Professional Documents
Culture Documents
A Modified Multi-Objective Genetic Algorithm For Two-Sided Assembly Line Re-Balancing Problem of A Shovel Loader
A Modified Multi-Objective Genetic Algorithm For Two-Sided Assembly Line Re-Balancing Problem of A Shovel Loader
To cite this article: Yahui Zhang, Xiaofeng Hu & Chuanxun Wu (2017): A modified multi-objective
genetic algorithm for two-sided assembly line re-balancing problem of a shovel loader, International
Journal of Production Research, DOI: 10.1080/00207543.2017.1402136
Article views: 16
A modified multi-objective genetic algorithm for two-sided assembly line re-balancing problem
of a shovel loader
Yahui Zhang, Xiaofeng Hu* and Chuanxun Wu
Institute of Intelligent Manufacturing and Information Engineering, School of Mechanical Engineering, Shanghai Jiao Tong
University, Shanghai, China
(Received 28 November 2016; accepted 31 October 2017)
This article deals with a real-life multi-objective two-sided assembly line rebalancing problem (MTALRBP) with modifi-
cations of production demand, line’s structure and production process in a Chinese construction machinery manufactur-
ing firm. The objectives are minimising the cycle time and rebalancing cost, considering some specific constraints
associated with the inevitable wait time, such as novel cycle time, idle time and balanced constraints. A modified non-
Downloaded by [University of Florida] at 03:14 26 November 2017
dominated sorting genetic algorithm II (MNSGA-II) is proposed to solve this problem. MNSGA-II employs some prob-
lem-specific designs for encoding and decoding, initial population, crossover operator, mutation operator and selection
operator. The great performance of MNSGA-II is demonstrated from two aspects: one is through the comparison
between the representative results and current situation in the production system in terms of some ALs’ performance
evaluation index, the other is utilising the comparison between the proposed MNSGA-II and two versions of initial
NSGA-II in terms of ratio, convergence and spread.
Keywords: two-sided assembly line rebalancing; MNSGA-II; multi-objective; multiple constraints; a shovel loader
assembly line
1. Introduction
Shovel loader is a type of construction equipment machinery and widely used in infrastructure and public facilities con-
struction. Assembly is the last and most important link for the production of shovel loaders. Assembly line (AL) of sho-
vel loaders is a typical two-sided line as shown in Figure 1, operators work in parallel on the two sides (left side (L)
and right side (R)) of the conveyor belt, and a pair of stations (e.g. stations 3 and 4) on the two sides is called as a
mated-station or a workstation, and each one is called a companion for the other. Compared with one-sided ALs, it has
several advantages: shorter line length, reduced throughput time, lower cost of tools and fixtures and less material han-
dling (Bartholdi 1993).
At present, there are many researches focused on two-sided assembly line balancing problems (TALBPs) (Make,
Rashid, and Razali 2017). TALBPs can be classified as two versions (Lee, Kim, and Kim 2001): type I (TALBP-I) is
minimising stations’ number with a given cycle time; type II (TALBP-II) is minimising cycle time, while the number of
stations is predetermined. As summarised in Table 1, most of current researches focused on TALBP-I and its extensions,
which are suitable for a new yet-to-be-built assembly line (Purnomo, Wee, and Rau 2013). However, driven by changes
of customer demand, core technology and market competition, ALs’ balance is destroyed constantly, product and tech-
nics even need to upgrade. Obviously, ALs balancing is usually conducted on reconfigured lines instead of first installed
lines in the real world (Falkenauer 2005). However, although TALBP-II is suitable for reconfigured ALs (Purnomo,
Wee, and Rau 2013), it does not consider changes of processes or rebalancing cost (Gamberini, Grassi, and Rimini
2006; Makssoud et al. 2015).
In recent years, assembly line rebalancing problems (ALRBPs) have been mentioned by some researchers, but
related researches are limited. Gamberini, Grassi, and Rimini (2006) proposed a multi-criteria single-pass heuristic algo-
rithm based on TOPSIS to deal with the single-model and one-sided manual ALRBP, where operators should be
retrained to perform new tasks from new assembly process. The objectives were minimising the expected completion
costs and maximising the degree of similarity between initial balanced line and rebalanced line. The latter was taken as
the measurement of rebalancing cost. Then, Gamberini et al. (2009) did an extended research. Grangeon, Leclaire, and
Norre (2011) proposed a heuristics sequencing approach based on operations’ transfer for a vehicle assembly line
Workstation
(mated-station)
conveyor belt
Simple TALBP-I Bartholdi (1993), Hu, Wu, and Jin (2008), Wu et al. (2008), Hu et al. (2010), Yang, Zhang, and
Zhu (2016)
TALBP-I with additional Kim, Kim, and Kim (2000), Baykasoglu and Dereli (2008), Özcan and Toklu (2009a), Özcan
constraintsa (2010), Özcan and Toklu (2010a), Özbakır and Tapkan (2011), Tapkan, Ozbakir, and
Baykasoglu (2012b), Wang et al. (2014), Yuan, Zhang, and Shao (2015), Chiang, Urban, and
Luo (2016), Tang et al. (2017)
TALBP-I with novel objectivesb Lee, Kim, and Kim (2001), Özcan and Toklu (2009b), Chutima and Chimklai (2012), Li, Tang,
and Zhang (2017)
TALBP-I with additional Khorasanian, Hejazi, and Moslehi (2013), Purnomo and Wee (2014), Tuncel and Aydin (2014),
constraints and novel objectives Li et al. (2014)
Fuzzy TALBP-I Özbakır and Tapkan (2010b), Tapkan, Özbakır, and Baykasoğlu (2012a)
Mixed-Model TALBP-I Özcan and Toklu (2009c), Simaria and Vilarinho (2009), Yuan et al. (2015), Buyukozkan et al.
(2016), Yılmaz et al. (2017)
Parallel TALBP-I Özcan, Gökçen, and Toklu (2010), Kucukkoc and Zhang (2015), Ağpak and Zolfaghari (2015),
Tapkan, Özbakır, and Baykasoğlu (2016)
Mixed-Model parallel TALBP-I Kucukkoc and Zhang (2016)
TALBP-II General Lee, Kim, and Kim (2001), Kim, Song, and Kim (2009), Purnomo, Wee, and Rau (2013),
Zhang, Hu, and Cheng (2014), Tang, Li, and Zhang (2016), Lei and Guo (2016)
Parallel Ağpak and Zolfaghari (2015), Sepahi and Naini (2016)
Additional Li, Tang, and Zhang (2016)
constraints
a
Additional constraints include stochastic, zoning constraints, positional constraints and synchronous task constraints, and etc.
b
Novel objectives include maximising work relatedness, work slackness, and line efficiency, minimising workload smoothness and etc.
rebalancing problem to minimise the number of stations, and maximise the load smoothing gradually. The number of
moved tasks was applied to account the rebalancing cost. Yang, Gao, and Sun (2013) studied a mixed-model assembly
line (MAL) rebalancing problem, considering the demand structure changes with seasonal shifts. A multi-objective
genetic algorithm was designed to solve it to minimise the number of stations, workload variations at each station for
different models and rebalancing costs. The total processing time of reassigned tasks was utilised to measure rebalancing
cost. Makssoud, Battaïa, and Dolgui (2013); Makssoud et al. (2015) developed an exact algorithm for small-scale
ALRBPs with the objective of minimising the number of modifications in the rebalanced line, and the number of moved
tasks was also applied to measure the rebalancing cost. Celik, Kara, and Atasagun (2014) proposed an ant colony opti-
misation algorithm for U-lines rebalancing problem with stochastic task times, and the objective was to minimise total
rebalancing cost which contains task transposition costs, workstation opening/closing costs and operating costs of work-
stations in a particular planning horizon. In addition, there are other studies which did not consider rebalancing costs
(Ağpak 2010; Rahman 2010; Oliveira et al. 2012; Antoine et al. 2014; Sikora, Lopes, and Magatão 2017). However, to
the best of our knowledge, there are no published papers for two-sided assembly line rebalancing problems (TALRBPs)
so far.
International Journal of Production Research 3
This paper deals with a multi-objective TALRBP of a shovel loader, which is the urgent problem of a Chinese con-
struction machinery manufacturing company named SDLG. In this problem, simple tasks reassignment is no longer able
to meet the increasing market demand. Thus, we design a rebalancing algorithm to achieve tasks’ reassignment based
on the process optimisation. The aim is to modify an existing balanced line to achieve the novel production targets
using minimum cycle time and rebalancing cost. The considered constraints are composed of basic and some specific
constraints related to inevitable wait time, such as novel cycle time, idle time and balanced constraints. A modified non-
dominated sorting genetic algorithm II (MNSGA-II) is proposed. It contains some original problem-specific designs for
encoding and decoding, determination of initial population, crossover, mutation and selection. The performance of
MNSGA-II is tested from two aspects: (1) the comparison of six representative rebalancing solutions and current bal-
anced solution of production system; (2) the comparison between the proposed MNSGA-II and two versions of initial
NSGA-II. To the best of our knowledge, few researches dealt with multi-objective two-sided assembly line rebalancing
problems (MTALRBPs) and this is the first time the proposed algorithm has been applied to solve MTALRBPs.
In the next section, we present the rebalancing problem of an industrial system. The problem-specific designed algo-
rithm for the MTALRBPs is described in Section 3, results and discussions are presented in Sections 4 and 5 provides
the conclusion and future research.
Downloaded by [University of Florida] at 03:14 26 November 2017
2. Rebalancing problem
The notations used in this paper are as follows:
i,l,v a task
j,p a workstation
M the total number of tasks
N the total number of workstations
I set of tasks, I={1, 2, . . ., i, . . ., M}
J workstations (mated-stations), J = {1, 2, . . ., j, . . ., N}
(j,k) a station of mated-station j and its operation direction k. k = 1, indicates a left station; k = 2, indicates a right station.
k′ the opposite side of k, i.e. k′ = 2, while k = 1 and vice versa
ti processing time of task i.
D(i) the preferred direction of task i, D(i)= L,R or E
Ct(j,k) completion time of station (j,k)
Ct(j, k′) completion time of station( j,k′)
Ct(j) completion time of workstation j
Sjk set of tasks assigned to station (j,k)
Sjk 0 set of tasks assigned to station (j,k′)
Sj Set of tasks assigned to workstation j
x0ijk 1, if task i is assigned to station (j,k) in the initial balancing; 0, otherwise.
y0ij 1, if task i is assigned to workstation j in the initial balancing; 0, otherwise.
xijk 1, if task i is assigned to station (j,k); 0, otherwise.
yij 1, if task i is assigned to workstation j; 0, otherwise.
ct Cycle time
RC rebalancing costs P
UITAL upper bound of idle time allowed by an AL, UITAL = 2N ct M Pti
i¼1
UITW Upper bound of idle time allowed by a workstation, UITW = 2 ct i2Sj ti
θ, σ small positive integers
φ, δ coefficient for the number of tasks transfer between and within workstations, which are decided by the manager
Figure 2 illustrates the precedence diagram of tasks, each task is associated with a label (D(i), ti). Initial balanced
solution with a cycle time of 1386 s according to current assembly line layout is shown in Figure 3. The white rectan-
gles represent tasks assigned to relevant stations and their numbers are placed on them. For each task, its start time and
finish time are shown alongside the bars. The shaded rectangles indicate idle time. The shaded rectangles between two
tasks are the inevitable wait time. Take workstation1 for example, tasks 1, 2, 3, 4, 5 and 6 are assigned to left station of
workstation1, while tasks 9, 7, 8, 10, 11 and 12 are assigned to right station. At this time, for task 1, x0111 = 1 and
y011 = 1; for task 9, x0912 = 1 and y091 = 1; … and so on. Since task1 is an immediate predecessor of task 9, i.e. task 9 can
only be executed after task 1 is finished. Thus, there are 62 (inevitable) idle time at the first position of right station and
task 9′ start time is 62 and finish time is 112.
Air tank
(B) Tasks 24 installation E 178
Air tank and pipe
combination Air pipe 24 installation E 240
25 installation E 120
(C) Tasks deletion Move task10 (Rear light assembly) from AL to a related subassembly line.
(D) Tasks addition Add new task 77 (Front axle oil pipe connection and exhaust nozzle installation, 210s) on the AL.
(E) Modified (1) Operation time of task19(Oil inlet and return pipes installation) from 313s down to 200s;
operation time (2) Operation time of task 45(Oil pipe connection) from 280s down to 240s.
International Journal of Production Research 7
Downloaded by [University of Florida] at 03:14 26 November 2017
2.2.3 Formulation for ZL50F shovel loader assembly line rebalancing problem
For a task pair (i, v) of a two-sided AL, suppose task v is an immediate predecessor of task i and task i is assigned to
station p, v is assigned to the companion of p, then task i can only be executed after task v is finished. Thus, there is
inevitable wait time may exist between tasks assigned to the same station. We use IWTi to express the inevitable wait
time caused by task i, which exists in the opposite side of the station where task i is assigned to. In order to improve
the AL efficiency, additional constraints related IWTi are proposed.
( )
X
IWTi ¼ max 0; ti tl (1)
l2MTi
where MTi represents the set of matched tasks of task i, which do not have precedence relationship with i and can be
assigned to different operation direction of i.
As shown in Figure 2, assume task i is 34, then MT34 = f 31, 32, 35, 36, 37,40; 33; 62g. We can see from Figure 3,
tasks {34, 38, 39, 43} are assigned to L station of workstation 5, while tasks {40, 41, 42} are assigned to R station of
workstation 5. Obviously, there is only one matched task of task 34 contained in the R side, i.e. task 40. At this time,
IWT34 = max {0, 123–210} = 0, i.e. there is no inevitable wait time in the R side of workstation 5 caused by task 34.
(1) New cycle time constraints
New cycle time constraints guarantee completion time of one workstation, which contains inevitable wait time, is
less than cycle time.
CtðjÞ ct; 8j 2 J (2)
X X
Ctðj; k 0 Þ ¼ tu þ IWTi ; 8j 2 J ; k = 1; k 0 ¼ 2 or k = 2; k 0 ¼ 1 (5)
u2Sjk 0 i2Sjk
X X
IWTi þ IWTu UITW ; 8j 2 J ; k = 1; k 0 ¼ 2 or k = 2; k 0 ¼ 1 (7)
Downloaded by [University of Florida] at 03:14 26 November 2017
i2Sjk u2Sjk 0
Min ct (10)
X X
2
XXX
Min RC ¼ / yij y0ij þ d xijk x0ijk yij y0ij (11)
i2I j2J i2I j2J k¼1
(2) Decoding
A heuristic decoding method, which can be called single workstation tasks’ assignment procedure (SWTAP) is
designed to guide tasks’ assignment of each single workstation.
Step1. Select a workstation j and record tasks moved into (that did not contain in j in the initial solution).
Step2. Get potential tasks’ exchange solutions (PTES), i.e. the schemes of tasks shifted between left and right sides
of j, and sort them in a deceasing order of the number of exchange tasks.
Take workstation 1 of Figure 5(b), for example, there are three E-type tasks (10, 6, 8), then the set of PTES contains
seven solutions in the order of {10}, {6}, {8}, {10, 6}, {10, 8}, {6, 8} and {10, 6, 8}. The first solution {10}
10 Y. Zhang et al.
represents shifting task 10 from the left station to right station, i.e. at this time, the solution of workstation 1 is assign-
ing tasks {1, 2, 11, 5} to the left station and tasks {3, 4, 7, 6, 8, 10} to the right station.
Step3. Select the first solution h of PTES, mark tasks of left station as L, tasks of right station as R, get assignment
sequence of tasks according to weights and related rules, gradually.
Weight 1: ti * ((the number of its immediate successors on j) + 1). Rank all the tasks according weight 1 in a
descending order.
Weight 2: Task ID. Rank all the remaining tasks according to weight 2 in an ascending order.
Weight 3: The number of MTi . Rank all the tasks according to weight 3 in an ascending order.
Weight 4: Random.
Step4. Assign tasks sequentially, calculate Ct (j) and idle time between tasks each time after assigning a task. If the
idle time exceeds MITALAP and MITCWAP, enumerate 100 times in a small range to reduce idle time. If the
idle time still exceeds MITALAP and MITCWAP, delete solution h from PTES and go back to Step3;
Downloaded by [University of Florida] at 03:14 26 November 2017
otherwise, go to Step5.
X
j1
MITALAP ¼ UITAL Idlep (12)
p¼1
X
MITCWAP ¼ 2 Tct ðti þ IWTi Þ (13)
i2Sj
PM
where Idlep means the total idle time contained in workstation p; Tct means target ct, Tct = i¼1 ti =ð2N 1Þ.
Step5. If Ct (j) is greater than Tct, delete solution h from PTES and go back to Step3; otherwise, save it as a feasi-
ble solution for workstation j, and go to the next workstation until assigning all the tasks.
① operation time of task, rank tasks in a descending order of their operation time.
② number of successors, rank tasks in a descending order of their successors.
③ number of immediate successors, rank tasks in a descending order of their immediate successors.
④ task ID, rank tasks in an ascending order of their ID.
⑤ random.
International Journal of Production Research 11
Step3. Assign tasks to the earliest workstation j (they can be assigned to) in order. If they are not E-type, assign
them to their preferred (L or R) side; otherwise, assign them to the side which has less completion time.
Step4. If Ct (j) is greater than Cd, go to the next workstation j + 1. If Ct (j + 1) is still greater than Cd, assign them
to a workstation of which the completion time is the closest to Cd.
P
Cd represents the theoretical average cycle time of initial balanced line, Cd ¼ M i¼1 ti =ð2NÞ.
In this paper, the fitness function is defined as the reverse form of the objective functions, because the traditional
genetic algorithm is designed for the maximum problems where the greater the fitness function, the better the solution.
Step5. Assign tasks which are not yet assigned using a procedure of heuristic rule named improved best-fit decreas-
ing (IBFD, Khuri, Schütz, and Heitkötter 1995), as described below.
Step5.1. Create sets RS, RC, and identify RSj, RCj of current workstation j.
Step5.2. Get workstation section [mini, maxi] where tasks of RCj can be assigned to, calculate Zi (Zi = maxi – mini).
Step5.3. If Zi = 0, assign i to mini; otherwise, rank tasks under rules, in the order of ① in ascending order of Z; ②
in a descending order of their operation time; ③ random.
Step5.4. Assign tasks to j one by one, and calculate Ct (j). If Ct (j) > aCT, withdraw the assignment and assign
next task. Repeat it until assigning the last task. PN
aCT is average cycle time of parent P1, aCT = j¼1 Ct(j)=N .
Step5.5. If tasks of RCj are not empty, add them to RCj+1 and assign them to next workstation. If the tasks of RCj+1
are not empty in the end, assign them to a workstation which has the least completion time.
Take P24 as example, assume h = 0.5 M = 12, L={1, 4, 5, 8, 11, 13, 14, 17, 19, 21, 23, 24}, child O1 is created as
shown in Figure 5(d) step1. Then tasks which are not yet assigned (marked with #) are assigned using IBFD, Figure 5(d)
step2 is the eventual structure of O1.
Offspring O2 can be generated by reversing the roles of parents P1 and P2.
In order to restrict crossover of similar individuals, we propose the formulation of similarity (S) and a certain thresh-
old (W). If the similarity of two individuals is less than threshold W, they cross directly. Otherwise, they can cross with
a random one-third probability and reduce one time of crossover.
S ¼ Count=M (14)
pffiffiffiffiffiffiffiffiffi
W ¼ 0:7 þ 0:25 g=G (15)
where Count is the number of the same genes of two individuals; g represents current generation iteratively; G is the
total generations sited in the initial population.
1400 S l7
1300
S l9
1200 S l8
cycle time
1100 S r7 7 workstations
8 workstations
1000 S r8
9 workstations
900 S r9
800
40 50 60 70 80 90 100
rebalancing cost
Figure 6 shows the Pareto frontier moves from the left corner of the graph towards the bottom right corner of the
graph. Obviously, the solution on the right corner has a minimum cycle time and a maximum rebalancing cost, while
the solution on the left corner is the opposite. Therefore, they can represent all the solutions interval of Pareto frontier.
Two solutions with different number of workstations is marked as Srz and Slz, respectively, r and l represent the position
in the graph and z is the number of workstations contained in the solution. Based on the trait of our rebalancing prob-
lem, the reasonable and appropriate values of z are 7, 8 and 9.
To illustrate the effectiveness of the proposed algorithm, six representative solutions Srz and Slz (z = 7, 8, 9) are chosen
to be compared with current balanced line in terms of cycle time, rebalancing cost and some general performance crite-
ria of ALs as follows.
(1) Line efficiency (LE)
The higher the LE means that the greater the production rate of ALs (when the workstation number is fixed), since
the idle time decreases with the increase of LE.
PM
ti
LE ¼ i¼1 100% (19)
2N ct
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
,
u
uX X 2
SI ¼ t ðSTmax STjk Þ 2
ð2N Þ (20)
j2J k¼1
P
where STjk ¼
jk ti , represents the total operation time of tasks which are assigned to station (j,k);
i2S
STmax = max STjk , is the maximum workload of all the stations.
(3) Completion time smoothness index (CSI)
Cycle time of an AL is determined by the longest completion time of stations. However, differing from one-sided
AL, the completion time of a station is not always equal to the workload of a station in two-sided AL due to the inevi-
table wait time. Thus, it is inappropriate to assess balance of two-sided ALs just using SI.
Referring to the definition of SI, the completion time smoothness index (CSI) of two-sided ALs is defined as shown
in Equation (21). Obviously, the smoother CSI is, the more stable and reliable the assembly line is.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
, ffi
u
uX X 2
CSI ¼ t ðCtmax Ctðj; kÞÞ2 ð2N Þ (21)
j2J k¼1
where Ctmax = maxfCt(j,k)g, is the maximum completion time of all the stations.
(4) Monthly production capacity (PC/m)
For an AL, one of the most important factors is the production capacity. Thus, we use monthly production capacity
(PC/m) as a performance measure of ALs. PC/m is equal to the total work time of a month divided by cycle time.
14 Y. Zhang et al.
Eventually, the detailed results of PC/m along with results of LE, SI and CSI are summarised in Table 3. All the
results are listed according to the number of workstations used and we do not directly analyse the merits and demerits
between them. The computational time is the average time of 10 runs. Obviously, the time is slightly longer because of
some special designs in the proposed algorithm.
Obviously, we can get some conclusions from Table 3: (1) Due to the modification of the production demand, line
structure and production process, tasks may transfer between and within workstations to get rebalance. Especially, the
modification of line structure, i.e. adding or removing workstations, will result in a greater number of tasks reassign-
ment. Thus, the related rebalancing costs of z = 7 and 9 are higher than z = 8. (2) Cycle time decreases with the increase
of workstations’ number. But when the number of workstations is fixed, a greater rebalancing cost leads to a smaller
cycle time. Take z = 9, for example, the rebalancing cost is 94 when the cycle time is 896, the rebalancing cost is 54
when the cycle time is 1118. (3) As we all know that a smaller cycle time means a greater production capacity and
fewer workstations means less production cost. Thus, the company can make decisions considering the trade-off
between costs and production capacity as we proposed. (4) The last and most important, all the Pareto solutions
obtained by the proposed MNSGA-II are better than the initial balanced line, no matter how many workstations are
used. For instance, when z = 7, although there is rebalancing cost due to tasks reassignment, cycle time is reduced obvi-
ously. SI and CSI are enhanced significantly, which means the idle time is reduced as much as possible and approxi-
Downloaded by [University of Florida] at 03:14 26 November 2017
mately the same amount of work is carried out in each workstation. Additionally, the line efficiency increases from 51%
to [75%, 93%] and the production capacity per month increases from 306 to [416, 510], which can fully meet the needs
of the market.
Cycle time (s) Rebalancing cost LE (%) SI CSI PC/m Computational time (s)
P
jAs j
Edx
x¼1
convergenceðAs Þ ¼ (23)
jAs j
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u K
uX fk ðxÞ fk ðyÞ 2
jA j
Edx ¼ min t (24)
y
k¼1
fkmax fkmin
where Edx is the Euclidean distance between solution x in As and solution y in A*, k represents objectives’ number of
the problem, k = 1,2,…, K. fkmax and fkmin are the maximum and minimum values of kth objective function in A*, respec-
tively.
The spread of the obtained Pareto optimal solution set (As) is used to measure the distribution of Pareto solutions in
As by calculating a relative distance between consecutive solutions. Obviously, a good distribution would make all dis-
tances Rdx equal to Rdand would make Rdf = Rdl = 0 (with existence of extreme solutions in the non-dominated set).
Thus, for this metric, zero is the best and represents a uniform distribution.
s 1j
jAP
Rdx Rd
Downloaded by [University of Florida] at 03:14 26 November 2017
Rdf þ Rdl þ
x¼1
SpreadðAs Þ ¼ (25)
Rdf þ Rdl þ ðjAs j 1ÞRd
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u K
uX fk ðxÞ fk ðx þ 1Þ 2
Rdx ¼ t (26)
k¼1
fkmax fkmin
where Rdf and Rdl are the Euclidean distances between the extreme solutions in A*and the boundary solutions in As.
Rdx is the Euclidean distance between consecutive solutions in As (x = 1,2 … |As – 1|) and Rd is the average Euclidean
distance of Rdx.
(1) To clarify the effect of our modified methods on the algorithm (NSGA-II), we propose an initial version 1 of
NSGA-II (INSGA-II-V1), which is designed using the same methods of encoding and decoding, initial popula-
tion, and mutation operator as the MNSGA-II. The differences are that INSGA-II-V1 does not consider the
crossover of similar individuals in the crossover operator and the two improved methods of selection operator.
The performance comparison of MNSGA-II and INSGA-II-V1 is shown in Table 4. Bold values mean the best val-
ues among the algorithms. In addition, the Pareto frontiers obtained by MNSGA-II and INSGA-II-V1 are shown in
Figure 7. Table 4 shows MNSGA-II produces great ratio and convergence which are far better than INSGA-II-V1 for
every size solutions (z = 7, 8, 9). In terms of spread metric, MNSGA-II is also superior to INSGA-II-V1, except for
z = 7. And it is observed from Figure 7, MNSGA-II generates more solutions than INSGA-II-V1, and all of them are
on the approximated true-Pareto frontiers.
(2) In order to prove the importance of our new model, i.e. the specific constraints related to the inevitable wait
time, such as novel cycle time, idle time and balanced constraints, we propose an initial version 2 of NSGA-II
(INSGA-II-V2), which is designed using the same methods of encoding, initial population, crossover operator,
mutation operator and selection operator as the MNSGA-II. The differences are that the decoding method of
INSGA-II-V2 uses Longest Processing Time (LPT) (Talbot and Patterson 1984). It is one of the basic heuristics
cycle time
cycle time
1100 1100 1100
which are commonly used to find better solutions of tasks assignment and do not contain specific constraints
related to the inevitable wait time.
The performance of MNSGA-II compared with INSGA-II-V2 is shown in Table 5. Numbers in bold represent the
best values among the algorithms. In addition, the Pareto frontiers obtained by MNSGA-II and INSGA-II-V2 are shown
in Figure 8. It is observed from Table 5 that MNSGA-II significantly outperforms INSGA-II-V2 in terms of ratio, con-
vergence and spread metric for all the size solutions (z = 7, 8, 9). As shown in Figure 8, MNSGA-II produces more
solutions than INSGA-II-V2, and all of them are located on the approximated true-Pareto frontiers.
cycle time
cycle time
future studies, we would like to develop a new heuristic algorithm or extend the proposed algorithm to solve more com-
plicated and realistic two-sided assembly line rebalancing problems, such as stochastic two-sided assembly line rebalanc-
ing problems with multi-constraints.
Acknowledgements
The authors are grateful to the anonymous reviewers whose valuable comments and suggestions helped a lot to improve this paper
and the editors for their kind and sincere supports.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This research was supported by the National Natural Science Foundation of China [grant number 51475303].
References
Ağpak, K. 2010. “An Approach to Find Task Sequence for Re-Balancing of Assembly Lines.” Assembly Automation 30 (4): 378–
387.
Ağpak, K., and S. Zolfaghari. 2015. “Mathematical Models for Parallel Two-sided Assembly Line Balancing Problems and Exten-
sions.” International Journal of Production Research 53 (4): 1242–1254.
Ahmadi, E., M. Zandieh, M. Farrokh, and S. M. Emami. 2016. “A Multi Objective Optimization Approach for Flexible Job Shop
Scheduling Problem under Random Machine Breakdown by Evolutionary Algorithms.” Computers & Operations Research 73:
56–66.
Antoine, M., B. E. H. Hind, T. André, and P. Jean-François. 2014. “Dynamic Rebalancing of an Assembly Line with a Reachability
Analysis of Communicating Automata.” IFIP International Conference on Advances in Production Management Systems, Vol.
438, Berlin Heidelberg, 597–604.
Bartholdi, J. J. 1993. “Balancing Two-Sided Assembly Lines: A Case Study.” International Journal of Production Research 31 (10):
2447–2461.
Baykasoglu, A., and T. Dereli. 2008. “Two-Sided Assembly Line Balancing Using an Ant-Colony-Based Heuristic.” International
Journal of Advanced Manufacturing Technology 36 (5–6): 582–588.
Buyukozkan, K., I. Kucukkoc, S. I. Satoglu, and D. Z. Zhang. 2016. “Lexicographic Bottleneck Mixed-Model Assembly Line Balanc-
ing Problem-Artificial Bee Colony and Tabu Search Approaches with Optimised Parameters.” Expert Systems with Applications
50 (15): 151–166.
Celik, E., Y. Kara, and Y. Atasagun. 2014. “A New Approach for Rebalancing of U-Lines with Stochastic Task times Using Ant Col-
ony Optimisation Algorithm.” International Journal of Production Research 52 (24): 7262–7275.
18 Y. Zhang et al.
Chiang, W. C., T. L. Urban, and C. Luo. 2016. “Balancing Stochastic Two-Sided Assembly Lines.” International Journal of Produc-
tion Research 54 (20): 6232–6250.
Chutima, P., and P. Chimklai. 2012. “Multi-Objective Two-Sided Mixed-Model Assembly Line Balancing Using Particle Swarm Opti-
misation with Negative Knowledge.” Computers & Industrial Engineering 62 (1): 39–55.
Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan. 2002. “A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II.” IEEE
Transactions on Evolutionary Computation 6 (2): 182–197.
Dou, J. P., J. Li, and C. Su. 2016. “Bi-Objective Optimization of Integrating Configuration Generation and Scheduling for Reconfig-
urable Flow Lines Using NSGA-II.” The International Journal of Advanced Manufacturing Technology 86 (5–8): 1945–1962.
Falkenauer, E. 2005. “Line Balancing in the Real World.” The International Conference on Product Lifecycle Management, Lumiere
Univeristy of Lyon, France, 360–370, July 1.
Gamberini, R., A. Grassi, and B. Rimini. 2006. “A New Multi-Objective Heuristic Algorithm for Solving the Stochastic Assembly
Line Re-Balancing Problem.” International Journal of Production Economics 102 (2): 226–243.
Gamberini, R., E. Gebennini, A. Grassi, and A. Regattieri. 2009. “A Multiple Single-Pass Heuristic Algorithm Solving the Stochastic
Assembly Line Rebalancing Problem.” International Journal of Production Research 47 (8): 2141–2164.
Goldberg, D. E. 1989. Genetic Algorithm in Search, Optimization, and Machine Learning. Reading, MA: Addison-Westley.
Gonçalves, J. F., and J. R. de Almeida. 2002. “A Hybrid Genetic Algorithm for Assembly Line Balancing.” Journal of Heuristics 8
(6): 629–642.
Grangeon, N., P. Leclaire, and S. Norre. 2011. “Heuristics for the Re-Balancing of a Vehicle Assembly Line.” International Journal
Downloaded by [University of Florida] at 03:14 26 November 2017
Oliveira, F. S., K. Vittori, R. M. O. Russel, and X. L. Travassos. 2012. “Mixed Assembly Line Rebalancing: A Binary Integer
Approach Applied to Real World Problems in the Automotive Industry.” International Journal of Automotive Technology 13
(6): 933–940.
Özbakır, L., and P. Tapkan. 2010. “Balancing Fuzzy Multi-Objective Two-Sided Assembly Lines via Bees Algorithm.” Journal of
Intelligent & Fuzzy Systems 21 (5): 317–329.
Özbakır, L., and P. Tapkan. 2011. “Bee Colony Intelligence in Zone Constrained Two-Sided Assembly Line Balancing Problem.”
Expert Systems with Applications 38 (9): 11947–11957.
Özcan, U. 2010. “Balancing Stochastic Two-Sided Assembly Lines: A Chance-Constrained, Piecewise-Linear, Mixed Integer Program
and a Simulated Annealing Algorithm.” European Journal of Operational Research 205 (1): 81–97.
Özcan, U., and B. Toklu. 2009a. “Multiple-Criteria Decision-Making in Two-Sided Assembly Line Balancing: A Goal Programming
and a Fuzzy Goal Programming Models.” Computers & Operations Research 36 (6): 1955–1965.
Özcan, U., and B. Toklu. 2009b. “A Tabu Search Algorithm for Two-Sided Assembly Line Balancing.” International Journal of
Advanced Manufacture Technology 43: 822–829.
Özcan, U., and B. Toklu. 2009c. “Balancing of Mixed-Model Two-Sided Assembly Lines.” Computers & Industrial Engineering 57
(1): 217–227.
Özcan, U., and B. Toklu. 2010. “Balancing Two-Sided Assembly Lines with Sequence-Dependent Setup times.” International Journal
of Production Research 48 (18): 5363–5383.
Özcan, U., H. Gökçen, and B. Toklu. 2010. “Balancing Parallel Two-Sided Assembly Lines.” International Journal of Production
Downloaded by [University of Florida] at 03:14 26 November 2017
Yuan, B., C. Zhang, and X. Shao. 2015. “A Late Acceptance Hill-Climbing Algorithm for Balancing Two-Sided Assembly Lines with
Multiple Constraints.” Journal of Intelligent Manufacturing 26 (1): 159–168.
Yuan, B., C. Zhang, X. Shao, and Z. Jiang. 2015. “An Effective Hybrid Honey Bee Mating Optimization Algorithm for Balancing
Mixed-Model Two-Sided Assembly Lines.” Computers & Operations Research 53: 32–41.
Zhang, Z., J. Hu, and W. Cheng. 2014. “An Ant Colony Algorithm for Two-Sided Assembly Line Balancing Problem Type-II.”
Advances in Intelligent Systems and Computing 213: 369–378.
Appendix 1
The crowding distance computation of NSGA-II
In the NSGA-II, crowding distance is executed after the non-dominated sorting of solutions in one population, which is used to sort
the solutions in the same Pareto front, in order to estimate of the density of solutions surrounding a particular solution in the popula-
tion. The crowding distance of one solution is the average side length of the cuboid which is formed by its nearest neighbours (solu-
tions) in the Pareto curve of the objective space, just as shown in Figure A1. As for a solution, the higher the crowding distances
value, the better the solution. To maintain the density of solution, boundary solutions (solutions with smallest and largest function val-
ues) are assigned an infinite distance value ensuring they are always selected. Take a non-dominated set P, for example, the procedure
of the crowding distance computation is described below.
Downloaded by [University of Florida] at 03:14 26 November 2017
Appendix 2
The true-Pareto frontier and its approximation
A multi-objective optimisation problem (MOP) is a scenario in which at least two objectives needed to be optimised simultaneously
(Chutima and Chimklai 2012). Assume that one problem contains k objectives and fi(X) represents the ith objective function (i = 1, 2,
… , k). For a solution, if there are no other solutions better than it considering all objectives in the feasible solution space, then this
solution is called a non-dominated (superior) solution or a Pareto optimal solution. Take a maximisation problem, for exmple, a deci-
sion vector (solution) X dominates a decision vector Y (written as X Y) if and only if:
fi ð X Þ fi ðY Þ; for alli 2 f1; 2; . . .; kg (A1)
and
All the non-dominated solutions in the population can be called a Pareto optimal set or efficient set. The collective region of the
points in the Pareto optimal set along a curve in the objective space are termed the Pareto optimal frontier (front) or efficient frontier
(front).
As shown in Figure A2, assume the shaded region is the feasible search region for a minimum problem, all the points (solutions)
on the lower boundary of the shaded area represent the true-Pareto solutions and they can be called true-Pareto set. The true-Pareto
set is unknown generally (Chutima and Chimklai 2012), thus, an approximation is needed. The approximated true-Pareto set is the
result of combining all final non-dominated solutions obtained from all the algorithms, applying Goldberg’s Pareto ranking technique
(Goldberg 1989) to the combined solutions, and the first frontier of the combined solutions is the approximated true Pareto frontier
(set).