5

Pseudocode 3 procedure VNS( )


i: =1;
procedureNEHPSS while i 2 do
:= ; : =NeighbourhoodSearch( , i );
Set the EDT rule as R1, the EDD rule as R2 and the MSS rule as R3; if Z ( ) < Z ( ) then
for i: =1 to 3 do := ;
Sort PSS orders by Ri to generate an initial sequence 0: =( 0 (1), 0 (2), …, 0 (n)); i: =1;
Sort the first two orders of 0, and the schedule with smaller total cost will be the partial Else
solution : =( (1), (2)); i : =i + 1;
for j: =3 to n do end if
Insert order 0 (j ) in all possible positions of , and the schedule with the smallest total end while
cost will be the partial solution : =( (1), (2), …, (j )); return
end for end
Add to ;
end for
: =best solution in ; The perturbation procedure is essential to ILS (Lourenço et al., 2010). It
return moves a solution away from a local optimum to a more promising re-
end gion in the solution space. Inspired by the exchange move and the
perturbation operator proposed by Naderi et al. (2010), this paper
presents an exchange-based perturbation procedure. The exchange
5.2. The ILS algorithm move is defined by selecting, at random, ( 3) different orders from
the current solution and rearranging these orders randomly. The pro-
ILS proposed by Lourenço, Martin, and Stützle (2010) has been cedure first performs the exchange move times on the current solu-
successively applied to various scheduling problems such as the vehicle tion to generate new solutions, and then selects the best one among
routing problem (e.g., Subramanian, Drummond, & Bentes, Ochi, & them as the perturbation solution. Pseudocode 6 provides the

258
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

perturbation procedure. 5.3. The SA algorithm

SA proposed by Kirkpatrick, Gelatt, and Vecchi (1983) is a sto-


Pseudocode 6
chastic algorithm inspired by an analogy between the physical an-
procedure Perturbation( , , ) nealing of solid matters and optimization problems. It utilizes the an-
:= ; nealing schedule and the acceptance criterion of Metropolis rule to
for i: =1to do escape from local optimums and achieve global optimization. As a
: =Exchange( , ); popular metaheuristic, SA has been applied to many combinatorial
Add to ;
optimization problems such as the travelling sales man problem (e.g.,
end for
: =best permutation in ;
Geng, Chen, Yang, Shi, & Zhao, 2011), the maintenance scheduling
return problem (e.g., Safaei, Banjevic, & Jardine, 2012).
end Different from the traditional SA algorithm, our proposed SA adopts
a neighbourhood structure similar to the presented VNS method. In
other words, the VNS is performed instead of simply generating one
After a new solution is generated by the local search procedure, a new solution for comparison in each iteration and repeating this process
decision needs to be made on whether to accept it as the current so- for a fixed iteration number at the same temperature. The reasons for
lution for the next iteration. We adopt a SA-like adopting the VNS in SA are two folds. First, we can take advantages of
acceptance criterion with a constant temperature similar to the one the VNS to expand the search scope of the solution space. Second, we
in (Osman & Potts, 1989). In this acceptance criterion, if is better believe that local search can be performed more fully if multiple
than , then it is accepted; otherwise, is accepted with a probability neighbourhood searches on the current solution are conducted before
of Pr ( Z ) = exp( Z T ) , where Z = Z ( ) Z ( ) and T is the applying the acceptance criterion. The acceptance probability of a non-
temperature value. Karabulut (2016) suggests that T should be in the improving solution is also Pr ( Z ) = exp( Z T ) as defined in the ILS
same scale of the objective value. Since our objective function is the algorithm. In the annealing schedule, the exponential annealing func-
total cost determined by the storage and tardiness time, we consider a tion is used here to decrease the temperature, i.e., T : = · T , where
temperature value defined by (0, 1) is the annealing rate. Pseudocode 9 shows the details of the
n ¯
[hi (STS ¯
pti ) + bi max{CTS di , 0}] proposed SA algorithm.
i=1
T = T0
100n (19)
where STS¯ and CTS
¯ are defined by Eqs. (17) and (18), respectively, and Pseudocode 9
T0 is a parameter to be calibrated. The acceptance criterion is shown in
procedure SA
Pseudocode 7, where Rand is a random number generated from a Set parameters and T0;
uniform distribution with the range [0, 1]. T: =formulation(24);
: =NEHPSS ;
*: = ;
Pseudocode 7 while termination criterion is not satisfied do
i: =1;
procedure AcceptanceCriterion( , , T) while i 2 do
if Z ( ) < Z ( ) then : =NeighbourhoodSearch( , i );
:= ; if Z ( ) < Z ( ) then
else ifRand exp( Z T ) then := ;
:= ; i: =1;
end if if Z ( ) < Z ( *) then
return *: = ;
end end if
else
if Rand exp( Z T ) then
Finally, a complete framework of the proposed ILS algorithm is pro- := ;
vided in Pseudocode 8. i: =1;
else
i : =i + 1;
Pseudocode 8 end if
end if
procedure ILS end while
Set parameters , and T0; T : = ·T ;
T: =formulation (24); end while
return *
0: =NEHPSS ;
: =VNS( 0); end
*: = ;
while termination criterion is not satisfied do
: =Perturbation( , , ); 5.4. The TS algorithm
: =VNS( );
if Z ( ) < Z ( *) then TS proposed by Glover (1986) is an extension of local neighbour-
*: = ; hood search. In TS, a tabu list is adopted to jump out of local optimums
end if and avoid circuitous search, and an aspiration criterion is used to in-
: =AcceptanceCriterion( , , T ); validate the tabu status of a move. TS also has extensive applications in
end while various of scheduling problems such as the vehicle routing problem
return *
(e.g., Cordeau & Maischberger, 2012) and the distributed permutation
end
flowshop scheduling problem (e.g., Gao et al., 2013).
The outline of the proposed TS is quite similar to that of the pro-
posed SA, except that a tabu list is involved and the temperature value

259
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Fig. 2. Means plot and LSD intervals for , and T0 in ILS.

is set as a constant defined by Eq. (19). The tabu list is used to record Pseudocode 10
the recently accepted neighbourhood moves which enter and exit the
end while
tabu list in first-in-first-out fashion. Tabu list size represents the number end while
of iterations that a neighbourhood move remains as tabu. We determine return *
the tabu list size depending on the order size n . More precisely, we set end procedure
the tabu list size to n , which denotes the maximum integer that is
not larger than n . The aspiration criterion used here is to allow a tabu
6. Computational evaluation
move if it is better than the current solution. The details of the proposed
SA algorithm are shown in Pseudocode 10.
In this section, we conduct simulation experiments to evaluate the
performance of the proposed ILS, SA and TS algorithms. These algo-
Pseudocode 10 rithms are implemented in MATLAB 2016b and run on a PC with
3.20 GHz Intel (R) Core (TM) i5-3470 CPU and 8G RAM.
procedure TS Since the PSSOSPTW is a novel scheduling problem and no bench-
Set parameterT0;
mark instances have been found in the existing literature, we randomly
T: =formulation (24);
: =NEHPSS ;
generate a set of test instances. For each order i , both the production
*: = ; time pti and service time sti are generated from a uniform distribution
TabuList: ={ }; with the range [1, 100], and both the unit storage cost hi and tardiness
while termination criterion is not satisfied do cost bi are generated from a uniform distribution with the range [1, 10].
i: =1; In reference to Potts and Van Wassenhove (1982), the centre of the time
while i 2 do
window ceni is generated from a uniform distribution with the range
: =NeighborhoodSearch( , i );
while ( TabuList )&(Z ( ) Z ( )) do
[PS (1 R 2), PS (1 + R 2)], where PS represents the earliest
: =NeighborhoodSearch( , i ); possible time for all orders to be delivered, is the tardiness factor and
end while R is the due date range factor. Since this paper assumes that the earliest
if Z ( ) < Z ( ) then allowable service time of each order is not earlier than its earliest
:= ; production completion time, i.e., ei pti , PS can be obtained from
i: =1; Proposition 1:
if Z ( ) < Z ( *) then
*: = ;
end if PS = max pti + min sti, min pti + sti
i N i N
else i N i N
if Rand exp{ (Z ( ) Z ( )) T } then
:= ; The time window width twi follows a uniform distribution with the
i: =1; range [ceni 100, ·ceni ], where represents the time window width
else range factor. Based on ceni and twi , the earliest allowable service time
i : =i + 1; ei = max{ceni twi 2, pti} and due date di = max{ceni + twi 2, ri + sti} .
end if
end if
The test instance set contains complete combinations of
if (i = 1) &( TabuList ) then n = {20, 30, 40, 50} , = {0.2, 0.3, 0.4} , R = {0.4, 0.6, 0.8} and
Add to the front of TabuList ; = {0.4, 0.6, 0.8} , and five instances for each combination. Hence, there
if TabuList > n then are 4 × 3 × 3 × 3 × 5 = 540 test instances in total.
Delete the permutation at the end of TabuList ; The termination criterion is set to n3 milliseconds elapsed CPU time
end if
for all test algorithms. The relative deviation index (RDI) is used for
end if
performance measure, and it is computed as follows:

260
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Fig. 3. Means plot and LSD intervals for and T0 in SA.

Z alg Z best The means plot and least significant difference (LSD) intervals (at
RDI = × 100
Z worst Z best 95% confidence level) for different levels of , and T0 in ILS are
shown in Fig. 2. As observed in Fig. 2(a), there is no statistically sig-
where Z alg is the objective value obtained by a given algorithm, Z best
nificant difference between = 3 and 4, but both choices yield statis-
and Z worst represent the best and worst solutions obtained by all com-
tically better results than = 5. According to Fig. 2(b), no statistical
pared algorithms, respectively. Note that if the best and worst solutions
difference is observed among different levels of and the same con-
take the same value, then RDI will be replaced by 0. Lower values of
clusion can be drawn for T0 by Fig. 2(c). This indicates that the ILS
RDI are always preferred.
algorithm is very robust with respect to in the range [20, 50] and T0 in
the range [0.1, 0.7], respectively. Based on the ANOVA results for the
6.1. Parameter calibration parameter calibration experiment of ILS, the final parameter settings in
ILS are = 3, = 30 and T0 = 0.1.
The performance of algorithms may be greatly influenced by their The means plot and LSD intervals for different levels of and T0 in
parameter values. In this section, we carry out a full factorial design in SA are depicted in Fig. 3. It is shown in Fig. 3(a) that no statistical
the Design of Experiments (DOE) approach (Montgomery, 2017) to significant difference exists between different levels of . This means
calibrate the parameters of the proposed algorithms. We take the that the SA algorithm is rather robust with respect to in the range
parameters as factors, and consider the levels of the parameters as [0.93, 0.99]. However, we can still consider = 0.97 to be slightly better
follows. For the ILS algorithm, is tested at 3 levels {3, 4, 5} , at 4 than the other choices. Fig. 3(b) demonstrates that T0 = 0.3 provides
levels {20, 30, 40, 50} and T0 at 4 levels T0 {0.1, 0.3, 0.5, 0.7} . For the SA statistically better results than the choices of 0.5 and 0.7, but there is no
algorithm, is tested at 4 levels {0.93, 0.95, 0.97, 0.99} and T0 at 4 levels statistically significant difference between 0.1 and 0.3. According to the
{0.1, 0.3, 0.5, 0.7} . For the TS algorithm, T0 is also tested at 4 levels ANOVA results for the parameter calibration experiment of SA, the final
{0.1, 0.3, 0.5, 0.7} . Therefore, ILS has a total of 48 combinations, SA 16 parameter settings used in SA are = 0.97 and T0 = 0.3.
and TS 4. A calibration instance set of 20 instances is generated as Fig. 4 shows the means plot and LSD intervals for different levels of
follows: one instance for each one of the 20 random combinations of n , T0 in TS. As seen in this figure, T0 = 0.1 and 0.3 yield statistically better
, R and , whose values are selected randomly from the aforemen- results than 0.7. Moreover, 0.3 gives a slightly better average perfor-
tioned sets of values. The generation of a calibration instance is the mance than 0.1. Hence, T0 = 0.3 is used in the TS algorithm.
same as that of a test instance. Each algorithm is run five times for each
calibration instance. 6.2. Algorithm comparisons
The obtained results are analysed by means of a multi-factor ana-
lysis of variance (ANOVA) technique, which is a powerful statistical In this section, we provide a comprehensive experimental evalua-
tool and has been extensively used in the scheduling literature in the tion and comparison of the proposed ILS, SA and TS algorithms. Each
past decade to successfully calibrate methods. Its three main hy- algorithm is run five times on each instance of the randomly generated
potheses, i.e., normality, homogeneity of variance and independence of test problem set.
residuals, have been checked and no significant deviations were found We first depict the overall performance of the test algorithms. Based
in the fulfilment of the hypotheses. The detailed ANOVA results and on the computational results, an analysis of variance is performed to
tables on the parameter calibration experiments of the proposed algo- validate the statistical significance of the observed differences in solu-
rithms are omitted here due to space limitations, but they are available tion quality. The analysis has a single factor which is the type of al-
upon request from the authors. Instead, we reproduce the means plot gorithm with 3 levels. The response variable is given by the RDI of each
with confidence intervals of the parameters in each algorithm. instance. The means plot and LSD intervals (at 95% confidence level)

261
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Table 2
ANOVA results for the experiment on the effects of instance factors.
Source Sum of squares Df Mean square F-ratio p-value

Main effects
A: n 19251.9 3 6417.29 12.10 0.0000
B: t 149.1 2 74.53 0.14 0.8690
C: R 5143.5 2 2571.73 4.85 0.0082
D: 368.1 2 184.05 0.35 0.7070

Interactions
AB 1267.8 6 211.30 0.40 0.8803
AC 5443.1 6 907.18 1.71 0.1162
AD 1937.8 6 322.97 0.61 0.7235
BC 1588.6 4 397.16 0.75 0.5592
BD 2096.4 4 524.11 0.99 0.4135
CD 938.8 4 234.69 0.44 0.7780

Residual 322588.8 608 530.57


Total 360773.8 647

Fig. 4. Means plot and LSD intervals for T0 in TS.

Fig. 6. Means plot for the interaction between algorithms and the problem size.

Fig. 5. Means plot and LSD intervals for the test algorithms.
To better examine the performance differences of the test problems
caused by the problem size and investigate its impact, we provide the
for various algorithms are shown in Fig. 5. It is observed from this corresponding means plot in Fig. 6. From this figure, the following
figure that no significant statistical difference exists between the ILS, SA results are observed: (1) The performance of SA and TS has a significant
and TS algorithms, indicating that all the proposed algorithms have decreasing trend when n increases, whereas the performance of ILS first
similar performance in solving the PSSOSPTW. decreases and then increases, indicating that the problem size has a
Next, we compare the performance of the proposed algorithms great effect on the performance of these algorithms when n [20, 50];
based on specific instance factors, which include the problem size, (2) When the problem size is relatively small (n = 20 ), TS gives the best
tardiness factor, due date range factor and time window width range performance; (3) Under medium problem sizes (n = 30, 40 ), SA yields
factor, and analyse the impacts of these instance factors. The ANOVA the best results; (4) When the problem size is relatively large (n = 50 ),
results for the experiment on the influences of instance factors are given ILS outperforms the other algorithms.
in Table 2. In this table, the magnitude of the F-ratio is a clear indicator To examine the interaction between the tardiness factor and the
of significance of the analysed factor on the response variable, and the performance of the algorithms, Fig. 7 gives the means plot for the
p-value smaller than 0.05 implies that the analysed factor has statisti- tardiness factor. Note that as the tardiness factor increases, the pro-
cally significant effects. According to Table 2, the problem size n results blems become more difficult to solve. As depicted in Fig. 7, it is inter-
in the largest F-ratio, meaning that this factor is the most important esting that the performance of all algorithms is not monotonic about the
factor that affects the performance of the proposed algorithms. Besides, tardiness factor. Compared with TS, ILS and SA are better at solving
the due date range factor R is observed to be the second most important difficult (t = 0.4 ) problems, meanwhile, SA shows superior performance
factor. However, the effect of the tardiness factor t and time window when the problems are relatively easy (t = 0.2 ).
width range factor on the test algorithms is not significant, and nei- Fig. 8 provides the means plot for the interaction between the due
ther is the interaction between any two factors among these factors. date range factor and the performance of the algorithms. From this

262
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Fig. 7. Means plot for the interaction between algorithms and the tardiness Fig. 9. Means plot for the interaction between algorithms and the time window
factor. width range factor.

PSSOSPTW problem.

7. Conclusion and future work

In this paper, the PSSOSPTW problem has been addressed for a PSS
provider with a manufacturing plant and a service centre. A mathe-
matical model for this problem is presented to minimize the total sto-
rage and tardiness cost. The studied problem is decomposed into two
subproblems to reduce the complexity due to its NP-hardness. The first
subproblem is to determine the optimal start and completion times of
production and service of PSS orders for a given sequence, and an idle
time insertion algorithm is proposed for this subproblem. The second
subproblem is to find the optimal sequence, and three new metaheur-
istics (ILS, SA and TS) are developed to achieve its goal. In the proposed
algorithms, an NEHPSS heuristic combined with three dispatching rules
is presented to generate initial solutions, and the idea of VNS is in-
corporated into these algorithms to perform efficient local search. For
ILS, a new perturbation procedure is also proposed. The performance of
these algorithms is compared by conducting simulation experiments,
Fig. 8. Means plot for the interaction between algorithms and the due date and the effects of some instance factors are analysed as well. The results
range factor. show that, the overall performance of the proposed algorithms is not
statistically significant different, indicating that they are all effective for
solving the PSSOSPTW problem. However, they can perform quite
figure, it is clear that the due date range factor has noticeable impact on differently for various instance factors.
the performance of these algorithms. Note that the greater the due date Future research can be conducted to the situation that a PSS pro-
range factor, the more scattered the due dates, and the easier the pro- vider has multiple manufacturing plants and service centres, which is
blems. Thus, when R increases from 0.4 to 0.8, the performance of all more realistic but more complex. Another potential research area is to
algorithms increases. consider uncertainties such as uncertain production and service times.
Finally, we check the influence of the time window width range Moreover, it is desirable to develop more time-saving heuristics or
factor. The corresponding means plot is shown in Fig. 9. As this factor metaheuristics for PSSOSPTW.
increases, the width of the time window may also increase, thus the
problems become easier to solve. It is interesting that the trend of each Acknowledgement
algorithm is different from that of each other for this factor. ILS per-
forms the best for relatively difficult problems ( = 0.4, 0.6), but it is This work was supported by the National Natural Science
outperformed by SA and TS for easy problems ( = 0.8). On the con- Foundation of China (Grant No. 71572020) , National Key Research and
trary, TS yields the worst results for difficult problems, but it becomes Development Program (Grant No. 2018YFB1701502), Fundamental
the most effective algorithm for easy problems. Research Funds for the Central Universities (Project No.
The above observations demonstrate that all proposed algorithms 2018CDJSK02PT08) and Start-up Project of High-level Talents
perform differently in various situations. Therefore, appropriate algo- Scientific Research in Chongqing Technology and Business University
rithms should be selected according to the characteristics of the (Project No. 950319002).

263
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Appendix

Proof of Theorem 1. Since PSS orders must be processed first by the manufacturing plant and then by the service centre, and the order processing sequence in
the production and service stages remains the same, the system operated by a PSS provider can be seen as a two-machine permutation flowshop. In the
presented mathematical model for the PSSOSPTW, let the unit storage cost hi = 0 , unit tardiness cost bi = 1 and the earliest allowable service time ei = 0 , the
studied problem can be regarded as the two-machine permutation flowshop scheduling problem with total tardiness minimization. It is well-known that the
latter problem is NP-hard (Pinedo, 2012), hence the PSSOSPTW problem is also NP-hard.
Proof of Proposition. Let the manufacturing plant produce products and service centre provide services as soon as possible. According to the timetabling
algorithm, the earliest service start time of order (n) is
STS (n ) = max{CTS (n 1) , max{e (n ) , CTP (n) }}

thus the following two inequalities hold:


STS (n ) CTS (n 1) = STS (n 1) + st (n 1) (A1)
and
STS (n ) max{e (n ) , CTP (n ) } (A2)
Again by the timetabling algorithm, the earliest service start time of order (1) is STS (1) = e (1) , and the earliest production completion time of
order (n) equals the sum of the production times of all orders, namely, CTP (n) = i N pti . Therefore, we can get from (A1) and (A2) that

STS (n) e (1) + st (j ) =e (1) + sti st (n ) ,


j Q {n} i N (A3)
and

STS (n ) max e (n ) , pti


i N (A4)
The earliest allowable service times of orders (1) and (n) are not less than the minimum value of those of all orders, namely, e (1) mini N ei and
e (n) mini N ei . In addition, the service time of order (n) is not greater than the maximum value of those of all orders, namely, st (n ) max i N sti .
Hence, (A3) and (A4) yield

STS (n ) min ei + sti max sti


i N i N
i N

and

STS (n ) max min ei , pti


i N
i N

Combing the above two equations, Eq. (17) is a lower bound of the service start time of order (n). Since CTS (n) = STS (n) + st (n) and the service
time of order (n) is not less than the minimum value of those of all orders, i.e., st (n) mini N sti , we have from (A3) and (A4) that

CTS (n ) e (1) + st (j) min ei + sti,


i N
j Q i N

and

CTS (n ) max e (n ) , pti + st (n ) max min ei, pti + min sti


i N i N
i N i N

Combing the above two equations, it is concluded that Eq. (18) is a lower bound of the service completion time of order (n ) .
Proof of Theorem. We proceed by contradiction. Assuming that there is idle time between orders (j) and (j + 1) in the production stage, namely,
STP (j + 1) > CTP (j) , we consider the following three cases:
Case 1: The production completion time of order (j) is earlier than its service start time, i.e., CTP (j) < STS (j) , and the production start time of
order (j + 1) is not later than the service start time of order (j) , i.e., STP (j + 1) STS (j) . In this case, the storage and total cost can be reduced by
delaying (STP (j + 1) CTP (j) ) time units the production start time of order (j) to eliminate idle time. Hence, the original schedule is not optimal in
Case 1.
Case 2: The production completion time of order (j) is also its service start time, namely, CTP (j) = STS (j) . In this case, from pt (j + 1) > st (j) and
STP (j + 1) > CTP (j) , we get
CTP (j + 1) = STP (j + 1) + pt (j + 1) > STS (j ) + st (j ) = CTS (j )

which indicates that the production completion time of order (j + 1) is later than the service start time of order (j) . Combing
d (j + 1) e (j) pt (j + 1) + st (j + 1) , STS (j) e (j) and CTP (j) = STS (j) , we have
e (j + 1) d (j + 1) st (j + 1) e (j ) + pt (j + 1) STS (j ) + pt (j + 1)
= CTP (j) + pt (j + 1) < STP (j + 1) + pt (j + 1) = CTP (j + 1)

which means that the production completion time of order (j + 1) is later than its earliest allowable service time. Therefore, it derives that

264
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

CTS (j + 1) = CTP (j + 1) + st (j + 1) >e (j ) + pt (j + 1) + st (j + 1) d (j + 1)

which denotes that order (j + 1) is tardy. To reduce the tardiness and total cost, we can advance the production start time of order (j + 1) to the
production completion time of order (j) , so that idle time can be eliminated. Thus, the original schedule is not optimal in Case 2.
Case 3: The production completion time of order (j) is earlier than its service start time, namely, CTP (j) < STS (j) , and the production start time
of order (j + 1) is later than the service start time of order (j) , namely, STP (j + 1) > STS (j) . In this case, the tardiness and total cost can be reduced
by implementing the following operations. Firstly, we can reduce the storage and total cost by delaying (STS (j) CTP (j) ) time units the production
start time of order (j) to reduce idle time, resulting in that the production completion time of order (j) equals its service start time, namely,
CTP (j) = STS (j) . Secondly, similar to Case 2, the tardiness and total cost can be further reduced by advancing the production start time of order
(j + 1) to the production completion time of order (j) to eliminate idle time. Therefore, the original schedule is not optimal in Case 3.
Proof of Corollary. From CTS (j ) = STS (j + 1) and pt (j + 1) st (j ) , we get
STS (j ) = CTS (j ) st (j ) = STS (j + 1) st (j ) STS (j + 1) pt (j + 1)

Subtracting both sides of the above inequality by CTP (j ) and applying the timetabling algorithm, we have
STS (j ) CTP (j ) STS (j + 1) CTP (j ) pt (j + 1) = STS (j + 1) CTP (j + 1)

namely, INV (j) INV (j + 1) .


By CTP (j) = STS (j) and the timetabling algorithm, we can get
CTP (j + 1) = CTP (j ) + pt (j + 1) = STS (j ) + pt (j + 1) (A5)
Combing st (j ) d (j ) e (j ) and d (j + 1) e (j ) pt (j + 1) + st (j + 1) , we have
max{st (j + 1) +e (j + 1) e (j ) , st (j ) +d (j + 1) d (j ) } d (j + 1) e (j ) pt (j + 1) + st (j + 1)

which yields that


e (j + 1) e (j ) + pt (j + 1) (A6)
and
st (j ) d (j ) pt (j + 1) + st (j + 1) d (j + 1) (A7)
Since STS (j ) e (j ) , we have from (A5) and (A6) that
CTP (j + 1) = STS (j ) + pt (j + 1) e (j ) + pt (j + 1) e (j + 1)

indicating that the completion time of order (j + 1) is not earlier than its earliest allowable service time. Then, from pt (j + 1) st (j ) and (A5), we
have
CTP (j + 1) STS (j ) + pt (j + 1) STS (j ) + st (j ) = CTS (j )

which indicates that the production completion time of order (j + 1) is not earlier than the service completion time of order (j) . Therefore, the
service start time of order (j + 1) is its production completion time, namely, STS (j + 1) = CTP (j + 1) . Combing (A5) and (A7), we can get
CTS (j + 1) d (j + 1) = STS (j + 1) + st (j + 1) d (j + 1)
= CTP (j + 1) + st (j + 1) d (j + 1)
= STS (j ) + pt (j + 1) + st (j + 1) d (j + 1)
STS (j ) + st (j ) d (j )
= CTS (j ) d (j )

namely, TAR (j + 1) TAR (j ) .

References and Operations Management, 18(6), 672–692.


Cheng, T. C. E., Chung, Y.-H., Liao, S.-C., & Lee, W.-C. (2013). Two-agent singe-machine
scheduling with release times to minimize the total weighted completion time.
Baines, T. S., Lightfoot, H. W., Benedettini, O., & Kay, J. M. (2009). The servitization of Computers & Operations Research, 40(1), 353–361.
manufacturing: A review of literature and reflection on future challenges. Journal of Chung, S. H., Chan, F. T. S., & Ip, W. H. (2011). Minimization of order tardiness through
Manufacturing Technology Management, 20(5), 547–567. collaboration strategy in multifactory production system. IEEE Systems Journal, 5(1),
Baines, T. S., Lightfoot, H. W., Evans, S., Neely, A., Greenough, R., Peppard, J., Roy, R., ... 40–49.
Wilson, H. (2007). State-of-the-art in product-service systems. Proceedings of the Cordeau, J.-F., & Maischberger, M. (2012). A parallel iterated tabu search heuristic for
Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 221(10), vehicle routing problems. Computers & Operations Research, 39(9), 2033–2050.
1543–1552. Framinan, J. M., & Perez-Gonzalez, P. (2018). Order scheduling with tardiness objective:
Behnamian, J., Ghomi, S. M. T. F., & Zandieh, M. (2010). Development of a hybrid me- Improved approximate solutions. European Journal of Operational Research, 266(3),
taheuristic to minimise earliness and tardiness in a hybrid flowshop with sequence- 840–850.
dependent setup times. International Journal of Production Research, 48(5), Frazzon, E. M., Albrecht, A., Pires, M., Israel, E., Kück, M., & Freitag, M. (2018). Hybrid
1415–1438. approach for the integrated scheduling of production and transport processes along
Beuren, F. H., Ferreira, M. G. G., & Miguel, P. A. C. (2013). Product-service systems: A supply chains. International Journal of Production Research, 56(5), 2019–2035.
literature review on integrated products and services. Journal of Cleaner Production, Gao, J., Chen, R., & Deng, W. (2013). An efficient tabu search algorithm for the dis-
47, 222–231. tributed permutation flowshop scheduling problem. International Journal of
Boehm, M., & Thomas, O. (2013). Looking beyond the rim of one's teacup: A multi- Production Research, 51(3), 641–651. https://doi.org/10.1080/00207543.2011.
disciplinary literature review of Product-Service Systems in Information Systems, 644819.
Business Management, and Engineering & Design. Journal of Cleaner Production, 51, Geng, X., Chen, Z., Yang, W., Shi, D., & Zhao, K. (2011). Solving the traveling salesman
245–260. problem based on an adaptive simulated annealing algorithm with greedy search.
Chen, Z. L. (2010). Integrated production and outbound distribution scheduling: Review Applied Soft Computing, 11(4), 3680–3689.
and extensions. Operations Research, 58(1), 130–148. Geng, X., Chu, X., Xue, D., & Zhang, Z. (2010). An integrated approach for rating en-
Chen, Z. L., & Pundoor, G. (2009). Integrated order scheduling and packing. Production gineering characteristics’ final importance in product-service system development.

265
Y. Zhang, et al. Computers & Industrial Engineering 133 (2019) 253–266

Computers & Industrial Engineering, 59(4), 585–594. Nawaz, M., Jr, E. E. E., & Ham, I. (1983). A heuristic algorithm for the m-machine, n-job
Glover, F. (1986). Future paths for integer programming and links to artificial in- flow-shop sequencing problem. Omega, 11 (1), 91–5.
telligence. Computers & Operations Research, 13(5), 533–549. Osman, I. H., & Potts, C. N. (1989). Simulated annealing for permutation flow-shop
Hendel, Y., & Sourd, F. (2007). An improved earliness–tardiness timing algorithm. scheduling. Omega, 17(6), 551–557.
Computers & Operations Research, 34(10), 2931–2938. Pan, Q. K., & Ruiz, R. (2012). Local search methods for the flowshop scheduling problem
Karabulut, K. (2016). A hybrid iterated greedy algorithm for total tardiness minimization with flowtime minimization. European Journal of Operational Research, 222(1), 31–43.
in permutation flowshops. Computers & Industrial Engineering, 98, 300–307. Pan, Q. K., Ruiz, R., & Alfaro-Fernández, P. (2017). Iterated search methods for earliness
Karimi, N., & Davoudpour, H. (2015). A branch and bound method for solving multi- and tardiness minimization in hybrid flowshops with due windows. Computers &
factory supply chain scheduling with batch delivery. Expert Systems with Applications, Operations Research, 80, 50–60.
42(1), 238–245. Pascual, R., Siña, M., Santelices, G., Román, M., & Droguett, E. L. (2017). Optimal channel
Kedad-Sidhoum, S., & Sourd, F. (2010). Fast neighborhood search for the single machine coordination in use-based product-service system contracts. International Journal of
earliness-tardiness scheduling problem. Computers & Operations Research, 37(8), Production Research, 5, 1–11.
1464–1471. Pinedo, M. L. (2012). Scheduling: Theory, algorithms, and systems. Springer Science &
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by Simulated Business Media.
Annealing. Science, 220(4598), 671. Potts, C. N., & Van Wassenhove, L. N. (1982). A Decomposition algorithm for the single
Kuo, T. C., & Wang, M. L. (2012). The optimisation of maintenance service levels to machine total tardiness problem. Operations Research Letters, 1(5), 177–181.
support the product service system. International Journal of Production Research, Safaei, N., Banjevic, D., & Jardine, A. K. (2012). Multi-threaded simulated annealing for a
50(23), 6691–6708. bi-objective maintenance scheduling problem. International Journal of Production
Lee, S., Yoo, S., & Kim, D. (2016). When is servitization a profitable competitive strategy? Research, 50(1), 63–80.
International Journal of Production Economics, 173, 43–53. Sawhney, M., Balasubramanian, S., & Krishnan, V. V. (2004). Creating growth with ser-
Li, K., Ganesan, V., & Sivakumar, A. (2005). Synchronized scheduling of assembly and vices. Mit Sloan Management Review, 45(2), 34.
multi-destination air-transportation in a consumer electronics supply chain. Shokohyar, S., Mansour, S., & Karimi, B. (2014). A model for integrating services and
International Journal of Production Research, 43(13), 2671–2685. product EOL management in sustainable product service system (S-PSS). Journal of
Li, N., & Jiang, Z. (2013). Modeling and optimization of a product-service system with Intelligent Manufacturing, 25(3), 427–440.
additional service capacity and impatient customers. Computers & Operations Subramanian, A., Drummond, L. M.d. A., Bentes, C., Ochi, L. S., & Farias, R. (2010). A
Research, 40(8), 1923–1937. parallel heuristic for the vehicle routing problem with simultaneous pickup and de-
Lin, W. C., Yin, Y., Cheng, S. R., Cheng, T. C. E., Wu, C. H., & Wu, C. C. (2017). Particle livery. Computers & Operations Research, 37(11), 1899–1911.
swarm optimization and opposite-based particle swarm optimization for two-agent Tseng, C. T., & Liao, C. J. (2008). A discrete particle swarm optimization for lot-streaming
multi-facility customer order scheduling with ready times. Applied Soft Computing, 52, flowshop scheduling problem. European Journal of Operational Research, 191(2),
877–884. 360–373.
Lindahl, M., Sundin, E., & Sakao, T. (2014). Environmental and economic benefits of Tukker, A. (2015). Product services for a resource-efficient and circular economy – A
Integrated Product Service Offerings quantified with real business cases. Journal of review. Journal of Cleaner Production, 97, 76–91.
Cleaner Production, 64(2), 288–296. Wang, G., & Cheng, T. C. E. (2007). Customer order scheduling to minimize total
Liu, X., Yang, T., Pei, J., Liao, H., & Pohl, E. A. (2019). Replacement and inventory control weighted completion time. Omega, 35(5), 623–626.
for a multi-customer product service system with decreasing replacement costs. Wang, K., Jiang, Z., Li, N., & Geng, N. (2013). Optimal production and admission control
European Journal of Operational Research, 273(2), 561–574. for a stochastic SOM system with demands for product and PSS. International Journal
Lockett, H., Johnson, M., Evans, S., & Bastl, M. (2011). Product Service Systems and of Production Research, 51(23–24), 7270–7288.
supply network relationships: An exploratory case study. Journal of Manufacturing Wang, K., Ma, W., Luo, H., & Qin, H. (2016). Coordinated scheduling of production and
Technology Management, 22(3), 293–313. transportation in a two-stage assembly flowshop. International Journal of Production
Lourenço, H. R., Martin, O. C., & Stützle, T. (2010). Iterated local search: Framework and Research, 54(22), 6891–6911.
applications. Handbook of Metaheuristics, 146, 363–397. Wu, C. C., Lin, W. C., Zhang, X. G., Chung, I. H., Yang, T. H., & Lai, K. J. (2019). Tardiness
Ma, H. L., Chan, F. T. S., & Chung, S. H. (2013). Minimising earliness and tardiness by minimisation for a customer order scheduling problem with sum-of-processing-time-
integrating production scheduling with shipping information. International Journal of based learning effect. Journal of the Operational Research Society, 70(3), 487–501.
Production Research, 51(8), 2253–2267. Wu, C. C., Liu, S. C., Lin, T. Y., Yang, T. H., Chung, I. H., & Lin, W. C. (2018). Bicriterion
Mahut, F., Daaboul, J., Bricogne, M., & Eynard, B. (2017). Product-service systems for total flowtime and maximum tardiness minimization for an order scheduling pro-
servitization of the automotive industry: A literature review. International Journal of blem. Computers & Industrial Engineering, 117, 152–163.
Production Research, 55(7), 2102–2120. Wu, C. C., Liu, S. C., Zhao, C., Wang, S. Z., & Lin, W. C. (2017). A multi-machine order
Marandi, F., & Fatemi Ghomi, S. M. T. (2019). Integrated multi-factory production and scheduling with learning using the genetic algorithm and particle swarm optimiza-
distribution scheduling applying vehicle routing approach. International Journal of tion. The Computer Journal, 61(1), 14–31.
Production Research, 57(3), 722–748. Xie, W., Jiang, Z., Zhao, Y., & Shao, X. (2014). Contract design for cooperative product
Massengale, A., Larson, M. C., Dai, C., & Jr, J. S. H. (2014). Integrated scheduling of service system with information asymmetry. International Journal of Production
production and rail transportation. Computers & Industrial Engineering, 74(1), Research, 52(6), 1658–1680.
240–256. Xu, J., Wu, C. C., Yin, Y., Zhao, C., Chiou, Y. T., & Lin, W. C. (2016). An order scheduling
Mladenović, N., & Hansen, P. (1997). Variable neighborhood search. Computers & problem with position-based learning effect. Computers & Operations Research, 74,
Operations Research, 24(11), 1097–1100. 175–186.
Montgomery, D. C. (2017). Design and analysis of experiments. John Wiley & Sons. Zhang, Y., Liu, S., Liu, Y., & Li, R. (2016). Smart box-enabled product–service system for
Naderi, B., Ruiz, R., & Zandieh, M. (2010). Algorithms for a realistic variant of flowshop cloud logistics. International Journal of Production Research, 54(22), 6693–6706.
scheduling. Computers & Operations Research, 37(2), 236–246.

266

You might also like