European Journal of Operational Research 233 (2014) 75–83

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Discrete Optimization

Optimal single machine scheduling of products with components

and changeover cost
Feng Zhou a,c, James D. Blocher a, Xinxin Hu a, H. Sebastian Heese b,⇑
a
Kelley School of Business, Indiana University, 1309 East 10th Street, Bloomington, IN 47405, USA
b
EBS University, Institute for Supply Chain Management, Konrad-Adenauer-Ring 15, 65187 Wiesbaden, Germany
c
California State University Stanislaus, College of Business Administration, One University Circle, Turlock, CA 95382, USA

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

Article history: We consider the problem of scheduling products with components on a single machine, where change-
Received 19 July 2012 overs incur fixed costs. The objective is to minimize the weighted sum of total flow time and changeover
Accepted 12 August 2013 cost. We provide properties of optimal solutions and develop an explicit characterization of optimal
Available online 22 August 2013
sequences, while showing that this characterization has recurrent properties. Our structural results have
interesting implications for practitioners, primarily that the structure of optimal sequences is robust to
Keywords: changes in demand.
Scheduling
Ó 2013 Elsevier B.V. All rights reserved.
Single machine
Components
Flow time
Changeover cost

1. Introduction We show that there exists a series of recurrent optimal se-

The problem of scheduling products with changeovers on a
common resource, like a single machine, exists in many manufac-
turing settings. For instance, a machine processes different types of
circuit boards by inserting components (cf. Coffman, Nozari, &
Yannakakis, 1989), or a paper machine is capable of producing
different grades of paper (cf. Magnanti & Vachani, 1990). One com-
mon feature of these problems is that products of the same type
are often grouped to form batches. A changeover is incurred when-
ever the machine is switched over to process a different batch of
products. The resulting sequencing and batching decisions are then
made based on a number of possible criteria, including for exam-
ple, total flow time, total changeover cost, number of late products,
etc.
This paper considers a setting where a manufacturer fabricates
multiple types of components on a single machine that are
ultimately assembled into identical products. A fixed cost is
incurred whenever there is a changeover from one component to
another. A component is immediately available for assembly after
its completion. Our objective is to determine an optimal sequence
(involving both sequencing and batching decisions) that minimizes
the weighted sum of total flow time and changeover cost.
⇑ Corresponding author. Tel.: +49 611 7102 2194; fax: +49 611 7102 10 2194.
E-mail addresses: fzhou@csustan.edu (F. Zhou), dblocher@indiana.edu
(J.D. Blocher), hu.xinxin@hotmail.com (X. Hu), sebastian.heese@ebs.edu
(H. Sebastian Heese).
quences, and we obtain closed-form solutions for the integer batch
sizes in these sequences. Based on the proven properties and the
existence of the recurrent optimal sequences, we provide an expli-
cit characterization of optimal sequences for any size problem. We
also show that when there is a changeover cost, rather than a
changeover time, the optimal sequence exhibits a constant-
batch-size pattern. This result is especially interesting given that
the algorithms of Coffman et al. (1989), who consider the same set-
ting but use changeover time rather than cost, lead to non-increas-
ing batch sizes in the optimal sequences.
The rest of the paper is organized as follows. In Section 2, we
review the related literature. In Section 3, we introduce the formu-
lation of the problem. Section 4 describes the major results about
the properties and a characterization of optimal sequences. Mana-
gerial implications of our solution are discussed in Section 5.
2. Literature review
There has been extensive research on changeover scheduling
problems with batching; we refer the reader to the surveys by
Potts and Kovalyov (2000) and Allahverdi, Ng, Cheng, and Kovalyov
(2008). Many of these papers refer to setups instead of change-
overs, although both terms can be used interchangeably. To avoid
confusion and maintain consistency, we use the term changeover
throughout the paper.
The existing literature on changeover scheduling with batching
can be divided into two main streams depending on whether or not

0377-2217/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2013.08.016
76 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83
the products are composed of components of multiple types. The
first research stream studies batching and/or sequencing decisions
for products without components. Most of the work in this stream
is conducted under a single machine setting. For instance, Santos
and Magazine (1985) and Dobson, Karmarkar, and Rummel
(1987) independently started this research stream by proposing
the same single-machine batching problem to minimize the total
flow time of one product. In both papers, a constant changeover
time is needed before the machine can process a different batch
of products. Dobson et al. (1987) also provide heuristics for the
multi-product case. Cheng and Kovalyov (2001) consider a sin-
gle-machine scheduling problem for multiple products with dead-
lines. A changeover cost is incurred when the machine is switched
to produce another product. The objective is to minimize total
changeover costs. A few papers address more complex manufac-
turing environments. Cheng, Kovalyov, and Chakhlevich (2004)
and Lin and Cheng (2005), for example, minimize the makespan
of a two-stage flowshop. The two papers differ in how they model
the second stage of production (dedicated machines or single flex-
ible machine) and where batching operations are performed (at the
first stage or the second stage).
The second research stream, which is more relevant to our
work, focuses on scheduling products with multiple component
types. A product is not completed until all of its components are
finished, which adds to the complexity of the problem. One set of
papers in this stream considers the problem of processing compo-
nents on parallel dedicated machines. In some cases, there is a sub-
sequent assembly stage, in which all of the components are
assembled into a final product. This problem is sometimes referred
to as the assembly scheduling problem with batching operations.
For instance, Kovalyov, Potts, and Strusevich (2004) study a two-
stage assembly scheduling problem minimizing the makespan. In
other cases, the assembly stage is ignored. For instance, Lin and
Cheng (2011) consider a batch scheduling model in a concurrent
open shop where operations of each product are independently
performed on their parallel dedicated machines. They consider
minimizing the maximum lateness, weighted number of tardy
products and total weighted completion time. In this paper, we fo-
cus on the problem of scheduling products with components on a
single bottleneck machine. We assume that there is ample capacity
at other resources, such as for example an assembly stage, and we
thus do not explicitly capture these non-bottleneck resources in
our model.
A second set of papers in this stream study the scheduling of
products with multiple component types on a single flexible ma-
chine. Papers in this set assume each product consists of two com-
ponents; one component is common to all products and the other
one is unique to each product. The assumption of changeover time
leads to a batching of the common components, but has no effect
on the unique components. Baker (1988) and Vickson, Magazine,
and Santos (1993) consider the same batching problem on a single
machine to minimize flow time, but use different assumptions
regarding the calculation of flow time for each product. Gerodimos,
Glass, and Potts (2000) and Lin (2002) study the same model as Ba-
ker (1988) and Vickson et al. (1993) but consider two due-date re-
lated objectives, minimizing the maximum lateness and the
number of late products. These papers also differ in how flow time
is computed. A similar model for a more complex environment is
studied by Cheng and Wang (1999) where they show minimizing
the makespan for a two-stage flowshop is NP-complete.
Finally, a third set of papers study the scheduling of products
with multiple component types, all of which are common to all
products. This is the single machine problem we address in this pa-
per. Both batching and sequencing decisions are needed for these
different common components and thus this problem is quite
complex. Coffman et al. (1989) started this research stream by
considering the problem of scheduling identical products with
two common component types on a single machine to minimize
flow time. A fixed changeover time is incurred whenever the ma-
chine is switched to process a batch of components of a different
type. They provide efficient algorithms to obtain optimal se-
quences. Cheng, Ng, and Yuan (2008) consider a model of schedul-
ing products with multiple operations on a single machine with
the objective of minimizing the total completion time. The opera-
tions of each product belong to different families. Each family of
operations can be grouped as a batch which requires a changeover
time. They provide some complexity results for the problem and
show that some special cases are solvable in polynomial time.
Most papers of these research streams assume changeovers in-
cur a time delay, but a few papers do consider the case where
changeovers could result in costs. As pointed out by Magnanti
and Vachani (1990), considering changeover cost instead of time
may be more appropriate in some settings, where changeovers
can be executed after the regular production hours without con-
suming resource time or when changeover time is small.
To the best of our knowledge, only Kella (1991) and Gim and
Han (1997) have studied the problem of scheduling products with
multiple common components on a single machine taking into ac-
count changeover cost. Both of these papers assume an infinite
horizon problem setting. Kella (1991) studies a setting where a
machine sequentially produces two types of fluid components
which are mixed to produce a final product. A fixed changeover
cost is incurred whenever the machine is switched from producing
one type of fluid component to the other. An optimal cyclic policy
is provided to minimize the long-run average cost that includes
both inventory holding cost and changeover cost. Gim and Han
(1997) address a variant of this problem by assuming N different
types of components are processed on a single machine and then
assembled into identical products. Both changeover time and cost
are incurred when the machine is switched to produce a different
component. They develop a heuristic based on a repetitive se-
quence for an infinite horizon problem to minimize total produc-
tion cost, which consists of changeover cost, inventory cost and
other costs. Our model can be viewed as a discrete optimization
version of Kella (1991). While Kella (1991) studies an infinite hori-
zon continuous optimization problem in which there is no
sequencing decision, we derive solutions for both the finite and
the infinite horizon problem, including the sequencing of the
components.
Our finding that there exists a series of recurrent optimal se-
quences is consistent with the optimality of a stationary periodic
policy suggested by Kella (1991). We find that when there is a
changeover cost, rather than a changeover time, the optimal se-
quence exhibits a constant-batch-size pattern. This structural in-
sight of non-increasing optimal batch sizes has also been
supported in Santos and Magazine (1985), Dobson et al. (1987),
Coffman and Yannakakis (1990), Mosheiov and Oron (2008), all
of which consider changeover time. This result is especially inter-
esting given that the algorithms of Coffman et al. (1989), who con-
sider the same setting but use changeover time rather than cost,
lead to non-increasing batch sizes in the optimal sequences.
3. Model framework
There are n identical products that need to be produced on a
single machine. While our structural results and related insights
extend to the case with more than two components (see Appendix
B), for expositional clarity throughout the main part of the paper,
we assume that each product is composed of two components.
Specifically, each product requires one component of type A with
processing time a > 0 and one component of type B with processing
time b > 0. For convenience, define T = b/a. Without loss of
 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83 77

generality, we assume that a 6 b, such that T P 1. The n products 4. Optimal sequences

can be produced in any sequence of the n type-A and n type-B
components.
In any given n-product sequence rn, a fixed changeover cost
c0 > 0 is incurred for every switch between the two types of compo-
nents. Let b(rn) denote the number of changeovers in rn, and note
that b 2 [1, 2n  1]. Let Fi(rn) denote the flow time for product
i = 1, . . . , n.
A component becomes available for assembly instantly after its
completion. Once all components are produced, a final product is
immediately assembled and shipped. We assume the assembly
stage has sufficient capacity, such that scheduling decisions at this
non-bottleneck resource are of negligible importance. We use (AB)s
to denote a block, or sequence, of s type-A components followed
by an equal number of type-B components. For example,
r4 = AAABBBBA = (AB)3(BA)1. Thus product completions occur
P4
after
the first three B and after the last A, yielding i¼1 F i ðr4 Þ
¼ ð3a þ bÞ þ ð3a þ 2bÞ þ ð3a þ 3bÞ þ ð4a þ 4bÞ ¼ 13a þ 10b and
b(r4) = 2. In another sequence, like ABBAABBA = (AB)1(BA)1(AB)1
(BA)1, the total flow time becomes (a + b) + (2a + 2b) + (3a +
3b) + (4a + 4b) = 10a + 10b. Apparently, changeovers help to reduce
the flow time of products but also incur some costs.
Flow time is an important criterion in production scheduling
because it represents the lead time of customer orders and also di-
The following three theorems allow us to characterize optimal
sequences for various values of n given the parameters a, b, and
c. Theorems 1 and 2 prove the recurring form and provide the inte-
ger sizes of the recurring blocks, while Theorem 3 describes how to
derive optimal sequences for problems where the number of prod-
ucts cannot be expressed through this recurring form. Some of the
proofs of the results provided in this section are based on the tech-
nical results in Lemmas 1–3, which are given in Appendix A.
Theorem 1. There exists a pair of integers (s⁄, k⁄) which minimize
sðs  1Þa þ kðk  1Þb þ 4c
Gðs; kÞ ¼ ; ð3Þ
2ðs þ kÞ

so that for n ¼ s þ hðs þ k Þ; h ¼ 0; 1; 2; . . . ; rn must have the form
{s⁄k⁄s⁄k⁄, . . . , s⁄k⁄s⁄}.
Proof. We first show that for any h = 0, 1, 2, . . . , there exist (s⁄,k⁄)
such that for nh = s⁄ + h (s⁄ + k⁄) products, an optimal sequence is
rnh ¼ fs k s k . . . s k s g.
By Lemma 1(c), the number of AB blocks in rn is m = m0 + m00 ,
0
where m0 is the number of ðABÞs blocks, m00 is the number of ðABÞs
00

rectly relates to work-in-process (WIP) inventory cost. As seen blocks, and js0  s00 j = 1 or 0. A similar0 conclusion holds for BA00
from the above examples, flow time can often be reduced by using blocks, which include l0 number of ðBAÞk blocks, l00 number of ðBAÞk
more changeovers, but these changeovers are costly. Therefore, the blocks, and jk0  k00 j = 0 or 1. By Lemma 3, there exist a ho and an no,
manufacturer needs to decide the appropriate number of change- such that
overs so as to minimize the total production cost, which includes
both the cost of changeovers and the costs associated with longer   0 00  
b rno ¼ m0o þ m00o þ lo þ lo ; max m0o ; m00o > ho þ 1; and
flow times (for example, work-in-process inventory costs). The  0 00 
objective thus is to find a sequence that minimizes the weighted max lo ; lo > ho :
sum of total flow time and changeover cost. Specifically, using
Using the indicator function 1(), we can let
w > 0 to denote the weight of the changeover cost relative to the    
total flow time, so ¼ s0o 1 m0o P m00o þ s00o 1 m00o > m0o and
! 0  0 00  00  00 0 
X
n ko ¼ ko 1 lo P lo þ ko 1 lo > lo :
rn ¼ arg min
r
Cðrn Þ ¼ arg min
r
F i ðrn Þ þ wc0 bðrn Þ : ð1Þ
n n
i¼1 Now, using Lemma 2, remove AB and BA blocks from rno until there
remain ho + 1 ðABÞso blocks and ho ðBAÞko blocks. The resulting
For expositional convenience, we use c = w c0 in the remainder of
sequence {sokosoko, . . . , sokoso} is optimal for the remaining so +
this paper.
ho(so + ko) products. Finally, Lemma 2(a) indicates that the form of
Similar to Coffman et al. (1989), it is easy to confirm that there
{sokosoko . . . sokoso} is also optimal for so + h(so + ko) products, where
is always an optimal sequence which is normal form, i.e., one that is
h 6 ho. That implies we can identify (s⁄, k⁄) for any h = 0, 1, 2, . . . .
a concatenation of the ðABÞsp (p = 1, 2, . . .) and ðBAÞkq (q = 1, 2, . . .),
We next justify that (s⁄, k⁄) must minimize G(s, k). For any (s, k),
where the AB blocks occur in every other sequence position inter-
consider n = s⁄ + (s + k)(s⁄ + k⁄) products. The optimal sequence is
leaved with the BA blocks. Thus, normal-form sequences are of the
form ðABÞs1 ðBAÞk1 ðABÞs2 ðBAÞk2 . . . or ðBAÞk1 ðABÞs1 ðBAÞk2 ðABÞs2 . . .. The
rn ¼ fs k s k . . . s k s g. An alternative sequence is rn = {s k s
sequence AAABBABB is not normal form because it is not composed k, . . . , s k s⁄}. We know that Cðrn Þ 6 Cðrn Þ. After removing the last

of blocks. Similarly, the sequence AAABBBAB is not normal form be- ðABÞs block from rn and rn, we get two sequences for
cause it is a concatenation of two sequential AB blocks. In the n0 = (s + k)(s⁄ + k⁄) products: ran0 ¼ fs k s k . . . s k g and
remainder of the paper, without loss of generality, we only con- rn0 ¼ fsksk . . . skg. From (2) and (1), we know Cðran0 Þ 6 Cðrbn0 Þ and
b

sider normal-form sequences. Cðran0 Þ Cðrbn0 Þ

n0 6 n0 , so
For notational convenience, we use sp and kq to denote the size
(number of products) of the pth AB block and the size of the qth BA   Pn0    
block, respectively. We omit the subscripts p and q when no confu- C ran0 i¼1 F i ran0 þ cb ran0
¼
sion might arise. n0 n0
1    
In general, the ith product in rn must belong to either an AB or a 1 0 ½s ðs  1Þa þ k ðk  1Þb 2c
¼ ðn þ 1Þða þ bÞ þ 2  þ  :
BA block. If it is the jth product in an s-size AB block, then Fi(rn) = 2 s þ k s þk
i(a + b) + (s  j)a. Similarly, if it is the jth product in a k-size BA
Hence, we have
block, then Fi(rn) = i(a + b) + (k  j)b. Using m to denote the number
 
of AB blocks in rn, we have s ðs  1Þa þ k ðk  1Þb þ 4c sðs  1Þa þ kðk  1Þb þ 4c
Xn Xm  6 ;
1 1 2ðs þ k Þ 2ðs þ kÞ
F i ðrn Þ ¼ nðn þ 1Þða þ bÞ þ sp ðsp  1Þa
2 2
i¼1 p¼1 which concludes our proof. h
bðr
Xn Þm
1 While the function in (3) is not necessarily well-behaved, the
þ kq ðkq  1Þb: ð2Þ following result provides closed-form expressions for the integer
q¼1
2
sizes of the recurring blocks that form an optimal sequence.
78 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83

 o  o
Theorem 2. Suppose so1 ; k1 ; . . . ; sot ; kt are (positive) integer min- is rn1 þ n2 ða þ bÞ þ so1 a < Cðrn1 Þ þ n2 ða þ bÞ þ so1 a ¼ C r0n2 .
C
o o o
imizers of (3), and so1 þ k1 6 so2 þ k2 6 . . . 6 sot þ kt , then
  o This is a contradiction to the assumed optimality of r0n2 .
ðs ; k Þ ¼ so1 ; k1 . More specifically,
To prove parts (a) and (b), consider the derivative of G(s, k) in (3)
(a) k⁄ = dKe or bK + (1 + 1/T) c, with respect to s,
 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiþ
ð1þ3TÞaþ ð1þ3TÞ2 a2 þ4ð4caÞðT 2 þTÞa
where K ¼ 2TðTþ1Þa
. kðk þ 1Þa þ kðk  1Þb þ 4c
G0s ðs; kÞ ¼ a  2
;
⁄ ⁄
(b) s = bX  k c or dX  k e, ⁄
ðs þ kÞ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
   
where X ¼ k ðk þ 1Þ þ k ðk  1ÞT þ 4c a
. which is increasing in s. Hence, given k⁄, by the first-order condi-
tion, we know that part (b) holds.
Due to the balance between AB and BA block sizes, we know
o
(k⁄  1)T 6 s⁄ 6 k⁄T + 1, so a necessary condition for a specific k to
Proof. We assume, w.l.o.g., t = 2, and denote X r ¼ sor þ kr and be optimal is
 o
G ¼ G sor ; kr , for r = 1, 2. In an optimal sequence, by Lemma 1(c),
the AB and BA blocks must satisfy the balancing conditions G0s ðkT þ 1; kÞ P 0 and G0s ðkT  T; kÞ 6 0:
l m
(s  1)a 6 kb and sa P (k  1)b, implying that sor ¼ ðX r  1Þ aþb b
Straightforward algebra then shows:
j k
o a   
and kr ¼ 1 þ ðX r  1Þ aþb . Clearly, if X1 = X2, then G0s ðk T þ 1; k Þ P 0 ) k P K
   o o  o o 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3þ
ðs ; k Þ ¼ s1 ; k1 ¼ s2 ; k2 . ð1 þ 3TÞa þ ð1 þ 3TÞ2 a2 þ 4ð4c  aÞðT 2 þ TÞa
For X1 < X2, we have so1 6 so2 . It suffices to prove that if ¼4 5
  2TðT þ 1Þa
rn2 ¼ so2 ko2 so2 . . . so2 ko2 so2 is optimal for n2 ¼ X 1 X 2 þ so2 products,
 o o o 
then rn1 ¼ s1 k1 . . . so1 k1 so1 is also optimal for n1 ¼ X 1 X 2 þ so1   
G0s ðk T  T; k Þ 6 0 ) k
products. qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
If so1 ¼ so2 , then both rn1 and rn2 are optimal, since their ð2T 2 þ T þ 1Þa þ ð1 þ 3TÞ2 a2 þ 4ð4c  aÞðT 2 þ TÞa
6 :
corresponding costs are equal due to (2), (1), and 2TðT þ 1Þa
 o  o   o
G so1 ; k1 ¼ G so2 ; k2 . Hence, we can let ðs ; k Þ ¼ so1 ; k1 . Hence, K 6 k 6 K þ 1þT

, so part (a) follows. h
T
If so1 < so2 , we can remove so2  so1 products from the last AB block
 o o 
in rn2 to get r0n1 ¼ so2 k2 . . . so2 k2 so1 . Again, based on (2), (1), and the According to Theorems 1 and 2, we can explicitly characterize
 o o  o o
fact that G s1 ; k1 ¼ G s2 ; k2 ¼ G , we know optimal sequences for all problems where the number of products
can be expressed as n = s⁄ + h(s⁄ + k⁄), which are recurrent optimal
Cðrn1 Þ ¼ C r0n1 sequences. When the planning horizon is infinite, it is straightfor-
ward to conclude that the optimal sequence is exactly recurrent,
1 1  
¼ n1 ðn1 þ 1Þða þ bÞ þ X 1 X 2 G þ so1 so1  1 a þ c: like {s⁄k⁄s⁄k⁄s⁄ . . .}.
2 2 The following result specifies the general structure of optimal
For n2 products, we compare rn2 , whose cost can be expressed in sequences (i.e., the number of changeovers, the sizes of AB and
term of the cost of r0n1 as BA blocks) for other n, i.e., s⁄ + h (s⁄ + k⁄) < n < s⁄ + (h + 1)(s⁄ + k⁄).
The determination of these optimal intermediate sequences is
1 o 
mainly based on the fact established in Lemma 1(c), which tells
Cðrn2 Þ ¼ C r0n1 þ s  so1 ½ðn1 þ 1Þ þ n2 ða þ bÞ
2 2 us that the sizes of AB and BA blocks in an optimal sequence need
1   to be balanced, i.e., any pair of AB and BA blocks in an optimal se-
þ so2  so1 so1 þ so2  1 a;
2 quence is composed of either v or v  1 products. Let hv denote the
      
with r0n2 ¼ so1 ko1 . . . so1 ko1 so1 þ 1 ko1 . . . so1 þ 1 ko1 so1 þ 1 , which has number of pairs with v products. For a given number of change-
  o   o
overs, this balance of block sizes must be maintained when the
X 2 þ 1  so2  s1 o
ðABÞs1 blocks, so2  s1 ðABÞs1 þ1 blocks, and X2
o

ko1
number of products is increased or decreased.
ðBAÞ blocks, so
1 o    Theorem 3. Define N = s⁄ + h(s⁄ + k⁄), where h = 0, 1, . . . . There exist
C r0n2 ¼ Cðrn1 Þ þ s  so1 ½ðn1 þ 1Þ þ n2 ða þ bÞ þ so2  so1 so1 a: two thresholds n1 and n2 satisfying N 6 n1 6 n2 6 N + (s⁄ + k⁄), so that
2 2

Since Cðrn1 Þ ¼ C r0n1 , comparison of these two expressions and (a) The optimal number of changeovers for n products, b⁄(n), is
the fact that so1 < so2 imply C r0n2 6 Cðrn2 Þ. Hence, if rn2 is optimal, determined as follows:
then r0n2 is also optimal, so that so2 ¼ so1 þ 1 and n2 = n1 + 1. 8
< 2h þ 1; if N < n 6 n1
>
We now show that rn1 is also optimal. If not, by Lemma 3 and
b ðnÞ ¼ 2h þ 2; if n1 < n < n2
Theorem 1, the optimality of r0n2 and the fact that >
: 
2h þ 3; if n2 6 n < N þ ðs þ k Þ
b r0n2 ¼ bðrn1 Þ ¼ 2X 2 þ 1 imply that there exists an optimal
(b) The optimal sequence for n products is characterized in Table 1,
sequence rn1 such that either bðrn1 Þ ¼ 2X 2 or b rn1 ¼ 2X 2  1, where the parameters v, s, and k satisfy (v  s  k)(v  s 
since n1 = n2  1. We can infer from Lemma 1(c) that if k  1) = 0.
o
bðrn1 Þ ¼ 2X 2 , then rn1 must have either so1 ðABÞs2 blocks or
o
k1 þ1
so1 ðBAÞ blocks, and if b rn1 ¼ 2X 2  1, then rn1 must have
o o Proof. The existence of n1 and n2 is guaranteed by Theorem 1 and
either X 1 ðABÞs2 blocks or X 1 ðBAÞk1 þ1 blocks. However, in both Lemma 3, so we focus on proving part (b). According to Lemma
o
cases, there exists at least one ðABÞs1 block. Increasing its size to so2 , 1(c), any pair of AB block and BA block in the optimal sequence
the resulting sequence has n2 products, and the corresponding cost can make either v or v  1 number of products.
 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83 79

Table 1
General structure of an optimal sequence for n – s⁄ + h(k⁄ + s⁄).

Parameters Blocks
Type (AB) (BA) (AB) (BA)
n v s k hv Size s vs v1k k
j k
N < n 6 n1 ðnþh1ÞðaþbÞþ2b (n  s)  h(v  1) hv + 1 hv h  hv h  hv
hðaþbÞþb
l m l m l m
n1 < n < n2 n ðv 1Þb ðv 2Þa n  (h + 1)(v  1) Number hv hv h + 1  hv h + 1  hv
hþ1 aþb aþb
j k
⁄ ⁄
n2 6 n < N + (s + k ) ðnþhÞðaþbÞþ2b (n  s)  (h + 1)(v  1) hv + 1 hv h + 1  hv h + 1  hv
ðhþ1ÞðaþbÞþb

Suppose there are v products generated by a given pair of AB Theorem 3 characterizes the structure of optimal sequences
block and BA block, and denote s as the optimal size of the AB block, when n – s⁄ + h (k⁄ + s⁄). Note that although n1 and n2 are not explic-
then s must minimize 12 sðs  1Þa þ 12 ðv  sÞðv  s  1Þb. Straight- itly specified, the optimal intermediate sequence could be easily
l m obtained by picking the sequence of the three candidate sequences
forward algebra shows s ¼ ðvaþb
1Þb
, and the optimal size of the BA
in Table 1 that has the smallest objective value. Furthermore, when
block is the corresponding recurring sequence is significantly long (i.e., h is
    large), the size of any AB (BA) block in the optimal intermediate se-
ðv  1Þb ðv  1Þa ðv  1Þa
v s¼ v  ¼ þ1¼ : quence is at most 1 unit different from s⁄(k⁄). This property enables
aþb aþb aþb
us not only to derive the closed-form expressions for n1 and n2, but
Similarly, suppose there are v  1 products generated, and also to convert them into the extra unit thresholds x1 and x2 estab-
denote k as the optimal size of the BA block, then we get lished in Proposition 1, which are independent of h. (Expressions for
l m
k ¼ ðvaþb
2Þa
, and the optimal size of the AB block is these thresholds are derived in the proof of the result.)
l m
v  1  k ¼ ðvaþb
2Þb
. Proposition 1. If n = s⁄ + h(k⁄ + s⁄) + x, where h P (k⁄ + s⁄),
Note that it is easy to check that either s = v  1  k or k = v  s. 0 < x < k⁄ + s⁄, an optimal sequence for the problem is characterized
(a) When N < n 6 n1, b⁄(n) = 2h + 1. There are h + 1 number of AB in Table 2, based on the two thresholds x1 and x2, which satisfy
blocks and h number of BA blocks. It is easy to see that v can be 0 < x1 6 x2 < k⁄ + s⁄.
  l m
determined by either ns h ¼ v or nðv 1kÞ
h ¼ v . Note that
Proof. The existence of x1 and x2 is guaranteed by Theorem 1 and
l n  sm  
ns ðv  1Þb Lemma 3. Since h P s⁄ + k⁄, for n = s⁄ + h(s⁄ + k⁄) + x products, we
¼v )v 1< 6 v ) n  hv 6 need to compare the following three sequences:
h h aþb
 
b
< n  hðv  1Þ ) n  1  h < ðv  1Þ h þ 6n1 (i) The best b-changeover sequence r0n . By Lemma 1(c), the
aþb
extra x products must be evenly distributed to either x AB
ðn  1Þða þ bÞ þ b ðn  1Þða þ bÞ blocks or x BA blocks. Specifically, (2) indicates that
) <v 6 þ 1;
hða þ bÞ þ b hða þ bÞ þ b  
s a 6 k b ) increase AB block’s size from s to s þ 1
  
    s a P k b ) increase BA block’s size from k to k þ 1
n  ðv  1  kÞ ðv  2Þb
¼ v ) n  hv 6 < n  hðv  1Þ ð4Þ
h aþb
 
b (ii) The best b + 1-changeover sequence r Still by Lemma 1(c),
1
n.
) n  1  2h < ðv  2Þ h þ 6n1h  
n0 = (h + 1)(s⁄ + k⁄) implies r1n0 ¼ fs k . . . s k g. We can derive
aþb ⁄ ⁄
rn from rn0 . If x P k , distribute x  k products to either
1 1
ðn  1Þða þ bÞ þ 2b
) <v x  k⁄ AB blocks or x  k⁄ BA blocks, following (4). If x 6 k⁄,
hða þ bÞ þ b
remove one product each from either k⁄  x AB blocks or
ðn  1Þða þ bÞ þ b k⁄  x BA blocks, according to
6 þ 1:
hða þ bÞ þ b    
ðs  1Þa 6 ðk  1Þb ) decrease BA block’s size from k to k  1
ns l m 
nðv 1kÞ ðs  1Þa P ðk  1Þb ) decrease AB block’s size from s to s  1
Hence, from either ¼ v or ¼ v , we conclude
h h j k ð5Þ
ðn1ÞðaþbÞþb ðn1ÞðaþbÞþb
hðaþbÞþb
<v 6 hðaþbÞþb
þ 1, which implies v ¼ ðn1ÞðaþbÞþb
hðaþbÞþb
þ1 .
(iii) The best b + 2-changeover sequence r Based on the opti-
2
n.
Furthermore, after excluding an (AB)s block, the remaining n  s
mal sequence for s⁄ + (h + 1) (s⁄ + k⁄) products, Lemma 1(c)
products form h pairs of AB and BA blocks. Denote hv as the number
implies that we remove (s⁄ + k⁄)  x products from either
of pairs which produce v units, then it is straightforward to see that
(s⁄ + k⁄)  x AB blocks or from (s⁄ + k⁄)  x BA blocks, accord-
hv = (n  s)  h(v  1), and there are h  hv pairs generating v  1
ing to (5).
units. Hence, the number of (AB)s is hv + 1, and the number of (BA)vs
is hv, the number of (AB)v1k is h  hv, the number of (BA)k is h  hv.
We summarize the above discussion on how to add and remove
(b) When n1 < n < n2, b⁄(n) = 2h + 2. There are h + 1 number of AB
units from a block into three cases: (s⁄  1)a 6 (k⁄  1)b 6 s⁄a 6 k⁄
blocks and h + 1 number of BA blocks. It is intuitive to know that
l m b, (k⁄  1)b 6 (s⁄  1)a 6 s⁄a 6 k⁄b, and (k⁄  1)b 6 (s⁄  1)a 6 k⁄
v ¼ hþ1 n
and hv = n  (h + 1)(v  1). The numbers of different AB b 6 s⁄a (the fourth case (s⁄  1)b 6 (k⁄  1) a 6 k⁄b 6 s⁄a is impos-
and BA blocks must be the ones shown in Table 1. sible due to a 6 b). We show how x1 and x2 can be characterized for
(c) When n2 6 n < N + (s⁄ + k⁄), b⁄(n) = 2 h + 3. There are h + 2 the first of the three cases, where (s⁄  1)a 6 (k⁄  1)b 6 s⁄a 6 k⁄b.
number of AB blocks and h + 1 number of BA blocks. Follow the (The other two cases are omitted for conciseness.) Denote
 
logic used in (a), we can verify the desired results. h C i ðxÞ ¼ C rin , i = 0, 1, 2. Based on (2), (1), (4) and (5), we have
80 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83

Table 2
Explicit characterizations of an optimal sequence for larger n = s⁄ + h(s⁄ + k⁄) + x, when h P (s⁄ + k⁄).
 1   þ1   
ðABÞs ðABÞs ðABÞs ðBAÞk 1
ðBAÞk ðBAÞk þ1

⁄ ⁄ ⁄ ⁄
If (s  1)a 6 (k  1)b 6 s a 6 k b
0 < x 6 x1 h+1x x h
x1 < x < x2, x < k⁄ h+1 k⁄  x h + 1  (k⁄  x)
x1 < x < x2, x P k⁄ h + 1  (x  k⁄) x  k⁄ h+1
x2 6 x < s⁄ + k⁄ h+2 s⁄ + k⁄  x h + 1  (s⁄ + k⁄  x)
⁄ ⁄ ⁄ ⁄
If (k  1)b 6 (s  1)a 6 s a 6 k b
0 < x 6 x1 h+1x x h
x1 < x < x2, x < k⁄ k⁄  x h + 1  (k⁄  x) h+1
x1 < x < x2, x P k⁄ h + 1  (x  k⁄) xk ⁄
h+1
x2 6 x < s⁄ + k⁄ s⁄ + k⁄  x h + 2  (s⁄ + k⁄  x) h+1
If (k⁄  1)b 6 (s⁄  1)a 6 k⁄b 6 s⁄a
0 < x 6 x1 h+1 hx x
x1 < x < x2, x < k⁄ k⁄  x h + 1  (k⁄  x) h+1
x1 < x < x2, x P k⁄ h+1 h + 1  (x  k⁄) x  k⁄
x2 6 x < s⁄ + k⁄ s⁄ + k⁄  x h + 2  (s⁄ + k⁄  x) h+1

( k ðk 1Þb  
 c þ ½s a  ðk  1Þbx if x 6 k practical implications. First, these results are intuitive and easy to
2
C 0 ðxÞ  C 1 ðxÞ ¼ k ðk 1Þb implement by production managers. Second, for large problems
  
s k ac 2
if x P k
the optimal sequence is robust with regard to demand changes.
8   s ðs 1Þa 
For example, imagine a production manager who has to decide a
< s ðk  1Þb  c  2
> if x 6 k production sequence based on early demand information, which
  
C 1 ðxÞ  C 2 ðxÞ ¼ ðs þ k Þðk  1Þb  ½ðk  1Þb  s ax then is updated after production is initiated due to, for example,
>
: 
 
c  s ðs þ2k 1Þa
if x P k
 the arrival of new orders or the cancelation of existing orders.
2
Our results imply that, with the exception of its ending portion
where minor adjustments might be required, the original sequence
 
 ð2s þ k Þðk  1Þb remains optimal even under the finalized demand information.
C 0 ðxÞ  C 2 ðxÞ ¼ ½s a  ðk  1Þbx þ
2 Third, in situations where the machine is required to be continu-
s ðs  1Þa ously running without any interruption for a long time (due to
  2c
2 technical constraints or in order to meet predictably high future
demand), the optimal production policy that minimizes the
Checking the first derivative with respect to x for the above expres-
weighted sum of flow time and changeover cost is a recurrent se-
sions, and noticing that C0(0) = min{C0(0), C1(0), C2(0)} and C2(k⁄ +
quence with constant batch sizes.
s⁄) = min{C0(k⁄ + s⁄), C1(k⁄ + s⁄), C2(k⁄ + s⁄)}, we determine x1 and x2
as follows: If C2(0)  C1(0) 6 0, then C1(x) P C2(x) for all x.
Comparing C0(x) and C2(x), ½C 0 ðxÞ  C 2 ðxÞ0x P 0 implies Appendix A. Additional technical results with proofs
 ðs 1Þa=2k ðk 1Þb=2s ðk 1Þb
x1 ¼ x2 ¼ 2cþs  
s aðk 1Þb
. If C2(0)  C1(0) > 0 then
x 6 k⁄ implies C2(x) > C1(x). Moreover, if C1(k⁄)  C0(k⁄) P 0, then Lemma 1. For n products, there exists an optimal sequence rn
x P k⁄ implies C1(x) P C0(x) and C2(k⁄) > C1(k⁄) P C0(k⁄). Hence, which has the following properties:
due to ½C 0 ðxÞ  C 1 ðxÞ0x P 0, comparing C0(x) and C2(x), we get
2cþs ðs 1Þa=2k ðk 1Þb=2s ðk 1Þb (a) If any two AB (or BA) blocks are exchanged,
x1 ¼ x2 ¼ s aðk 1Þb
. On the other hand, if C1(k⁄)   the resulting
⁄ ⁄
< C0(k ), then x < k implies that we need to compare C0(x) and sequence is still optimal. Moreover, if b rn is even, then
C1(x) and x P k⁄ implies that we need to compare C1(x) and C2(x). switching all AB blocks with all BA blocks will result in an opti-
   ðs 1Þa=2s ðk 1Þb mal sequence.
Thus, x1 ¼ ck ðk 1Þb=2
, x2 ¼ cþs þ k. h    
s aðk 1Þb  
s aðk 1Þb (b) If a 6 b, and there are m AB blocks, then m ¼ b rn =2 . That
is, the number of AB blocks is never less than the number of
BA blocks.
5. Conclusion
(c) The sizes of the different types of blocks must satisfy
(s  1)a 6 kb and sa P (k  1)b, the sizes of the same type of
We study the problem of scheduling identical products with
blocks differ at most by 1, and there are at most 3 different
components on a single machine where changeovers incur fixed
block sizes.
costs. The objective is to minimize the weighted sum of total flow
time and changeover cost. We prove several structural results
which ultimately lead to an explicit characterization of optimal se-
quences for given problem parameters. We show that the optimal Proof.
sequences are composed of constant batch sizes (i.e., sizes of inter-
Pn
mediate batches differ at most by one), except for the beginning (a) For any rn, from (1), i¼1 F i ðrn Þ depends on neither the
and the ending batches. We also show that there exists a series sequence of the AB blocks, nor the sequence of the BA blocks.
of recurrent optimal sequences, and we obtain closed-form expres- If b(rn) is even, then switching all of the AB blocks with all of
P
sions for the integer batch sizes in these recurrent optimal the BA blocks leaves ni¼1 F i ðrn Þ unchanged.
sequences. (b) For any rn, each BA block must follow an AB block, so that
The existence of recurrent optimal sequences as well as the con- jbo  2mj 6 1. If bo  2m = 1 and a 6 b, then that sequence
stant-batch-size property in the optimal sequence have important begins and ends with a BA block, and by (1), its flow time
 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83 81

does not increase when changing the last BA block to an AB
block, and then interchanging all of the remaining BA blocks
with the AB blocks.
(c) Let ron denote a minimum flow-time sequence with bo
changeovers, and in ron , consider two different types of
 
Proof. We first prove the monotonicity of b rn in n. Based on
Lemma 1(c), w.l.o.g., we assume that a rn has m0 (AB)s blocks, m00
 
(AB)s1 blocks, and b rn  ðm0 þ m00 Þ (BA)k blocks. Denote ronþ1 as
 
the best sequence for n + 1 products with b rn changeovers.
rnþ1 can be constructed from rn by adding the extra product to
o 

blocks, (AB)s and (BA)k. Because ron is flow-time minimal, either an AB block or a BA block. If m00 = 0, the extra product goes
(s, k) must minimize 12 sðs  1Þa þ 12 kðk  1Þb, which requires to either an (AB)s block when sa 6 kb or a (BA)k block when sa > kb;
sa P (k  1)b and kb P (s  1)a. Suppose there are ns prod- otherwise, by Lemma 1(c), the extra product must go to an (AB)s1
ucts in the AB blocks and nk = n  ns products in the BA block, implying (s  1)a 6 kb.
blocks. It is easy to show, through induction on m, that the We now prove the result by contradiction. Take the case where
minimization problem        
m00 > 0. Suppose b rnþ1 < b rn , then C rnþ1 < C ronþ1 . Let m⁄

denote the total number of AB blocks in sequence rnþ1 ; s the
X
m
1
min sp ðsp  1Þa largest size of all AB blocks, and k⁄ the largest size of all BA blocks.
s1 þ...þsm ¼ns 2  
p¼1 Since Lemma 1(b) shows that m⁄ 6 m0 + m00 and b rnþ1  m 6
 
b rn  ðm0 þ m00 Þ, we can infer from (2) that either s⁄ P s or
is solved by sp = dns/me, for p = 1, . . . , m0 , and sp = bns/mc, for k⁄ P k. (i) If s⁄ P s, we can remove one product from an ðABÞs block
p = m0 + 1, . . . , m, where m0 = (ns mod m). Hence, the sizes of the AB in sequence rnþ1 . Denote the resulting sequence as ron , then (2),
blocks can differ at most by 1. The same logic holds for the BA blocks,        
and (1), and s⁄ P s imply C rnþ1  C ron P C ronþ1  C rn )
so there are at most 4 different block sizes. Assume, w.l.o.g., these  o  
C rn < C rn , which is a contradiction to the assumption of an
block sizes are s and s  1 for the AB blocks and k and k  1 for the
optimal sequence. (ii) If s⁄ < s, then k⁄ > k. Remove one product
BA blocks, where s – k. Since ron is the best bo-changeover sequence,
it must have the smallest total flow time, so (2) and (1) imply that from a ðBAÞk block in rnþ1 to get ron . Then, we have
       
1
sðs  1Þa þ 12 ðk  1Þðk  2Þb ¼ 12 ðs  1Þðs  2Þa þ 12 kðk  1Þb. Thus, C rnþ1  C ron  C ronþ1  C rn ¼ ðk  1Þb  ðs  1Þa < 0,
2
any transformation between (AB)s(BA)k1 blocks and (AB)s1(BA)k which is a contradiction to m00 > 0 ) (s  1)a < kb ) (k⁄  1)b 
blocks keeps the sequence’s flow-time optimality, so we can always (s  1)a > (k⁄  1)b  kb P 0. The argument for the case where
construct an optimal bo-changeover sequence which has at most 3 m00 = 0 follows similar logic.
 
different block sizes. h We next show that limn!1 b rn ¼ 1. Suppose that there exists
   
an N such that n P N ) b rn ¼ b rN ¼ bN , and that there are mN
AB blocks and lN BA blocks in rN and rn . Let sp ðnÞ denote the size of
the pth AB block in rn and let kq ðnÞ denote the size of the qth BA
  an optimal sequence rn has m AB blocks, and con-

Lemma 2. Suppose

sider mo 6 b rn  m. Then we have the following properties: block. It is true that limn!1 sp ðnÞ ¼ 1 and limn!1 kq ðnÞ ¼ 1.
Denote r0n as the best (bN + 2)-changeover sequence, then
Pn    Pn  0
(a) After removing any mo AB blocks and any mo BA blocks and i¼1 F i rn  i¼1 F i rn < 2c must hold for all n P N.
resorting the remaining blocks such that the AB blocks and For any n0 P N, we can create an alternative (bN + 2)-changeover
the BA blocks are alternating (to ensure normal form), the sequence r00n for n(> n0 ) products from the best (bN + 2)-changeover
resulting sequence is optimal for the remaining number of sequence for n0 -product r0n0 as follows. (i) The sizes of the last AB
products. block and the last BA block are the same as in sequence r0n0 , i.e.,
 
(b) If b rn ¼ 2m, then after removing any (mo  1) AB blocks and 00 0
s00mN þ1 ðnÞ ¼ s0mN þ1 ðn0 Þ and klN þ1 ðnÞ ¼ klN þ1 ðn0 Þ. (ii) Since both rn and
mo BA blocks and resorting the remaining blocks such that the 
AB blocks and the BA blocks are alternating, the resulting rn0 have bN = mN + lN changeovers, we can adjust the sizes of the first
sequence is optimal for the remaining number of products. mN AB blocks and the first lN BA blocks according to the difference
between rn and rn0 . That is, s00p ðnÞ  s0p ðn0 Þ ¼ sp ðnÞ  sp ðn0 Þ and
00 0  
kq ðnÞ  kq ðn0 Þ ¼ kq ðnÞ  kq ðn0 Þ. Obviously, we must have
00
Proof. It suffices to prove that the conclusion holds for mo = 1. limn!1 s00p ðnÞ
¼ 1 and limn!1 kq ðnÞ
¼ 1, implying there exists an
n o
00 00
n00 > n0 such that max kq ðn00 Þjq ¼ 1; . . . ; lN P klN þ1 ðn00 Þ þ 1 þ 2c
b
.
(a) According to Lemma 1(a), we can focus on removing the Then, according to Lemma 1(c), we can improve sequence r00n00 by
first two blocks (AB)s(BA)k in an optimal sequence rn . The moving one unit from the largest BA block to the last BA block, and
resulting sequence is ron0 , where n0 = n  (s + k). Lemma 1(a) the resulting sequence’s cost must be at least b + 2c less than the cost
and (b) suggest that there exists a rn0 starting with an Pn00    Pn00  0  Pn00   
of r0000 . That is, i¼1 F i rn00  i¼1 F i rn00 > i¼1 F i rn00
AB block. Correspondingly, putting (AB)s(BA)k in front of hP 00 n   i
n 00
rn0 , we get a ron . Clearly, from (2) and (1), we have  i¼1 F i rn00  b  2c > 2c. This is a contradiction to the previous
 o     P   P  
C rn  C rn0 ¼ C ron  C rn0 , implying Cðron0 Þ 6 Cðrn0 Þ,

conclusion that ni¼1 F i rn  ni¼1 F i r0n < 2c holds for all n P N.
so that ron0 must also be optimal.
  Appendix B. Extension to products with n components
(b) If b rn ¼ 2m, then Lemma 1(a) and (b) suggest that we can
always make the optimal sequence start with a (BA)k block.
Then, removing this initial block and applying the logic in Suppose A1, A2,    , Am are the m components of a product and
(a), we can directly verify the conclusion. h their corresponding processing times satisfy a1 6 a2 6    6 am.
Similar to the 2-component product setting, we use
(AuAu+1    Av)n to represent a sequence that produces Au-type to
  Av-type components ascending in subscripts and each component
Lemma 3. If rn is an optimal sequence with b rn changeovers, then type by n units, where 1 6 u 6 v 6 m.
   
there exists an optimal sequence rnþ1 such that b rnþ1 P b rn . In

For expositional simplicity, we consider 2m types of blocks, de-
 
addition, limn!1 b rn ¼ 1. noted as (Am1, Am)u and (Am, Am1)u, u = 0, 1,    , m  1. (Am1, Am)u
82 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83

type (or (Am, Am1)u type) blocks have the following features: (1) treating ðRm1
v ¼uþ1 av Þ as one consolidated component’s processing
they start with Am 1-type (or Am-type) component; (2) all Am-type
time and Rm
v ¼uþ1;v –m1 av as another’s, we can apply a logic similar
(or Am 1-type) components are produced in a single batch; (3) an-
other u types of single-batch produced components are the first u to the one used for the 2-component product to derive the optimal s
ones in the list excluding Am-type (or Am1-type); and (4) the and k as
remaining (m  u  1) types of components are produced in two 2 3
batches which are located both before and after Am-type (or Am ðx  1Þ Rm
v ¼uþ1;v –m1 av
s ¼ 6
6 m1
7and
7
1-type) batch. The notation below
6 Rv ¼uþ1 av þ Rm v ¼uþ1;v –m1 av 7
 x
Asm1 ; Am ¼ ðAm1 Þs ðA1    Au Þx ðAuþ1    Am2 Þs ðAm Þs 2 m1
3
u
ðx  1Þ Rv ¼uþ1 av
ðAm Þxs ðAuþ1    Am2 Þxs ðAm1 Þxs ; ¼6
k 6 m1
7:
7
x 6 Rv ¼uþ1 av þ Rm v ¼uþ1;v –m1 av 7
ðAkm ; Am1 Þu k x k
¼ ðAm Þ ðA1    Au Þ ðAuþ1    Am2 Þ ðAm1 Þ k

 
ðAm1 Þxk ðAuþ1    Am2 Þxk ðAm Þxk ; Hence, we have F xu Asm1 ; Am ¼ F xu Akm ; Am1 .

indicates how x units of products are produced respectively by The Lemma below is established to show that, similar to 2-com-
(Am1, Am)u and (Am, Am1)u type blocks, where the superscript x rep- ponent product problem, the optimal sequence for an m-compo-
resents the block size and the superscripts s and k the batch sizes of nent product is also built on some basic blocks according certain
a corresponding block type. We next explore the important proper- rules.
ties of these block types.
First, block types (Am1, Am)u and (Am, Am1)u have the change- Lemma 4. For any n units of m-component products, there exists an
overs of (2m  2  u). Hence, corresponding to u = 0, 1,    , m  1, optimal sequence which has the following properties:
we have the sequences with the changeovers from (2m  2) to
(m  1). If u < m  1, the starting component type is the same as 1. It starts with an Am1-type component.
the ending one. If u = m  1, then s = k = x. That is, 2. Its blocks must be in the type of (Am1, Am)u, u = 0, 1,    , m  1.
 x x  x n 3. The block sizes of (Am1, Am)u and (Am1, Am)v differ at most by 1,
Am1 ; Am m1 ¼ ðAm1 Þx ðA1    Am2 Þx ðAm Þx and Am ; Am1 m1 ¼
where 0 6 u 6 v < m  1.
ðAm Þx ðA1    Am2 Þx ðAm1 Þx . The ending component is different 4. The batch sizes of (Am1, Am)u and (Am1, Am)v differ at most by 1,
from the starting one. It is interesting to notice that block type where 0 6 u 6 v 6 m  1.
(Am1, Am)0 is the combination of the block types (Am1, Am)m1 5. The sequence remains optimal if removing any (Am1, Am)u type
and (Am, Am1)m 1, and that (Am, Am1)0 the combination of block, where 0 6 u < m  1.
(Am, Am 1)m1 and (Am1, Am)m1.
Second, the total flow time of producing x units by block type.
(Am1,Am)m1 or (Am,Am1)m1 includes both basic flow time
1
xðx þ 1ÞðRv av Þ and the flow time penalty Proof. Consider an optimal schedule with a given number of
2
changeovers. The least possible number of changeovers is
  1
F xm1 Anm1 ; Am ¼ xðx  1ÞðRv –m av Þ; or (m  1), if producing all components of the same type in one single
2  n
batch. It is easy to check that Anm1 ; Am m1 must be optimal for
x
 n  1  n
F m1 Am ; Am1 ¼ xðx  1ÞðRv –m1 av Þ; (m  1) changeovers. Anm1 ; Am m1 starts with an Am1-type com-
2
ponent and ends with an Am-type component. By breaking the sin-
which bears similarity to the flow time calculation for the 2-compo- gle batch into two and moving one batch to the end, the resulting
nent product, if we regard (Rv–mav) and (Rv–m 1av) as the process- sequence has one more changeover. There are at most (m  1) sin-
ing time of the consolidated components. It is easy to see that gle batches to break, since at least one single batch must remain to
   
F xm1 Anm1 ; Am 6 F xm1 Anm ; Am1 . In fact, it is most cost efficient to bridge those two batches of the same type components. As a result,
consolidate all component types, except for the Am type. we can get sequences with m to (2m  2) changeovers. Straightfor-
Third, for u = 0, 1,    , m  2, the total flow time of using an  n
ward algebra verifies that both Asm1 ; Am and
(Am1, Am)u type (or (Am, Am1)u type) block to produce x units with 2m2b
 n
a batch size of s (or k) includes again the basic flow time Akm ; Am1 are optimal for the changeovers
1 2m2b
2
xðx þ 1ÞðRav Þ and the flow time penalty
b = m,m + 1,    , 2m  2, where
  1   1
F xu Asm1 ; Am ¼ xðx  1Þ Ruv ¼1 av þ sðs  1Þ Rm1
v ¼uþ1 av 2 3
2 2 ðn  1Þ Rm
v ¼2m1b;v –m1 av
1 s ¼ 6 7and
þ ðx  sÞðx  s  1Þ Rm v ¼uþ1;v –m1 av ; or 6 m1 7
2 6 Rv ¼2m1b av þ Rm v ¼2m1b;v –m1 av 7
1   1 2 3
F xu ðAkm ; Am1 Þ ¼ xðx  1Þ Ruv ¼1 av þ ðx  kÞðx  k  1Þ Rm1
v ¼uþ1 av m1
ðn  1Þ Rv ¼2m1b av
2 2 ¼6 7:
1 k 6 m1 7
þ kðk  1Þ Rm v ¼uþ1;v –m1 av :
6 Rv ¼2m1b av þ Rm v ¼2m1b;v –m1 av 7
2
    Note that the starting component type is the same as the ending one.
Hence, we have F xu Asm1 ; Am ¼ F xu Axs
m ; Am1 . That is, the blocks  n  n
 s x  xs x For Asm1 ; Am , both are of Am1-type, and for Akm ; Am1 ,
Am1 ; Am u and Am ; Am1 u are equivalent in terms of total flow 2m2b 2m2b

time. This implies that the two can be exchanged when necessary. both are of Am-type. For more than 2m  2 changeovers, it is clear that
 x  s the optimal sequence can be composed by (Am1,Am)u blocks, where
Recall that Asm1 ; Am 0 can be decomposed to Asm1 ; Am m1 and
 xs xs  s x  s x u = 0, 1,    , m  1. Hence, Results 1 and 2 follow.
Am ; Am1 m1 , so the sequence Am1 ; Am 0 Am1 ; Am u must also
 s s  x In the optimal sequence with (Am1, Am)u blocks
be equivalent to the sequence Am1 ; Am m1 Axs
m ; Am1 u (u = 0, 1,    , m  1), there is at most one (Am1, Am)m1 block,
 xs xs
Am ; Am1 m1 in terms of total flow time. More interestingly, by which, if it exists, must be at the end of the sequence. The order of
 F. Zhou et al. / European Journal of Operational Research 233 (2014) 75–83 83

the other (Am1, Am)u blocks (u = 0, 1,    , m  2) does not matter. References

Without loss the generality, we suppose existence of an (Am1,
Am)m1 block, as well as of another two blocks (Am1, Am)u and Allahverdi, A., Ng, C., Cheng, T., & Kovalyov, M. (2008). A survey of scheduling
problems with setup times or costs. European Journal of Operational Research,
(Am1, Am)w with 0 6 u 6 w < m  1. Suppose further that the block 187(3), 985–1032.
sizes for (Am1, Am)m1, (Am1, Am)u and (Am1, Am)w are x, y, and z, Baker, K. (1988). Scheduling the production of components at a common facility. IIE
respectively. Then the batch sizes are x, sy, and sz. It is easy to Transactions, 20(1), 32–35.
Cheng, T., & Kovalyov, M. (2001). Single supplier scheduling for multiple deliveries.
calculate that the flow time penalty for the last y  sy number of Annals of Operations Research, 107(1–4), 51–63.
 sy y
products in Am1 ; Am u is 12 ðy  sy Þðy  sy  1Þ Rm
v ¼1;v –m1 av , and
Cheng, T., Kovalyov, M., & Chakhlevich, K. (2004). Batching in a two-stage flowshop
 z z with dedicated machines in the second stage. IIE Transactions, 36(1), 87–93.
the flow time penalty for the last z  sz products in Asm1 ; Am u is Cheng, T., Ng, C., & Yuan, J. (2008). Single-machine scheduling of multi-operation
jobs without missing operations to minimize the total completion time.
1
2 ðz  sz Þðz  sz  1Þ Rm
v ¼1;v –m1 av . Hence, optimality requires European Journal of Operational Research, 191(2), 320–331.
Cheng, T., & Wang, G. (1999). Scheduling the fabrication and assembly of
that y  sy and z  sz differ at most by 1. On the other hand, an components in a two-machine flowshop. IIE Transactions, 31(2), 135–143.
equivalent optimal sequence can be created by converting Coffman, E., Nozari, A., & Yannakakis, M. (1989). Optimal scheduling of products
 sy y z  y  zsz z with two subassemblies on a single machine. Operations Research, 37(3),
Am1 ; Am u and ðAsm1 z
; Am Þw to Ays
m ; Am1 u and Am ; Am1 w ,
y
426–436.
 x x
and moving them after Am1 ; Am m1 . Then, by calculating the flow Coffman, E., & Yannakakis, M. (1990). Batch sizing and job sequencing on a single
 x machine. Annals of Operations Research, 26(1), 135–147.
time penalty for the products in Axm1 ; Am m1 , for the last sy Dobson, G., Karmarkar, U., & Rummel, J. (1987). Batching to minimize flow times on
 ysy y one machine. Management Science, 33(6), 784–799.
products in Am ; Am1 u and for the last sz products in Gerodimos, A., Glass, C., & Potts, C. (2000). Scheduling the production of two-
 zsz z
Am ; Am1 w , we can similarly get that the difference between x, component jobs on a single machine. European Journal of Operational Research,
120(2), 250–259.
sy, and sz is at most 1. Correspondingly, we can conclude Results 3 Gim, B., & Han, M.-H. (1997). Economic scheduling of products with N components
and 4. on a single machine. European Journal of Operational Research, 96(3), 570–577.
Finally, Result 5 can be argued similarly as in 2-component Kella, O. (1991). Optimal manufacturing of a two product mix. Operations Research,
39(3), 496–501.
product setting. h Kovalyov, M., Potts, C., & Strusevich, V. (2004). Batching decisions for assembly
production systems. European Journal of Operational Research, 157(3), 620–642.
Based on the above lemma, we can show our main conclusion Lin, B. (2002). Fabrication scheduling on a single machine with due date constraints.
European Journal of Operational Research, 136(1), 95–105.
below. The proof for this result closely follows the logic presented Lin, B., & Cheng, T. (2005). Two-machine flowshop batching and scheduling. Annals
for the 2-component product case and thus is omitted. of Operations Research, 133(1-4), 149–161.
Lin, B., & Cheng, T. (2011). Scheduling with centralized and decentralized batching
policies in concurrent open shops. Naval Research Logistics, 58(1), 17–27.
Theorem 4. There exists a vector of integers (s⁄, k⁄, u⁄) 2 Magnanti, T., & Vachani, R. (1990). A strong cutting plane algorithm for production
{(s, k, u)ju = 0, 1,    , m  2}, where the pair (s, k) minimizes scheduling with changeover costs. Operations Research, 38(3), 456–473.
Mosheiov, G., & Oron, D. (2008). A single machine batch scheduling problem with
sðs  1ÞðRv –m av Þ þ kðk  1ÞðRv –m1 av Þ bounded batch size. European Journal of Operational Research, 187(3),
Gu ðs; kÞ ¼
2ðs þ kÞ 1069–1079.
Potts, C., & Kovalyov, M. (2000). Scheduling with batching: A review. European
2skðRv 6u av Þ þ 2ð2m  2  uÞc Journal of Operational Research, 120(2), 228–249.
þ ; Santos, C., & Magazine, M. (1985). Batching in single operation manufacturing
2ðs þ kÞ
s þk
systems. Operations Research Letters, 4(3), 99–103.

so that either h Asm1 ; Am blocks make an optimal sequence for Vickson, R., Magazine, M., & Santos, C. (1993). Batching and sequencing of
u components at a single facility. IIE Transactions, 25(2), 65–70.
 s þk
⁄ ⁄
h(s + k ) units of products or h Asm1 ; Am blocks plus an
u
s þk
s
Am1 ; Am block make an optimal sequence for s⁄ + h(s⁄ + k⁄) units
m1
of products, where h = 0, 1, 2,    .