Complex & Intelligent Systems (2021) 7:191–202

https://doi.org/10.1007/s40747-020-00184-x

ORIGINAL ARTICLE

Multitasking scheduling with batch distribution and due date

assignment
Xinrui Xu1 · Guangqiang Yin2 · Chunyu Wang2

Received: 2 June 2020 / Accepted: 25 July 2020 / Published online: 31 August 2020
© The Author(s) 2020

Abstract
This study addresses the multitasking scheduling problems with batch distribution and due date assignment (DDA). Compared
with classical scheduling problems with due date-related optimization functions, the job due dates are decision variables rather
than given parameters. The jobs completed are distributed in batches, and the sizes of all batches are identical, which may be
bounded or unbounded. The jobs in every batch are scheduled one by one. Each batch incurs a fixed cost. Under multitasking
environment, it allows the machine to put an uncompleted job on hold and turn to another uncompleted job. The goal is
to identify the optimal primary job sequence, the optimal job due dates, and the optimal batch production and distribution
strategy so that one of the following two optimization functions is minimised: the total cost composed of the earliness penalty,
DDA cost, tardiness penalty and batch distribution cost, and the total cost composed of the earliness penalty, weighted number
of late jobs, DDA cost and batch distribution cost. We devise efficient exact algorithms for the problems we consider, and
perform numerical experiments to check how multitasking affects the scheduling cost or value, the results of which can assist
decision-makers to justify the extent to put to use or refrain from multitasking.

Keywords Multitasking · Scheduling · Due date assignment · Batch distribution

Introduction tasking, and verify the effects of multitasking on the schedule

criteria through computational experiments. Subsequently,
The problem we investigate in this study covers three momen- the research on this line has attracted increasing attention.
tous sub-areas of scheduling research, i.e., multitasking Hall et al. [13] introduce two different multitasking schedul-
scheduling, scheduling with DDA, and batch distribution ing problems, in which the first one addresses alternate period
scheduling. All of these three sub-areas have been widely processing and the second one investigates shared process-
investigated in the literature. In the remaining part of this ing. Ji et al. [15] consider the identical parallel-machine
section, we briefly review some related research from these scheduling problem with slack due-window assignment
sub-areas. (DWA) under multitasking, in which the slack due windows
The research about multitasking scheduling is initialized are machine dependent. Hall et al. [19] consider the multi-
by Hall et al. [12], where the processing of a chosen job tasking scheduling problem with common DDA such that all
can be suspended by other uncompleted jobs. The authors the jobs have a common due date. Xiong et al. [24] investigate
demonstrate that the solution algorithms for some classi- the unrelated parallel-machine scheduling problem under
cal scheduling criteria are more complex in terms of time multitasking, and present an exact mathematical-based pro-
complexity than the corresponding problems without multi- gramming for the problem with the goal of minimising the
total completion time. Li et al. [18] and Wang et al. [23]
B Chunyu Wang address multitasking scheduling with two competing agents,
[email protected]
and study the computational complexity and present exact
1 School of Computer Science and Engineering, University of algorithms for the problems, in which every agent requires to
Electronic Science and Technology of China, Chengdu process its respective jobs and wishes to minimise its respec-
610054, China tive optimization function that is related to its respective jobs.
2 School of Information and Software Engineering, University As for due window assignment scheduling, Zhu et al. [34]
of Electronic Science and Technology of China, Chengdu address the multitasking scheduling problem with DDA and
610054, China

123
192 Complex & Intelligent Systems (2021) 7:191–202

a rate-modifying activity (RMC), and Zhu et al. [33] fur- What is noteworthy is that, as far as our information goes,
ther generalize the model to the case with multiple RMCs, in there is no study addressing the batch delivery schedule in
which the RMC may alter the process speed of the machine. the field of multitasking setting. This study tries to fill up
In the aforementioned studies, Ji et al. [15], Liu et al. [19] this gap, the contributions of which to the literature can be
and Zhu et al. [34] address the due date-related optimiza- summarized as follows.
tion functions, in which the jobs’ due dates or time windows
require to be decided by the decision-makers in the decision- • We introduce a new scheduling model which simul-
making process along with the primary job sequence. This taneously considers batch delivery, multitasking and
kind of scheduling problems is referred to scheduling with DDA with the goal to identify the optimal primary job
DDA or due window assignment in the literature, which has sequence, the optimal job due dates, and the optimal batch
been extensively studied in the area of just-in-time schedul- production and delivery strategy such that two objective
ing and has been receiving widespread attention during the functions considered are minimised.
past four decades. The literature on this topic is abound, the • We demonstrate that the problem with the first optimiza-
interested readers are referred to the review papers by Gordon tion function may be solved in polynomial time and that
et al. [9,10], and Kaminsky and Hochbaum [11]. Neverthe- the problem with the second optimization function is
less, in contrast to our study, all the aforementioned studies pseudo-polynomial time solvable.
as well as the review papers concentrate only on the problem • We perform computational experiments to assess how
of how to process the jobs while neglecting the problem of multitasking affects the scheduling cost or value, the
scheduling job delivery. results of which can assist decision-makers to justify the
Recently, some researchers pay attention to the study extent to put to use or refrain from multitasking.
of DDA (due window assignment) and batch distribution
scheduling. Chen [4] address the common DDA and batch We organize the study as follows. The next section for-
distribution scheduling, and devise an algorithm with polyno- mally describes the studied problem, followed by which the
mial running time to minimise the weighted sum of tardiness preliminary analysis and several properties about the optimal
penalty, earliness penalty, batch distribution cost and DDA schedule for the studied problems are presented. The subse-
cost. Shabtay [22] considers the unrestricted DDA and batch quent section focuses on developing the solution procedures
distribution scheduling with acceptable lead-times, in which for solving the problems. The computational experiments
the jobs can be assigned different due dates, and prove that to verify the effects of multitasking on the objective func-
the considered problem is N P-hard and demonstrate that tion values are shown before the concluding section. The last
several special cases may be solved in polynomial time. Yin section provides the conclusions and future studies.
et al. [29] consider the common DDA and batch distribu-
tion scheduling with a RMC to minimise the weighted sum
of earliness penalty, holding cost, tardiness penalty, batch Problem description
distribution cost and DDA cost, and demonstrate that sev-
eral special cases are polynomial time solvable. Yin et al. This section formally describes the studied problems. Table 1
[26,28] focus on the common DDA and batch distribution summarizes the main notation used throughout this study.
scheduling, where we can reduce the job processing times To be precise, a set of n jobs N = {J1 , . . . , Jn } needs
by assigning a certain amount of resources to process the to be scheduled on a machine, which may schedule at most
jobs. Mor and Mosheiov [20] and Yin et al. [30,32] study the one job at a time. Every job J j , j = 1, . . . , n, has a non-
DDA and batch distribution scheduling with two competing negative processing time p j and a non-negative due date d j .
agents. Yin et al. [27] consider the common DWA and batch Different from the classical assumption that the due date d j
distribution scheduling problem such that the due windows is pre-defined, it is a decision variable in this study.
of all jobs are identical, which specify the earliest and latest As in Hall et al. [12] and Wang et al. [23], we allow
delivery date, and show that the problem with the goal to uncompleted jobs to suspend the jobs being processing, i.e., it
minimise the weighted sum of earliness cost, window loca- permits the machine to suspend an uncompleted job and turn
tion cost, window size cost, holding cost, tardiness cost, and to another uncompleted job. We regard the job scheduled in
batch distribution cost is polynomial time solvable. For more anytime as the primary job, and the other uncompleted jobs
results on DDA and batch distribution scheduling, the inter- as the waiting jobs, which may interrupt the processing of
ested readers are recommended to read the recent papers by the primary job. We assume that each job may be handled
Ahmadizar and Farhadi [1], Agnetis et al. [2,3], Gong et al. as a primary job only once and immediately after another
[8], Kovalyov et al. [16], Li et al. [17], Xu et al. [25] and Yin primary job completing its processing.
et al. [31], and the review papers by Chen [5] and Hall and In the execution of a primary job, two types of times, say
Potts [14]. interruption time and switching time, are incurred, in which

123
Complex & Intelligent Systems (2021) 7:191–202 193

Table 1 Notation used in this study time of job Jk . The switching time for reviewing the waiting
Notation Meaning jobs of job J j is defined as ϕ(|W j |), which only depends on
the number of waiting jobs. As a consequence, it is evident
n The number of jobs that the remaining processing time of job Jk after interrupting
J = {J1 , J2 , . . . , Jn } The set of jobs r , r = 1, . . . , n − 1, primary jobs equals
pj The processing time of job
Jj f k (r − 1) − α f k (r − 1) = (1 − α)r pk . (1)
dj The due date of job J j
rk
j
The remaining processing Thus, the time that job Jk suspends the (r + 1)th primary job
time of job Jk equals
immediately when job J j
becomes a primary job
α f k (r ) = α(1 − α)r pk . (2)
f k (r ) The remaining processing
time of job Jk after
interrupting We distribute the completed jobs to the customers in
r , r = 1, . . . , n − 1, batches, and process the jobs in a batch sequentially. We
primary jobs refer to the sum of the processing times of all the jobs in a
Wj The set of jobs that are batch to the processing time of the batch. The sizes of all
waiting for processing batches are identical, which may be bounded or unbounded.
when job J j becomes a
primary job By bounded batch size, we mean that there may be at most
ϕ(x) A nondecreasing function
b jobs in each batch. In this study, we mainly concentrate on
on variable x ∈ [0, +∞) the case with bound batch size, and show how to extend the
with ϕ(0) = 0 results to the unbounded case. We distribute each batch to the
Dj The distribution time of job customers immediately after the last job in it completing its
J j , which equals the processing, and each distribution incurs a fixed cost ϑ. We
completion time of the
call the corresponding time as the delivery time of the jobs
last job in the batch that
contains job J j in the batch.
E j = max{d j − D j , 0} The earliness of job J j If we distribute a job to its customer before its due date, an
T j = max{D j − d j , 0} The lateness of job J j earliness penalty is incurred which depends on how early it
Uj The lateness indicator, is (E j ). In addition, if we distribute a job to its customer after
where U j = 1 if D j > d j its due date, a lateness penalty is incurred, which depends on
and U j = 0 otherwise how late it is (T j ) or whether it is late (U j =0 or 1). A job is
α ∈ (0, 1) The interruption rate of all referred to be early if U j = 0, and late otherwise.
the jobs The goal is to find (1) the optimal primary job sequence,
b The batch size which (2) the optimal due dates, (3) the optimal number of
indicates that any batch
batches m, and (4) the optimal batch distribution partition
can contains at most b
jobs (B1 , B2 , . . . , Bm ) such that the optimization function
ϑ The distribution cost per   
batch 1 = μE j + ηT j + γ d j + mϑ
μ The unit earliness cost J j ∈N
γ The unit lateness cost
or
η The unit DDA cost
  
ωj The cost of job J j being late 2 = μE j + ω j U j + γ d j + mϑ

n
P= pj The total processing time of J j ∈N
j=1 all jobs
is minimised. In the remaining part of this study, we refer to
the problems of minimizing the objective functions 1 and
2 as problems P1 and P2, respectively. In addition, as for
the former represents the time during which the primary job’s problem P2, we assume that the jobs which would be late
waiting jobs interrupt its processing, and the latter measures are not processed.
the time spent in the inspecting its waiting jobs. Specially, The following practical example concerning a two-level
given any primary job J j and each waiting job Jk ∈ W j , the supply chain, which involves a steel manufacturer and a set of
time that job Jk suspends the processing of job J j is defined orders from several customers, motivates the studied prob-
j
as αrk , which is proportional to the remaining processing lems, where each order consists of producing numbers of

123
194 Complex & Intelligent Systems (2021) 7:191–202

various medium to small steel coils. In this example, the man- d j > D j , the change in the optimization function value
ufacturer is referred to as a single “machine” and the orders of 1 is −(γ + μ)(d j − D j ) < 0.
are referred to as “jobs”. A sufficient number of transport (2) γ ≤ η. Assume there exists a solution S such that there
vehicles are available to distribute the completed orders to the is a job J j with d j > 0. In what follows, we prove that
customers, where the transport vehicles have fixed capacity, a shift of d j to the left such that d j = 0 can only reduce
and the cost per distribution is fixed. The processing require- the optimization function value of 1 . When d j ≥ D j ,
ment dictates that the orders containing in the same batch are the change in the optimization function value of 1 is
processed contiguously and the distribution date of a batch ηD j − (γ d j + μ(d j − D j )) ≤ 0. When d j < D j , the
equals the completion time of the last order in the batch. To change in the optimization function value of 1 is ηD j −
reduce the setup times that are needed to perform some clean- (γ d j + η(D j − d j )) = (η − γ )d j ≤ 0. 

ing operations, or remove a previous container and install a
new one when the manufacturer switches processing from The proposition above demonstrates that an optimal solu-
one type of steel coil to another type of steel coil, multitask- tion to each of the problems exists where the earliness of each
ing is permitted, which allows unfinished orders to seize the job equals zero. It follows that the optimization functions 1
production resources and suspend the orders under process- and 2 reduce to
ing. The steel manufacturer will negotiate with customers
to set the due dates for completing their orders. To reduce 
ξ D j + mϑ (3)
the operating cost and improve the overall satisfaction of the
J j ∈N
customers, the steel manufacturer requires to determine an
effective way to allocate its services over time to perform the
and
orders of the customers in a timely and cost-effective manner.
This situation can be modeled as the studied models.   
ω j U j + γ D j + mϑ, (4)
J j ∈N

Structure property analysis

respectively, in which ξ = min{γ , η}. As a consequence, it
This section provides some structure properties on the opti- is beneficial to process the primary jobs consecutively from
mal solution for problems P1 and P2. time 0 (exclude the interruption and switching times).
Given a job sequence, we let J[ j] be the jth primary job in In what follows, we limit our attention to solutions
the sequence, and let Bk and |Bk | represent the set of primary fulfilling the properties above. Now we analyze the formula-
jobs scheduled in the kth batch and the number of primary tion of the optimization function 1 of a given schedule.
jobs in Bk (k = 1, 2, · · · ), respectively. The following lemma Given any such primary job sequence S, which is par-
illustrates the property about assignment due dates. titioned into m batches B1 = {J[1] , J[2] , . . . , J[h 1 ] }, . . .,
Bm = {J[h m−1 +1] , J[h m−1 +2] , . . . , J[n] }, in which h k is the
Proposition 3.1 There is an optimal solution to the problem number of primary jobs contained in the first k batches for
so that k = 0, 1, 2, . . . , m, with h 0 = 0 and h m = n. Thus, for each
j = 1, . . . , n, job J[ j] ’s completion time is equal to
(i) d j = D j if γ < η and d j = 0 otherwise for all j =
1, 2, . . . , n when objective function is 1 ;   

j n 
j
(ii) d j = D j if γ D j < ω j and d j = 0 otherwise for all C[ j] = p[l] + 1 − (1 − α) j p[l] + ϕ(n − l), (5)
j = 1, 2, . . . , n when the objective function is 2 . l=1 l= j+1 l=1

Proof It is similar to the proof of Yin et al. [30]. For com- in which the first term is the sum of the total processing times
pleteness, we provide a brief proof for (i). The proof for (ii) of the first j primary jobs, the second term denotes the sum
is similarly. To be precise, we require to address two cases. of the interruption times that the last n − j jobs interrupt the
first j primary jobs, and the third term gives the sum of the
(1) γ < η. Assume there exists a solution S such that there switching times during processing the first j primary jobs.
is a job J j with d j = D j , i.e., d j < D j or d j > D j . In By Eq. (5), the distribution time D[ j] of job J[ j] such that
what follows, we prove that a shift of d j to the right (resp., h k−1 + 1 ≤ j ≤ h k for some k = 0, 1, . . . , m is
left) if d j < D j (resp., d j > D j ) such that d j = D j can
only reduce the optimization function value of 1 . When 
hk   
n 
hk
d j < D j , the change in the optimization function value D[ j] = p[l] + 1 − (1 − α)h k p[l] + ϕ(n − l), (6)
of 1 is (γ − η)(D j − d j ) < 0. On the other hand, when l=1 l=h k +1 l=1

123
Complex & Intelligent Systems (2021) 7:191–202 195

and thus the optimization function 1 of schedule S can be According to Eq. (7), the difference between the optimization
formulated as function values of solutions S and S  is
 k−1 
1 = ξ D j + mϑ  
J j ∈N pi (n − h k−1 ) + 1 − (1 − α)hl |Bl |

m 
hk l=1
=ξ D[ j] + mϑ k 
 
k=1 j=h k−1 +1 + p j (n − h k ) + 1 − (1 − α)hl |Bl |

m 
hk 
hk l=1
=ξ p[l] + (1 − (1 − α) ) hk  k−1 
 
k=1 j=h k−1 +1 l=1 − p j (n − h k−1 ) + 1 − (1 − α)hl |Bl |
l=1

n 
hk
k 
p[l] + ϕ(n − l) + mϑ  
l=h k +1 l=1
+ pi (n − h k ) + 1 − (1 − α)hl |Bl |

m  
k−1 
l=1
  
=ξ (n − h k−1 )Pk + Pk (1 − (1 − α)hl )|Bl | = ( pi − p j ) h k − h k−1 − 1 − (1 − α)h k |Bk |
k=1 l=1
= ( pi − p j )(1 − α)h k |Bk | > 0.

hk
+ (n − h k−1 ) ϕ(n − j) + mϑ, (7)
As a consequence, solution S  is no worse than S, as required.
j=h k−1 +1


h k
in which Pk = j=h k−1 +1 p[ j] represents the processing The proposition above demonstrates that both problems
time of batch Bk . P1 and P2 allow an optimal solution so that the primary
As for the optimization function 2 , given any primary jobs are proceeded in terms of the SPT (Shortest Processing
job sequence S with n e early jobs, in which the early jobs are Time first) rule. In what follows, we re-arrange the jobs in
partitioned into m e batches B1 = {J[1] , J[2] , . . . , J[h 1 ] }, . . ., terms of the SPT rule so that d1 ≤ d2 ≤ · · · ≤ dn .
Bm e = {J[h m e −1 +1] , J[h m e −1 +2] , . . . , J[n e ] }, then the opti-
mization function 2 of schedule S can be formulated as
 Dynamic programming algorithms
2 = (γ D j + ω j U j ) + mϑ
J j ∈N
me 
 k−1 
  This section devises dynamic programming (DP) algorithms
=γ (n − h k−1 )Pk + Pk 1 − (1 − α)hl |Bl | for problems P1 and P2 which relies on the properties given
k=1 l=1 in Lemmas 3.1 and 3.2. Specially, we prove that problem P1

hk  is polynomial time solvable, whereas problem P2 may be
+ (n − h k−1 ) ϕ(n − j) + w j + mϑ solved in pseudo-polynomial time.
J j ∈L
j=h k−1 +1
Let us begin with devising a polynomial-time forward DP
(8)
algorithm, denoted as Algorithm M D BT , for problem P1.
The procedure of the algorithm is as follows. For each j =
in which L stands for the set of late jobs in solution S.
0, 1, . . . , n, we let H j stand for the set of states encoding the
Based on Eqs. (7) and (8), one can draw the following
partial solutions for the first j jobs J1 , . . . , J j . Each state in
conclusion.
H j is in the form of (a, g) such that
Proposition 3.2 There is an optimal solution to each of the
problems P1 and P2 so that the maximum processing time • There is exactly a primary jobs contained in the last batch
among the primary jobs scheduled in any batch Bk , k = of the partial solution.
1, 2, . . ., is equal to or less than the minimum processing • The optimization function value of the partial solution is
time among the primary jobs scheduled in batch Bk+1 . g.

Proof We provide the proof for problem P1. The proof for Algorithm M D BT begins with an initial tag set F0 =
problem P2 is analogous. Assume there exists a solution S {(0, 0)}. For each j = 1, . . . , n, the tag set F j is con-
with two continuous batches Bk and Bk+1 such that pi > p j structed from tag set F j−1 . To be precise, for each tag
for some primary jobs Ji ∈ Bk and J j ∈ Bk+1 . We create (a, g) ∈ F j−1 that corresponds to a partial solution for the
a new solution S  from S by exchanging primary jobs Ji jobs J1 , . . . , J j−1 , to create a new solution by appending
and J j while keeping the other primary jobs being the same. the next unassigned job J j to the end of the above partial

123
196 Complex & Intelligent Systems (2021) 7:191–202

solution, one requires to take the following two cases into batch size. By the analysis in Case 1, the value that increases
account. in the optimization function 1 due  to this operation in this

j a
Case T1: Start a new batch and append job J j to it. What case is zero if j < n, and ξ(a+1) l= j−a pl + l=0 ϕ(l)
is noteworthy is that we take into account the contributions otherwise. Thus, if a < b, one can construct a new tag (a +
of the jobs to the optimization function value in batches, i.e., 1, g) and add it into g = g if j < n and
the contributions of all jobs contained in a batch are added  F j , in whicha
j
g = g + ξ(a + 1) l= j−a pl + l=0 ϕ(l) otherwise.
to the optimization function value of 1 immediately when
we complete the processing of all the jobs contained in the What is noteworthy is that not all the generated tags during
batch. Thus, the contributions of all the jobs contained in the the procedure above will form a complete better schedule.
previous batch to the optimization function value of 1 in Thus, it is necessary to find the tags that can be discarded for
this case equals further consideration by the following dominant rule.

Lemma 4.1 Given any tags (a, g) and (a  , g  ) in F j such that

 
j−1
ξ (n − ( j − 1 − a)) pl + a a ≤ a  , and g ≤ g  , one can discard the latter tag.
l= j−a
Proof Given the conditions stated in the lemma, it is evident
 
n
that every feasible extension of the partial solution corre-
1 − (1 − α) j−1 pl
sponding to the latter one is also feasible for the partial
l= j
solution corresponding to the former one, and leads to a com-

a−1  plete solution whose objective function value is equal to or
+(n − ( j − 1 − a)) ϕ(n − j + 1 + l)
less tan that of the solution constructed by using the same
l=0
extension to the former one. 

according to Eq. (7), in which j − 1 − a gives the number of
j−1
primary jobs processed before the previous batch, l= j−a pl According to the above analyses, we provide the following
gives the total processing times of all the jobs contained in the procedure for Algorithm MDBT :
n
previous batch, and l= j pl measures the total processing Algorithm MDBT
times of all the primary jobs processed after the previous Step 1. Re-arrange the jobs in terms of the SPT rule so that
batch. Therefore, the value that increases in the optimization p1 ≤ · · · ≤ pn .
function 1 due to this operation in thi Step 2. Set F0 = {(0, 0)}.
s case is Step 3. Construct tag set F j from F j−1
For j = 1 to n do
 
j−1
F j ← ∅;
ξ (n − ( j − 1 − a)) pl + a
For each (a, g) ∈ F j−1 , do
l= j−a
/* Correspond to Case 1 */
 
n  j−1
1 − (1 − α) j−1 pl F j ← F j ∪ {(1, g + ξ((n − ( j − 1 − a)) l= j−a pl
n
l= j + a(1 − (1 − α) j−1 ) l= pl +
a−1 j

a−1  (n − ( j − 1 − a)) l=0 ϕ(n − j + 1 + l)) + ϑ)};
+(n − ( j − 1 − a)) ϕ(n − j + 1 + l) + ϑ. /* Correspond to Case 2 */
l=0 If 1 ≤ a < b then
F j ← F j ∪ {(a + 1, g)}, in which g = g if j < n
As a consequence, one can construct a new tag j a
and g = g + ξ(a + 1)( l= j−a pl + l=0 ϕ(l))
  
j−1 otherwise;
1, g + ξ (n − ( j − 1 − a)) pl Endif
l= j−a /* Discard nondominated states from F j */
 
n Given any tags (a, g) and (a  , g) in F j such that a ≤

+a 1 − (1 − α) j−1 pl a , discard the latter tag from F j ;
l= j Given any tags (a, g) and (a, g  ) in F j such that g ≤
 
a−1 g , discard the latter tag from F j ;
+(n − ( j − 1 − a)) ϕ(n − j + 1 + l) + ϑ Endfor
l=0
Step 4. Return the optimal optimization function value
and add it into F j . min{g|(a, g) ∈ Fn }.
Case T2: Append job J j to the last batch. To ensure the
feasibility, there must be 1 ≤ a < b by the restriction on the

123
Complex & Intelligent Systems (2021) 7:191–202 197

Theorem 4.2 Problem P1 may be solved by Algorithm Specially, we need to enumerate all possible values of the
MDBT with running time O(nb). number of early jobs in the complete optimal solution. Given
each such value e and each j, e, j = 0, 1, . . . , n, we let H(e, j)
Proof Due to the fact that all possible scheduling choices for represent the set of tags encoding the partial solutions for jobs
each job J j are addressed in Step 3, Algorithm 1 can find an J1 , . . . , J j so that the number of early jobs in the complete
optimal schedule. In Step 1, we execute a sorting procedure optimal solution is exactly e. Each tag in H(e, j) is in the form
requiring n log n time. In Step 3, since there are b different of (x, a, t1 , t2 , g) such that
values of a, there are at most b different combinations of
{a, g} according to the dominant rule. For any state, at most • there is exactly x primary jobs processed after the first
two new states can be constructed. As a consequence, going batch of the partial solution,
through n iterations, Step 3 may be implemented in O(nb) • there is exactly a primary jobs in the first batch of the
time, which completes the proof. 
 partial solution,
• the total processing time of the primary jobs processed
In what follows, we use the following example to demon-
after the first batch of the partial solution is t1 ,
strate Algorithm MDBT .
• the total processing time of the primary jobs processed
Example 4.3 Let n = 3, α = 0.1, μ = η = 3, γ = 1, ϑ = in the first batch of the partial solution is t2 ,
10, p1 = 30, p2 = 20, p3 = 15, b = 2, and ϕ(x) = x for
all x ∈ [0, +∞). We implement Algorithm MDBT to solve in which g is defined similarly as that in Algorithm M D BT .
this example in the following way. The algorithm begins with an initial tag set H(e,0) =
{(0, 0, 0, 0, 0)}. For each j = 1, · · · , n, the tag set H(e, j)
Step 1: The jobs are already re-arranged in the LPT order. is generated from tag set H(e, j−1) . To be exact, given any tag
Step 2: Set F0 = {(0, 0)} and ξ = min{γ , η} = 1. (x, a, t1 , t2 , g) ∈ F(e, j−1) that corresponds to a partial solu-
Step 3: For j = 1, we generate a tag (1, ϑ) = (1, 10) tion for the jobs J1 , . . . , J j−1 , to generate a new schedule
from the tag (0, 0) ∈ F0 . Thus, F1 = {(1, 10)}. by appending the next unassigned job J j to the start of the
above partial solution, one require to take the following two
For j = 2, we generate two tags (1, 10 + 3 p1 + 0.1( p2 + cases into account.
p3 ) + 3ϕ(2) + ϑ) = (1, 119.50) and (2, 10) from the tag
(1, 10) ∈ F1 . Thus, F2 = {(1, 119.50), (2, 10)}. Case L1: assign job J j as a late job. The value that
For j = 3, we generate two tags (1, 119.5+2 p2 +(1−(1− increases in the optimization function 2 due to this
α)2 ) p3 + 2ϕ(1) + ϑ) = (1, 174.35) and (2, 119.50 + 2( p2 + operation is ω j . Thus, one can construct a new tag
p3 + ϕ(1))) = (2, 191.50) from the tag (1, 119.50) ∈ F2 ; (x, a, t1 , t2 , g + w j ) and add it into H(e, j) .
and a tag (1, 10 + 3( p1 + p2 ) + 2(1 − (1 − α)2 ) p3 + Case L2: assign job J j as an early job. There must be
3(ϕ(1) + ϕ(2)) + ϑ) = (1, 184.70) from the tag (2, 10). a + x < e in this case, and we require to further address
During the elimination process, (1,184.70) is deleted. Thus, two subcases.
F2 = {(1, 174.35), (2, 191.50)}, and the optimal solution Subcase L21: Start a new batch and append job J j
value is 174.35 and the optimal primary job sequence is to the new batch. The contributions of all jobs in the
(J1 , J2 , J3 ), where each single job forms a batch. previous batch to the objective function value of 2
What is noteworthy is that Algorithm MDBT with a slight in this case is
amendment by deleting the constraint on a can be applied for 
 
solving problem P1 with unbound batch size. In terms of this γ (x + a)t2 + a 1 − (1 − α)e−x t1
modification, there are at most n possible values for a. Thus,
problem P1 with unbounded batch size may be solved in 
a−1
+(x + a) ϕ(x + l)
O(n 2 ) time.
l=0
Now we turn to the solution algorithm for problem P2.
It’s worth noting that Algorithm MDBT with directly amend- by Eq. (8). Thus, the value that increases in the opti-
ments cannot be used to solve problem P2 since we cannot mization function of 2 due to this operation in this
determine which jobs would be scheduled after the current case equals
batch. Instead, we devise a backward dynamic programming
algorithm, denoted as MDB L , that appends a job immedi-
ately before the partial schedule. To keep things simple, in the
sequel we re-arrange the jobs so that p1 ≥ p2 ≥ · · · ≥ pn .

123
198 Complex & Intelligent Systems (2021) 7:191–202

 /* Correspond to Case L22 */

 
γ (x + a)t2 + a 1 − (1 − α)e−x t1 If a < b then
H(e, j) ← H(e, j) ∪ {(x, a + 1, t1 , t2 + p j , g)},

a−1 in which g = g if j < n and
a−1
+(x + a) ϕ(x + l) + ϑ. g = g + γ (x + a)(t2 + l=0 ϕ(x + l)) oth-
l=0 erwise;
Endif
As a consequence, one can construct a new tag
Endif
 Endfor
x + a, 1, t1 + t2 , p j , g + γ ((x + a)t2 . /* Discard nondominated tags from H(e, j) */
  Given any tags (x, a, t1 , t2 , g) and (x  , a, t1 , t2 , g)
+a 1 − (1 − α)e−x t1 + (x + a) in H(e, j) such that x ≤ x  , discard the
a−1 latter one from H(e, j) ;
ϕ(x + l) + ϑ)
l=0 Given any tags (x, a, t1 , t2 , g) and (x, a  , t1 , t2 , g)
in H(e, j) such that a ≤ a  , discard the
and add it into H(e, j) . latter one from H(e, j) ;
Subcase L22: Append job J j to the first batch. In Given any tags (x, a, t1 , t2 , g) and (x, a, t1 , t2 , g)
this case, there must be a < b by the restriction on in H(e, j) such that t1 ≤ t1 , discard the
the batch size. The value that increases in the opti- latter one from H(e, j) ;
mization function 2 due  to this operation is zero if Given any tags (x, a, t1 , t2 , g) and (x, a, t1 , t2 , g)
a−1
j < n, and γ (x + a) t2 + l=0 ϕ(x + l) , oth- in H(e, j) such that t2 ≤ t2 , discard the
erwise. Thus, if a < b, one can construct a new latter one from H(e, j) ;
tag (x, a + 1, t1 , t2 + p j , g) and add it into H(e, j) , Given any tags (x, a, t1 , t2 , g) and (x, a, t1 , t2 , g  )
in which g = g if j < in H(e, j) such that g ≤ g  , discard the
a−1  n and g = g + γ (x +
latter one from H(e, j) ;
a) t2 + l=0 ϕ(x + l) otherwise.
Endfor
Endfor
To cut down the number of states, one can use the follow-
Step 4. Return the optimal optimization function value
ing dominant rule.
min{ f A |(x, e − x, t1 , t2 , g) ∈ H(e,n) , e = 0, 1, · · · , n}.
Lemma 4.4 Given any tags (x, a, t1 , t2 , g) and (x  , a  , t1 , t2 ,
Theorem 4.5 Problem P2 may be solved by Algorithm
g  ) in F j such that x ≤ x  , a ≤ a  , t1 ≤ t1 , t2 ≤ t2 , and
MDB L with running time O(n 3 b P 2 ).
g ≤ g  , one can discard the latter one.
Proof It is analogous to the proof of Theorem 3.2, in which
Proof It is similar to the proof Lemma 3.1. 

the only differences lie in that: x, t1 and t2 take at most n, P
According to the analyses above, we provide the following and P different values, respectively. 

procedure for Algorithm MDB L .
Similarly, Algorithm MDB L with a slight amendment by
Algorithm M D B L deleting the constraint on a can be applied for solving prob-
Step 1. Re-arrange the jobs so that p1 ≥ · · · ≥ pn . lem P2 with unbound batch size, which runs in O(n 4 P 2 )
Step 2. Let H(e,0) = {(0, 0, 0, 0, 0)} for all e = 0, 1, · · · , n. time.
Step 3. Construct tag set H(e, j) from H(e, j−1)
For e = 0 to n do
For j = 1 to n do Computational experiments
H(e, j) ← ∅;
For each (x, a, t1 , t2 , g) ∈ H(e, j−1) , do This section assesses how multitasking affects the scheduling
/* Correspond to Case L1 */ cost or value through computational experiments. In terms of
H(e, j) ← H(e, j) ∪ {(x, a, t1 , t2 , g + w j )}; the parameter settings in Hall et al. [12] and Wang et al. [23],
/* Correspond to Case L2 */ we generate different sets of parameter values as follows.
If x + a < e then
/* Correspond to Case L21 */ • We selected n from the set {50, 60, . . . , 120} for problem
H(e, j) ← H(e, j) ∪ {(x + a, 1, t1 + t2 , p j , g + P1 and from the set {5, 10, 15, 20} for problem P2.
γ ((x + a)t2 + a(1 − (1 − α)e−x )t1 + • We randomly selected p j from the uniform distribution
a−1
(x + a) l=0 ϕ(x + l)) + ϑ)}; [10, 50].

123
Complex & Intelligent Systems (2021) 7:191–202 199

• We randomly selected η and γ from the uniform distri- P1, and the L L avg value for α = 0.15, 0.10 and 0.05,
bution [1, 10]. respectively, increases 301.63%, 196.56% and 112.32%
• We randomly selected ϑ from the uniform distribution on average compared to that for α = 0.01 for problem
[20, 100], and b from the discrete uniform [2, n]. P2. The reason behind this is that the interruption time
• We randomly selected ω j from the uniform distribution that the waiting jobs suspend the primary jobs becomes
[b P/n, P/2]. large when α increases.
• We selected α from the set {0.01, 0.05, 0.10, 0.15}. • The affect of the number of jobs on the L Tavg and L L avg
values is obviously. From Tables 1 and 2, we can see
The reason for the setting of n is as follows: the pre- that the instances with more larger n induce more lager
liminary computational test indicates that Algorithm MDBT cost of multitasking for most problem instances. For
takes 2 h on average for solving the instances with n = 120 of instance, the average cost of multitasking ranges from
problem P1, and Algorithm MDBl takes more 2 h on average 98.15 to 143.61% when (α, ϕ(x)) = (0.10, − 0.05x)
for solving the instances with n = 20 of problem P2. The and ranges from 98.50 to 116.67% when (α, ϕ(x)) =
average running times of the algorithms on testing instances (0.10, − 0.05x), which increases approximately in pro-
with different numbers of jobs are depicted in Tables 2 and portion to the number of jobs.
3, respectively. • Negative L Tavg or L L avg value is possible, implying
Given any combination of the parameters above, we per- that we can get some revenue from multitasking. For
form two computational experiments, i.e., ϕ(x) = 0.05x and instance, when n = 5, we can get 0.01% revenue from
ϕ(x) = − 0.05x. And for any combination of n and α, 30 multitasking when (α, ϕ(x)) = (0.05, − 0.05x), and
instances were randomly chosen and the average results were get 0.05% revenue from multitasking when (α, ϕ(x)) =
reported. (0.01, − 0.05x) for problem P2. However, this result
We apply the algorithm for each problem by setting α = 0 does not continue as the number of jobs increases.
for solving the corresponding problem without multitasking. • The affect of multitasking for problem P1 is more larger
The developed DP algorithms are coded in MATLAB and the than that for problem P2. This is due to the fact that mul-
numerical experiments are performed on a notebook com- titasking will affect the completion times of jobs, and the
puter with a 3.80-GHz CPU and 16 GB memory. optimization function of problem P1 is more related to
Let 1 and 1 be the optimal optimization function val- the completion times of jobs compared to that of problem
ues to the multitasking problem P1 and the corresponding P2.
problem without multitasking, respectively, and L Tavg be the
 −
mean value of 1 1 over the 50 chosen instances for any
1
combination of n and α, which demonstrates the average
Conclusions
multitasking cost or value. In a similar way, we let 2 and
2 be the optimal optimization function values to the multi-
In this study, we investigate a novel scheduling model that
tasking problem P2 and the corresponding problem without
coinstantaneously involves batch distribution, multitasking
multitasking, respectively, and L L avg be the mean value of
2 −2 and DDA. In the developed model, the job due dates are to be
2
over the 50 chosen instances. decided by the decision-makers in the decision-making pro-
The results for problems P1 and P2 are summarized in cess. The completed jobs are distributed to their customers
Tables 2 and 3, from which we have the following observa- in batches, and the sizes of all batches are identical, which
tions. may be bounded or unbounded. The jobs contained in each
batch are scheduled sequentially. In addition, it permits the
• As expected, the affect of multitasking on the scheduling machine to suspend an uncompleted job and turn to another
cost or value increases as the α value increases, especially uncompleted job. The goal is to identify the optimal primary
for problem P1. For instances, when ϕ(x) = 0.05x, the job sequence, the optimal job due dates, and the optimal batch
L Tavg value for α = 0.15, 0.10 and 0.05 respectively production and delivery strategy so that one of the following
increases 54.02%, 39.12% and 28.17% on average com- optimization functions are minimised: the total cost com-
pared to that for α = 0.01 for problem P1, and the L L avg posed of the earliness penalty, DDA cost, tardiness penalty
value for α = 0.15, 0.10 and 0.05 respectively increases and batch distribution cost, and the total cost composed of
512.17%, 158.37% and 158.71% on average compared the earliness penalty, weighted number of late jobs, DDA
to that for α = 0.01 for problem P2. In addition, when cost and batch distribution cost. We prove that the problem
ϕ(x) = − 0.05x, the L Tavg value for α = 0.15, 0.10 and with the first optimization function may be solved in polyno-
0.05, respectively increases 96.81%, 84.05% and 43.47% mial time and that the problem with the second optimization
on average compared to that for α = 0.01 for problem function is pseudo-polynomial time solvable. We also con-

123
 200

Table 2 The results for problem P1

n Avg. time (s) Cost of multitasking under (α, ϕ(x))
(0.15, 0.05x) (%) (0.10, 0.05x) (%) (0.05, 0.05x) (%) (0.01, 0.05x) (%) (0.15, − 0.05x) (%) (0.10, − 0.05x) (%) (0.05, − 0.05x) (%) (0.01, − 0.05x)

123
50 28.01 112.61 98.15 88.50 64.31 107.23 98.50 70.46 52.15
60 34.36 116.10 103.05 95.35 69.94 109.51 105.14 76.13 54.21
70 491.22 130.22 112.96 104.58 83.12 114.18 108.65 85.41 55.26
80 975.08 131.51 123.64 109.61 87.86 116.93 107.12 87.24 60.32
90 1059.74 139.92 126.78 121.74 94.26 119.16 113.51 91.02 62.15
100 1372.05 144.21 131.54 125.16 101.34 123.38 115.26 90.35 63.98
110 1555.82 147.29 139.92 126.30 105.21 126.19 114.98 96.81 68.22
120 1851.00 151.34 143.61 132.71 116.34 130.54 116.67 97.42 70.13

Table 3 The results for problem P2

n Avg. time (s) Cost of multitasking under (α, ϕ(x))
(0.15, 0.05x) (%) (0.10, 0.05x) (%) (0.05, 0.05x) (%) (0.01, 0.05x) (%) (0.15, − 0.05x) (%) (0.10, − 0.05x) (%) (0.05, − 0.05x) (%) (0.01, − 0.05x)

5 10.02 0.16 0.04 0.05 0.01 0.04 0.00 − 0.01 − 0.05

10 130.66 2.98 2.05 1.49 0.83 1.08 0.77 0.44 0.12
15 1175.31 7.32 5.94 5.68 3.21 6.13 4.01 3.38 1.23
20 6870.21 14.18 10.92 9.67 5.42 9.48 7.19 6.18 3.29
Complex & Intelligent Systems (2021) 7:191–202
Complex & Intelligent Systems (2021) 7:191–202 201

duct numerical experiments on randomly chosen instances to 2. Agnetis A, Aloulou MA, Kovalyov MY (2017) Integrated produc-
assess how multitasking affects the scheduling cost or value, tion scheduling and batch delivery with fixed departure times and
inventory holding costs. Int J Prod Res 55(20):6193–6206
the results of which can assist decision-makers to justify the 3. Agnetis A, Aloulou MA, Fu LL (2016) Production and interplant
extent to put to use or refrain multitasking. batch delivery scheduling: dominance and cooperation. Int J Prod
As for future research, the following topics are interesting Econ 182:38–49
and necessary. 4. Chen ZL (1996) Scheduling and common due date assignment with
earliness-tardiness penalties and batch delivery costs. Eur J Oper
Res 93:49–60
• Providing the computational complexity status of the 5. Chen ZL (2010) Integrated production and outbound distribution
scheduling: review and extensions. Oper Res 58(1):130–148
problem P2, i.e. whether it is N P-hard.
6. Cheng BY, Leung JYT, Li K, Yang SL (2015) Single batch machine
• Investigating the model with other due date assignment scheduling with deliveries. Naval Res Logist 62:470–482
methods. 7. Gao K, Huang Y, Sadollah A, Wang L (2020) A review of energy-
• Extending the model to other machine setting, for exam- efficient scheduling in intelligent production systems. Complex
Intell Syst 6:237–249
ple, flowshop or parallel-machine setting.
8. Gong H, Tang L, Leung JYT (2016) Parallel machine scheduling
• Investigating the model in a uncertain or dynamic envi- with batch deliveries to minimize total flow time and delivery cost.
ronment, for example in the present of random processing Naval Res Logist 63:492–502
times, unexpected machine breakdown, etc. 9. Gordon V, Proth JM, Chu C (2002) A Survey of the state-of-the-art
of common due date assignment and scheduling research. Eur J
• Extending the model to the case with energy efficiency,
Oper Res 139:1–25
consumption, or cost as constraints or objectives (Gao et 10. Gordon V, Strusevich V, Dolgui A (2012) Scheduling with due date
al. [7]). assignment under special conditions on job processing. J Sched
• Developing effective intelligent optimization algorithms 15:447–456
11. Kaminsky P, Hochbaum D (2014) Due-date quotation models and
for solving large-scale problem, such as Tabu search,
algorithms. In: Leung JY-T (ed) Handbook of scheduling: algo-
genetic algorithm, or genetic programming (Nguyen et rithms, models and performance analysis. CRC Press, Boca Raton,
al. [21]). pp 20:1–20:22
12. Hall NG, Leung JYT, Li CL (2015) The effects of multitasking on
operations scheduling. Prod Oper Manag 24:1248–1265
Acknowledgements We thank the editor, associate editor, and two 13. Hall NG, Leung JYT, Li CL (2016) Multitasking via alternate and
anonymous referees for their helpful comments on earlier versions of shared processing: algorithms and complexity. Discrete Appl Math
our paper. This paper was supported in part by the National Natural 208:41–58
Science Foundation of China under grant number 71971041, by the 14. Hall NG, Potts CN (2003) Supply chain scheduling: batching and
Outstanding Young Scientific and Technological Talents Foundation of delivery. Oper Res 51:566–584
Sichuan Province under grant number 2020JDJQ0035, and in part by 15. Ji M, Zhang W, Liao L, Cheng TCE, Tan Y (2019) Multitasking
the Science and Technology Project of Education Department of Jiangxi parallel-machine scheduling with machine-dependent slack due-
Province under grant number 180375. window assignment. Int J Prod Res 57:1667–1684
16. Kovalyov MY, Oulamara A, Soukhal A (2015) Two-agent schedul-
ing with agent specific batches on an unbounded serial batching
Compliance with ethical standards machine. J Sched 18:423–434
17. Li F, Chen ZL, Tang L (2017) Integrated production, inventory
Conflict of interest On behalf of all authors, the corresponding author and delivery problems: complexity and algorithms. INFORMS J
states that there is no conflict of interest. Comput 29(2):232–250
18. Li S, Chen R, Tian J (2019) Multitasking scheduling problems
Open Access This article is licensed under a Creative Commons with two competitive agents. Eng Optim. https://doi.org/10.1080/
Attribution 4.0 International License, which permits use, sharing, adap- 0305215X.2019.1678609
tation, distribution and reproduction in any medium or format, as 19. Liu M, Wang SJ, Zheng FF, Chu CB (2017) Algorithms for the
long as you give appropriate credit to the original author(s) and the joint multitasking scheduling and common due date assignment
source, provide a link to the Creative Commons licence, and indi- problem. Int J Prod Res 55:6052–6066
cate if changes were made. The images or other third party material 20. Mor B, Mosheiov G (2011) Single machine batch scheduling with
in this article are included in the article’s Creative Commons licence, two competing agents to minimize total flowtime. Eur J Oper Res
unless indicated otherwise in a credit line to the material. If material 215:524–531
is not included in the article’s Creative Commons licence and your 21. Nguyen S, Mei Y, Zhang M (2017) Genetic programming for pro-
intended use is not permitted by statutory regulation or exceeds the duction scheduling: a survey with a unified framework. Complex
permitted use, you will need to obtain permission directly from the copy- Intell Syst 3:41–66
right holder. To view a copy of this licence, visit http://creativecomm 22. Shabtay D (2010) Scheduling and due date assignment to mini-
ons.org/licenses/by/4.0/. mize earliness, tardiness, holding, due date assignment and batch
delivery costs. Int J Prod Econ 123:235–242
23. Wang D, Yu Y, Yin Y, Cheng TCE (2020) Multi-agent scheduling
References problems under multitasking. Int J Prod Res. https://doi.org/10.
1080/00207543.2020.1748908
1. Ahmadizar F, Farhadi S (2015) Single-machine batch delivery 24. Xiong X, Zhou P, Yin Y, Cheng TCE, Li D (2019) An exact
scheduling with job release dates, due windows and earliness, tar- branch-and-price algorithm for multitasking scheduling on unre-
diness, holding and delivery costs. Comput Oper Res 53:194–205 lated parallel machines. Naval Res Logist 66:502–516

123
202 Complex & Intelligent Systems (2021) 7:191–202

25. Xu S, Bean JC (2016) Scheduling parallel-machine batch opera- 31. Yin Y, Wang Y, Cheng TCE, Wang D, Wu CC (2016) Two-agent
tions to maximize on-time delivery performance. J Sched 19:583– single-machine scheduling to minimize the batch delivery cost.
600 Comput Ind Eng 92:16–30
26. Yin Y, Cheng TCE, Cheng SR, Wu CC (2013) Single-machine 32. Yin Y, Yang Y, Wang D, Cheng TCE, Wu CC (2018) Integrated
batch delivery scheduling with an assignable common due date production, inventory, and batch delivery scheduling with due date
and controllable processing times. Comput Ind Eng 65:652–662 assignment and two competing agents. Naval Res Logist 65:393–
27. Yin Y, Cheng TCE, Hsu CJ, Wu CC (2013) Single-machine batch 409
delivery scheduling with an assignable common due window. 33. Zhu ZG, Liu M, Chu CB, Li JL (2019) Multitasking scheduling with
Omega 41:216–225 multiple rate-modifying activities. Int Trans Oper Res 26:1956–
28. Yin Y, Cheng TCE, Wu CC, Cheng SR (2013) Single-machine com- 1976
mon due-date scheduling with batch delivery costs and resource- 34. Zhu ZG, Zheng FF, Chu CB (2017) Multitasking scheduling prob-
dependent processing times. Int J Prod Res 51(17):5083–5099 lems with a rate-modifying activity. Int J Prod Res 55:296–312
29. Yin Y, Cheng TCE, Xu D, Wu CC (2012) Common due date assign-
ment and scheduling with a rate-modifying activity to minimize
the due date, earliness, tardiness, holding, and batch delivery cost.
Publisher’s Note Springer Nature remains neutral with regard to juris-
Comput Ind Eng 63:223–234
dictional claims in published maps and institutional affiliations.
30. Yin Y, Li D, Wang D, Cheng TCE (2018) Single-machine serial-
batch delivery scheduling with two competing agents and due date
assignment. Ann Oper Res. https://doi.org/10.1007/s10479-018-
2839-6

123