You are on page 1of 17

Applied Mathematical Modelling 48 (2017) 423439

Contents lists available at ScienceDirect

Applied Mathematical Modelling


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

Rules-based heuristic approach for the U-shaped assembly


line balancing problem
Ming Li a,b, Qiuhua Tang b,, Qiaoxian Zheng c, Xuhui Xia c, C.A. Floudas d
a
College of Science, Wuhan University of Science and Technology, Wuhan 430065, Hubei, China
b
College of Mechanics and Automation, Wuhan University of Science and Technology, Wuhan 430081, Hubei, China
c
School of Computer Science and Information Engineering, Hubei University, Wuhan 430062, Hubei, China
d
Department of Chemical and Biological Engineering, Princeton University, Princeton 08544, NJ, USA

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

Article history: The type-2 U-shaped assembly line balancing problem is important for many just-in-time
Received 22 August 2015 manufactures, but an ecient algorithm is not available at present. Thus, in this study, a
Revised 18 September 2016
novel heuristic approach based on multiple rules and an integer programming model is
Accepted 31 December 2016
proposed to address this problem. In the proposed approach, three rules are systematically
Available online 10 January 2017
grouped together, i.e., task selection, task assignment, and task exchange rules. The su-
Keywords: cient conditions for implementing the exchange rules are proposed and proved. Thirteen
Heuristic approach small or medium scale benchmark issues comprising 63 instances were solved, where the
Local search algorithm computational results demonstrate the eciency and effectiveness of the proposed method
Task exchange rule compared with integer programming. The computational results obtained for 18 examples
U-shaped assembly line balancing problem comprising 121 instances demonstrate that the task exchange rules signicantly improve
the computational accuracy compared with the traditional heuristic. Finally, 30 new stan-
dard instances produced by a systematic data generation process were also solved effec-
tively by the proposed approach. The proposed heuristic approach with multiple rules can
provide a theoretical basis for other local search algorithms, especially for addressing is-
sues such as the U-Shaped assembly line balancing problem.
2017 Published by Elsevier Inc.

1. Introduction

An assembly line is a manufacturing system where components are assembled consecutively on unidirectional stations
until the nal products are completed. These stations are connected via a conveyor belt. Each station repeatedly performs a
subset of tasks within a limited time period, which is denoted as the cycle time. The tasks are performed within a certain
time period called the task time. The sum of the task times assigned to a station is called the station time. Partial precedence
constraints, represented by a precedence graph, are introduced to specify the sequence of tasks. In the example shown in
Fig. 1, 10 assembly tasks are involved, where the numbers over the circles represent the task times and the directed arcs
represent the assembly precedence constraints.
Assembly lines can be classied into two general groups: traditional straight assembly lines and U-shaped assembly lines.
In the traditional straight assembly line, all the operators work along the unidirectional line. However, a U-shaped assembly
line is divided into two sub-lines: the entrance sub-line and exit sub-line, and thus an operator may either perform tasks on


Correspondence author.
E-mail addresses: tangqiuhua@wust.edu.cn, 8045872@qq.com (Q. Tang).

http://dx.doi.org/10.1016/j.apm.2016.12.031
0307-904X/ 2017 Published by Elsevier Inc.
424 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

2 5 4
2 4 7
2 3 1
4 3 6 6 8 9 10
1 3 5 6

Fig. 1. Precedence graph.

Fig. 2. Task assignments on two types of assembly lines under the same cycle time.

Fig. 3. The optimal solutions of UALBP-1 and UALBP-2.

one of the two sub-lines or on both sub-lines simultaneously. Due to the precedence constraints among tasks in a U-shaped
assembly line, each task can be implemented either on the entrance sub-line when all its predecessors have been nished,
or on the exit sub-line after all its successors have been complemented. For convenience, a task with fullled predecessors
or successors is called the entrance task or exit task, respectively.
Compared with straight assembly lines, U-shaped lines have many benets such as superior assignment exibility [1],
smaller work space requirements [2], a reduced work-in-process inventory, improved equipment utilization, and lower ma-
terial handling costs [3]. As shown in Fig. 2, the 10 tasks restricted by the precedence relations in Fig. 1 will be performed
on a straight or U-shaped assembly line. If the cycle time is set to 12, four stations are required by the straight line, but
only three stations are needed by the U-shaped line. In addition, task 6 can only be assigned to station III in the straight
line, whereas it can be the entrance task and exit task simultaneously in the U-shaped line. Compared with straight lines,
U-shaped lines can assign more candidate tasks to their stations and obtain a higher balancing eciency [1]. This conclusion
was veried by Toksar et al. [4], Baykasoglu and Dereli [5], zcan and Toklu [6], and Kara et al. [7]. Thus, the exibility of
task assignment can greatly increase the productivity of a U-shaped line.
The problem of assigning all tasks to an ordered sequence of stations on a U-shaped assembly line, where the precedence
constraints are satised and some performance measures are optimized, is called the U-shaped assembly line balancing
problem (UALBP), which was rst proposed and modeled by Miltenburg and Wijngaard in 1994 [8]. The UALBP is classied
into three versions [9]: (a) UALBP-1, which aims to minimize the number of stations within a given cycle time; (b) UALBP-2,
which minimizes the cycle time in a given number of stations; and (c) UALBP-E, which maximizes the line eciency within
a variable cycle time and at a variable number of stations. UALBP-1 applies primarily in the design phase for an assembly
line, UALBP-E is considered to obtain operational improvements in the line when necessary, but UALBP-2 is encountered
frequently when it is necessary to adjust the production rhythm in the manufacturing process.
In addition, UALBP-2 is far more dicult to solve. Denote the cycle time in UALBP-1 and the station number in UALBP-2
for the same instance as C and |J |, respectively. Assuming that the optimal values of the two problems are marked as |J |
and C , then C C because the two problems have the same station number. When C = C, the two problems are equivalent
(as shown in Fig. 3(a)). However, when C < C, each of the stations in UALBP-2 has less space to hold tasks (as shown in
Fig. 3(b)), so a more sophisticated control approach is required to assign tasks, which increases the complexity of solving
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 425

Fig. 4. Precedence graph for ve tasks.

UALBP-2. Hence, in this study, a novel heuristic approach with excellent local search performance is proposed for solving
UALBP-2.
Fig 3 is expected to be appended here and Fig 4 should be moved to the front of Fig 4. (a) (b) Fig. 3 The optimal solutions
of UALBP-1 and UALBP-2.
In the past decade, many variants of UALBP have been proposed to meet production demands, such as the mixed-model
UALBP [1013], stochastic UALBP [1417], multi-objective UALBP [18,19], parallel UALBP [20,21], and other UALBP variants
[4]. These problems are highly complex, but the UALBP is still a challenging issue for researches. Furthermore, its solutions
can be further adapted to more realistic problems.
Many algorithms have been proposed to address the UALBP and its variants, including exact methods [22], heuristics [6],
meta-heuristics [14,11,23], and other methods [24,25]. Among these algorithms, heuristics have the advantage of rapid speed
but the disadvantage of low computational accuracy. However, the quality of their solutions can be enhanced by using local
search techniques. This method is also utilized often for solving SALBP-2 (the type 2 simple assembly line balancing prob-
lem) [2628,6,29]. Among the local search techniques, an effective method involves exchanging the tasks between different
stations, where the key to exchanging tasks is meeting the precedence constraints among them. Compared with SALBP-2 in
straight assembly lines, the conditions for exchanging tasks in U-shaped assembly lines are much more complex and dicult
because there are two sub-lines: an entrance sub-line and exit sub-line, and more conditions must be considered to meet
the precedence constraints. In this study, the conditions dened above are probed and formalized further with mathematical
proofs, thereby providing theoretical support for local search algorithms when solving the UALBP.
The reminder of this paper is organized as follows. Section 2 introduces the mathematical model of UALBP-2.
Section 3 proposes the heuristic approach with multiple rules for UALBP-2. Section 4 presents the results of the compu-
tational studies, and our conclusions are given in Section 5.

2. Problem statement

2.1. Notations

The notations used in the study are given as follows:


I Set of tasks, I = {i|i = 1, 2, , n}.
J Set of stations, J = {j|j = 1, 2, , m}.
ti Processing time for task i, i I.
P Immediate predecessor matrix. For given i1 ,i2 I, if task i1 is the immediate predecessor of task i2 , then Pi2 ,i1 = 1;
otherwise, Pi2 ,i1 = 0.
Q Predecessor matrix. For given i1 ,i2 I, if task i1 is the predecessor of task i2 , then Qi2 ,i1 = 0; otherwise, Qi2 ,i1 = 1.
C Cycle time for the U-shaped assembly line.
nc Iteration time for the heuristic approach, nc = 1, 2, , NC.
1 n
Cm Mean station time, Cm = m i=1 ti .
Wj Set of tasks assigned to station j, j J.
V Set of assigned tasks.
UV Set of unassigned tasks.
STj Station time for station j, j J.
K Set of symbols representing entrance or exit tasks, K = {k|k = 1, 2}. If the task is the entrance task, then k = 1;
otherwise, k = 2.
xijk Binary variables. For given i I, j J, k K, if task i is assigned to stationjas an entrance or exit task, then xijk = 1;
otherwise, xijk = 0.

2.2. Assumptions

As demonstrated by zcan and Toklu [6], UALBP-2 is formalized based on the following assumptions.
Task times are deterministic and independent of the assignment.
The precedence relationships among tasks are known.
Parallel stations and work in a process buffer are not allowed.
The walking times of operators within a station are ignored.
There is no uncertainty in UALBP-2.
426 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

2.3. Mathematical model

The UALBP-2 aims to minimize the cycle time with a given number of stations. The mathematical model of UALBP-2 is
presented as follows:

min C (1)
subject to

ti xi jk C, j = 1, 2, . . . , m, (2)
iI kK

xi jk = 1,i = 1, 2, . . . , n, (3)
kK jJ

 
xi2 j2 1 xi1 j1 1 , if Pi2 ,i1 = 1, (4)
j2 J j1 J

  
j1 xi1 j1 1 j2 xi2 j2 1 +M j3 xi2 j3 2 , if Pi2 ,i1 = 1, (5)
j1 J j2 J j3 J

 
xi1 j1 2 xi2 j2 2 , if Pi2 ,i1 = 1, (6)
j1 J j2 J

  
j2 xi2 j2 2 j1 xi1 j1 2 +M j3 xi1 j3 1 , if Pi2 ,i1 = 1, (7)
j2 J j1 J j3 J

xi jk {0, 1}, i I, j J, k K. (8)


The objective represented in Eq. (1) is minimizing the cycle time. Constraint (2) limits the station time so that each
station time is no more than the cycle time. Constraint (3) ensures that each task is assigned to exactly one station as an
entrance task or an exit task. Constraints (4)(7) are precedence constraints, where task i1 is the immediate predecessor
of task i2 . Constraints (45) ensure that if a task is an entrance task, then its preceding tasks are also entrance tasks, and
the preceding task is assigned to the earlier stations. Similarly, Constraint (6) ensures that if a task is an exit task, then its
succeeding tasks are also exit tasks and the succeeding task is assigned to the earlier stations. Constraint (8) restates the
denitions of binary variables. Note that M is a large constant in Eqs. (5) and (7).

3. Heuristic approach to UALBP-2

In the rules-based heuristic approach, three types of rules are proposed to solve the UALBP-2: task selection, task assign-
ment, and task exchange rules. The task assignment rule is used to balance the station loads by ensuring that the station
times are close to the mean station time. The task exchange rule, which exchanges tasks in the current station with those
in the following stations, is effective for achieving more balanced station loads at these stations. The procedure for the
heuristic approach is presented as follows.
1. Initialize variables, set parameters, and set the index for the current station jj = 1
2. for j=jj to m-1 do
3. Generate the candidate task set CA.
4. Assign tasks to station j according to the task selection rule and task assignment rule.
5. if the average of the assigned station times is less than the mean station time Cm, do
6. Implement the task exchange rule for station j and stop assigning tasks to station j
7. else
8. Stop assigning tasks to station j
9. end if
10. end for
11. Assign the remaining tasks to station m and obtain a feasible solution.
12. Calculate the current best cycle time, Cbest, and update the upper/lower bound of the station time, Cmax\Cmin.
13. for nc = 1 to NC do
14. Search the set of stations for station times that are greater than Cmax and then set jj as the smallest index of these stations.
15. Execute steps 29.
16. if the maximum number of current station times is less than Cbest, do
17. Update Cbest, Cmax, and Cmin.
18. else
19. Go to step 22.
20. end if
21. end for
22. Output the best cycle time and the best solution, and stop.
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 427

3.1. Task selection rule

In the task selection rule, the improved ranked positional weight technique [30] is applied to choose a task for station j
(j = 1, 2, , m 1). The selection weight for candidate task i is calculated by

ti + Qi1 ,i ti1 i CA f
i1 UV
Weighti =  , (9)
ti + Qi,i1 ti1 i CAb
i1 UV

where CAf (CAb) denotes the set of entrance (exit) candidate tasks.
 
 
CA f = i| Pi,i1 = Pi,i2 , i I , (10)
i1 I i2 V

 
 
CAb = i| Pi1 ,i = Pi2 ,i , i I . (11)
i1 I i2 V

In addition, CA is denoted as the candidate task set and CA = CAf CAb.


Utilizing the task selection rule, the candidate task with the largest selection weight is chosen rst and the tasks with
relatively small processing times remain. Thus, the station times may be increased for the unassigned stations and the global
search ability can be enhanced using this rule.

3.2. Task assignment rule

In our algorithm, the task selection rule is only used to choose a task, but the decision about assigning the task is made
by the task assignment rule. In the task assignment rule, a well-balanced station load is achieved for station j (j = 1, 2, ,
m 1), which is dened by the following bound condition.

3.2.1. Bound condition


Let Cmax /Cmin denote the upper/lower bound of all the station times and we give three bound conditions, as follows:

ST j + ti Cmax , (12)

ST j + ti > Cmin , (13)


ti + ti > j Cm, (14)
i V

where task i is the task chosen by the task selection rule. It should be noted that if task i is allocated to station j, then
station j without/with task i is called the old/new station, respectively. Then, Conditions (12) and (13) bound the maximum
and minimum of the new station time respectively. Condition (14) limits the lower bound of the sum of the assigned task
times. Eqs. (12) and (13) are called bound condition (a), which ensures the balance of the solution, and Eqs. (12) and (14) are
called bound condition (b), which ensures that the cycle time is no more than Cmax . A station that meets bound condition
(a) or (b) is called a well-balanced station.
The values of Cmax and Cmin are initialized with Eq. (15) and updated by Eq. (16).

Cmax = Cm + , Cmin = 2Cm Cmax , (15)

Cmax max{ST j } dt, Cmin 2Cm Cmax , if max{ST j } < Cbest . (16)
jJ jJ

In Eq. (15), is a parameter and its value is set by the method proposed in Section 4.1. In Eq. (16), dt is constant during
the initiation of the solving process and it is reduced by half if a better solution is not obtained in the current generation.
In addition, the algorithm terminates if dt is less than 1.
In order to achieve a well-balanced station load, two assignment rules are proposed in the task assignment rule.
428 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

3.2.2. Assignment rules


Two assignment rules are used to assign task to the current station: the direct assignment rule and preferential assign-
ment rule.

Direct assignment rule. Assign the chosen task i directly to station j when Eq. (17) is satised.

ST j + max ti < Cmin . (17)


i CA

Preferential assignment rule. Skip the task selection rule, select the tasks that satisfy Eq. (18), and assign the task with the
longest task time (denoted by i ) among them as a preferential candidate task according to one of the following three
conditions, where the priority order is (i), (ii), and (iii).

ST j + max ti Cmax or ST j + min ti Cmax . (18)


i CA i CA

(i) Task i , i CA, satises bound condition (a).


(ii) Task i , i CA, satises bound condition (b).
(iii) Task i , i CA, satises Eq. (13).

If all the candidate tasks do not satisfy two assignment rules, the process of assigning task to station j is terminated. In
addition, if this station is not well balanced, then the following task exchange rule is employed to improve it.

3.3. Task exchange rule

In the task assignment rule, the preferential assignment rule is used mainly to assign the last task on the current station,
so the capacity to obtain a well-balanced station load is limited. However, the capacity can be enhanced if we partially
exchange the tasks assigned to this station with the unassigned tasks. During the process for exchanging tasks, the moved
tasks should satisfy sucient conditions such that the precedence constraints are not violated. For a U-shaped assembly
line, these constraints require that predecessors of an entrance task or successors of an exit task must be assigned rst
before the entrance and exit tasks are assigned.
In the task exchange rule, two types of tasks are moved: an emigrant task that moves out of the task set for the current
station and an immigrant task that moves into the task set of the current station.
It is obvious that when more tasks are moved, the probability of obtaining a well-balanced station load becomes higher
but at the cost of a longer computational time. To achieve a trade-off between the solution quality and computational time,
three pairs of immigrant and emigrant tasks are set, i.e., (1, 1), (2, 1), and (1, 2), and thus exchange rule 11, exchange rule
21, and exchange rule 12 are produced, respectively.
In the three exchange rules mentioned above, situations where the immigrant task(s) are moved into the entrance sub-
line and the emigrant task(s) are moved from the exit sub-line, or the immigrant task(s) are moved from the exit sub-line
and the emigrant task(s) are moved from the entrance sub-line, occur repeatedly. Thus, we rst give the conditions for
exchange tasks in this situation in Lemma 1. For convenience, the movement paths of the immigrant task(s) and emigrant
task(s) are considered to be non-crossing in this situation.

Lemma 1. The exchange can be performed if the movement paths of the immigrant tasks and emigrant tasks are non-crossing
and the relative precedence constraints are satised.

The proofs of Lemma 1 and the following three theorems are given in the Appendix.

3.3.1. Exchange rule 1-1


A task is called a dual-port task if its predecessors and successors have been fully assigned, i.e., it is both an entrance
and exit task. If an original entrance (exit) task is not a dual-port task, it is called a pure entrance (exit) task. The sets of
pure entrance tasks, pure exit tasks, and dual-port tasks are denoted as Wfj , Wbj , and Wej , respectively, where
 
   
W fj = i| Qi,i1 = Qi,i2 , Qi1 ,i = Qi2 ,i , i W j , (19)
i1 I i2 V i1 I i2 V

 
   
Wbj = i| Qi,i1 = Qi,i2 , Qi1 ,i = Qi2 ,i , i W j , (20)
i1 I i2 V i1 I i2 V

 
   
Wej = i| Qi,i1 = Qi,i2 , Qi1 ,i = Qi2 ,i , i W j . (21)
i1 I i2 V i1 I i2 V
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 429

For pure entrance tasks, pure exit tasks, and dual-port tasks, the emigrant tasks are classied into three categories: Of,
Ob, and Oe, respectively, which are described as follows:
 

Of = i| Pi1 ,i = 0, i W f j , (22)
i1 W f j

 

Ob = i| Pi,i1 = 0, i W b j , (23)
i1 W b j

Oe = W e j . (24)
Similarly, the immigrant tasks belong to three categories: If, Ib, and Ie, respectively, as follows:
 
 
If = i| Pi,i1 = 0, Pi2 ,i = 0, i UV , (25)
i1 UV i2 UV

 
 
Ib = i| Pi,i1 = 0, Pi2 ,i = 0, i UV , (26)
i1 UV i2 UV

 
 
Ie = i| Pi,i1 = 0, Pi2 ,i = 0, i UV . (27)
i1 UV i2 UV

According to Lemma 1 and Eqs. (19)(27), the sucient conditions for implementing exchange rule 11 are given as
follows.

Theorem 1. Let Out = Of Ob Oe, In = If Ib Ie. Exchange rule 11 can be implemented for any task, i1 Out, i2 In, if
one of the following conditions is satised:

(1.1) i1 Of, i2 If and Qi2 ,i1 = 1;


(1.2) i1 Ob, i2 Ib and Qi1 ,i2 = 1;
(1.3) i1 Of, i2 Ib Ie, or i1 Ob, i2 If Ie, or i1 Oe, i2 In.

3.3.2. Exchange rule 21


Unlike exchange rule 11, two tasks are moved out of a station in exchange rule 21. During the movement process, after
one task (denoted as i1 ) is moved out, another new emigrant task can be generated, which is added to the emigrant task
set. The second emigrant task (denoted i2 ) is chosen from this new set.
The task denoted as i is a new emigrant task if it is the immediate predecessor of the rst emigrant task (denoted as
i1 , i1 Of), and task i1 only has one immediate successor in the entrance sub-line of station j. We denote O f i1 as their set,
where
 

O f i1 = i|Pi1 ,i = 1, Pi2 ,i = 1, i1 O f, i W f j . (28)
i2 W f j

Task i1 and the task in set O fi1 can be moved out of station j together as two entrance tasks.
Similarly, the set of new emigrant tasks for the rst emigrant task i1 (i1 Ob) is dened as follows:
 

Obi1 = i|Pi,i1 = 1, Pi,i2 = 1, i1 Ob, i W b j . (29)
i2 W b j

Task i1 and the task in set Obi1 can be moved out of station j together as two exit tasks.
It is important to note that when the rst emigrant task i1 in set Oe is moved as a pure entrance task or a pure exit one,
its immediate predecessors and successors in the dual-port task set will be transformed into pure entrance and exit tasks,
respectively.
Illustrative example Assume that tasks 1, 2, 3, 4, and 5 are in set Oe, and their precedence graph is shown in Fig. 4. If task
3 is moved out of station j as a pure entrance task, then its immediate predecessors, i.e., tasks 1 and 2, are transformed
into pure entrance tasks, and its immediate successors, i.e., tasks 4 and 5, are transformed into pure exit tasks (Fig. 5). This
conclusion is also valid when task 3 is moved out of a station as a pure exit task.
430 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

Entrance

Exit

Fig. 5. Single distribution graph for the four tasks in Oe.

Entrance

Exit

Fig. 6. Set graph for three types of tasks.

If emigrant task i1 (i1 Oe) is moved as a pure entrance (exit) task, then some new entrance (exit) tasks can be gener-
f
ated. We denote Oei (Oebi ) as the set of these new entrance (exit) tasks, where
1 1
 


Oeif = i|Pi1 ,i = 1, i1 Oe, i W e j i|Pi1 ,i = 1, Pi,i2 = 1, i1 Oe, i W f j , (30)
1
i2 W f j

 


Oebi1 = i|Pi,i1 = 1, i1 Oe, i W e j i|Pi,i1 = 1, Pi2 ,i = 1, i1 Oe, i W b j . (31)
i2 W b j

f
In the rst sub-set of Oei in Eq. (30), all of the successors of task i (i Wej ) are assigned to station j according to the
1
denition of set Wej . Thus, task i has no successors on the entrance sub-line of station j if the rst emigrant task i1 (i1
Oe), i.e., one of its successors, is moved out of the station as an entrance task and its other successors are regarded as pure
exit tasks. Thus, task i can be moved out of a station along with i1 as another emigrant task. A similar approach can be
used to explain the rst sub-set of Oebi in Eq. (31).
1
Illustrative example Assume that task i1 Oe, i2 Wej , i2 is the immediate predecessor of task i1 , and i3 and i4 are the
immediate successors of i2 . Then, task i2 can be moved out of station j along with i1 if tasks i3 and i4 are regarded as pure
exit tasks, as shown in Fig. 6.
Based on the sets above, the sucient conditions for implementing exchange rule 21 are given as follows:

Theorem 2. Exchange rule 21 can be implemented if one of the following conditions is satised for any i1 Out, i3 In.
f
(2.1) (i1 Of and i2 O f O fi1 Oe or i1 Oe and i2 Oei ) and i3 Ib Ie, or (i1 Ob and i2 Ob Obi1 Oe or i1 Oe
1
and i2 Oebi ) and i3 If Ie, or i1 Oe and i2 Oe.
1
f
(2.2) (i1 Of and i2 O f O fi1 or i1 Oe and i2 Oei ), i3 If, Qi3 ,i1 = 1 and Qi3 ,i2 = 1.
1
(2.3) (i1 Ob and i2 Ob Obi1 or i1 Oe and i2 Oebi ), i3 Ib, Qi1 ,i3 = 1 and Qi2 ,i3 = 1.
1
(2.4) i1 Of, i2 Ob Oe, i3 If and Qi3 ,i1 = 1.
(2.5) i1 Ob, i2 Of Oe, i3 Ib and Qi1 ,i3 = 1.
(2.6) i1 Of, i2 Ob, i3 Ie and (Qi3 ,i1 = 1orQi2 ,i3 = 1).

3.3.3. Exchange rule 12


Similar to exchange rule 21, some new immigrant tasks in UV are generated after one is moved out of UV, and these
tasks should be added to the set of immigrant tasks. Another immigrant task will be chosen from this new set.
Assume that the rst immigrant task i2 is in set If. The task denoted as i is a new immigrant task if it is the immediate
successor of task i2 and task i2 only has one immediate predecessor in the entrance sub-line of station j. Denote I f i2 as their
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 431

set, where
 

I f i2 = i|Pi,i2 = 1, Pi,i = 1, i2 I f, i UV . (32)
i UV

If the rst immigrant task i2 is in set Ib, denote Ibi2 as the set of added immigrant tasks, where
 

I bi2 = i|Pi2 ,i = 1, Pi ,i = 1, i2 Ib, i UV . (33)
i UV

If the rst immigrant task is i2 Ie, then no added immigrant tasks can be found in UV because all of its predecessors
and successors have been assigned to a station.
Based on the sets dened above, the sucient conditions for implementing exchange rule 12 are proposed in
Theorem 3.

Theorem 3. Exchange rule 12 can be implemented if one of the following conditions is satised for any i1 Out, i2 In, i3
UV.

(3.1) i1 Ob Oe, i2 If and i3 I f I fi2 Ie, or i1 Of Oe, i2 Ib and i3 Ib Ibi2 Ie, or i2 Ie and i3 Ie.
(3.2) i1 Of, i2 If, i3 I f I fi2 , Qi2 ,i1 = 1 and Qi3 ,i1 = 1.
(3.3) i1 Of, i2 If, i3 Ib Ie and Qi2 ,i1 = 1.
(3.4) i1 Ob, i2 Ib, i3 Ib Ibi2 , Qi1 ,i2 = 1 and Qi1 ,i3 = 1.
(3.5) i1 Ob, i2 Ib, i3 If Ie and Qi1 ,i2 = 1.
(3.6) i1 Oe, i2 If, i3 Ib and (Qi2 ,i1 = 1 or Qi1 ,i3 = 1).

During the execution of the task exchange rules given above, the exchange rule will be terminated if the station time for
one exchanged station load is satised with the bound condition.

4. Computational experiments

In order to test the performance of the proposed heuristic approach, we performed two types of computational experi-
ments with different scales, where one type focused on small-scale and medium-scale examples, whereas the other consid-
ered large-scale examples. The proposed heuristic approach was programmed in MATLAB R2010a and all the experiments
are performed using a PC with an Intel (R) Core (TM) i5-3337 U @ 1.80 GHz processor and 4.00 GB of RAM. Before solving
these examples, the parameter setting experiments were conducted rst to determine the parameters for the rules-based
heuristic approach.

4.1. Parameter setting experiments

Two parameters are employed in the proposed approach: and dt. Parameter dt mainly affects the computational time
of the algorithm and the other parameter mainly inuences the quality of the initial solution. The inuence of dt is small,
so it was set as a constant with a value of 10.
For all UALBPs, the main target is obtaining a solution with a high balance ratio, where this ratio is determined by
parameter for the initial solution. Thus, we can use a higher balance ratio to estimate the initial value of , and vice versa.
The method is given as follows.
Let denote the balance ratio of the initial solution, which is calculated with

iI ti m Cm Cm
= = = . (34)
m Cmax m.(Cm + ) Cm +
The value of is obtained using
Cm (1 )
= . (35)

The optimal balance ratio for most benchmark problems is no less than 90%, so we set 10 values for the balance ratio ,
i.e., 90%, 91%, 92% and 99%. To obtain the best value among the 10 parameters, we solved four benchmark cases described
by Scholl [31] and the computational results are shown in Table 1.
In Table 1, Columns 1 shows the 10 values of parameter . Columns 29 show the cycle time (CT) and computational
time (CPU) obtained for the four benchmark cases, where Heskiaoff 288 represents the case source and the given num-
bers of tasks and stations.
As shown in Table 1, 39 out of 40 invariable cycle times are obtained, which shows that the value of balance ratio is
insensitive to the algorithm. This conclusion is also supported by the standard deviations in Table 2.
As shown in Table 2, the standard deviations of the computational time in three cases (Heskialff, Kilbridge & Wester, and
Tong) were all less than 1, but the standard deviation was greater for Arcus2. These results show that the computational
432 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

Table 1
Computational results obtained in the parameter setting experiments.

Heskiaoff 288 Kilbridge & Wester 456 Tong 7012 Arcus2 11120

CT CPU (s) CT CPU (s) CT CPU (s) CT CPU (s)

90% 129 1.4926 92 0.9133 295 4.2735 7558 72.825


91% 129 1.4122 92 0.7038 295 4.8893 7558 54.867
92% 129 1.3555 92 0.7721 295 3.8931 7558 48.235
93% 129 1.6981 92 0.7630 295 4.9186 7558 41.918
94% 129 1.6283 92 0.7238 295 4.5038 7558 37.806
95% 129 1.4298 92 0.7750 295 4.2636 7558 30.591
96% 129 1.3684 92 0.7432 295 3.5205 7558 24.774
97% 129 1.4030 92 0.7097 297 1.7623 7558 19.517
98% 129 1.2314 92 0.6552 295 3.9106 7558 13.701
99% 129 1.2216 92 0.6274 295 3.2299 7558 9.032

Table 2
Standard deviations of the cycle time and computational time.

Case Heskiaoff 288 Kilbridge & Wester 456 Tong 7012 Arcus2 11120

Item CT CPU (s) CT CPU (s) CT CPU (s) CT CPU (s)

STDEV 0 0.1520 0 0.0783 0.6325 0.9317 0 19.8522

time was sensitive to the value of in some cases. In addition, the computational time required to solve Arcus2 decreased
as increased according to Table 1, which shows that the computational time may be decreased by increasing , although
this may be accompanied by a worse solution (Tong 7012, = 97%). According to the analysis of these results, the value
of was set to 95%.

4.2. Effectiveness experiments

In the rst group of effectiveness experiments, 13 small or medium scale examples comprising 63 instances from Sawyer
(1970), Kilbridge and Wester (1962), Tonge (1961), and Arcus (1963) were solved using integer programming (in Section 2.3)
and the rules-based heuristic approach. Integer programming was solved using GAMS CPLEX. The results obtained by integer
programming and the heuristic approach are compared in Table 3.
In Table 3, columns 1 and 2 represent the problem source and the number of stations. Column 3 shows the round
required to reach the mean station time for each instance (C = Cm
), which is called the ideal cycle time. The results
obtained by integer programming are presented in columns 4, 5, and 6, which comprise the optimal cycle time (OPT),
computational time (CPU, seconds), and the mean computational time (MCPU). Columns 912 show the results obtained by
the rules-based heuristic approach, which comprise the best cycle time (CT), computational time, mean computational time,
and percentage excess cycle time (PECT%) compared with integer programming solved by GAMS CPLEX. The computational
time for GAMS CPLEX was limited to 3600 s and in column 5 denotes that the CPU time exceeded 3600 s, so the cycle
time obtained might not have been the optimal. In column 12, the percentage excess cycle time was set to 0 if the solution
obtained by the heuristic approach was better than that using integer programming.
As shown in Table 3, most of the optimal solutions could be obtained by integer programming for the 29 small instances
(n 30), except one (Buxey-29-9), i.e., 96.6%. However, the proportion decreased to 41.2% for the 34 medium instances (30
< n 70), where the downward trend increased with the numbers of tasks. These results show that integer programming
can mainly be used to solve small instances and it has diculty solving medium or large problems (n > 70) in a short
period of time.
As shown in column 9, 21 optimal solutions were obtained by the rules-based heuristic approach for 29 small instances,
i.e., 72.4%. The mean percentage excess cycle time for the 29 results was only 0.58%. The maximal percentage excess cy-
cle time was 3.13% (Rosenberg-25-4), but the largest excess cycle time was only 2 (Heskiaoff-28-9). For the 34 medium
instances, 18 results solved by the rules-based heuristic approach were less than or equal to those obtained by integer pro-
gramming, i.e., 52.9%, and the improvement was 18.9%. The average of all the percentage excess cycle times was only 0.64%.
These statistical results demonstrate that the proposed rules-based heuristic approach could effectively solve the small and
medium examples.
The mean computational times in columns 6 and 11 show that the rules-based heuristic approach had a slight advantage
compared with integer programming for ve examples (Bowmax-8, Jackson-11, Mitchell-21, Rosenberg-25 and Sawyer-30),
but it was much better for the other examples. Moreover, the standard deviation of the computational time was also cal-
culated for each instance, where the maximum was 0.810 (Tonge-70) and the minimum was only 0.019 (Bowmax-8). These
results demonstrate that the computational time was stable using the heuristic approach for 63 instances. Therefore, the
proposed heuristic approach can solve small and medium examples with high and stable speed.
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 433

Table 3
Comparison of the results obtained using integer programming and the rules-based heuristic approach.

Example m C Integer programming Rules-based heuristic approach

OPT CPU MCPU Two Rules Three Rules

CT CPU CT CPU MCPU PECT

Bowmax-8 2 38 38 0.725 0.432 38 0.029 38 0.036 0.058 0


3 26 26 0.284 26 0.055 26 0.071 0
4 19 20 0.288 27 0.015 20 0.068 0
Jackson-11 2 23 23 0.181 0.216 23 0.028 23 0.027 0.107 0
3 16 16 0.276 17 0.091 16 0.136 0
4 12 12 0.190 12 0.121 12 0.158 0
Mitchell-21 4 27 27 0.309 0.35 27 0.238 27 0.259 0.244 0
5 21 21 0.365 22 0.296 21 0.101 0
6 18 18 0.377 18 0.314 18 0.373 0
Rosenberg-25 4 32 32 0.401 0.75 33 0.219 33 0.265 0.345 3.13
5 25 25 1.121 25 0.143 25 0.105 0
6 21 21 0.663 23 0.193 21 0.427 0
7 18 18 0.814 18 0.354 18 0.434 0
8 16 16 0.726 16 0.429 16 0.496 0
Heskiaoff-28 6 171 171 0.468 17.10 171 1.161 171 1.267 0.923 0
7 147 147 0.550 152 0.292 147 1.182 0
8 128 128 83.024 129 0.821 129 0.822 0.78
9 114 114 0.823 117 0.629 116 0.576 1.75
10 103 108 0.642 109 0.274 108 0.766 0
Buxey-29 5 65 65 0.783 858.56 69 0.226 66 0.501 0.542 1.54
6 54 54 13.14 56 0.336 55 0.596 1.85
7 47 47 1.038 50 0.241 48 0.549 2.13
8 41 41 1.568 44 0.248 41 0.594 0

9 36 37 39 0.155 38 0.447 2.70
10 33 33 1534.868 35 0.285 34 0.565 3.03
Sawyer-30 5 65 65 0.404 7.00 65 0.790 65 0.695 0.608 0
6 54 54 24.725 55 0.800 54 0.170 0
7 47 47 0.681 48 0.821 47 0.729 0
8 41 41 2.207 41 0.710 41 0.836 0
Lutz1-32 6 2357 2366 912.395 3009.50 2448 0.218 2404 0.491 0.341 1.61

7 2020 2030 2092 0.299 2092 0.370 3.05

8 1768 1782 1830 0.179 1816 0.380 1.91
9 1572 1592 3335.109 1644 0.265 1622 0.155 1.88

10 1414 1442 1474 0.300 1474 0.309 2.22
Gunther-35 5 97 97 0.649 2.72 98 0.969 97 1.091 0.848 0
6 81 81 1.662 83 0.557 82 0.729 1.23
7 69 69 4.868 71 0.651 70 0.754 1.45
8 61 61 3.690 62 0.699 61 0.818 0
Kilbridge&Wester-45 6 92 92 1.018 1.97 95 0.609 92 0.412 1.216 0
7 79 79 0.819 83 0.881 79 1.900 0
8 69 69 4.515 73 0.888 69 0.390 0
9 62 62 1.384 66 0.769 62 1.609 0
10 56 56 2.143 59 0.885 56 1.768 0

Hahm-53 5 2806 2808 2183.33 2948 0.232 2823 1.569 1.169 0.53

6 2338 2340 2420 0.358 2416 1.211 3.25

7 2004 2008 2082 0.511 2013 1.706 0.25
8 1754 1792 109.191 1840 0.436 1827 1.206 1.95
9 1559 1775 7.491 2193 0.113 1775 0.151 0

Warnecke-58 10 155 156 3210.20 162 0.917 155 2.663 2.283 0

11 141 142 149 0.955 143 2.589 0.70

12 129 131 136 0.758 131 2.293 0
13 120 120 481.651 127 0.945 122 2.151 1.67

14 111 113 115 1.272 112 2.140 0

15 104 106 110 0.875 105 2.497 0

16 97 99 103 1.032 98 2.104 0

17 92 94 96 1.339 94 1.826 0

Tonge-70 10 351 352 3600 359 1.538 352 3.873 2.606 0

11 320 321 323 2.026 320 3.282 0

12 293 294 298 2.186 295 2.899 0.340

13 270 271 275 1.138 273 2.145 0.74

14 251 253 255 1.450 252 2.457 0

15 234 237 237 1.255 236 2.0 0 0 0

16 220 221 232 1.240 222 1.515 0.45
434 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

The three rules in the proposed approach comprise task selection, task assignment, and task exchange rules. Among these
rules, the task exchange rule is most important. In order to verify its key role, we solved 63 small or medium instances and
58 large ones using the rst two rules and all three rules, respectively. The cycle times and computational times obtained
are shown in columns 7 and 8 in Table 3, and columns 4 and 5 in Table 4, respectively. The large instances came from ve
examples, i.e., Lutz2-89, Mukherjee-94, Arcus2-111, Bartholdi-148, and Scholl-297 [32].
As shown in columns 7 and 9 in Table 3, 47 of the results solved using the three rules were better than those solved
using two rules, and the improvement rate was 74.6%. The improvement rate increased to 87.9% for 58 large instances
(Table 4). The maximum CPU time using two or three rules was no more than 70 s, so the very small increase using the three
rules is acceptable. These results indicate that the task exchange rules embedded in the rules-based heuristic approach can
signicantly improve the solution performance at acceptable computational costs, and thus none of them can be excluded
from the rules-based heuristic approach.
In fact, the task selection rules select the task with a longer task time for the initial stations in the rules-based heuristic
approach, which improves the global balance of all the stations. Thus, the task assignment rules assign suitable tasks to
the current station and improve this station into a well-balanced and local optimal state. The task exchange rules improve
the current station by exchanging tasks between this station and the following stations, thereby further enhancing the local
optimization. In addition, the bound conditions make the current station well balanced by bounding the scope of the task
times and pointing out the direction of optimization. All of these rules are integrated in a systematic and complementary
manner, and the integrated system guarantees that good performance is obtained by the proposed approach.
Five examples comprising 58 instances were solved by the heuristic approach to test its effectiveness with large scale
problems, and the results are shown in Table 4. This table also shows the ideal cycle time calculated for every instance,
which we used to verify the accuracy of the heuristic approach.
In Table 4, columns 1, 2, and 3 represent the problem source, number of stations, and the round required to reach the
mean station time, respectively. The results obtained by the heuristic approach are shown in columns 610, i.e., the best
cycle time, percentage excess cycle time compared with C (%), average of the percentage excess cycle time (MP.E.C.T, %),
computational time (seconds), and mean computational time.
As shown in columns 68, 15 ideal cycle times were solved by the rules-based heuristic approach, i.e., 25.9%. The average
and maximum of all the percentage excess cycle times were 0.49% and 3.85%, respectively. The results in column 9 show
that the maximum computational time was less than 69.36 s. The average of all the computational times was only 20.67 s.
These statistical results demonstrate that the proposed approach can solve large scale problems in an effective and rapid
manner.
In order to further test the performance of the algorithm, we solved 30 new standard instances comprising 15 large
instances (100 tasks) and 15 very large instances (10 0 0 tasks) from Otto et al. [33], and the results are shown in Table 5.
In Table 5, columns 14 represent the source number, optimal number of stations, the given cycle time in the SALBP-1
data, and the round required to reach the mean station time, respectively. Columns 56 show the best cycle time and
computational time (seconds) for the proposed algorithm, respectively. The last two columns show the percentage excess
cycle time compared with C and C (P.E.C.T1 and P.E.C.T2, %), respectively, where the value was set to 0 when the cycle time
obtained was less than C .
As shown in columns 5 and 7, all of the best cycle times solved by the proposed heuristic approach were less than the
given cycle time C . These results support the conclusion given in Section 1 and they also demonstrate the effectiveness of
the proposed approach. Compared with the ideal cycle time C , the average percentage excess cycle time was only 0.18%. In
particular, the optimal solutions were all found for 15 very large instances. The computational times for 15 large instances
were still low and their average was only 12.64 s. The computational times for 15 very large instances were long and the
average was 2860.42 s, but the times were also short with respect to their size. These statistical results also demonstrate
the effectiveness of the proposed approach at solving large and very large scale problems.

5. Conclusions

In this study, we modeled integer programming for UALBP-2 and proposed a novel rules-based heuristic approach, which
embeds task selection, task assignment, and task exchange rules. In order to ensure that the task exchange process does not
violate the precedence constraints, we proposed and proved the sucient conditions for their implementation.
The rules-based heuristic approach was tested with 63 small and medium scale instances, 58 large scale instances,
and 30 brand new standard instances for UALBP-2. Using the proposed approach, most of these instances were solved to
optimality at acceptable computational costs, and we obtained better results within 70 s for the remainder compared with
the 3600 s required by the integer programming solver. In addition, due to the restriction imposed by an increasingly smaller
bound on the task times, the task selection rules enhance global optimization by selecting a task with a longer task time
for the initial stations, where the task assignment rules and task exchange rules aim to achieve a well-balanced and local
optimum by assigning suitable tasks to the current station and exchanging tasks between the current and following sta-
tions. The systematic and complementary integration of these rules ensures that there is a balance between exploration and
exploitation in the heuristic approach, thereby improving the eciency and effectiveness when solving UALBP-2. The high
computational speed and precision of the proposed algorithm are useful for applications by improving the productivity and
facilitating real-time adjustments by rms.
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 435

Table 4
Results obtained by the heuristic approach for large scale problems.

Example m C Heuristic approach

Two rules Three rules

CT CPU CT P.E.C.T MP.E.C.T CPU MCPU

Lutz2-89 13 38 40 1.906 38 0 0.81 2.88 2.68


14 35 37 1.922 35 0 2.89
15 33 33 2.735 33 0 2.78
16 31 31 2.567 31 0 2.62
17 29 30 2.384 30 3.45 2.45
18 27 29 1.722 27 0 2.73
19 26 27 2.416 27 3.85 2.32
20 25 27 1.714 25 0 2.65
21 24 24 2.594 24 0 2.76
Mukherjee-94 14 301 307 3.671 302 0.33 0.96 8.12 6.36
15 281 284 6.118 283 0.71 6.92
16 263 267 3.653 265 0.76 6.31
17 248 252 3.013 251 1.21 5.71
18 234 237 4.554 236 0.85 6.05
19 222 224 4.029 224 0.90 4.44
20 211 223 2.847 212 0.47 5.37
21 201 211 4.467 205 1.99 8.07
22 192 203 2.036 193 0.52 4.16
23 183 193 1.639 185 1.09 7.54
24 176 186 2.465 179 1.70 7.24
Arcus2-111 9 16,711 17,208 34.03 16,718 0.04 0.45 31.38 14.00
13 11,570 11,812 28.96 11,576 0.05 21.01
14 10,743 10,776 40.12 10,775 0.30 16.58
15 10,027 10,423 12.92 10,050 0.23 17.15
16 9400 9836 7.723 9418 0.19 16.00
17 8847 8955 25.85 8862 0.17 15.58
18 8356 8650 12.29 8377 0.25 14.58
19 7916 8143 15.66 7955 0.49 10.66
20 7520 7802 9.994 7558 0.51 11.96
21 7162 7388 11.87 7211 0.68 9.55
22 6837 6937 18.23 6861 0.35 9.83
23 6540 6663 15.98 6590 0.76 8.24
24 6267 6340 17.40 6335 1.08 8.06
25 6016 6330 2.530 6086 1.16 5.47
Bartholdi-148 7 805 806 19.04 805 0 0 20.77 17.42
8 705 712 12.00 705 0 20.23
9 626 628 16.38 626 0 11.33
10 564 564 14.30 564 0 15.87
11 513 520 10.27 513 0 13.68
12 470 475 10.03 470 0 10.51
13 434 437 11.12 434 0 24.60
14 403 405 12.58 403 0 22.35
Scholl-297 21 3317 3490 10.13 3328 0.33 0.27 35.65 48.08
23 3029 3189 10.27 3034 0.17 53.13
25 2787 2931 11.32 2790 0.11 47.36
27 2580 2695 25.45 2582 0.08 69.36
29 2402 2529 14.11 2406 0.17 48.79
31 2247 2366 10.67 2256 0.40 43.26
33 2111 2203 16.08 2114 0.14 56.53
35 1991 2011 41.62 1993 0.10 62.17
37 1883 1983 13.43 1888 0.27 45.60
39 1787 1882 14.91 1792 0.28 40.66
41 1699 1779 17.18 1702 0.18 58.57
43 1620 1706 13.87 1623 0.19 36.72
45 1548 1620 12.98 1551 0.19 53.30
47 1483 1562 8.671 1488 0.34 42.92
49 1422 1496 15.44 1438 1.13 34.72
50 1394 1458 12.13 1398 0.29 40.51
436 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

Table 5
Results obtained by the heuristic approach for 30 new standard instances.

Example m C C Heuristic approach

Best CPU P.E.C.T1 P.E.C.T2

100_1 23 10 0 0 988 994 10.25 0 0.61


100_2 21 10 0 0 965 969 11.96 0 0.41
100_3 20 10 0 0 972 973 11.14 0 0.10
100_4 24 10 0 0 974 977 13.86 0 0.31
100_5 22 10 0 0 973 976 13.48 0 0.31
100_6 22 10 0 0 981 982 16.05 0 0.10
100_7 26 10 0 0 964 967 13.88 0 0.31
100_8 24 10 0 0 973 975 13.89 0 0.21
100_9 23 10 0 0 993 996 14.56 0 0.30
100_10 22 10 0 0 955 963 10.54 0 0.84
100_11 24 10 0 0 977 984 11.30 0 0.2
100_12 25 10 0 0 983 989 9.94 0 0.61
100_13 24 10 0 0 978 983 9.76 0 0.51
100_14 20 10 0 0 991 993 17.46 0 0.20
100_15 24 10 0 0 960 965 11.53 0 0.52
10 0 0_1 135 10 0 0 997 997 2976.41 0 0
10 0 0_2 137 10 0 0 998 998 3011.42 0 0
10 0 0_3 136 10 0 0 10 0 0 10 0 0 2825.03 0 0
10 0 0_4 138 10 0 0 996 996 2905.91 0 0
10 0 0_5 135 10 0 0 997 997 3098.34 0 0
10 0 0_6 141 10 0 0 10 0 0 10 0 0 2941.64 0 0
10 0 0_7 136 10 0 0 999 999 2783.45 0 0
10 0 0_8 138 10 0 0 999 999 2712.14 0 0
10 0 0_9 134 10 0 0 10 0 0 10 0 0 2689.54 0 0
10 0 0_10 140 10 0 0 996 996 2742.25 0 0
10 0 0_11 134 10 0 0 998 998 2942.63 0 0
10 0 0_12 134 10 0 0 994 994 2880.97 0 0
10 0 0_13 131 10 0 0 999 999 2932.86 0 0
10 0 0_14 136 10 0 0 10 0 0 10 0 0 2764.83 0 0
10 0 0_15 136 10 0 0 994 994 2698.93 0 0

The heuristic approach has advantages in terms of its high computational speed, but the uncertainty of its computational
accuracy limits its applications. Many heuristic factors and rules have been proposed to improve the computational accuracy,
but these factors and rules are mostly based on experience, which may be limited. Furthermore, testing of the algorithm
depends on the results obtained when solving benchmarks and the testing process is incomplete. Thus, it would be nec-
essary to generate vast amounts of instances to gain experience of heuristic factors and extract rules from them according
to machine learning techniques, where the design target for heuristics will vary among different assembly line balancing
problem examples. Developing an algorithm based on big data analytics may enhance the accuracy of the calculations and
a vast number of instances can be used to more fully test the performance of the algorithm. These issues will be studied in
future research.

Acknowledgments

This study was supported by the National Natural Science Foundation (Nos. 51275366, 151305311, and 50875190), the
Specialized Research Fund for the Doctoral Program of Higher Education (20134219110 0 02), and the Fund Project of Hubei
Provincial Department of Education (No. D20161104), China.

Appendix

Lemma 1. The exchange can be performed if the movement paths of the immigrant tasks and emigrant tasks are non-crossing
and the relative precedence constraints are satised.

Proof. Assume that only two tasks, i.e., emigrant task i and immigrant task i , are moved and their movement paths are
non-crossing. Let the emigrant task be an entrance task and the immigrant task is an exit task. In Fig. 7, the emigrant task
in station j and the immigrant task in UV are shown, and two dotted arrows indicate their directions of movement.

Emigrant task i is an entrance task, so it has no successors in the entrance sub-line, i.e., all of its successors are in UV
or in the exit sub-line. In a similar manner, task i has no successors in set UV and it can be moved to the exit sub-line of
station j, i.e., its successors are assigned to the exit sub-line. Then, task i is not the successor of task i and task i is not the
predecessor of task i, so the move does not violate the predecessor constraints. Using a similar approach, the condition for
more than one emigrant tasks or immigrant tasks is also valid.
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 437

Entrance

Exit

Fig. 7. Movement directions for tasks i and i .

Theorem 1. Let Out = Of Ob Oe, In = If Ib Ie. Exchange rule 11 can be implemented for any task i1 Out, i2 In, if
one of the following conditions is satised:

(1.1) i1 Of, i2 If and Qi2 ,i1 = 1;


(1.2) i1 Ob, i2 Ib and Qi1 ,i2 = 1;
(1.3) i1 Of, i2 Ib Ie, or i1 O_b, i2 I f Ie, or i1 Oe, i2 In.

Proof. In condition (1.1), since i1 Of, i2 If, so task i1 has no successors in the entrance sub-line of station j, and task i2
has no predecessors in the unassigned task set UV. The condition of Qi2 ,i1 = 1 ensures that task i1 is not the predecessor
of task i2 . Therefore, all the successors of task i1 are not in the entrance sub-line, and all the predecessors of task i2 are
assigned to the entrance sub-line, which means that the movements of two tasks do not violate the precedence constraints,
so 11 exchange rule can be implemented.

In condition (1.2), i1 Ob, i2 Ib, so task i1 has no predecessors in the exit sub-line of station j, and task i2 has no
successors in UV. The condition of Qi1 ,i2 = 1 ensures that task i1 is not the successor of task i2 . Therefore, all the predecessors
of task i1 are not in the exit sub-line and all the successors of task i2 are in the exit sub-line, which means that the
movements of two tasks do not violate the precedence constraints, so 11 exchange rule can be implemented.
In condition (1.3), task i2 can be moved to station j as a pure exit task when i1 Of, i2 Ib Ie. Thus, the movement
paths of task i1 and i2 are non-crossing, and similar conclusions can also be reached based on conditioni1 Ob, i2 If Ie
and i1 Oe, i2 In. Therefore, exchange rule 11 can be implemented according to Lemma 1.

Theorem 2. Exchange rule 21 can be implemented if one of the following conditions is satised for any i1 Out, i3 In.
f
(2.1) (i1 Of and i2 O f O fi1 Oe or i1 Oe and i2 Oei ) and i3 Ib Ie, or (i1 Ob and i2 Ob Obi1 Oe or i1
1
Oe and i2 Oebi ) and i3 If Ie, or i1 Oe and i2 Oe.
1
f
(2.2) (i1 Of and i2 O f O fi1 or i1 Oe and i2 Oei ), i3 If, Qi3 ,i1 = 1 and Qi3 ,i2 = 1.
1
(2.3) (i1 Ob and i2 Ob Obi1 or i1 Oe and i2 Oebi ), i3 Ib, Qi1 ,i3 = 1 and Qi2 ,i3 = 1.
1
(2.4) i1 Of, i2 Ob Oe, i3 If and Qi3 ,i1 = 1.
(2.5) i1 Ob, i2 Of Oe, i3 Ib and Qi1 ,i3 = 1.
(2.6) i1 Of, i2 Ob, i3 Ie and (Qi3 ,i1 = 1 or Qi2 ,i3 = 1) .

Proof. In condition (2.1), if i1 Of, i2 O f O fi1 Oe and i3 Ib Ie, then emigrant tasks i1 and i2 can be moved as
pure entrance tasks, and immigrant task i3 can be moved as a pure exit task. Their movement paths are non-crossing.
Similar conclusions can also be reached based on the other conditions in condition (2.1). Therefore, exchange rule 21 can
be implemented according to Lemma 1.
f
In condition (2.2), if i1 Of, i2 O f O fi1 (or i1 Oe, i2 Oei ) and i3 If, then three tasks can be moved as pure
1
entrance tasks. All the successors of task i1 and i2 are not in the entrance sub-line and all the predecessors of task i3 are
assigned to station according to , 28), and (25). Conditions Qi3 ,i1 = 1 and Qi3 ,i2 = 1 ensure that task i1 and i2 are not the
predecessors of task i3 . Then, all the predecessors of task i3 have been assigned to the entrance sub-line, so exchange rule
21 can be implemented.
The conclusion based on condition (2.3) can be proved in a similar manner to that for condition (2.2).
In condition (2.4), if i1 Of, i3 If and Qi3 ,i1 = 1, then two tasks can be exchanged according to the proof of (1.1) in
Theorem 1. In addition, conditions i2 Ob Oe, i3 If also ensure that two tasks can be exchanged according to the proof
of (1.3) in Theorem 1. Therefore, three tasks can be exchanged according to Lemma 1.
In condition (2.5), conditions i1 Ob, i3 Ib and Qi1 ,i3 = 1 ensure that two tasks can be exchanged according to the proof
of (1.2) in Theorem 1. In addition, conditions i2 Of Oe, i3 Ib also ensure that two tasks can be exchanged according to
the proof of (1.3) in Theorem 1. Therefore, three tasks can be exchanged according to Lemma 1.
In condition (2.6), emigrant task i3 can be regarded as a pure entrance task or a pure exit task. If it is regarded as a pure
entrance task, then the other conditions i1 Of, i2 Ob and Qi3 ,i1 = 1 ensure that three tasks can be moved according to
438 M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439

(2.4) in Theorem 2. If it is regarded as a pure exit task, then the other conditions i1 Of, i2 Ob and Qi2 ,i3 = 1 ensure that
three tasks can also be moved according to (2.5) in Theorem 2.

Theorem 3. Exchange rule 12 can be implemented if one of the following conditions is satised for any i1 Out, i2 In, i3
UV.

(3.1) i1 Ob Oe, i2 If and i3 I f I fi2 Ie, or i1 Of Oe, i2 Ib and i3 Ib Ibi2 Ie, or i2 Ie and i3 Ie.
(3.2) i1 Of, i2 If, i3 I f I fi2 , Qi2 ,i1 = 1 and Qi3 ,i1 = 1.
(3.3) i1 Of, i2 If, i3 Ib Ie and Qi2 ,i1 = 1.
(3.4) i1 Ob, i2 Ib, i3 Ib Ibi2 , Qi1 ,i2 = 1 and Qi1 ,i3 = 1.
(3.5) i1 Ob, i2 Ib, i3 If Ie and Qi1 ,i2 = 1.
(3.6) i1 Oe, i2 If, i3 Ib and (Qi2 ,i1 = 1 or Qi1 ,i3 = 1).

Proof. In condition (3.1), if i1 Ob Oe, i2 If and i3 I f I fi2 Ie, then emigrant task i1 can be regarded as a pure
exit task, and immigrant tasks i2 and i3 can be regarded as pure entrance tasks. Their movement paths are non-crossing.
Similar conclusions can also be reached based on the other conditions in condition (3.1) . Therefore, exchange rule 12 can
be implemented according to Lemma 1.

In condition (3.2), i1 Of, i2 If and i3 I f I fi2 , so emigrant task i1 is a pure entrance task, and immigrant tasks i2
and i3 are pure entrance tasks. Therefore, all the successors of task i1 are not in the entrance side and all the predecessors
of tasks i2 and i3 are assigned to station. The conditions of Qi2 ,i1 = 1 and Qi3 ,i1 = 1 ensure that task i1 is not the predecessor
of task i2 and i3 . Therefore, all the predecessors of task i2 and i3 are assigned to the entrance sub-line, so exchange rule 12
can be implemented.
In condition (3.3), conditions i1 Of, i2 If and Qi2 ,i1 = 1 ensure that two tasks can be exchanged according to the proof
of (1.1) in Theorem 1. In addition, condition i3 Ib Ie ensures that task i3 can be moved to station j as a pure exit task.
Therefore, three tasks can be exchanged according to Lemma 1.
In condition (3.4), i1 Ob, i2 Ib and i3 Ib Ibi2 , so task i1 is a pure exit task, and immigrant tasks i2 and i3 can be
regarded as pure exit tasks. Therefore, all the predecessors of task i1 are not in the exit sub-line, and all the successors of
tasks i2 and i3 are assigned to a station. The conditions of Qi1 ,i2 = 1 and Qi1 ,i3 = 1 ensure that task i1 is not the successor of
tasks i2 and i3 . Therefore, all the successors of tasks i2 and i3 are assigned to the exit sub-line, so exchange rule 12 can be
implemented.
In condition (3.5), conditions i1 Ob, i2 Ib and Qi1 ,i2 = 1 ensure that two tasks can be exchanged according to the proof
of (1.2) in Theorem 1. In addition, condition i3 If Ie ensures that task i3 can be moved to station j as a pure entrance
task. Therefore, three tasks can be exchanged according to Lemma 1.
In condition (3.6), condition i1 Oe can ensure that emigrant task i1 is a pure entrance task or a pure exit task. If task
i1 is regarded as a pure entrance task, then the other conditions i2 If, i3 Ib and Qi2 ,i1 = 1 ensure that three tasks can be
moved according to (3.3) in Theorem 3. If task i1 is regarded as a pure exit task, then the other conditions i2 If, i3 Ib
and Qi1 ,i3 = 1 ensure that three tasks can also be moved according to (3.5) in Theorem 3.

References

[1] H. Gkcen, K. Agpak, A goal programming approach to simple U-line balancing problem, Eur. J. Oper. Res. 171 (2006) 577585.
[2] R. Hwang, H. Katayama, A multi-decision genetic approach for workload balancing of mixed-model U-shaped assembly line systems, Int. J. Prod. Res.
47 (14) (2009) 37973822.
[3] Y. Kara, T. Paksoy, C.T. Chang, Binary fuzzy goal programming approach to single model straight and U-shaped assembly line balancing, Eur. J. Oper.
Res. 195 (2) (2009) 335347.
[4] M.D. Toksar, S.K. Isleyen, E. Gner, .F. Baykoc, Simple and U-type assembly line balancing problems with a learning effect, Appl. Math. Model. 32
(12) (2008) 29542961.
[5] A. Baykasoglu, T. Dereli, Simple and U-type assembly line balancing by using an ant colony based algorithm, Math. Comput. Appl. 14 (1) (2009) 112.
[6] U. zcan, B. Toklu, A new hybrid improvement heuristic approach to simple straight and U-type assembly line balancing problems, J. Intell. Manuf. 20
(2009) 123136.
[7] Y. Kara, C. zgven, N. Yaln, Y. Atasagun, Balancing straight and U-shaped assembly lines with resource dependent task times, Int. J. Prod. Res. 49
(21) (2011) 63876405.
[8] G.J. Miltenburg, J. Wijngaard, The U-line balancing problem, Manag. Sci. 40 (10) (1994) 13781388.
[9] C. Becker, A. Scholl, A survey on problems and methods in generalized assembly line balancing, Eur. J. Oper. Res. 168 (2006) 694715.
[10] A. Hamzadayi, G. Yildiz, A genetic algorithm based approach for simultaneously balancing and sequencing of mixed-model U-lines with parallel
workstations and zoning constraints, Comput. Ind. Eng. 62 (2012) 206215.
[11] S.M. Kazemi, R. Ghodsi, M. Rabbani, R. Tavakkoli-Moghaddam, A novel two-stage genetic algorithm for a mixed-model U-line balancing problem with
duplicated tasks, Int. J. Adv. Manuf. Technol. 55 (2011) 11111122.
[12] N. Manavizadeh, N. Hosseini, M. Rabbani, F. Jolai, A simulated annealing algorithm for a mixed model assembly U-line balancing type-1 problem
considering human eciency and just-in-time approach, Comput. Ind. Engineering 64 (2) (2013) 669685.
[13] M. Rabbani, S.M. Kazemi, N. Manavizadeh, Mixed model U-line balancing type-1 problem: a new approach, J. Manuf. Syst. 31 (2012) 131138.
[14] M. Bagher, M. Zandieh, H. Farsijani, Balancing of stochastic U-type assembly lines: an imperialist competitive algorithm, Int. J. Adv. Manuf. Technol. 54
(1-4) (2011) 271285.
[15] A. Baykasoglu, L. zbakr, Stochastic U-line balancing using genetic algorithms, Int. J. Adv. Manuf. Technol. 32 (1-2) (2007) 139147.
[16] U. zcan, T Kellegz, B. Toklu, A genetic algorithm for the stochastic mixed-model U-line balancing and sequencing problem, Int. J. Prod. Res. 49 (6)
(2011) 16051626.
[17] T.L. Urban, W.C. Chiang, An optimal piecewise-linear program for the U-line balancing problem with stochastic task times, Eur. J. Oper. Res. 168 (3)
(2006) 771782.
M. Li et al. / Applied Mathematical Modelling 48 (2017) 423439 439

[18] M.H. Alavidoost, H. Babazadeh, S.T. Sayyari, An interactive fuzzy programming approach for bi-objective straight and U-shaped assembly line balancing
problem, Appl. Soft Comput. J. 40 (2016) 221235.
[19] A. Nourmohammadi, M. Zandieh, R. Tavakkoli-Moghaddam, An imperialist competitive algorithm for multi-objective U-type assembly line design, J.
Comput. Sci. 4 (2013) 393400.
[20] I. Kucukkoc, D.Z. Zhang, Coping with model variations on parallel u-shaped assembly line congurations, IFAC Proc. Vol. (IFAC-Papers Online) 48 (3)
(2015) 20302035.
[21] I. Kucukkoc, D.Z. Zhang, Balancing of parallel U-shaped assembly lines, Comput. Oper. Res. 64 (C) (2015) 233244.
[22] A. Fattahi, S. Elaoud, S. Azer, M. Turkay, A novel integer programming formulation with logic cuts for the U-shaped assembly line balancing problem,
Int. J. Prod. Res. 52 (5) (2014) 13181333.
[23] I. Sabuncuoglu, E. Erel, A. Alp, Ant colony optimization for the single model U-type assembly line balancing problem, Int. J. Prod. Econ. 120 (2) (2009)
287300.
[24] S. Avikal, R. Jain, P.K. Mishra, H.C. Yadav, A heuristic approach for U-shaped assembly line balancing to improve labor productivity, Comput. Ind. Eng.
64 (4) (2013) 895901.
[25] H. Gkcen, K. Agpak, C. Gencer, E. Kizilkaya, A shortest route formulation of simple U-type assembly line balancing problem, Appl. Math. Model. 29
(2005) 373380.
[26] N. Hamta, S.M.T.F. Ghomi, F. Jolai, U. Bahalke, Bi-criteria assembly line balancing by considering exible operation times, Appl. Math. Model. 35 (12)
(2011) 55925608.
[27] S.D. Lapierre, A. Ruiz, P. Soriano, Balancing assembly lines with tabu search, Eur. J. Oper. Res. 168 (3) (2006) 826837.
[28] C. Otto, A. Otto, How to design effective priority rules: example of simple assembly line balancing, Comput. Ind. Eng. 69 (2014) 4352.
[29] Z. Qiaoxian, L. Ming, L. Yuanxiang, T. Qiuhua, Station ant colony optimization for the type 2 assembly line balancing problem, Int. J. Adv. Manuf.
Technol. 66 (2013) 18591870.
[30] W.B. Helgeson, D.P. Birnie, Assembly line balancing using the ranked positional weight technique, J. Ind. Eng. 12 (6) (1961) 394398.
[31] A. Scholl, Balancing and Sequencing of Assembly Lines, second ed., Physica, Heidelberg, 1999.
[32] A. Scholl, C. Becker, State-of-the-art exact and heuristic solution procedures for simple assembly line balancing, Eur. J. Oper. Res. 168 (2006) 666693.
[33] A. Otto, C. Otto, A. Scholl, Systematic data generation and test design for solution algorithms on the example of SALBPGen for assembly line balancing,
Eur. J. Oper. Res. 228 (1) (2013) 3345.

You might also like