Clemson University

TigerPrints
All Dissertations Dissertations

8-2017

Balancing and Sequencing of Mixed Model

Assembly Lines
Anas Alsayed Alghazi
Clemson University, [email protected]

Follow this and additional works at: https://tigerprints.clemson.edu/all_dissertations

Recommended Citation
Alghazi, Anas Alsayed, "Balancing and Sequencing of Mixed Model Assembly Lines" (2017). All Dissertations. 2022.
https://tigerprints.clemson.edu/all_dissertations/2022

This Dissertation is brought to you for free and open access by the Dissertations at TigerPrints. It has been accepted for inclusion in All Dissertations by
an authorized administrator of TigerPrints. For more information, please contact [email protected].
BALANCING AND SEQUENCING OF MIXED MODEL
ASSEMBLY LINES

A Dissertation
Presented to
the Graduate School of
Clemson University

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
Industrial Engineering

by
Anas Alsayed Alghazi
August 2017

Accepted by:
Dr. Mary E. Kurz, Committee Chair
Dr. David Neyens
Dr. Kevin Taaffe
Dr. Joshua Summers
 ABSTRACT

Assembly lines are cost efficient production systems that mass produce identical

products. Due to customer demand, manufacturers use mixed model assembly lines to

produce customized products that are not identical. To stay efficient, management

decisions for the line such as number of workers and assembly task assignment to stations

need to be optimized to increase throughput and decrease cost. In each station, the work to

be done depends on the exact product configuration, and is not consistent across all

products.

In this dissertation, a mixed model line balancing integer program (IP) that

considers parallel workers, zoning, task assignment, and ergonomic constraints with the

objective of minimizing the number of workers is proposed. Upon observing the limitation

of the IP, a Constraint Programming (CP) model that is based on CPLEX CP Optimizer is

developed to solve larger assembly line balancing problems. Data from an automotive

OEM are used to assess the performance of both the MIP and CP models. Using the OEM

data, we show that the CP model outperforms the IP model for bigger problems. A

sensitivity analysis is done to assess the cost of enforcing some of the constraint on the

computation complexity and the amount of violations to these constraints once they are

disabled. Results show that some of the constraints are helpful in reducing the computation

time. Specifically, the assignment constraints in which decision variables are fixed or

bounded result in a smaller search space. Finally, since the line balance for mixed model

is based on task duration averages, we propose a mixed model sequencing model that

ii
minimize the number of overload situation that might occur due to variability in tasks times

by providing an optimal production sequence. We consider the skip-policy to manage

overload situations and allow interactions between stations via workers swimming. An IP

model formulation is proposed and a GRASP solution heuristic is developed to solve the

problem. Data from the literature are used to assess the performance of the developed

heuristic and to show the benefit of swimming in reducing work overload situations.

iii
 DEDICATION

To my parents for their support, encouragement and prayers.

To my wife for her care, endurance, and endless love.
To my children for being the joy of my life.

iv
 ACKNOWLEDGMENTS

First and foremost, all praise and thanks to Allah the most merciful and the most

gracious for his guidance and blessings he bestowed upon me and my family.

I would like to express my deepest gratitude to my dissertation advisor, Dr. Mary

Beth Kurz for her guidance and support through this journey. I am honored to have her as

my advisor, her support in my research is unmatched by anything else. The dedication and

attention she pays to her students and their work is amazing. She will always be one of my

great role models. I would like also to show my gratitude to my committee members Dr.

Joshua Summers, Dr. Kevin Taaffe, and Dr. David Neyens for their valuable advice and

comments. I am also grateful for the valuable comments and support that our Department

Chair Dr. Cole Smith has provided to me in this journey.

Special thanks are extended to the faculty members and staff in the Department of

Industrial Engineering. I am thankful for Martin Clark and Brandon March for the technical

support they provided to me. Words of appreciation are due to my colleagues in the

Department of Industrial Engineering and to my friends. Special thanks to Bryan Pearce

for the fruitful discussions we used to have about the line balancing problem. Thanks to

him and to the team at BMW for effort they made in collecting data.

I am thankful to my parents Alsayed Alghazi and Afaf Bahjat for their guidance

and prayers, my brothers and sisters for their unconditional love and support. I am thankful

to my lovely wife and friend Afaf Effat who has been always on my side during times of

serenity and stress. Without her love and support this work would have been much harder.

v
I cannot thank her enough for putting up with me during my studies. Finally, I am thankful

to my children Albatool, Omar, Sofana, and Suhyla for being the gifts that always reminded

me what is important.

vi
 TABLE OF CONTENTS

Page

TITLE PAGE ........................................................................................................ i

ABSTRACT ......................................................................................................... ii

DEDICATION .................................................................................................... iv

ACKNOWLEDGMENTS ................................................................................... v

LIST OF TABLES .............................................................................................. ix

LIST OF FIGURES ..........................................................................................xiii

CHAPTER

1. INTRODUCTION ......................................................................................... 1

Assembly line balancing problem ............................................................ 1

Constraint programming background ...................................................... 7
Integer programming and constraint programming ............................... 12
Contribution and research questions ...................................................... 14
Dissertation organization ....................................................................... 16

2. MIXED MODEL LINE BALANCING WITH PARALLEL STATIONS,

ZONING CONSTRAINTS, AND ERGONOMICS ........................... 18

Introduction ............................................................................................ 18
Problem description ............................................................................... 21
Related work .......................................................................................... 33
Motivation .............................................................................................. 46
The mathematical model ........................................................................ 47
The constraint programming model ....................................................... 53
Computational experiments ................................................................... 58
Conclusion ............................................................................................. 62

vii
Table of Contents (Continued)
Page

3. A CASE STUDY ......................................................................................... 64

Introduction ............................................................................................ 64
The line balancing constraints ............................................................... 65
Experimental configuration ................................................................... 67
Results .................................................................................................... 68
Discussion .............................................................................................. 73
Conclusion ............................................................................................. 79

4. MIXED MODEL SEQUENCING PROBLEM ........................................... 81

Introduction ............................................................................................ 81
Mixed model sequencing versus car sequencing ................................... 82
Problem description ............................................................................... 84
Related work .......................................................................................... 86
Model formulation ................................................................................. 98
Mathematical model............................................................................. 100
Solution approach ................................................................................ 103
Computational experiments ................................................................. 110
Conclusion ........................................................................................... 121

5. CONCLUSION AND FUTURE WORK .................................................. 123

REFERENCES .......................................................................................... 126

APPENDIX ................................................................................................ 132

viii
 LIST OF TABLES

Table Page

1. Example data for the line balancing problem ................................................... 30

2. Assignment constraints ..................................................................................... 32

3. Summary of related literature ........................................................................... 44

4. Line balancing problem’s notations.................................................................. 49

5. Number of tasks in each band ........................................................................... 58

6. Test instances .................................................................................................... 59

7. Solution time under time limit .......................................................................... 60

8. IP solution time under no time limit ................................................................. 61

9. Test data sets ..................................................................................................... 67

10. Sensitivity to maximum ergonomic score in Band 1 ........................................ 68

11. Sensitivity to maximum ergonomic score in Band 26 ...................................... 69

12. Sensitivity to maximum ergonomic score in Band 47 ...................................... 69

13. Band 1 sensitivity to constraints ....................................................................... 70

14. Band 26 sensitivity to constraints ..................................................................... 70

15. Band 47 sensitivity to constraints ..................................................................... 71

16. Number of IP violations in Band 1 ................................................................... 72

17. Number of IP violations in Band 26 ................................................................. 72

18. Number of IP violations in Band 47 ................................................................. 72

ix
List of Tables (Continued)

Table Page

19. Example data for the mixed model sequencing problem.................................. 92

20. Mixed model sequencing problem’s notations ............................................... 100

21. Processing time for the 9 types of engines ..................................................... 111

22. Demand plans for 46 instances ....................................................................... 112

23. IP results for two versions of Gurobi .............................................................. 114

24. IP vs GRASP Block 1 results ......................................................................... 116

25. IP vs GRASP Block 2 results ......................................................................... 117

26. Effect of swimming on Block 1 overload situations ...................................... 119

27. Effect of swimming on Block 2 overload situations ...................................... 120

x
 LIST OF FIGURES

Figure Page

1. Types of assembly lines ...................................................................................... 4

2. Propagation and arc consistency ....................................................................... 10

3. Product zone map ............................................................................................. 23

4. Worker interference example ........................................................................... 23

5. Option-based joint precedence graph ............................................................... 30

6. Mounting position for each task ....................................................................... 31

7. Gantt chart with tasks at early start times ......................................................... 31

8. An optimal solution to the example .................................................................. 32

9. Example for a worker/station profile ................................................................ 48

10. Band 1 ergonomic score histogram .................................................................. 74

11. Band 1 mounting positions histogram .............................................................. 74

12. Band 26 ergonomic score histogram ................................................................ 75

13. Band 26 mounting positions histogram ............................................................ 76

14. Band 26 ergonomic score histogram ................................................................ 77

15. Band 47 mounting positions histogram ............................................................ 78

16. Movement in the side-by-side policy ............................................................... 93

17. Movement in the skip policy ............................................................................ 94

18. Movement with swimming allowed ................................................................. 95

xi
List of Figures (Continued)

Figure Page

19. Worker movement in an optimal solution for the example .............................. 96

20. Movement of a worker in a station ................................................................... 98

21. Greedy randomized algorithm for a minimization problem ........................... 105

22. Insertion operator ............................................................................................ 106

23. Exchange (swap) operator .............................................................................. 106

24. Inversion operator ........................................................................................... 106

25. Local search algorithm ................................................................................... 107

26. Computing the objective function by evaluating a sequence ......................... 108

xii
 CHAPTER ONE

1 INTRODUCTION

1.1 Assembly line balancing problem

An assembly line is a flow-oriented production system in which final products are

assembled from components while flowing through a sequence of serially aligned

production units known as stations. Production is initiated with a job which is a workpiece

that is launched down the line. This workpiece flows through stations using a transportation

system such as a conveyor belt. At each station a number of operations, known as tasks,

are performed on the workpiece. Each task requires a certain amount of time to be

performed in addition to other requirements like tools. Due to technological constraints or

the physical structure of the product, there are relations between tasks that necessitate

finishing some tasks before starting others. These constraints are captured in a precedence

graph in which each task is represented by a node and each precedence relation is

symbolized by a directed arc.

The assembly line balancing problem (ALBP) is the decision problem of finding a

feasible line balance; in other words, assigning each task to one station such that the

precedence constraints and any additional restrictions are fulfilled. For the case of paced

assembly line, the cycle time determines the maximum time a workpiece can spend at each

station. The line balance is feasible only if the total sum of task durations at any given

station, named the station load, does not exceed the cycle time. In addition, if the station

load is less than the cycle time then the idle time for each cycle would be the difference

between the station load and the cycle time. The word “balancing” stemmed from the fact

1
that minimizing the total idle time across all station would yield a balanced workload across

all stations which can be achieved by minimizing the number of workstations. This

problem was named the simple assembly line balancing problem (SALBP) by Baybars

(1986). The assumptions for the SALBP are: production of one product, paced line with

fixed cycle time, deterministic operation times, no assignment restrictions other than the

precedence constraints, serial line layout with one sided stations, and identical stations in

terms of machines and workers.

1.1.1 Classification

Different objectives can be used in the decision problem of assembly line

balancing; the objective used should reflect the strategic goals of the organization. While

using a cost/profit type of objective is tempting, it may not be the best choice because of

the errors that may occur when estimating either the cost of operating the line over a long

period of time or the profit from selling the final products. For this reason, other objectives

are used which define four versions of the SALBP. The first version utilizes the objective

of minimizing the number of stations for a given fixed cycle time which is referred to as

SALBP-1. The second version utilizes the objective of minimizing the cycle time for a

given fixed number of stations which is named SALBP-2. The third version is SALBP-E

which maximizes the line efficiency, and the fourth version is SALBP-F which searches

for a feasible solution given a fixed number of station and cycle time. (Scholl, 1999)

Traditionally the research in the assembly line balancing field was focused on the

SALBP with all of its restricting assumptions. However, recently a lot of research is

generalized because real life line balancing problems require more than the set of limiting

2
assumptions of the SALBP. In the simple problem, other practical consideration are

overlooked in the model such as having more than one product, parallel work stations and

tasks, zoning restrictions, processing alternatives, stochastic processing times and U

shaped assembly lines. According to Baybars (1986), the attempts to extend the SALBP

by integrating practical issues to make the problem more realistic are categorized under the

general assembly line balancing problem (GALBP). In order to structure the research in

the field of ALBP in an attempt to close the gap between academic discussion and practical

applications, Boysen et al. (2007) introduced a tuple notation similar to the one adopted in

the machine scheduling field so that any ALBP can be easily characterized by it. This tuple

notation characterizes the problem according to: precedence graph characteristics, station

and line characteristics, and objectives. Using this notation more than half a billion

variations of the ALBP can now be easily characterized. This notation also accommodates

more realistic ALBP after the structure of SALBP have been well studied and more

research is going to be geared towards the applicable realistic line balancing problems.

1.1.2 Mixed Model Assembly Lines

Assembly lines in which one product is assembled are categorized as single model.

If more than one model or version of the product is assembled on the same line, the

assembly line is categorized either as mixed-model or multi-model depending on how these

different models are intermixed on the assembly line. For random inter-mixing of models

the problem of balancing the line is referred to as the mixed model assembly line balancing

problem (MALBP). The different line types are shown in Figure 1 in which different

models are represented by different geometric shapes. Usually models may differ from

3
each other in size, color, material or any other specifications such that their production may

require new tasks, different task times, or different precedence constraints. However, most

of the time all models contain identical (or interchangeable) parts and usually their

production process is similar. The difference could be as simple as options that are either

present or absent from the desired model. (Becker & Scholl, 2006)

Figure 1: Types of assembly lines

1.1.3 Mixed Model Line Balancing and Sequencing

In managing the mixed-model assembly line, several decision problems with

different planning horizons need to be solved. For the long term horizon, the decision on

the assignment of tasks to workstations (line balancing) needs to be taken. This includes

other decisions like determining the number of workstations, station workload, production

rate or equivalently, the cycle time. All of these decisions are part of the MALBP. As in

the single model problem, there are different versions of the MALBP depending on the

4
objective sought. However, the mixed model problem is much harder than the single model

as station utilizations’ need to be balanced with respect to each model or even across all

models. Thus, single model objectives are not adequate in this case. In addition, the cycle

time for each of the mixed models is different from the one used in the single model case

and it is not considered as an upper bound to the station time because average task durations

are used for the mixed model case. So if there are some models with station times that are

significantly smaller than the cycle line, it is expected that for some other models the station

time will exceed the cycle time. This in turn will lead to stations with idle time or work

overload depending on the sequence of models production.

For mixed model line management, the long term decision of determining the line

balance and the desired production rate is not enough. This is due to the inefficiencies that

result from the variation of station utilizations with respects to the models, which is a result

of using averages of task durations across all models to come up with line balance. To deal

with this, a short term decision problem is used to find a sequence of models that satisfies

the demand over the production period and optimize some given objective. This problem

is named the mixed model sequencing problem (MMS). This problem is considered short

term since it arises every production shift which could be daily or weekly. The balancing

and sequencing problems are interrelated and each depends on the other. The actual

performance of a line balance that is achieved by using task averages depends on the order

of the production sequence. If models with high task times are consecutive in a production

sequence some stations will require more than the cycle time to finish the assembly tasks

assigned to it. This will cause a work overload that disturbs the assembly line and might

5
cause it to stop incurring costs. One way to avoid work overload is to alternate high and

low demanding models such that work overload is minimized. Mixed model sequencing

takes into account worker movement in stations and uses different model task duration to

determine the best production sequence that minimizes the work overload.

6
1.2 Constraint programming background

Operations Research (OR) has been the traditional field of research for solving

combinatorial problems like scheduling. In OR, solutions are achieved by solving simple

mathematical models that heavily exploit the combinatorial structure of the scheduling

problem to get the best performance. So “OR” is more about achieving a high level of

efficiency by using the proper solution algorithm. However, simplifying the problem by

imposing non-practical assumptions and removing constraints is unavoidable in order to

solve a practical combinatorial problem like scheduling using the classical OR models. In

some cases, this makes the solutions to these simplified models uninteresting because they

are not applicable as a solution to the real problem. On the other hand, artificial intelligence

(AI) research in scheduling tends to explore more general scheduling models and solution

paradigms. AI algorithms focus on providing general algorithms to solve a wide range of

problems. Because of the emphasis on general algorithms and when compared to OR

algorithms, AI algorithms may perform poorly on some problems. Constraint programming

was introduced to combine the benefits of OR and AI such that we have algorithms that

are both efficient and can be applied to a wide variety of problems (Baptiste et al. 1995).

Constraint programming (CP) is a declarative programming paradigm that was

initially influenced by computer science, or specifically artificial intelligence (AI), and

programming languages. However, recently operations research (OR) has been a major

influence to CP. The interface and modeling aspect of CP is close to the computer science

field but recently the solution strategies and efficient algorithms in CP has been influenced

by OR. Generally, CP is used to solve what is called the constraint satisfaction problem

7
(CSP). An instance of the CSP is composed of a set of variables, a domain for each variable

that defines the values to which the variable may be assigned, and a set of constraints that

define the relation between the values of one or several variables. Solving a CSP involves

assigning values to variables such that all constraints are satisfied. If a feasible solution is

found by assigning values from the domain to each variable, the problem is said to be

satisfiable. If no assignment of values to variables from their domains that satisfy all

constraints, the problem is said to be unsatisfiable.

Sometimes it is not useful to only get a feasible solution; instead we want to find a

solution that is optimal with respect to a certain criterion. This problem is called the

constraint optimization problem (COP) which is simply a CSP with an objective function.

CSP can be modified to solve COP by creating an objective variable that represents the

objective function. When an initial feasible solution is found, a new objective constraint is

added that forces the (new) objective variable to be better than the initial solution found.

This is repeated every time a new better feasible solution is found until the problem

becomes unsatisfiable in which the last feasible solution found is the optimal solution to

the COP. It should be noted that there are different approaches to tackle the CSP. One

approach is using IP techniques such as the cutting plane method, and the Branch and

Bound algorithm (B&B). Another approach is by using metaheuristics such as Simulated

Annealing (SA), Tabu Search (TS), and Genetic Algorithms (GA). The term constraint

programming in general means a computer implementation to solve the CSP but it has been

also used in the literature to denote implementing these algorithms in a conventional logic

programming language. Enhancements to logic programming languages have been made

8
to overcome some of the insufficiencies and new languages have been developed such as

Constraint Handling in PROLOG (CHIP). During this time, some authors started using the

term constraint logic programming (CLP) in the literature in place of CP which might be

confusing. Another approach to tackle CSP was implementing CP using general purpose

programming languages like C++ or a special declarative language such as the OPL

language. Several commercial packages provide a graphical interface environment for CP

such as IBM ILOG CPLEX that is based on the OPL language or the Fico Express

optimization suite which is based on the Mosel language. Recently, algebraic modeling

languages such as AMPL and AIMMS have been extended to allow CP. There are also

many free CP modeling languages that can be considered medium level modelling

languages that usually require a low-level CP solver. The area of CSP research has received

more attention lately with the advancement in computer hardware which has made efficient

CP algorithms possible (Brailsford et al. 1999).

CP variables may have different types of domains, such as integer, logical, or any

non-numerical domain. Also, unlike linear programming, different types of constraints can

be used for each type of variable such as arithmetic and logical constraints. The solution

procedure for CP involves a combination of domain reduction, constraint propagation, and

search. To demonstrate these techniques, a CSP can be presented as a graph in which the

nodes are variables represented by their domain and the arcs are the constraints between

these variables. We use examples from Fromherz (2001) to demonstrate domain reduction

and constraint propagation techniques. Domain reduction is the direct application of

constraints on a variable’s domain. For example if integer variable X has a domain [0,10]

9
and there is a direct constraint X>3, the domain of X is reduced to [4,10]. If we have another

integer variable Y with the domain [0,10] that is connected to X via a constraint X<Y-3,

then the changes in the domain of X are propagated to change the domain of Y to be [7,10].

The change in Y’s domain is propagated again to change the domain of X to [4,7]. This

process is called domain propagation and results in what is called arc consistency in which

inconsistent values are removed from the variables’ domains along the arcs of the graph.

The process of constraint propagation and arc consistency are illustrated in Figure 2.

Figure 2: Propagation and arc consistency

The main aim of constraint propagation and domain reduction is to remove values

from variable’s domains that will certainly lead to infeasible solutions in hopes to reduce

the search space. Usually constraint propagation is denoted as an incomplete procedure as

not all inconsistent values are removed from the variables’ domains. If all domains are

reduced to one variable then a feasible solution might be found as a result of constraint

10
propagation only but a feasible solution is not guaranteed. In most cases search is usually

needed to supplement constraint propagation. The search procedure enumerates

assignments of values to variables, when a variable is fixed during search constraint

propagation is used to reduce the domains of variables that are not fixed in order to reduce

the search space.

The search in CP is usually done as a depth-first tree search with domain filtering.

It is a constructive algorithm meaning that each variable is assigned a value incrementally

until a feasible solution is found or a constraint is violated which implies infeasibility. If a

constraint is violated the last variable assignment is undone and the variable is assigned

another value from its domain. If there is still no feasible solution backtracking occurs in

which the search backtracks to a previously assigned variable. The search continues until

all branches in the search tree are explored. The search efficiency is affected by the way

the search tree is built in terms of variable ordering; also value ordering is an important

factor. There are several heuristics that can be used to choose a better variable and value

ordering in order to minimize backtracking and increase the search efficiency. In addition,

there are special heuristics that can be used to exploit the structure of some problems and

in turn increase the search efficiency by improving domain filtering which is the process

of removing variables values that will not be part of any feasible solution. Specialized

algorithms are developed for different constraints to help with domain filtering. Also,

evolutionary algorithms, dominance rules, large neighborhood search, and machine

learning are examples of techniques that are incorporated in CP search mechanism.

11
1.3 Integer programming and constraint programming

Research in the OR field developed a mathematical programming framework with

tools such as linear programming (LP) and integer programming (IP) to solve

combinatorial optimization problems, while research in the AI field and specifically

constraint satisfaction and logic programming along with some of OR ideas developed the

CP framework. What is common between IP and CP is that their solution method is based

on a tree search in which branching is used to explore the search space. The main difference

between the two is how the tree is explored and how the branching decisions are made at

each node of the search tree. For example, in IP the branching is done on a node using LP

sub-problems that are generated by relaxing the integrality constraints. The node is

fathomed in case the LP sub-problem has no feasible solution, or has an integer solution

that is better than the best solution found, or has a solution that is worse than the best

solution found. If the solution to the sub-problem is better than the best solution found yet

does not satisfy the integrality constraint a new branching step is required. This process is

repeated until the search terminates when it is proven that there is no feasible solution

available that is better than the best solution found. On the other hand, CP uses domain

reduction and constraint propagation at each node to reduce the domain of each variable.

The node is fathomed if domain reduction reduced all domains such that they are empty,

and the search continues to enumerate all remaining possible solutions after further domain

reduction and constraint propagation steps are done. The search terminates when all nodes

have been fathomed. It should be noted that the constraint propagation algorithm used is a

key factor in the efficiency of the CP (Jain & Grossmann 2001).

12
 Hooker (2002) compared optimization to CP and gave several examples on the

differences. The first is the word “programming” in CP which refers to computer

programming since the problem is usually formulated using a declarative modeling

language. Usually the modeling language “programming” in CP gives the user great control

over the search procedure using a lower level language like Java or C++. However the

word “programming” in linear programming is not related to computer programming as

the name originated when George Dantzig developed linear programming formulations to

solve planning problems (or programs) in the US Air Force back in 1946. Another

distinction is the use of inference in both CP and IP. In CP inference is done to reduce the

search space directly using domain reduction and constraint propagation. On the other

hand, cutting planes in IP might be considered an inference tool that creates better

relaxations and speed up the search indirectly.

13
1.4 Contribution and research questions

Much of the research done in the line balancing problem literature does not consider

most of the real life restrictions that are facing decision makers managing assembly lines.

In well-known car manufacturing plants, assembly line balancing is still done manually

over several days since no applicable tool is available for them. The GALBP started to

appear in the literature in which the limited assumptions of the SALBP are relaxed so that

it becomes more applicable in real life instead of being just in academic research papers.

However, research in the GALBP usually concentrates on one generalization and extends

the model in that direction only. In this research, the MALBP with parallel stations, zoning

constraints, and assignment restriction is extended by adding more detailed task assignment

constraints, worker interference constraints, and ergonomics. For task assignment

constraints, instead of having two constraints for station/worker incompatibility, the

extended model has four constraints for tasks that need to be done by either the same

worker, or different workers and tasks the need to be done in the same station, or in

different stations. In the extended model, tasks are assigned to workers instead of fixed

workplaces that are tied to certain mounting positions. This way each worker can be

assigned any task regardless of the mounting position. To avoid worker interference, a new

constraint is added to avoid the overlap of any two tasks that share the same mounting

position in any station. Furthermore, unproductive movement is avoided by limiting each

worker from being assigned tasks with opposite mounting positions. Finally, ergonomics

are added to the extended MALBP model so that worker health is not endangered by

assigning him too many ergonomically demanding tasks that might lead to a lifelong injury.

14
 In this research, an IP is presented for our extended MALBP model with parallel

stations, zoning constraints, assignment restrictions, and ergonomics. A CP model is

proposed to solve the problem using the scheduling module of the CPLEX CP optimizer.

To the best of our knowledge, there is no published research that uses CP to solve the

MALBP with parallel stations, zoning constraints, assignment restrictions, and

ergonomics.

Subsequently, an extended sensitivity analysis is done using data from an

automotive OEM to measure the cost of adding the different constraints on the processing

time needed to reach the optimal solution. The problems’ constraints are divided into hard

constraints that make up the assembly line balancing problem, and soft constraints that can

be relaxed without violating the basic assumptions of the line balancing problem.

Experiments are done to build intuition about the relative cost of modeling and forcing

some of the realistic constraints in the mixed model line balancing problem.

Because the mixed model line balancing is done based on task average times, the

related problem of MMS is studied to minimize any disturbance to the assembly line that

might be caused by work overload in any station. An IP is proposed with the objective of

minimizing the work overload situations. The assumption for the proposed MMS model is

based on the skip policy in which the utility worker works exclusively on any work piece

whenever an overload situation is foreseeable while the normal worker skips this work

piece to the next one in sequence. This problem is discussed in the literature with

assumption of closed stations. That is no worker can swim past his station’s boundaries to

15
complete a task if needed. In this research, we extend the work done by assuming open

stations and develop a heuristic to solve the problem.

The research questions addressed in this dissertation are as follows:

1- Can we model the mixed model line balancing problem with parallel stations,

zoning, assignment, and ergonomic constraints as a scheduling CP? If so how

efficient is the CP model when compared to the IP in terms of computation time?

2- How sensitive are the IP and CP models’ computation time to the maximum

ergonomic score limit? What is the impact of removing some of the constraints on

the solution quality and computation complexity?

3- Is there a more efficient way to solve the mixed model sequencing problem? What

is the benefit of considering workers swimming?

1.5 Dissertation organization

This dissertation is organized as follows. Chapter 1 introduces the dissertation by

defining the assembly line balancing problem and presenting the different classification of

the problem. This is followed by introducing the mixed model assembly lines and how are

they different from the single and multi-model assembly lines. Then, the interrelated

problems of mixed model balancing and sequencing are defined. Constraint programming

background and its relation to integer programming is given next. Chapter 1 concludes

with research contribution and organization.

Chapter 2 presents the problem of mixed model line balancing with parallel

stations, zoning constraints, and ergonomics. The problem description is given first using

an example. Then a literature review on related work is presented followed by the

16
motivation for this research. The mathematical model along with the integer program is

presented followed by the proposed constraint programming model. The performance of

the proposed models is shown in the computation experiments section along with

discussion of the results.

In Chapter 3, a case study is presented in which sensitivity analysis is done on both

the proposed models to assess the cost of enforcing some constraints on the computation

complexity along with the amount of violations occurred when a constraint is disabled.

Chapter 4 introduces the problem of mixed model sequencing and how is it different

from the car sequencing problem. This is followed by a review in related work. Next,

problem description is given and illustrated by an example. Model formulation is presented

in the next section followed by the mathematical integer programming model.

17
 CHAPTER TWO

2 MIXED MODEL LINE BALANCING WITH PARALLEL

STATIONS, ZONING CONSTRAINTS, AND ERGONOMICS

2.1 Introduction

An assembly line is a flow-oriented production system that is composed of

productive areas called stations arranged in a serial manner. The final product starts as a

work piece that is launched down the line passing by all stations on some kind of

transportation system such as a conveyor belt. At each station, a number of assembly

operations are performed on the work piece. The amount of time each work piece is within

the boundaries of a station depends on the conveyor speed, which is also defined by the

interval between launching work pieces, is called the cycle time. Generally, the total

amount of assembly operations done on the work piece by a single operator at any given

station cannot exceed this cycle time. In addition, the assembly operations required to be

done on the work piece have some precedence constraints because of technological and

organizational conditions. The decision problem of assigning the different assembly

operations to stations with respect to some objective and satisfying the cycle time limitation

and the precedence requirement between operations is referred to as the assembly line

balancing problem.

Initially, assembly lines were used as a cost efficient way for mass production of

identical products. However, in order to respond to customer needs, companies added the

option to customize the products they offer. Automation and multipurpose machines made

it easier to produce customized products on the same line in a cost effective way similar to

18
the mass production counterparts. Yet these assembly lines are associated with

considerable investment costs, which makes the configuration of these lines an important

factor to still make it cost efficient. The configuration of the line includes management

decisions that are related to setting system capacity, such as the cycle time, as well as task

assignment to stations.

Most of the research in the configuration planning of assembly systems is focused

on the problem of line balancing. Since most of the work in the literature is based on

simplifying assumptions, the research was named simple assembly line balancing problem

(SALBP) in the review paper by Baybars (1986). Some later studies extended the simple

problem by incorporating practical aspects such as parallel stations and zoning constraints;

this type of research is categorized under the general assembly line balancing name in the

survey by Becker and Scholl (2006). Despite this effort in extending the problem to be

more realistic, there is still a gap between academic papers and practical implementation.

To this day, many automotive assembly organizations still line balance manually. Boysen

et al. (2007) tried to explain the reasons behind this gap between the academia and practice.

The first reason is that researchers have not considered the true real world problem yet.

The second reason is that the problem was studied but no satisfactory solution was found

to it. Finally, the third reason is that scientific results could not be translated into something

practical. The classification scheme proposed in Boysen et al. (2007) was the first step to

close this gap between academia and real life.

The assembly line balancing problem is not just the long term decision of efficiently

designing the assembly line by determining the number of stations required, as re-balancing

19
is needed periodically when production processes, overall demand or demand mix changes.

Choosing the optimization objective impacts the resulting solution and depends on the

strategic goals of the company. Instead of using a cost/profit type of an objective, a

surrogate objective is often used to maximize the line utilization or minimize the number

of workers needed.

Becker and Scholl (2009) consider the problem of balancing assembly lines with

variable parallel workplaces in which the objective is to minimize the number of

workplaces needed. A workplace is the combination of a worker at a station, in the case of

multiple workers per station. The model proposed is based on the simple assembly line

assumption of mass production of a homogenous product. In this work, we extend the

model of Becker and Scholl (2009) to the mixed model environment and account for

ergonomic constraints, usually ignored in line balancing models. We develop both an

integer programming (IP) model and a constraint programming (CP) model for the problem

at hand and assess the performance of both models using data from industry.

In this chapter, the problem description is given first with details about the different

constraints that define the problem at hand. This is followed by an example to illustrate the

different constraints of the problem and the solution structure in section 2.2. Next, a

literature review on related work is given in section 2.3 followed by the motivation behind

this work in section 2.4. In section 2.5 the mathematical program is given followed by the

constraint programming model in section 2.6. Computational experiments and results

discussion are presented in section 2.7. Finally, concluding remarks and future possible

research directions are given in section 2.8.

20
2.2 Problem description

We describe the decision problem that is intended to solve the realistic mixed model

line balancing problem by an example. To achieve realism, we incorporate several

conditions and limitations that are faced in an industry’s real assembly line. The constraints

that make a line balancing problem applicable in real world industrial setting are explained

in detail in Falkenauer (2005). Our problem is based on the assembly line balancing

problem with flexible parallel workstations that was introduced by Becker and Scholl

(2009) and an OEM’s environment in which the objective is to minimize the total number

of workers on the line subject to several constraints.

The characteristics of the problem that are similar to the characteristics of the

SALBP are: serial line layout, fixed cycle time (paced line), and deterministic task times

that are less than the cycle time. In addition, to capture realism the following characteristics

are included: mixed model assembly, parallel workers, zoning constraints, assignment

restrictions and ergonomic risk constraints.

2.2.1 Mixed model environment

In a mixed model environment, different base models known as variants that may

include a variety of options are assembled on the same line. Since some automakers provide

their customers with the flexibility in customizing the cars they order with the options they

want, these automakers usually operate under a make-to-order environment. The number

of options/variants combinations lead to a huge number of possible customized cars to be

assembled on the same line. The installation of different options typically leads to

21
variations in assembly task times; for example the installation of a power liftgate requires

a different amount of time when compared to the manual liftgate installation.

There are two primary approaches to deal with variation in task durations when it

comes to enforcing the cycle time constraint. The first is enforcing the cycle time on the

average task times, that is the sum of average task times assigned to a worker should not

exceed the cycle time. The drawback of this method is having some models exceed the

cycle time at some stations. The second method is enforcing the cycle time on all model

variations. The drawback of this method is that it requires a higher cycle time which is

translated into a lower production rate. Boysen et al. (2008) reviews the different

approaches used to handle the mixed model problem in the literature. In this chapter, the

first approach is used in which the average task time is guaranteed to be less than the cycle

time. In addition, because of the huge number of models that could be configured,

automakers use option-based forecasts instead of figuring out the forecasts for each

possible model. The option-based precedence graph introduced by Boysen et al. (2009)

will be used in this research based on the proportion 𝑝ℎ of products (cars) that require task ℎ

with duration 𝑑ℎ . This probability is derived for a production run using the forecasted

demand of the volume of cars that will require the execution of that task divided by the

total production volume. The weighted task time for task ℎ, 𝑑̅ℎ is calculated as

follows 𝑑̅ℎ = 𝑝ℎ 𝑑ℎ ∀ ℎ.

2.2.2 Zoning constraints

The zoning constraints are used to avoid worker interference in a station and reduce

non-productive walking time between mounting positions. Each task to be executed by any

22
worker has to be done in a specific location on the product. This attribute is called the

mounting position and it is one of the different positions available in a product zone map.

Figure 3 shows an example of a product zone map in which there are nine different

mounting positions.

Figure 3: Product zone map

Each task ℎ has a mounting position 𝑞ℎ from the 9 positions on the product zone in

which the task takes place. Mapping each task to a mounting position helps in prohibiting

the overlap in the scheduling of two tasks ℎ, 𝑙 with the same mounting position 𝑞ℎ = 𝑞𝑙 at

any station. Figure 4 shows an example where interference might happen when two

workers execute tasks that share the same mounting positions at the same time.

Figure 4: Worker interference example

23
2.2.3 Parallel workers

One operator per station is assumed in the simple assembly line balancing

problem’s literature. However, it is common to have more than one operator working

within the same station in big product assembly lines such as cars, trucks, and big machines.

More specifically in the automobile industry, the assembly of cars involves using relatively

large parts and thus many stations have enough space to accommodate multiple workers

who are able to work at the same time. We assume that each worker is not fixed to one

work area inside the station and can be assigned different tasks that have different mounting

positions. Furthermore, we provide detailed scheduling of the tasks assigned to each

worker to avoid conflicts between workers in the same station and indicate if the line

balance is actually feasible. Infeasibility could occur even if the sum of all tasks assigned

within a station is less than the cycle time. This is because there exist precedence relations

between tasks that could lead to idle time if one operator is waiting for another to finish a

task.

2.2.4 Assignment restrictions

Some tasks are restricted to be assigned to some stations due to different reasons.

For example, a task that requires a resource that is available only on some stations or a task

that is done below the car and requires the car to be elevated. Different assignment

restrictions are needed in order to have a flexible tool that can be useful for real life line

balancing. These restrictions can be categorized as follows

24
2.2.4.1 Resource constraints

This constraint is used when a task requires a certain resource such as a tool that is

not available in all stations. For example, some tools are heavy and expensive to move

between stations so they will be fixed to certain stations. Tasks that require this tool should

be only assigned to this station. For a task ℎ that require a certain resource not available in

all stations there is a set of eligible stations 𝑈ℎ that contains all stations this task can be

assigned to. If a tool is available in only one station, tasks that need this tool are assigned

to this station in the preprocessing step.

2.2.4.2 Adjacency constraints

This constraint is for tasks that need to be executed consecutively by the same

worker. For example, a task for picking up a part must be followed by a task to assemble

it. In this case, the worker is not open to any other tasks since he has the part in his hands.

Adjacency constraints are treated in the preprocessing phase by combining the adjacent set

of tasks into one task.

2.2.4.3 Same worker constraints

This constraint is for tasks that need to be executed by the same worker but does

not need to be done consecutively. An example would be inspection or quality checks that

should be done by a worker after finishing a task. Since the inspection may involve several

tasks, the same worker constraint is used instead of the adjacency constraint. All tasks

constrained to be done by the same worker are collected in sets labeled by 𝑅 𝑠𝑤 .

25
2.2.4.4 Incompatible worker constraint

This constraint is needed when a set of tasks are not allowed to be assigned to the

same worker. This constraint is mainly used to prohibit unproductive walking by limiting

the assignment of tasks with opposite mounting positions to the same worker. An example

for the product zone in Figure 3 would be to limit any worker from being assigned a pair

of tasks ℎ, 𝑙 with mounting positions {(𝑞ℎ , 𝑞𝑙 )} = {(1,9), (7,3), (4,6)}. That is, if a worker

is assigned a task with mounting position 1 the assignment of all tasks with mounting

position 9 is prohibited for him since it would be unproductive to walk all the way from

product zones 1 to 9. All tasks that are incompatible worker are collected in sets labeled

by 𝑅 𝑛𝑤 .

2.2.4.5 Same station constraint

This constraint is used to indicate that a set of tasks needs to be done on the same

station but not necessarily done by the same worker. An example would be a large piece

that needs to be installed by two workers via two different tasks. Both tasks need to be

done on the same station so that each worker would be assigned one of them. All tasks

that need to be assigned to the same station are collected in sets labeled by 𝑅 𝑠𝑠 .

2.2.4.6 Not the same station

This constraint prohibits a set of tasks from being assigned to the same station. An

example for using this constraint would be for a task to lubricate a movable part and task

to install the seat. These two tasks should be assigned to different stations to avoid the

danger of soiling the seat with the lubrication material. Another example would be a gluing

26
task that needs time for sitting before executing another follower task. All tasks that are

station incompatible are collected in sets labeled by 𝑅 𝑛𝑠 .

2.2.5 Ergonomics risks constraints

The ergonomic risks are treated in two different ways, directly and indirectly. An

indirect way of treating ergonomic risk is divided to station-related and worker-related

methods. An example for dealing with an indirect station related ergonomic risk is the

relation between tasks and the position of the car on the line. Some tasks that require

installing parts above the car need the car to be tilted to avoid injury for the worker

executing it. These tasks should be restricted to station in which the car can be tilted.

Furthermore, tackling the worker interference issue and unproductive walking are

considered to be indirect worker level ergonomic risk constraints. On the other hand, the

direct application of ergonomic risk mitigation can be accomplished through the use of a

weighted ergonomic risk score 𝑔̅ℎ associated with each task ℎ to make sure that each

worker is assigned tasks with a total weighted ergonomic risk score less than a predefined

maximum recommended level 𝐺.

2.2.6 Pre-processing

The pre-processing phase is important to prepare the input before it is used with the

mathematical program. For the mixed model problem, this involves computing the

weighted task durations and the weighted ergonomic scores. Furthermore, in order to

reduce the complexity of the solution techniques, some of the available information is

useful in lowering the number of decision variables by fixing some of them. Also, it is

possible to remove unnecessary constraints by reducing the set of variables the constraint

27
applies to. This helps in getting a tighter feasible solution space and helps in reducing the

combinatorial explosion that happened with these type of problems. The pre-processing

phase involves the following:

- Reduce the number of tasks by combining the tasks in an adjacency set into one

task.

- Defining the feasible stations set 𝐹 for all tasks via the use project management

techniques to calculate the early start and late finish for each task.

- Defining the eligible stations set 𝑈 for all tasks that require a specific resource

that is not available in all stations. If this resource is available in one station only,

all tasks that require it are pre-assigned to this station.

- Calculating the weighted task duration 𝑑̅ and the weighted ergonomic score 𝑔̅ for

all tasks.

- Populating the assignment constraints sets 𝑅 𝑠𝑤 , 𝑅 𝑛𝑤 , 𝑅 𝑠𝑠 , and 𝑅 𝑛𝑠 .

2.2.7 An example

Consider the example instance with 12 tasks and the option-based task precedence

graph in Figure 5. Assume that 10 cars are to be assembled, each with different options that

require different tasks. The volume and proportion of each task are given in Table 1 along

with the task duration and ergonomic score. The weighted values of both the task duration

and the ergonomic score are calculated based on the volume of cars that require that task.

It should be noted that both the ergonomic and cycle time constraints are enforced on the

weighted (average) values only; that is the total weighted task duration and ergonomic

28
scores are less than the cycle time and the ergonomic risk maximum recommended level

respectively. Figure 6 shows the mounting position for each task, which is the location in

which the task will take place on the car. Since the number of stations is predefined, the

early start and late finish for each task are calculated and shown in the Gantt chart (Figure

7) with each task scheduled at its early start time. The Gantt chart helps reduce the feasible

station set for each task, for example tasks 1, 6, and 12 can only be assigned to stations 1,

2, and 3 respectively. The assignment constraints are given in Table 2. Since tasks 7a and

7b are in an adjacency set, they are joined together as task 7 in the preprocessing phase.

The “incompatible worker” assignment constraint is used in this example to avoid

unproductive walking between opposite mounting positions. For example a worker who is

assigned a task with mounting position 1 will not be assigned another task with mounting

position 4 and vice versa but he can be assigned tasks with mounting positions 2 and 3. An

optimal solution to this example is shown in Figure 8 in which all 6 potential workers are

assigned tasks. In Figure 8, the tasks are represented by blocks in which the length denotes

the weighted duration of the task and the height of the block denotes the weighted

ergonomic score. Since task 6 has a high weighted ergonomic score it is the only scheduled

task for worker (2,1). The scheduling of tasks to workers is important for reasons other

than avoiding worker interference since just assigning tasks to workers might yield an

infeasible solution because of the precedence relations. For example worker (1,1) has an

unavoidable idle gap since he had to wait for task 1 to finish before starting task 5 because

of the precedence relation between the two tasks.

29
Figure 5: Option-based joint precedence graph

Table 1: Example data for the line balancing problem

Mounting Weighted Weighted

Task Volume Proportion Duration ErgoS
position Duration ErgoS
1 10 1 1 50 50 50 50
2 10 1 2 30 40 30 40
3 3 0.3 3 80 70 24 21
4 5 0.5 3 50 40 25 20
5 7 0.7 4 70 80 49 56
6 10 1 3 70 90 70 90
7a 10 1 2 20 40 20 40
7b 10 1 2 20 30 20 30
8 8 0.8 1 30 30 24 24
9 10 1 2 50 70 50 70
10 2 0.2 3 70 80 14 16
11 5 0.7 3 50 60 35 42
12 10 1 4 30 30 30 30

30
Figure 6: Mounting position for each task

Figure 7: Gantt chart with tasks at early start times

31
Table 2: Assignment constraints

Assignment Tasks
Constraint
Adjacency set (7a,7b)
Same worker (9,12)
Same station (4,5)
Not the same worker {(ℎ, 𝑙): 𝑞ℎ , 𝑞𝑙 = (1,4), (2,3)}
Not the same station (10,11)

Figure 8: An optimal solution to the example

32
2.3 Related work

The basic mixed model assembly line balancing problem (MALBP) is based on the

same assumptions that are used for the SALBP. The main difference is having multiple

models each with its own precedence graph and task times. In the MALBP literature, two

basic approaches are used to model and solve the problem: reduction to single model

problem and horizontal balancing (Becker and Scholl, 2006).

The MALBP can be transformed to the SALBP by the use of a joint precedence

graph. This graph is constructed by averaging the processing times that varies across

models taking into account the probability of having each model in the expected model

mix. The new joint precedence graph will have tasks with the expected processing times.

This way the mixed model balancing problem is reduced to a single model case and

traditional single model techniques can be used to solve it. With the increase of product

variety in some production fields such as car manufacturing, reliable estimation of each

model is becoming harder. In the paper by Pil and Holweg (2004) the number of model

varieties that are available from different car manufacturers is shown; for some German

car manufacturers, the theoretical variations can reach over 1024 different models. With this

huge number of models, getting an estimation for the demand of each model is impossible.

Boysen et al (2009) propose the use of the occurrence of the options instead of the model

in determining the task times of the joint precedence graph. However, this requires the

assignment of tasks to options and not just models. Van Zante-de Fokkert and De Kok

(1997) review the two approaches of transforming the mixed model problem into a single

33
model: the combined precedence diagram approach and the adjusted task processing time

approach.

Gokcen and Erel (1997) develop a binary goal programming (GP) model for the

MALBP. The authors claim that they are the first to apply the multiple criteria decision

making approach to the MALBP which gives the decision maker a more realistic approach

and the ability to use different conflicting objectives in the model. Although the optimal

solution may not be identified through this solution procedure, a satisfactory result that

compromises the different conflicting objectives is provided. In addition, Gokcen and Erel

(1998) propose a binary integer programming model for the MALBP along with some

computational properties of the model. They claim that their model is superior to the other

in the literature in terms of decision variables numbers and constraints. Furthermore, a

shortest route formulation for the MALBP is presented by Erel and Gokcen (1999). The

formulation is based on the shortest route formulation for the single model problem, thus

the reduction of the mixed model to a single model by the combined precedence graph is

required for this formulation.

Sawik (2002) compares two approaches to solve the combined balancing and

scheduling problem for the flexible assembly line. In the monolithic approach, the

balancing and assembly decisions are made simultaneously using a mixed integer program.

For the hierarchical approach, the station workloads are balanced first and then the

assembly process is scheduled by solving a permutation flow shop problem. Öztürk et al.

(2013) also study the problem of balancing and scheduling in mixed model assembly lines

34
with parallel stations. They formulate the problem as a mixed integer program and then

propose a decomposition scheme to be applied for large scale applications.

Some research introduces the multi-objective (MO) assembly line balancing

problem such as the multi-objective approach proposed by Kara et al. (2011) to solve the

MALBP for model mixes that have precedence conflicts. They develop a binary

mathematical model with a single objective and then extend the model to incorporate three

objectives into a single objective model. Mahdavi et al. (2009) propose a fuzzy multi-

objective linear programming model for solving multi-objective MALBP. Their model

makes use of a two-phase linear program: the first using a max-min approach and the

second using the max-min solution as a lower bound to maximize the composite

satisfaction degree.

2.3.1 Exact methods

Most of the literature on exact solution methods is focused on the SALBP as only

a few propose exact methods for the MALBP. To use the single model solution methods

on the mixed model problem, the mixed model need to be reduced to a single model by the

use of combined precedence graph, after which modified single model techniques can be

applied. Swell and Jacobson (2012) present an exact algorithm for the SALBP named the

branch, bound and remember algorithm. The authors claim that this algorithm manages to

find the optimal solution for all combined benchmark problems of Hoffmann, Talbot, and

Scholl. The proposed algorithm is a hybrid method that combines branch and bound with

dynamic programming. Scholl and Becker (2006) review state of the art exact and heuristic

methods used to solve the SALBP. An important point to note is that although solving the

35
average SALBP guarantees that the line balance will not violate the cycle time on average,

the optimal solution might still have significant inefficiencies when put to practice. Scholl

(1999) point out that there are some necessary modifications that need to be done on any

SALBP exact solution when applying it to a MALBP. This could be done by adding a

secondary objective for smoothed stations loads. However, for branch and bound based

algorithms, all optimal solutions have to be considered and the node fathoming procedure

needs to be updated. In addition, the search should not be limited to maximum station loads

as the secondary objective may prefer otherwise.

Ege et al. (2009) propose a branch and bound algorithm to solve the assembly line

balancing problem with station paralleling. Boysen and Fliedner (2008) propose a two

stage graph algorithm (Avalanche) that is able to solve GALBP with relevant constraints.

The authors claim that this approach can be easily modified to include extensions such as

parallel work stations, processing alternatives, zoning restrictions, stochastic processing

times and U-shaped assembly lines.

Vilà and Pereira (2014) study the assembly line worker assignment and balancing

problem. They provide an exact enumeration procedure that is based on the branch-bound

and remember algorithm presented by Swell and Jacobson (2012). The authors develop

several lower bounds, reductions, and dominance rules for the problem.

Wilhelm and Gadidov (2004) develop two models to address different tooling

requirements in the multiple product assembly system design problem. The authors

propose a branch and cut approach that employs a facet generation procedure to generate

cuts in the search tree by exploiting some special structures of the problem. The authors

36
demonstrate the use of the approach through experiments and compare it to the available

solution procedures.

One of the assumptions for the MALBP states that identical tasks have to be

assigned to the same station for all models. The problem can be decomposed into several

SALBP (one for each model) by relaxing this assumption. Bukchin and Rabinowitch

(2006) relax the assumption of assigning identical tasks across different models to the same

station, which allows a common task to be assigned to different stations for different

models. The authors develop an integer program formulation and an exact solution method

based on a backtracking branch and bound algorithm. In their model, some costs are

associated with assigning an identical task to different stations. Hence, they needed to

modify the objective function as the goal is to minimize the total cost associated with

balancing and not just minimizing the number of stations. However, according to Becker

and Scholl (2006), this relaxation is not desired in practice for several reasons such as the

complex production control, facility requirements, loss of specialization effects and setup

inefficiencies.

2.3.2 Constraint programming models

There has not been much research in using CP to solve the assembly line balancing

problem in the past although it is one of the most successful tools in solving combinatorial

problems. Bockmayr and Pisaruk (2001) propose a hybrid approach for solving the SALBP

by combining CP and IP. The main contribution in their paper is developing a branch and

cut solver for the SALBP and to show how it can make use of the CP solver in pruning the

search tree. Pastor et al. (2007) compare the performance of impulse variables based

37
models, step variable based models, CP models, and IP models in solving the SALBP with

different objectives. The authors conclude that the CP formulation performs better and is

faster than IP but the best results were for the impulse variable based models. Recently

Öztürk et al. (2013b) propose a MIP and a CP model to solve the flexible mixed model

assembly line balancing and sequencing problems. They compare the performance of the

MIP, the complete and various decomposition methods, and CP. According to their

computational study, the CP outperforms all the other approaches over all sizes of the test

instances. To the best of our knowledge there is no other published research that uses CP

in solving the mixed model assembly line balancing problem with parallel station, zoning

constrains and ergonomics.

2.3.3 Heuristics and meta-heuristics

The bulk of the research in the solution procedures for the MALBP is based on

heuristics and meta-heuristics. The main reason for that is the complexity of the problem

as the SALBP is just a bin packing problem when the precedence constraints are removed.

The bin packing problem is a known NP-hard problem, thus the MALBP is NP-hard since

it is just a SALBP with more than one model. More details on the complexity of the

problem is given in Scholl (1999).

2.3.3.1 Heuristics:

Matanchai and Yano (2001) propose an approach to solve the MALBP by using an

objective that helps achieving a better short term workload stability. Based on this

objective, the authors develop a heuristic solution procedure that is based on filtered beam

search in which feasible subsets of tasks are constructed at each station. For every station,

38
several feasible subsets with the best objectives are retained. Potential subsets at

subsequent stations are then constructed by branching off the subset with the best objective.

If the process is stopped because of infeasibilities, backtracking is used. After a feasible

solution is found, an improvement procedure is used to reassign tasks between stations to

improve the objective function.

Bukchin et al. (2002) design a three-stage heuristic to solve the mixed model

assembly line design problem in a make to order environment. The heuristic minimizes the

number of stations given a predetermined cycle time through stages: balancing using

combined precedence graph, balancing for each model given constraints from the first

stage, and neighborhood search. The authors relax the assignment constraint by allowing

identical tasks across models to be assigned to different stations.

Hop (2006) solve the fuzzy MALBP in which processing time is fuzzy. The author

propose a heuristic that aggregates fuzzy time, and used the combined precedence diagram

to transfer the problem into a fuzzy SALBP. In addition, the author develop new

approximated fuzzy arithmetic operations to calculate fuzzy numbers and then formulated

the problem as a mix-integer program. The heuristic proposed is based on a flexible

exchange sequence procedure to assign tasks into workstations.

Tonelli et al. (2013) propose a mixed integer program model that is solved by an

iterative heuristic which aims at solving a number of aggregate planning problems in a

mixed model production environment. The iterative heuristic is used to solve linear relaxed

problems and reduced mixed integer program problems. The proposed optimization

39
approach combines the use of mixed integer program solver with a rolling horizon

decomposition heuristic to solve practical problems.

McMullen and Frazier (1997) present a heuristic to solve a stochastic MALBP with

task paralleling. The authors compared different task selection rules through simulation

experiments. Sparling and Miltenburg (1998) propose an approximate solution algorithm

for the U-shaped MALBP. The heuristic is based on a branch and bound algorithm applied

to the combined precedence graph, followed by a smoothing algorithm to smooth the

balance.

2.3.3.2 Meta-heuristics:

Tsujimura et al. (1995) are the first to use a genetic algorithm (GA) with a GALBP

in which processing time is fuzzy. Falkenauer (1998) propose in his book a modified

version of the GA named the grouping GA in which he shows the advantages of applying

it to grouping problems including the assembly line balancing problem. Simaria and

Vilarinho (2004) develope an iterative GA based procedure and applied it to the MALBP

with parallel stations. The authors claim that their procedure accommodates other

extensions such as zoning constraints and workload balancing. Recently, Sivasankaran and

Shahabudeen (2013) use GA to solve the MALBP without converting the problem to a

SALBP. That is, they used the original task times for each model in determining the line

balance. In addition to the work presented, GA was used to solve the SALBP in many

studies. More details on the use of GA to solve assembly line balancing problem is given

in the survey by Tasan and Tunali (2008).

40
 Vilarinho and Simaria (2002) develop a two-stage procedure using simulated

annealing (SA) to solve the MALBP with zoning constraints and parallel workstations. The

first stage searches for a sub-optimal solution to the primary objective which is minimizing

the number of stations while the second stage deals with the secondary objective of

smoothing workload across stations. Özcan and Toklu (2009) propose a SA algorithm to

solve the mixed model two-sided line balancing problem by considering two objectives:

minimizing the line length and minimizing the number of workers. Two performance

measures are used: maximizing the weighted line efficiency and minimizing the weighted

smoothness index. Manavizadeh et al. (2013) develope a three-stage SA algorithm to solve

the mixed model U-line assembly line balancing problem in a Just in Time (JIT) production

system. The stages are: solving the balancing problem to determine the number of stations,

solving a worker assignment problem, and designing an alert system that is based on the

Kanban system to balance the work in process inventory.

Vilarinho and Simaria (2006) revisit the same problem that they studied in (2002)

but using an ant colony optimization (ACO) algorithm that searches for solutions in which

the workload among workstations is smoothed. The authors claim that the results obtain

using ACO were better than results obtained in (2002) using SA. Simaria and Vilarinho

(2009) develop another ACO for the two sided MALBP in which two ants work together

to build a feasible balancing solution with respect to zoning, capacity, side and

synchronism constraints. McMullen and Tarasewich (2003) use a heuristic that is based on

ACO to solve the problem of MALBP with constraints such as parallel workstations and

stochastic task times. The authors claim that the solution method developed exploits the

41
properties of the assembly line balancing problem and produces good solutions in a

reasonable time.

Bock (2008) uses clustered Tabu search as a base for his distributed search

approach to solve the integrated problem of mixed model line balancing, personal, and

process planning. The idea behind this approach is to use the distributed search technique

on a network of computers in order to solve complex problems.

Furthermore, hybrid heuristics have been used to solve the mixed model assembly

line balancing problem. Noorul Haq et al. (2006) propose a hybrid GA approach to solve

the MALBP in which the modified ranked positional solution is used as an initial solution

to reduce the search space. Akpınar and Bayhan (2011) develop a hybrid GA algorithm to

solve the MALBP with parallel workstations and zoning constraints. To overcome the

GA’s shortage of exploring the search space effectively, the approach sequentially

hybridized three well known heuristics with GA. Akpınar et al. (2013) solve the MALBP

with parallel workstations, zoning constraints, and sequence dependent setup times

between tasks using a ACO-GA hybrid algorithm. The authors aimed at combining the

power of diversification in the ACO with the power of intensification in the GA.

Meta-heuristics are also used in several multi-objective assembly line balancing

work. McMullen and Frazier (1998) use SA to solve a multi-objective MALBP with

parallel workstations. In addition to the traditional performance objectives, the authors

study the effect of a combination of several search objectives on performance measures

such as cycle time and design cost. McMullen and Tarasewich (2006) develop a technique

that is derived from ACO to solve the multi-objective GALBP. Several objectives are

42
addressed simultaneously such as crew size, system utilization, and system design cost.

Recently, Chutima and Chimklai (2013) present a special PSO algorithm to solve the two-

sided multi-objective MALBP. The primary objectives include mated stations and number

of stations and the conflicting secondary objectives are workload relatedness and

smoothness.

The summary of the extensions and methodology used in related literature is shown

in Table 3.

43
Table 3: Summary of related literature

Researchers (Year) Mixed Eligible Tooling Task Parallel Worker- Ergonomics Methodology
model stations (Worker/Station) stations interference
compatibility zoning
Akpınar & Bayhan
(2011)    Hybrid GA

AkpıNar et al. (2013)

   ACO-GA

Becker & Scholl (2009)

    B&B,
Heuristics
Bock (2008)
  TS

Bukchin et al. (2002)

 Heuristics

Bukchin et al. (2006)

 B&B,
Heuristics
Chutima & Chimklai
(2012)  PSO

Ege et al. (2009)

  B&B,
Heuristics
Erel (1999)
 Shortest path
model
Gokcen & Erel (1997)
   GP

Gokcen & Erel (1998)

  Binary IP

Haq et al. (2006)

 Hybrid GA

Hop (2006)
  Fuzzy binary
LP
Kara (2011)
 MO GP

Mahdavi et al. (2009)

 MO fuzzy LP

Manavizadeh et al.
(2013)  SA

McMullen & Frazier

(1997)   Heuristics

44
McMullen &
Tarasewich (2003)   ACO

McMullen & MO ACO

Tarasewich (2006)
Otto & Scholl (2011)
 Heuristics

Özcan Toklu (2009)

 SA

Öztürk et al. (2013)

   Decomposition

Öztürk et al. (2013b)

   CP

Sawik (2002)
 IP

Simaria & Vilarinho

(2004)    GA

Simaria & Vilarinho

(2009)   ACO

Sivasankara &
Shahabudeen (2013)  GA

Tsujimura et al. (1995) GA

Vilà & Pereira (2014) B&B

Vilarinho & Simaria

(2002)    SA

Vilarinho & Simaria

(2006)    ACO

Wilhelm & Gadidov

(2004)   B&C,
Heuristics

45
2.4 Motivation

The main motivation for this work is the fact that there is no applicable model that can be

used to solve real life assembly line balancing problems. There is a gap between the research and

the industry in the area of line balancing. Falkenauer (2005) discussed how the SALBP solution

methods discussed in the literature has little or no application to the real world assembly line

balancing problem. Furthermore, even the reported extensions in the general assembly line

balancing problem literature are not applicable to real problems in the industry. The reason is that

each extension usually tackles a generalization in one direction, while another research might

consider other generalizations in another direction. What the industry really needs is a way to deal

with all of these generalizations at the same time. The main characteristics needed to capture

realism and make the solution applicable are summarized as follows:

- The majority of real life line balancing takes place at an already built assembly plants so

the number of stations is fixed and re-balancing is what is needed.

- Each station has its own identity with the tooling and restrictions of the tasks that can be

done on it.

- There are zoning constraints that should be considered in line balancing.

- Since the assembly plants are already built, elimination of stations is not feasible.

- Smoothing the workload across stations should be an important objective of the line

balance.

- Multiple operators usually work in a station, so a schedule is needed to increase utilization

and remove idle time.

- Station level ergonomic risks should be considered and might lead to assignment

restrictions.

46
 Also incorporating worker level ergonomic risks such as avoiding work area interference

between workers should be incorporated.

- In a multi-product line, care should be taken in computing the average task times,

horizontal balancing should be applied, and handling of drifting operations.

To the best of our knowledge, no published research has considered all of these

characteristics simultaneously. Thus there is a need for a solution approach that takes most if not

all of these realistic aspects into consideration.

2.5 The mathematical model

The integer program formulation for this problem is based on the formulation by Becker

and Scholl (2009) for the single model variable parallel workstations assembly line balancing

problem. In our model, it is assumed that tasks have a duration that is less than the cycle time so

there is no need to divide tasks between stations. Also, mounting positions are not assigned to

workers so several workers in the same station can be assigned tasks that share the same mounting

position. Detailed scheduling is used to make sure that worker interference does not occur.

47
 Each product assembled on the line starts with a work-piece that is launched down the line

and have assembly tasks done on it until the work-piece exits the final station as a finished product.

The work-piece enters the first station at time 𝑡 = 0 and leaves the last station at time 𝑡 = 𝐾𝑐. For

each of the 𝑖 = 1, … 𝐾 stations in the line there are 𝑗 = 1, … , 𝑊 potential workers available, so

worker (𝑖, 𝑗) represents the potential worker 𝑗 at station 𝑖. The binary variable 𝑦𝑖,𝑗 is associated

with each potential worker (𝑖, 𝑗) and is equal to 1 if the potential worker is assigned and 0

otherwise. Figure 9 shows an example profile for a line with 3 stations and 4 workers, the greyed

blocks represent the assignment of potential worker (𝑖, 𝑗) when 𝑦𝑖,𝑗 = 1. In this example 3, 2, and

4 workers are assigned to stations 1, 2, and 3 respectively. The model’s notations are shown in

Table 4.

Figure 9: Example for a worker/station profile

48
Notations:
Table 4: Line balancing problem’s notations

𝐾 Maximum number of stations indexed 𝑖 = 1,2,3, … , 𝐾

𝑊 Maximum Number of potential workers per station indexed 𝑗 = 1,2,3, … , 𝑊
𝑁 Number of tasks indexed ℎ, 𝑙 = 1,2,3, … , 𝑁
𝑐 Cycle time
𝐴𝑖 Starting time of station 𝑖 𝐴𝑖 = (𝑖 − 1) ∗ 𝑐
𝑡ℎ𝑠 Early start time for task ℎ
𝑓 Late finish for task ℎ
𝑡ℎ
𝑞ℎ Mounting position of task ℎ
𝑝ℎ Proportion of models that will require task ℎ in the production run
𝑑ℎ Duration of task ℎ
𝑑̅ℎ Weighted duration of task ℎ
𝑔ℎ Ergonomic score for task ℎ
𝑔̅ℎ Weighted ergonomic score for task ℎ
𝐺 Ergonomic risk score limit
𝐹ℎ The set of feasible stations that task ℎ is assignable to
𝑈ℎ The set of stations with the required resources to perform task ℎ
𝑂ℎ The set of immediate predecessors of task ℎ
𝑅 𝑠𝑠 The set of tasks that must be done on the same station
𝑅 𝑛𝑠 The set of tasks that cannot be done on the same station
𝑅 𝑠𝑤 The set of tasks that must to be done by the same worker
𝑅 𝑛𝑤 The set of tasks that cannot be done by the same worker

Decision Variables:

1 𝑖𝑓 𝑡𝑎𝑠𝑘 ℎ 𝑖𝑠 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑 𝑡𝑜 𝑠𝑡𝑎𝑡𝑖𝑜𝑛 𝑖 𝑎𝑛𝑑 𝑤𝑜𝑟𝑘𝑒𝑟 𝑗

𝑥𝑖𝑗ℎ = {
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

1 𝑖𝑓 𝑝𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 𝑤𝑜𝑟𝑘𝑒𝑟 𝑗 𝑎𝑡 𝑠𝑡𝑎𝑡𝑖𝑜𝑛 𝑖 𝑖𝑠 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑

𝑦𝑖𝑗 = {
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

1 𝑖𝑓 𝑡𝑎𝑠𝑘 ℎ 𝑖𝑠 𝑒𝑥𝑒𝑐𝑢𝑡𝑒𝑑 𝑏𝑒𝑓𝑜𝑟𝑒 𝑡𝑎𝑠𝑘 𝑙

𝑣ℎ𝑙 = {
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑠ℎ = 𝑠𝑡𝑎𝑟𝑡𝑖𝑛𝑔 𝑡𝑖𝑚𝑒 𝑜𝑓 𝑡𝑎𝑠𝑘 ℎ

49
Objective:
𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = ∑ ∑ 𝑦𝑖𝑗
𝑖 𝑗

The objective is to minimize the total number of workers assigned.

Subject to the following constraints:

 Task assignment to worker: each task has to be assigned to only one worker.

∑ ∑ 𝑥𝑖𝑗ℎ = 1 ∀ℎ (C2.1)
𝑖∈𝐹𝑆ℎ 𝑗

 Cycle time: the total weighted duration assigned to a worker should not exceed the

cycle time.

∑ 𝑥𝑖𝑗ℎ 𝑑̅ℎ ≤ 𝑐 𝑦𝑖𝑗 ∀(𝑖 ∈ 𝐹ℎ , 𝑗) (C2.2)

ℎ

 Station time: each task assigned to a worker should be scheduled between the

worker’s station start and finish times.

𝑠ℎ ≥ ∑ ∑ 𝑆𝑖 𝑥𝑖𝑗ℎ ∀ℎ (C2.3)
𝑖∈𝐹ℎ 𝑗

𝑠ℎ + 𝑑̅ℎ ≤ ∑ ∑(𝑆𝑖 + 𝑐)𝑥𝑖𝑗ℎ ∀ℎ (C2.4)

𝑖∈𝐹𝑆ℎ 𝑗

 Precedence relations: a task can only start when all of its predecessors are finished.

𝑠ℎ + 𝑑̅ℎ ≤ 𝑠𝑙 ∀ ℎ, 𝑙 ∈ 𝑂𝑙 (C2.5)

 Tasks overlap: all tasks assigned to a worker should not overlap.

𝑣ℎ𝑙 + 𝑣𝑙ℎ ≥ 𝑥𝑖𝑗ℎ + 𝑥𝑖𝑗𝑙 − 1 ∀𝑗 , ℎ ≠ 𝑙 , 𝑎𝑛𝑑 𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 (C2.6)

50
 𝑓
𝑠ℎ + 𝑡ℎ ≤ 𝑠𝑙 + (1 − 𝑣ℎ𝑙 )(𝑡ℎ − 𝑡ℎ𝑠 ) ∀ℎ ≠ 𝑙 (C2.7)

 Worker interference: tasks that share the same mounting position should not

overlap.

𝑣ℎ𝑙 + 𝑣𝑙ℎ ≥ ∑(𝑥𝑖𝑗ℎ + 𝑥𝑖𝑗𝑙 ) − 1 ∀𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 , ℎ ≠ 𝑙 , 𝑎𝑛𝑑 𝑞ℎ = 𝑞𝑙 (C2.8)

𝑗

 Ergonomic risk: the total weighted ergonomic risk for each worker should be within

limits.

∑ 𝑥𝑖𝑗ℎ 𝑔̅ℎ ≤ 𝐺 ∀(𝑖 ∈ 𝐹ℎ , 𝑗) (C2.9)

ℎ

 Assignment restrictions: pairs of tasks with same (station/worker) or not the same

(station/worker) restrictions.

∑ 𝑥𝑖𝑗ℎ = ∑ 𝑥𝑖𝑗𝑙 ∀𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑠𝑠 (C2.10)

𝑗 𝑗

𝑥𝑖𝑗ℎ = 𝑥𝑖𝑗𝑙 ∀ (𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 , 𝑗) 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑠𝑤 (C2.11)

∑(𝑥𝑖𝑗ℎ + 𝑥𝑖𝑗𝑙 ) ≤ 1 ∀𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑛𝑠 (C2.12)

𝑗

𝑥𝑖𝑗ℎ + 𝑥𝑖𝑗𝑙 ≤ 1 ∀ (𝑖 ∈ 𝐹ℎ ∩ 𝐹𝑙 , 𝑗) 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑛𝑤 (C2.13)

 Resource requirement: tasks that require a resource should only be assigned to

eligible stations.

𝑥𝑖𝑗ℎ = 0 ∀𝑖 ∉ 𝑈ℎ , 𝑗, 𝑎𝑛𝑑 ℎ (C2.14)

51
  Breaking symmetry.

𝑦𝑖𝑗 ≥ 𝑦𝑖(𝑗+1) ∀ (𝑖, 𝑗 ≤ 𝑊 − 1) (C2.15)

 Variables domain.

𝑥𝑖𝑗ℎ , 𝑦𝑖𝑗 , 𝑣ℎ𝑙 ∈ {0,1},

∀𝑖, 𝑗, ℎ, 𝑎𝑛𝑑 𝑙 (C2.16)
+
𝑠ℎ ∈ ℤ

The original model formulation proposed by Becker and Scholl (2009) did not provide a

good lower bound for the integer program. The introduction of the cycle time constraint (C2.2)

provides a good lower bound for the integer program that is equivalent to the capacity bound (LB1)

presented in their paper. The lower bound achieved by this constraint gives the lowest possible

number of workers needed to perform all tasks which is equivalent to rounding up the sum of all

weighted task durations divided by the cycle time.

∑ℎ 𝑑̅ℎ⁄
𝐿𝐵 = ⌈ 𝑐⌉ (2.1)

In addition, the assumption that all tasks that share the same mounting position are to be

assigned to the same worker is relaxed in this work. Any task can be assigned to any worker

regardless of its mounting position. Reducing the unproductive waking from a position to another

is achieved by prohibiting the same worker from being assigned tasks with non-contiguous

mounting positions. To avoid worker interference, constraint (C2.8) is added to make sure that no

two tasks that share the same mounting position overlap in time in any station. Constraints (C2.10)

and (C2.13) are for the same station and same worker assignment restrictions constraints. The

52
original model uses the worker incompatible and station incompatible constrains only. Constraint

(C2.14) helps in reducing the feasible space by fixing some decision variable by the use of the

eligible station set for each task. Since all workers are identical, constraint (C2.15) reduces the

problem’s feasible region by removing the symmetry from the problem in an effort to reduce the

computation time. For example, the constraint enforces the assignment of worker 1 before

assigning worker 2.

2.6 The constraint programming model

Since it is anticipated that the mathematical programming model will face difficulties in

finding a solution for real life sized problem instances, we propose a constraint programming (CP)

model. CP is proven to be a useful technique for solving scheduling problems efficiently. Several

tools exist for developing CP models; ILOG CPLEX optimization studio is one of the most used

for scheduling problems because of the ILOG Scheduler extension that was developed specifically

with several constraint propagation and search algorithms that exploit the structure of scheduling

problems. ILOG Scheduler is used in this work to model the mixed model line balancing problem.

Before presenting the formulation we first introduce some basic variable types used in the ILOG

Scheduler environment. After that some specialized constraints used in the model are introduced.

These specialized constraints state some relation between different variables in the model.

Although these relations could be modeled using traditional logical constraints; the use of the

specialized constraint increases the efficiency of the solver by exploiting the structure of the

problem through the use of specialized filtering algorithm.

53
2.6.1 CP variables and special constraints

2.6.1.1 Interval variables

Interval variables represent an interval of time in which a task or an activity is carried out

whose position in time is unknown and need to be scheduled. This decision variable is

characterized by a start value, an end value and a size. Interval variables can be optional; that is

not all variables would be considered in the solution. This is usually helpful in modeling optional

tasks that doesn’t need to be executed, tasks that can be executed on alternative resources, and

tasks with alternative modes of operation. If an interval variable is optional it can be present or

absent. An absent variable is not considered in any constraint or expression it is involved in.

2.6.1.2 Interval sequence variables

The interval sequence variable is defined over a set of interval variables and its value

represents the ordering of the tasks in that set. Absent interval variables are not considered in this

ordering. All the permutations of the intervals in the set are possible values for the interval

sequence decision variable. It should be noted that this variable does not enforce any constraint on

the relative position between interval variables but there are several constraints that can be used

on this decision variable.

2.6.1.3 Specialized constraints

 Alternative constraint:

The 𝑎𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒 constraint is used to create a constraint between an interval variable 𝑎

and a set of optional interval variables {𝑏1 , … , 𝑏𝑛 } such that if interval 𝑎 is present then exactly

one of the intervals {𝑏1 , … , 𝑏𝑛 } is present. It is indicated as 𝑎𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒(𝑎, {𝑏1 , … , 𝑏𝑛 })

54
  The no overlap constraint:

The 𝑛𝑜𝑂𝑣𝑒𝑟𝑙𝑎𝑝 constraint facilitates modeling disjunctive resources and operates on

sequence variables to define a chain of non-overlapping intervals. The intervals in the chain are

scheduled such that any interval in the chain ends before the start of the next interval in the chain.

 Precedence relation constraint:

There are several constraints that can be used to restrict the relative position of the interval

variables. One of these special constraints is the 𝑒𝑛𝑑𝐵𝑒𝑓𝑜𝑟𝑒𝑆𝑡𝑎𝑟𝑡 constraint that enforces the

precedence relation between two interval variables such that the predecessor interval ends before

the start of the successor interval.

 Logical constrains:

The presence status of interval variables can be controlled by using the logical

constraint 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑎) . For example, interval 𝑎 can be forced to be present if interval 𝑏 is

present by using the following logical constrain:

𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑎) ≥ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑏)

2.6.2 The constraint programming model formulation

Decision variables:

𝑇𝑎𝑠𝑘ℎ : Interval variable associated with task ℎ with a size equal to 𝑑̅ℎ

𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 : Interval variable associated with task ℎ that is performed by worker 𝑗 on

𝑓
station 𝑖, limited to lie within [𝑡ℎ𝑠 , 𝑡ℎ ] ∩ [𝐴𝑖 , 𝐴𝑖 + 𝑐]

𝑊𝑜𝑟𝑘𝑒𝑟𝑖,𝑗 : Sequence variable associated with worker 𝑗 at station 𝑖

Objective:

Minimize ∑𝑖 ∑𝑗 max 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 )

ℎ

55
 Subject to:

 Task assignment to worker: defining the alternative interval variable such that each task

interval variable has an alternative variable for each worker at each station. Only one of

the alternatives will be present; in other words each task is assigned to only one worker.

𝑎𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒(𝑇𝑎𝑠𝑘ℎ , 𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛∀𝑖,∀𝑗 ) ∀ℎ (CP2.1)

 Precedence relations: a task can start when all of its predecessors are finished.

𝑒𝑛𝑑𝐵𝑒𝑓𝑜𝑟𝑒𝑆𝑡𝑎𝑟𝑡(𝑇𝑎𝑠𝑘ℎ , 𝑇𝑎𝑠𝑘𝑙 ) ∀ ℎ, 𝑙 ∈ 𝑂𝑙 (CP2.2)

 Tasks overlap: tasks that are assigned to one worker should not overlap.

𝑛𝑜𝑂𝑣𝑒𝑟𝑙𝑎𝑝(𝑊𝑜𝑟𝑘𝑒𝑟𝑖,𝑗 ) ∀𝑖, 𝑗 (CP2.3)

 Worker interference: tasks that share the same mounting position should not overlap.

𝑛𝑜𝑂𝑣𝑒𝑟𝑙𝑎𝑝(𝑇𝑎𝑠𝑘ℎ ) ∀ℎ: 𝑞ℎ = 1,2,3, … , 𝑄 (CP2.4)

• Ergonomic risk: the total weighted ergonomic risk for each worker should be within limits.

∑ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 ) 𝑔̅ ≤ 𝐺 ∀ 𝑖, 𝑗 (CP2.5)
ℎ

• Assignment restrictions: pairs of tasks with same (station/worker) or not the same

(station/worker) restrictions.

∑ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 ) =
𝑗
∀ 𝑗 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑠𝑠 (CP2.6)
= ∑ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛𝑙,𝑖,𝑗 )
𝑗

𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 ) =
∀ 𝑖, 𝑗, 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑠𝑤 (CP2.7)
= 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛𝑙,𝑖,𝑗 )

56
 ∑(𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 )
𝑗

+ ∑ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛𝑙,𝑖,𝑗 )) ∀𝑗 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑛𝑠 (CP2.8)

𝑗

≤1

𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 )
∀ 𝑖, 𝑗, 𝑎𝑛𝑑 (ℎ, 𝑙) ∈ 𝑅 𝑛𝑤 (CP2.9)
+ 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛𝑙,𝑖,𝑗 ) ≤ 1

• Resource requirement: tasks that require a resource should only be assigned to eligible

stations.

𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 ) == 0 ∀𝑖 ∉ 𝑈ℎ , 𝑗, 𝑎𝑛𝑑 ℎ (CP2.10)

• Breaking symmetry:

max 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗+1 )
ℎ
∀ (𝑖, 𝑗 ≤ 𝑊 − 1) (CP2.11)
≤ max 𝑝𝑟𝑒𝑠𝑒𝑛𝑐𝑒𝑂𝑓(𝑇𝑎𝑠𝑘𝐴𝑠𝑠𝑖𝑔𝑛ℎ,𝑖,𝑗 )
ℎ

57
2.7 Computational experiments

We perform computation experiments to compare the performance of the CP and IP in

solving industrial sized assembly line balancing problems. Data from the automotive industry is

used to define a test bed. The data belongs to 3 different bands of actual auto assembly lines,

Details on the size of each band before and after the preprocessing phase is provided in Table 5.

Four instances generated from each band’s data with different number of tasks (50, 100, 150, and

all tasks) and different number of stations. Details on test instances from all bands are provided on

Table 6. The same data is used in the dissertation by Pearce (2015).

The IP model is coded in AMPL and solved using Gurobi 6 solver. The CP is coded under

the ILOG CPLEX optimization studio 12.6 and solved using the CPLEX CP Optimizer engine.

The computer has an Intel i7-3770 CPU with 3.4 GHz clock speed and 32 GB of memory. The

solution time is limited to one hour; if the optimal solution is not provably found, the best solution

is recorded along with the time at which the best solution was reached. For CP, the lower bound

is similar to the IP one which is equivalent to LB in (5.1). The search in CP is terminated when the

solution is equal to the lower bound. Since CP search is based on random number generation, 10

runs were done for each instance with different seeds and the average solution time is recorded.

The default inference level is set to “extended” in the CP along with the precedence, interval

sequence, and no overlap inference levels.

Table 5: Number of tasks in each band

Band Original number of tasks Number of tasks after pre-processing

Band 1 395 209
Band 26 317 174
Band 47 408 178

58
Table 6: Test instances

Instance Number of tasks Number of stations

B1.1 50 5
B1.2 100 10
B1.3 150 10
B1.4 209 13
B26.1 50 5
B26.2 100 6
B26.3 150 10
B26.4 174 10
B47.1 50 5
B47.2 100 7
B47.3 150 11
B47.4 178 12

Table 7 summarizes the results of both the IP and CP subject to a one hour time limit. Since

10 runs were executed for each instance in CP, the number of runs the given solution (number of

workers) was found is also shown. As expected the IP faced difficulties as the size of the instance

increases. The optimal solution is found only in 3 small sized instances while no feasible solutions

were found on the largest instance of each band within the one hour time limit. On the other hand,

the CP managed to find optimal solutions in 10 out of the 12 instances with only a small gap in the

solution of the other two instances. In CP instance B1.3 two different solutions were found for

different seed runs, of which one is optimal. It is clear that Band 1 instances were difficult for both

the CP and IP solvers. Table 8 presents the processing time needed by the IP solver to reach the

optimal solution for different instances under no time limits. Results show that CP clearly

outperforms the IP in terms of processing time making it more applicable in the industry. Table 8

shows the time it takes the IP to reach the optimal solution when there is no time limit, in which

some instances utilized all memory before finding an optimal solution.

59
Table 7: Solution time under time limit

IP (3600 sec time limit) CP (3600 sec time limit)

Min Max Avg SD
Instance LB Solution Solution Solution
# found solution solution solution Of solution
(# of workers) time (sec.) (# of workers)
time (sec.) time (sec.) time (sec.) time (sec.)
B1.1 7 7 34.36 7 10 34.92 35.03 34.96 0.03
B1.2 12 15 3208 13 10 317.13 345.21 327.24 7.61
16 6 509.245 3970.34 2263.94 1439.80
B1.3 16 - -
17 4 442.61 505.984 468.03 30.48
B1.4 22 - - 23 10 1122.64 3347.13 1684.64 649.39
B26.1 5 5 13.47 5 10 34.11 34.68 34.30 0.23
B26.2 9 11 2030 9 10 100.47 110.16 104.98 3.19
B26.3 12 20 2552 12 10 446.15 807.44 534.73 116.33
B26.4 13 - - 13 10 291.23 485.13 354.43 56.83
B47.1 6 6 17.83 6 10 28.41 28.47 28.45 0.02
B47.2 11 12 720 11 10 135.54 137.15 136.38 0.50
B47.3 16 - - 16 10 492.02 717.79 555.20 69.22
B47.4 18 - - 18 10 805.72 972.31 842.93 51.25

60
Table 8: IP solution time under no time limit

Instance IP time to optimal solution

(days, HH:MM:SS)
B1.1 00:00:34
B1.2 1 days, 21:43:19
B1.3 1 days, 19:01:49
B1.4 -
B26.1 00:00:13
B26.2 10:28:52
B26.3 12:37:00
B26.4 -
B47.1 00:00:17
B47.2 05:05:43
B47.3 3 days, 01:19:43
B47.4 -

61
2.8 Conclusion

In this chapter, the model by Becker and Scholl (2009) is extended to the mixed model

environment with additional constraints from Falkenauer (2005) to improve the model’s

applicability to real mixed model assembly lines. First, we explain each constraint incorporated

into the model and illustrate the problem at hand with an example. Then, we propose an IP

formulation to solve the mixed model assembly line balance with parallel stations, zoning

constraints, and ergonomics. Since the problem is NP-hard, we develop a scheduling CP model to

solve industrial sized assembly line problems.

To test the model’s applicability to real assembly lines, the data of three bands from an

automotive assembly line were used. The results show the limitation of the IP, in which only

instances with 50 tasks were solved to optimality within the allowed 1 hour of processing time.

On the other hand, the CP model’s results were optimal for 10 out of the 12 instances under the

same time limit. For example, the CP managed to balance 178 tasks over 12 stations for band 47.

The computational results show the potential of CP scheduler as a tool to solve real life sized

assembly line balancing problems.

This work can be extended in future research by incorporating the idea of horizontal

balancing and model sequencing to the mixed model line balancing problem. If a worker is

expected to have a notably different amount of time to work on different models, his station might

not be able to manage the work overload caused by having several models with time consuming

tasks since only the average task duration satisfies the cycle time constraint. There are two ways

to address this problem: horizontal balancing and model sequencing. Horizontal balancing makes

the workload of each station (with respect to models/options) as uniform as possible over time

regardless of the product sequence. This can be seen as a robust mixed model line balance that

62
minimizes work overload regardless of the sequence used. On the other hand, this problem can

also be avoided by solving a sequencing problem that prohibits having multiple demanding models

in a consecutive sequence and thus avoids any work overload. This can be achieved by developing

an integrated model that solves real life sized line balancing / sequencing problems simultaneously.

63
 CHAPTER THREE

3 A CASE STUDY

3.1 Introduction

This chapter explores the models developed for the assembly line balancing problem

presented in chapter two through an increased sensitivity analysis using the case study data

supplied by an automaker. In generating the data for the case study, the research team learned that

some data is relatively easier or harder to obtain and maintain, and some constraints may not be

discovered until a potential solution is presented that violates the constraints. Requiring an OEM

to maintain large databases that are needed for only application to line balancing may increase the

potential cost. In practice, satisfying the late-discovered constraints can sometimes be

accomplished by a small permutation of the current solution, or the decision maker may determine

that this or another constraint can be violated. The nature of the mathematical programming

techniques used in this dissertation is to enforce all constraints provided, even if the decision maker

considers some of them to be less-critical. The motivation is to build intuition about the relative

cost of modeling, data collection and maintenance required to enforce some of the realistic

constraints in the mixed model line balancing problem.

Some of the line balancing constraints define the problem at a fundamental level and cannot

be ignored or relaxed. These constraints are items such as assigning each task to one worker, cycle

time, task precedence relations, and tasks overlap constraints. On the other hand, some other

constraints might be amenable to relaxation or removal such as worker interference (mounting),

ergonomic risk, the worker/station compatibility assignment constraints, and the eligibility

constraints. It maybe that some elements that are described as constraints are in fact preferences,

and can be accommodated partially. If such constraints, when included in the model, are

64
computationally difficult, it may be acceptable to remove them and manually accommodate them

later. Moreover, some constraints may not be acknowledged until a balance is proposed, at which

time a manual readjustment is made, instead of reformulating and resolving. By studying the

effects of relaxing or removing some of these constraints on the model behavior, we gain more

insight on the complexity each constraint contributes to the model so that decision makers can

make better use of it. In the following section a classification of the different line balancing

constraints is given.

3.2 The line balancing constraints

We classify the line balancing constraints as hard or soft constraints. A constraint is labeled

“hard” if removing it will either violate any of the SALBP assumptions or fundamentally affect

the way the problem is modeled. On the contrary, a soft constraint will not violate any of the

SALBP assumptions or fundamentally affect the model if removed.

3.2.1 Hard constraints

1. Task to worker assignment:

This constraint ensures that every task is assigned to one worker only.

2. Cycle time:

This constraint ensures that the total expected task durations assigned to a worker does not

exceed the cycle time.

3. Station time:

This constraint defines each station’s time window so tasks scheduled on that station are

limited to start at or after the start of that station and finish before or at the finish time of that

station.

65
 4. Precedence relation:

The precedence relation between assembly tasks is a hard constraint that cannot be relaxed.

Without the precedence constraint the problem can be transformed into a bin-packing problem in

which the goal is to assign (pack) as many tasks to every station (bin) such that the number of

stations is minimized.

5. Tasks overlap:

This constraint prohibits two tasks that are assigned to a worker from overlapping in time.

That is, the worker should finish a task before starting another one.

3.2.2 Soft constraints

1. Same mounting position worker interference (mounting):

This constraint prohibits schedules in which two workers work on the same mounting

position at the same time. That is, no two tasks that share the same mounting position overlap in

time. This constraint can be relaxed assuming that the interference can either be avoided by

manually changing the station’s schedule or by communication between workers in the same

station. These constraints are either enforced or not.

2. Task assignment restrictions constraints:

This includes same worker/station and not same worker/station constraints in addition to

the task/station eligibility constraints. These constraints are either enforced or not.

3. Ergonomic risk:

This constraint limits the maximum possible weighted ergonomic score for the tasks

assigned to any worker. This constraint can take different levels of maximum ergonomic score, or

be completely eliminated.

66
3.3 Experimental configuration

The experiments use data provided by an OEM for 3 bands in their assembly line. The

summary of the data sets used with the number of tasks along with the number of different

constraints is given in Table 9. The experiments are conducted using both the IP and CP models

developed previously.

It should be noted that because of the IP’s limitation, the maximum number of tasks is

limited to 100 in order to reach a solution in reasonable time. The solution time of the IP is limited

to a day and 3600s for the CP; if no optimal solution is found the best solution is recorded along

with the time it took the IP or CP to reach that solution. The CP experiments are run 10 times for

each instance and the average is recorded along with the number of times the solution was found.

Table 9: Test data sets

Band # Number Task assignment constraints Mounting

tasks of Same Same Not Not Eligibility zone
stations/ station worker same same constraints
workers station worker
Band 1 50 5/25 0 9 0 102 8 640
Band 26 75 5/25 2 2 0 88 34 1896
Band 47 100 7/35 0 0 2 332 14 650

There are two main experiments to assess the models’ sensitivity. The first experiment

varies the maximum weighted ergonomic score allowed for each worker and recording both the

solution and computation time. The second experiment is conducted by disabling a subset of

constraints and recording the solution along with the computation time. Because the CP

performance is based on the random number used, each CP instance is run for 10 times with

random number seeds from 0 to 9. Both of these experiments are done on an i7-4790 CPU with 32

GB on Windows 10 platform. For the IP, AMPL Version 20161220 is used along with Gurobi

7.0.2 solver. For the CP, IBM ILOG CPLEX Optimization Studio Version 12.7.0 is used.

67
3.4 Results

3.4.1 Experiment 1: Sensitivity to maximum ergonomic score

In this experiment, the maximum ergonomic score allowed is varied depending on the

tasks’ ergonomic scores for each band. Table 10, Table 11, and Table 12 show results for bands 1,

26, and 47 respectively. In these tables, the maximum ergonomic score used is shown along with

the IP’s lower bound (LB) on the solution (number of workers), the IP solution (optimal is bolded),

the computation time to reach this IP solution, the CP solution(s) (optimal is bolded), the number

of time the CP solution is found, and the average CP computation time.

The range of possible maximum ergonomic scores for the different bands increases as the

problem size increases. This range might be limited by the possible number of workers in each

station. It should be noted that for each band there is a setting for the maximum ergonomic score

in which the IP fails to reach the optimal solution even though values below and above this setting

allows optimal solutions to be found. For these hard instances, the CP finds the optimal solution

in some of the runs with increased computational time. This indicates that the problem becomes

harder for some imposed maximum ergonomic levels as the solution pool becomes smaller.

Results also show that relaxing the maximum ergonomic score does not always help in reducing

the computation time.

Table 10: Sensitivity to maximum ergonomic score in Band 1

Exp Max IP CP
(Band1) Ergo. LB Solution Time Solution # found Avg time
(HH:MM:SS) (HH:MM:SS)
E1.1 155 9 9* 00:03:59 9 10 00:00:49
E1.2 160 8 8 00:00:45 8 10 00:00:50
E1.3 165 8 8 00:01:01 8 10 00:00:53
E1.4 170 8 8 00:00:33 8 10 00:00:47
7 7 00:38:56
E1.5 175 7 8 00:00:11
8 3 00:00:38
E1.6 180 7 7 00:00:33 7 10 00:01:33
E1.7 No limit 7 7 00:00:07 7 10 00:00:36
* Optimal solution is bolded

68
Table 11: Sensitivity to maximum ergonomic score in Band 26

Exp Max IP CP
(Band26) Ergo. LB Solution Time Solution # found Avg time
(HH:MM:SS) (HH:MM:SS)
E26.1 150 12 12 00:05:43 12 10 00:00:57
10 3 00:23:54
E26.2 170 10 11 00:02:38
11 7 00:00:54
E26.3 180 10 10 00:09:20 10 10 00:01:08
E26.4 200 9 9 00:10:07 9 10 00:00:57
E26.5 250 7 8 00:00:49 7 10 00:09:00
E26.6 350 7 7 00:05:13 7 10 00:01:22
E26.7 No limit 7 7 00:08:26 7 10 00:01:08

Table 12: Sensitivity to maximum ergonomic score in Band 47

Exp Max IP CP
(Band Ergo. LB Solution Time Solution # found Avg time
47) (HH:MM:SS) (HH:MM:SS)
E47.1 90 32 32 06:47:13 32 10 00:02:17
E47.2 100 29 29 08:03:06 29 10 00:02:17
26 5 00:02:18
E47.3 110 26 27 05:32:34
27 5 00:02:26
E47.4 120 24 24 05:50:04 24 10 00:02:11
E47.5 130 22 22 08:02:31 22 10 00:02:14
E47.6 140 21 21 05:21:45 21 10 00:02:14
E47.7 150 19 19 10:03:30 19 10 00:02:25
E47.8 160 18 18 03:19:44 18 10 00:02:10
E47.9 170 17 17 07:05:31 17 10 00:02:12
E47.10 180 16 16 10:39:56 16 10 00:02:13
E47.11 200 15 15 05:12:00 15 10 00:02:13
E47.12 300 11 11 09:02:07 11 10 00:02:15
E47.13 500 11 11 11:32:06 11 10 00:02:15
E47.14 No limit 11 11 07:51:35 11 10 00:02:18

69
3.4.2 Experiment 2: Sensitivity to constraints

In this experiment, the effect of disabling one (or more) soft constraints on the solution

quality and computation time is studied. For each band, a set of 8 different instances is generated.

The first instance is created by disabling all three types of constraints, namely the ergonomics,

mounting and assignment constraints. In the next three instances, only one type of constraints is

disabled. The following three instances study the interaction between the different constraints by

disabling two of them at a time. The last instance is the problem with no constraints disabled for

comparison.

Table 13, Table 14, and Table 15 summarize the constraints sensitivity results for Bands

50, 26, and 47 respectively.

Table 13: Band 1 sensitivity to constraints

Exp Constraints IP CP
Ergonomics Mounting Assignment LB Solution Time Solution Avg time
(Max=165) (HH:MM:SS) (HH:MM:SS)
S1.1    7 7 00:00:08 7 00:00:36
S1.2    8 8 00:03:08 8 00:00:37
S1.3    7 7 00:01:05 7 00:00:34
S1.4    7 7 00:00:10 7 00:00:37
S1.5    8 8 07:37:56 8 00:00:40
S1.6    8 8 00:00:51 8 00:00:52
S1.7    7 7 00:00:07 7 00:00:36
S1.8    8 8 00:01:01 8 00:00:53

Table 14: Band 26 sensitivity to constraints

Exp Constraints IP CP
Ergonomics Mounting Assignment LB Solution Time Solution Avg time
(Max=200) (HH:MM:SS) (HH:MM:SS)
S26.1    7 7 00:04:59 7 00:00:55
S26.2    9 9 02:28:42 9 00:00:56
S26.3    7 8 00:03:58 7 00:01:21
S26.4    7 7 00:00:51 7 00:00:55
S26.5    7 10 00:04:08 7 00:01:02
S26.6    9 9 01:07:20 9 00:00:52
S26.7    7 7 00:08:26 7 00:01:09
S26.8    9 9 00:10:07 9 00:00:57

70
Table 15: Band 47 sensitivity to constraints

Exp Constraints IP CP
Ergonomics Mounting Assignment LB Solution Time Solution Avg time
(Max=200) (HH:MM:SS) (HH:MM:SS)
S47.1    11 11 12:34:17 11 00:02:07
S47.2    15 15 11:53:30 15 00:02:05
S47.3    11 11 26:06:38 11 00:02:12
S47.4    11 11 03:32:52 11 00:02:04
S47.5    15 16 10:56:52 15 00:02:04
S47.6    15 15 01:17:59 15 00:02:02
S47.7    11 11 07:36:31 11 00:02:08
S47.8    15 15 05:12:00 15 00:02:13

There is an obvious trend in computation time increase for both the IP and CP as the

number of tasks increase from 50 to 75, then to 100 across the different bands. Removing the

assignment constraints in Band 1 increased the IP time to reach an optimal solution drastically. In

addition, the IP was unable to find an optimal solution within a day after removing the assignment

constraints in bands 26 and 47. On the contrary, solving the problem with only the mounting

constraints increased the computation time in band 47; no optimal solution was reached in band

26. It should be noted that the effect of disabling all constraints in CP is minimal as the standard

deviation in computational time is less than 10 secs across all 3 bands.

Table 16, Table 17, Table 18 show the number of IP violations in the line balance incurred

by disabling the constraints for bands 1, 26, and 47 respectively. We set the maximum ergonomic

score for each band as shown in the 2nd column for those experiments in which the ergonomic

constraints are included. In the “Number of violations” column, the total number of mounting and

assignment constraints is listed. It should be noted that it is not possible to violate all constraints

at the same time, as the total number of constraints represent the number of ways a violation may

occur.

71
Table 16: Number of IP violations in Band 1

Exp Constraints Max Number of Violations

Ergonomics Mounting Assignment Ergonomic Mounting worker Assignment
(Max=165) Score interference(119) (640)
V1.1    299.77 12 36
V1.2    163.95 10 32
V1.3    259.96 0 31
V1.4    326.52 13 0
V1.5    164.56 0 27
V1.6    164.47 5 0
V1.7    215.62 0 0
V1.8    164.47 0 0

Table 17: Number of IP violations in Band 26

Exp Constraints Max Number of Violations

Ergonomics Mounting Assignment Ergonomic Mounting worker Assignment
(Max=200) Score interference(1896) (126)
V26.1    385.42 67 36
V26.2    199.83 51 33
V26.3    -* - -
V26.4    331.25 48 0
V26.5    - - -
V26.6    197.95 42 0
V26.7    403.97 0 0
V26.8    199.98 0 0
* No optimal solution obtained.

Table 18: Number of IP violations in Band 47

Exp Constraints Max Number of Violations

Ergonomics Mounting Assignment Ergonomic Mounting worker Assignment
(Max=200) Score interference(650) (362)
V47.1    456.59 40 41
V47.2    199.97 17 28
V47.3    407.03 0 35
V47.4    365.64 23 0
V47.5    - - -
V47.6    199.79 27 0
V47.7    418.28 0 0
V47.8    199.49 0 0

72
3.5 Discussion

In this section, we discuss the sensitivity analysis results obtained for each band. The

distribution of both the ergonomic score and mounting positions are shown for each band followed

by a comment on the results.

3.5.1 Band 1 (50 tasks):

The histogram for the ergonomic scores of tasks in band 1 is shown in Figure 10. A high

ergonomic score means the task is physically demanding and might cause injury to the worker

performing it. Results for band 1 - experiment 1 are shown in Table 10. The average ergonomic

score for band 1 is 24.29. The first “maximum ergonomic score” level used for this band is 155,

in which an optimal solution is found with 9 workers. Increasing the maximum ergonomic score

to 160 reduced the number of workers in the optimal solution to 8. The IP struggles when the

maximum ergonomic score is increased to 175 as the lower bound is reduced to 7 workers but the

best solution found within the time limit is 8. The CP managed to find the optimal solution in 7

out of 10 replications. The combinatorial problem of finding a combination of tasks with a total

score of 175 assigned to 7 workers and conforming to the rest of constraints seems to be harder

than other levels of maximum ergonomic score. The ergonomic constraint is not binding if a

maximum score limit of 180 or more are used as the solution to the problem is 7 workers which is

the lower bound to the problem.

73
 Band 1
30
Frequency

20
10
0
10 20 30 40 50 60 70 80 90 100
Ergonomic score

Figure 10: Band 1 ergonomic score histogram

The mounting positions used for tasks in band 1 are shown in Figure 11. Mounting position

0 is for tasks that are sub assembled on the side of the station. The tasks are distributed on the four

corners of the car; this is reflected on the lower count of mounting position constraints and

violations shown in Table 16. Also, this explains why this band behaved differently when the

mounting position constraints are the only constraints enforced as shown in Table 13. The solution

time did not change significantly as it did in the other two bands. The biggest impact on solution

time occurred when both the mounting and ergonomics constraints are enforced. The IP solution

time increased from a minute to over 7 hours. The CP solution time did not change and is consistent

across the different instances of experiment 2.

Band 1
15
Frequency

10
5
0
0 1 2 3 4 5 6 7 8 9
Mounting positions

Figure 11: Band 1 mounting positions histogram

74
3.5.2 Band 26 (75 tasks):

Ergonomic scores of band 26 tasks are shown in the histogram of Figure 12. The average

ergonomic score for band 26 is 22.53. Unlike band 1, band 26 histogram shows a trend in the

relation between the number of tasks and ergonomic score. Results for band 26 - experiment 1 is

shown in Table 11. Starting maximum ergonomic score for experiment 1 is 150; solving the

problem gives an optimal line balance with 12 workers. Increasing the maximum ergonomic score

to 170 makes the problem hard to solve for the IP. The number of workers are reduced to 11 but

the lower bound is 10. The IP is unable to reach the optimal solution within the time limit. The CP

managed to find an optimal line balance with 10 workers in 3 out the 10 replications. In addition,

the IP was unable to reach the optimal when the maximum score limit is increased to 250 while

the CP managed to get to the optimal in all 10 replications. Increasing the maximum ergonomic

score to 350 or more makes this constraint a non-binding constraint.

Band 26
20
Frequency

15
10
5
0
10 20 30 40 50 60 70 80 90 100
Ergonomic score

Figure 12: Band 26 ergonomic score histogram

The tasks in band 26 mostly belongs to two mounting position as shown in Figure 13. This

is the reason behind having so many mounting position constraints. It took the IP 5 minutes to

reach the optimal solution without enforcing any of the soft constraints. Adding the ergonomics

constraints increased the time to reach the optimal for the IP. The huge number of mounting

75
constraints affected the IP as expected; no optimal solution is found by the IP within the time limit.

By enforcing the assignment constraint only the IP solution time is reduced to less than a minute.

By enforcing all soft constraints, the IP is able to find an optimal solution in 10 minutes. Disabling

the mounting constraints only led to an increase in the IP’s computation time. In addition, by

disabling the assignment constraints, the IP is unable to reach the (relaxed) optimal solution within

the time limit.

Band 26
60
Frequency

40
20
0
1 2 3 4 5 6 7 8 9
Mounting positions

Figure 13: Band 26 mounting positions histogram

The number of band 26 violations recorded for experiment 2 is shown in Table 17. The

highest number of violations occurred when no soft constraints are enforced. Among the 75 tasks,

there are 67 worker interference cases in the line balancing schedule. This number might look

small when compared to the number of mounting constraints, but it is not possible to violate all of

the constraints at the same time and this number is large relative to the number of tasks. The

number of assignments violation is also large relative to the total number of tasks. That is, almost

half of the tasks violated the assignment constraint. Enforcing either the ergonomics, the

assignment, or both constraints led to a decrease in mounting worker interference violations.

76
3.5.3 Band 47 (100 tasks)

The histogram for the ergonomic scores of band 47 is shown in Figure 14. The average

ergonomic score for this band is 28.29. The scores for this band are almost uniform with more

tasks having a low ergonomic score. This is reflected on the wide range of maximum ergonomic

scores for which the IP is able reach an optimal solution. The results of band 47 - experiment 1 is

shown in Table 12; the maximum ergonomic score for this band is varied between 90 and 500.

The lower maximum ergonomic score used (90) yielded a line balance with 32 workers while the

highest level used (300) resulted in a line balance with only 11 workers. This range gives the

decision maker a good room to change the maximum ergonomic score based on the balance of

cost/workers’ health risks. Again for one level of maximum ergonomic score (110), the IP is unable

to reach the optimal solution within the time limit. However, the CP reached the optimal solution

in 5 out of the 10 replications. The time of CP computations is more consistent when compared to

the time of IP computations.

Band 47
40
Frequency

30
20
10
0
10 20 30 40 50 60 70 80 90 100
Ergonomic score

Figure 14: Band 26 ergonomic score histogram

The mounting positions used in band 47 are shown in Figure 15. In this band, all mounting

position are used except for mounting position 5. The results for band 47 - experiment 2 are shown

in Table 15. Solving the problem with no soft constraints enforced yielded a line balance with 11

workers. It took the IP over 12 hours to reach the optimal solution; perhaps there are many alternate

77
solutions, and symmetry breaking constraints would help. When the mounting position constraints

are the only soft constraints enforced, the IP computation time is doubled when comparing to the

solution time when no constraints are enforced. For the case when the mounting position

constraints are disabled and the assignment and ergonomic constraints are enforced, the IP

computation time is reduced from 12 hours to 1 hour. Also, enforcing only the assignment

constraints reduced the IP time to reach the optimal solution. On the other hand, the IP is unable

to reach an optimal solution within the time limit when the assignment constraints are removed

and the other two soft constraints are enforced. Enabling all constraints reduced the computation

time to less than half the case with no constraints enforced.

Band 47
20
Frequency

15
10
5
0
1 2 3 4 5 6 7 8 9
Mounting positions

Figure 15: Band 47 mounting positions histogram

The number of constraint violations recorded for each instance of band 47 - experiment 2

is shown in Table 18. When disabling all soft constraints, the number of mounting and assignment

violations amount to less than half the number of tasks on this band which is still high. In this line

balance, one of the workers is assigned tasks with a total weighted ergonomic score of 456.59. By

applying only the ergonomic constraints, both the mounting and assignment violations are reduced.

78
3.6 Conclusion

In this chapter, two sensitivity analysis experiments are conducted using data from an

automaker. The first experiment studies the effect of changing the maximum ergonomic score limit

on the solution quality and computational time. Results show that the problem becomes harder for

some levels of maximum ergonomic level limit. For these instances, the IP failed to reach the

optimal solution within the allowed time. However, the optimal solution is found in some of the

CP runs for each hard instance. The decision maker should be aware of this when deciding on the

maximum ergonomic limit. The second experiment studies the effect of disabling a subset of

constraints on the solution quality and computational time. Results show that the assignment

constraint helps the IP reach the optimal solution within the time limit. By disabling the assignment

constrain the IP is unable to reach the optimal in two bands and the computation effort required to

reach the optimal increased in one band. Furthermore, using only the mounting constraint led to a

huge computational effort to reach the optimal in one band while the optimal is not even reached

in one band.

These three bands, while from an operating OEM and hence observed and not designed for

this study, provide varied characteristics that may be useful in future research. The bands have

different distributions of ergonomic scores and different types of mounting positions that may

provide some hypotheses for future research. For example, a further study of the impact of

ergonomic score distribution can be conducted assuming all parts are in the same mounting

position and only one worker can be assigned to each station; this would extend the concept

presented earlier of a two-dimensional bin packing problem enhanced with precedence constraints.

As a result of such a study, individual tasks could be targeted for redesign or increased attention,

based on their impact on the overall line balance. As another example of a future study, we observe

79
that two bands studied here had disjoint mounting positions while the third had contiguous

mounting positions. Further, in band 26, the mounting positions were such that no single person

could work in more than one, while the potential for a person to work in more and more mounting

positions was moderate in band 1 and extreme in band 47. The types of creative assignments of

tasks to workers in the bands with more potential for work being done in adjacent mounting

positions may merit further investigation.

Exploring other soft constraints that are just a preference and may not have significant

technical implications (e.g., some precedence relations) maybe a good direction for future

research. By identifying which of these constraints are hard on the IP, the computation time can

be reduced without affecting the quality of the results obtained. In addition, since the results of

this study is not consistent, further investigation is needed.

80
 CHAPTER FOUR

4 MIXED MODEL SEQUENCING PROBLEM

4.1 Introduction

Mixed model assembly lines are used to produce a variety of customized products in a cost

efficient flow mass production environment. Customers have the privilege of selecting the options

they would like to be included in their product and the manufacturer will have to manage

significant product variety as a result. In order to keep costs down, manufacturers utilize the

worker/machine flexibility that is available in the mass production environment to jointly produce

different unique products in an intermixed fashion on the same line. After dealing with the long

term problem of line balancing, another short term problem is faced which is the sequencing

problem. The sequencing problem is solved to answer the question of how to sequence the given

different models for production in a production run. Different production sequences will have

different economic impacts that are related to worker utilization or material usage. The sequencing

problem in the literature is introduced using two different general objectives, namely work

overload minimization and leveling part usage. Work overload occurs when several work intensive

models are sequenced in a consecutive manner, which can be avoided by choosing a sequence in

which models that require much work alternate with others that require less work. Furthermore,

since different options usually require different parts or material, the model sequence is an

important factor in parts logistics. In this work we will be concentrating on the work overload

minimization objective. See Boysen et al. (2009a) for a survey on level scheduling research.

According to Boysen et al. (2009a), there are two different approaches in the literature to

handle the work overload problem:

81
  The mixed model sequencing problem: In this approach, a detailed schedule of the

production run is sought taking into account operation times, worker movement, station

length, station borders, and other line characteristics.

 The car sequencing problem: This approach avoids the troubles of collecting data that is

needed for the detailed mixed model sequencing by minimizing the work overload in an

indirect manner by formulating a set of sequencing rules. These rules are to limit the

occurrence of certain time consuming options in a given length of order sequence. The

objective is to minimize the violations to these rules.

In the following section a review on the comparison between the two approaches is given

followed by a literature review on related work. The subsequent section illustrates the problem

description with an example. This is followed by model building and mathematical formulation.

Next, the proposed solution approach is shown followed by computational experiments.

4.2 Mixed model sequencing versus car sequencing

The two approaches to handle the work overload problem in a mixed model assembly line

are the mixed model sequencing problem (MMS) and the car sequencing (CS) problem. Although

the two problems may have the same business goal, they are based on two different mathematical

formulations with different objective functions. The MMS objective directly minimizes the work

overload while the CS problem has an indirect objective in which the number of sequencing rules’

violations is minimized. These rules are generated using different approaches that aim to avoid

having several labor-intensive car models in consecutive order which in turn causes work overload.

CS is a more aggregate approach for finding a production sequence when compared with MMS. It

assumes that different models can be distinguished by either having the option or not. For each

option 𝑜, CS uses a sequencing rule 𝐻𝑜 : 𝑁𝑜 to limit the occurrence of that option to at most 𝐻𝑜 in

82
any subsequence of 𝑁𝑜 models. For example, a rule 2:5 for the option sunroof means that among

any subsequence of five models only two of them may have the option sunroof.

Golle et al. (2014) conducted experiments to quantify the difference in solution quality

between the two approaches. They used different approaches to generate the rules from the

literature such as the Bolat and Yano sequencing rule which assumes only one option at each

station. Since the Bolat and Yano rule is limited and was found to exclude feasible solutions from

consideration, Golle et al. (2014) also tested the multiple sequencing rule approach that ensures

that no feasible sequence violates a sequencing rule. In addition, the authors considered the three

most commonly used objective functions in the literature. Also, several weighting factors from the

literature were used. Tests were done on a large number of randomly generated instances. The

results showed that although the objectives of the two approaches are positively/ linearly

correlated, the MMS solution quality outperformed the CS counterpart. The authors claim that the

solution of the CS problem is not even competitive across all test instances and the different

objectives used. According to the authors, the CS results in solutions with at least 23% more work

overload when compared to the MMS problem. They managed to decrease this gap to 15% by

considering their weighting factor for CS. Furthermore, they noted that the difference in quality

increases with the use of inadequate rules and objective functions. They concluded that the use of

surrogate objective function in CS aggregate too much data which makes it less useful in finding

a sequence that truly minimizes the work overload. Following this result, we will be using the

MMS approach.

83
4.3 Problem description

In mixed model assembly lines, different models that have a variety of customizations are

assembled on the same line. Since these models require different processing times at each station

depending on the custom options required, usually some of them will require an amount of time

that is higher than the cycle time while others will have a processing time which is less than the

cycle time. If several of the models with the high processing times are produced in a consecutive

order, workers at some station will have a hard time finishing the required task before reaching the

end of the station which will result in the inability to return to the beginning of the station when a

new work piece enters the station, or in other words, the worker will fall behind. This shift in a

worker’s position will accumulate as long as models that require a processing time that is higher

than the cycle time are sequenced after each other. Work overload occurs when the worker is

unable to finish working on the work piece within the boundaries of the station. If work overload

occurs, several reactions are possible including but not limited to:

- The assembly line is stopped until all work is done on work pieces across all stations.

- Utility workers are used either to help workers or to take over the work on the work piece

that is causing the overload.

- The unfinished work is left to be done off line in a special station after the work piece exists

the last station of the line.

- The worker increases his/her production speed at the risk of quality issues.

To avoid any of the costly reactions to work overload, mixed model sequencing attempts to

find a sequence such that cars with options that require a high processing time alternates with other

cars that require less processing time at each station. The use of utility workers has been studied

in the literature as one of the methods to handle work overload but the detail of how utility workers

84
are used differed. The assumption that is usually used in the literature was named the side-by-side

policy by Boysen et al. (2011). In this assumption the utility worker is called whenever an overload

situation is anticipated. The utility worker is assumed to support the regular worker by working

with him side by side and doubling his processing speed. This way the worker will only require

half the duration of the work overload amount to finish processing the required tasks on the car. It

is assumed that the utility worker help starts at a position in the station such that the work required

at that station is done by the right hand border of the station assuming that the flow of the assembly

is from left to right. Boysen et al. (2011) argue that the assumption of side-by-side policy is

restrictive and not realistic because, for example, work piece size may allow only one worker, the

assumption of doubling speed is not realistic and dependent on the interaction between the

workers, and utility workers cannot simply show up exactly when needed. As an alternate solution,

they propose the skip policy which they claim is being used by major European car manufacturers.

In the skip policy, the worker calls the utility worker whenever a work overload is expected. The

utility worker then takes over the work on the work piece that is expected to cause the overload

while the normal worker skips it and starts working on the next work piece coming after it. The

utility worker is usually a group leader who is responsible for several stations and who is cross

trained to be able to handle all tasks that are assigned to any of his stations. The difference between

the two policies is illustrated with an example in the following section.

Stations on a production line might be closed in the sense that no operators can cross their

boundaries to an adjacent station. This limitation is necessary if the working area require specific

environmental conditions such as heating champers or paint shops. On the contrary, workers can

cross borders of open stations but the range of how far they can go into an adjacent stations is

always limited. The restriction might be related to a limit on the range of a power tool or the

85
window of the material handling system. Furthermore, the areas of adjacent stations may overlap

but workers from the two stations are not allowed to interfere with each other by working on the

same work-piece at the same time. Open stations are useful in mixed model production because it

possible for a worker who finished working early on a work-piece to “swim” (move upstream) to

the previous station so he can have more time to work on labor-intensive models. With a proper

production sequence, the worker can avoid any work overload or the need of the utility worker

help.

4.4 Related work

Yano & Rachamadugu (1991) investigate the mixed model sequencing problem with the

objective of minimizing work overload. First, the authors propose an optimal algorithm for the

case of one station and two products. Then they propose a greedy heuristic that is based on branch

and bound to solve the problem with multiple stations. They tested the proposed heuristic on data

provided by an automobile company and compared the heuristic results against the results the

company get using their own procedure. Yano & Rachamadugu (1991) claim that the proposed

heuristic reduces total work overload by 55% on average.

Bolat & Yano (1992a) develop a procedure to minimize utility work for the MMS problem

with one station, two types of products, and no buffers. They prove that the proposed spacing rule

simple procedure is optimal if the load in the system is sufficiently high or low. In addition, they

develop error bounds on the algorithm if the load conditions are not satisfied. For a moderately

loaded system, Bolat & Yano (1992a) develop a greedy algorithm that is optimal under certain

processing time requirements. For cases in which the optimality is not guaranteed, the authors

develop a dynamic program in addition to a variation of the greedy algorithm proposed. Results

show that the greedy algorithm outperforms the spacing rule procedure over some parameter

86
values. The authors note that in the class of problems for which the solutions are still far from

optimal require a combination of different procedure to be solved. Moreover, Bolat & Yano

(1992b) introduce a surrogate objective to the utility work that yields a scheduling problem that is

easier to solve. The objective is a car sequencing rule of not having a number of cars with an option

in any sequence of a given length.

Tsai (1995) considers the problem of mixed model sequencing for one station with the

objectives of minimizing both utility work and the risk of line stoppage. The author lists two

different management philosophies in handling work overload in U.S. and Japan. In the U.S. the

utility worker is deployed to finish work left undone by the primary worker. On the other hand in

Japan the operator pushes a stop button when he is unable to finish his work. A proof that the

problem of minimizing either objective is NP hard in the strong sense is given first. Assuming that

product processing times can take only one of two distinct values, Tsai (1995) proposes an optimal

algorithm that minimizes both objectives in O(log N) computation time. Furthermore, the objective

value from the single station case can be used as a lower bound for a multiple stations/processing

times case.

Sarker & Pan (1998) consider a mixed model assembly line with both open and closed

stations. They study the problem of minimizing the utility and idle time costs that are incurred due

to various line parameters along with the sequencing of different models. The authors develop two

different models for the open-station and closed-station systems in order to find the optimal

parameters and sequences for each system. The models are able to provide the best parameters that

would minimize the total cost. The results show that for a given line length, the open-station system

incurred less cost when compared to the closed-station counterpart.

87
 Scholl et al. (1998) study the MMS problem with fixed launch rate, closed stations, and the

objective of minimizing the work overload. They propose a reformulation of the problem to

consider patterns in subsequences that can be put together to make the complete sequence. The

basic idea is to decompose the problem of finding the complete sequence to problems of finding

appropriate patterns and then combining them to get the solution. They define a pattern to be a

subsequence in which the worker resets his position to the station border after completing all the

work required for that subsequence. Scholl et al. present a method to solve the problem by first

generating patterns based on a pattern based vocabulary building strategy using the column

generation technique and then searching for a complete sequence (solution) using tabu search.

Furthermore, the authors show that the vocabulary building procedure of the method is not affected

by the number of products to be sequenced while other phases of the method will be affected by

the computational complexity of solving for a longer sequence.

Bautista & Cano (2008) propose a single station sequencing heuristic algorithm with two

types of products that is based on the heuristics proposed by Yano & Rachamadugu (1991) and

Bolat & Yano (1992a). They also extend their heuristic to account for multiple products and

multiple stations. Results show a satisfactory performance in terms of quality and speed.

Cano-Belmán et al. (2010) use a hyper heuristic to solve the problem of MMS with closed

stations (without links). The proposed hyper heuristic uses scatter search as high level

metaheuristic along with a low level priority rule constructive heuristic. The authors compare the

two proposed hyper heuristics against other algorithms from Bautista and Cano (2008), improved

best solution using local search, and exact solution using CPLEX. They find that improvement

methods are needed to get good quality solutions that outperform other heuristics in comparison.

88
 Bautista & Cano (2011) extend the models of Yano and Rachamadugu (1991) with the

objective of maximizing the total work completed in the assembly line, and Scholl (1998) with the

objective of minimizing the work overload. The proposed models take into consideration the link

between neighboring stations. This is done by adding a new constraint that limits the starting

position of the work done on a work piece form starting while work is not finished in the previous

station. The variables of the model are based on the work piece instead of a period. The authors

introduce an interruption rule for the management of interruptions and work intensification and

propose two new models based on the rule to make the stations more autonomous. Furthermore,

they propose a solution that they named the bounded dynamic programming to solve the given

sequencing problem. Computational results on 225 instances form the literature show the fast

performance of the proposed method compared to the CPLEX solver. Next, the authors present a

case study from the Nissan powertrain plant in Barcelona to test the new proposed solution method

on real world assembly lines. The solution and lower bounds of the proposed approach were better

compared to CPLEX but results were not guaranteed to be optimal.

Bautista et al. (2012) present two models for mixed model sequencing problem for

assembly lines with stations in series and homogeneous processors in parallel extending the models

in Bautista & Cano (2011). The difference between the two models is the use of absolute time

scale versus the use of relative time scale. The same assumptions that the line can be stopped and

operation interruption can happen anytime while retaining the work in progress still hold. In

addition, it is assumed that a worker who did not complete all the work at the station is allowed to

leave the unit so that the unfinished work can be completed later. Bautista et al. (2012) define an

alternate surrogate for workload minimization which is minimizing the variation of work overload

rates across all stations on the line. They propose different approaches to avoid work overload

89
concentrations at each workstation and at specific time instants of the work day. These approaches

are regulating the work required, the work completed, or the overload at a station throughout the

workday. They apply the proposed approaches to a set of examples representing different scenarios

from the powertrain production line of a Nissan plant in Barcelona. Finally, they compare the

different approaches and conclude that models with regularity functions produce better results

when compared to previous models.

Boysen et al. (2011) study utility work organization in the MMS problem with closed

stations. They propose a more realistic problem in which the objective is to minimize the work

overload situations. In the literature, the side-by-side policy is often assumed in which a utility

worker helps the regular worker finish the remaining work whenever the regular worker is unable

to finish the work assigned to him within the stations’ boundaries. In the proposed policy, a normal

worker skips working on a work-piece whenever work overload is expected and the utility worker

starts working alone on that work-piece (calling this the skip policy). In addition, Boysen et al.

(2011) develop a branch and bound procedure along with different heuristics to solve the problem.

Finally, the authors show the advantage of the proposed skip policy by comparing it to the

traditional side-by-side policy.

Tamura et al. (2011) consider the MMS problem with the objective of minimizing the line

stoppage time. They propose a formulation for the problem and derive a relationship between the

line stoppage time and makespan. The authors develop a branch and bound algorithm and show

the limitation of solving numerical examples using commercial mathematical programming

packages. They conclude that heuristics are needed to solve larger problems.

Gujjula et al. (2011) develop a heuristic that is based on Vogel’s approximation to solve

the MMS problem. The motivation for their work is to find a method to solve real-life large

90
problems. They use two other heuristics from the literature to benchmark the results of the test

scenarios; the first is a combination of a priority rule-based method with lower bounds

determination, and the second is an adjusted version of the heuristic proposed by Bautista and

Cano (2008). Finally, Gujjula et al. (2011) show the performance of the proposed heuristic by

comparing the results to the results of the benchmark heuristics.

Cortez, & Costa (2015) study the problem of MMS with a heterogeneous workforce which

is motivated by assembly lines in sheltered work centers for the disabled. Based on heterogeneous

workforce, task processing time is both model and worker dependent. The authors formulate the

problem and propose a constructive heuristic method that is based on a simplified and

approximated models for the problem. Furthermore, Cortez, & Costa (2015) propose a meta-

heuristic that is based on the greedy randomized adaptive search procedure (GRASP). The authors

show the speed of proposed heuristics by comparing it against the solution obtained using CPLEX.

Bautista et al. (2015) consider the mixed model sequencing in which the objective is to

minimize the work overload. They propose to consider the work pace factor in the model by the

use of variable processing times with the objective of minimizing the work overload or increasing

the completed work. The activity of workers is set by using a function that represents the periods

of adaptation and fatigue throughout the working day. The authors use a case study from the Nissan

powertrain plant in Barcelona to analyze the performance of the proposed model using the Gurobi

solver. The results show that a very small increase in the activity of the operators can reduce the

work overload or completely avoid it.

Recently, Mosadegh et al. (2016) study the problem of MMS with one station, stochastic

processing times, and the objective of minimizing the expected total work overload. The authors

find that their proposed dynamic programming approach is not guaranteed to find the optimal

91
solution because of the recursive computation required to calculate the work overload. To

overcome this problem, the authors reformulated the problem as a shortest path problem. They

propose a heuristic that is based on Dijkstra’s algorithm to solve the problem and show the quality

of the solution obtained by comparing to results from other methods.

4.4.1 An example

Consider an assembly line with 3 stations, in which each station has a length of 13 time

units (TU) which is the time a work-piece spends in the station. Assume that every 10 TU a new

work-piece is launched down the assembly line (cycle time) and that workers are not allowed to

swim past the station borders. Furthermore, assume that the processing time for each model at each

station is as shown in Table 19 and the demand is 2 units of models 1 & 2 and 1 unit of models 3

& 4. Let the sequence be (model 1, model 1, model 2, model 2, model 3, and model 4).

Table 19: Example data for the mixed model sequencing problem

Time required Time required Time required

Model Demand
at Station 1 at Station 2 at Station 3
1 2 13 8 11
2 2 10 12 13
3 1 8 10 9
4 1 9 11 10

In the figures, the production sequence starts from the bottom and ends at the top of the

figure. Figure 16 shows the position and movement of both the regular and utility workers for the

side-by-side policy. The horizontal solid blocks represent the movement of the regular worker

when he is working on a work-piece while the diamond filled blocks represent the movement of

both the regular and utility workers when they are working together on a work-piece. The dashed

diagonal lines represent the return movement of the regular worker to start working on the next

work-piece.

92
 The outlined diamond blocks represent the time in which both the utility worker and the

normal worker are both working side-by-side on the work piece. The time it takes the regular

worker to reposition is assumed to be instantaneous.

Figure 16: Movement in the side-by-side policy

Figure 17 shows the movement for the skip policy case, resulting in an overload usage

count of 3 compared to 5 in the side by side policy for the same given sequence. The checkerboard

blocks represent the time the utility worker is working at the regular pace alone on the work-piece.

93
Figure 17: Movement in the skip policy

If it is assumed that a regular worker can swim across his station’s border to the previous

station, he can start working on the coming work piece in the station that precedes his station if he

is ready and if the worker in the previous station has finished the required tasks on that work piece

early. The number of overload usage can be reduced for the given sequence if swimming is allowed

as in Figure 18.

94
Figure 18: Movement with swimming allowed

Even if swimming of workers is not allowed, an optimal model sequence to the example

reduce the number of overload usage count to 1. The detailed worker movement diagram for an

optimal solution is shown in Figure 19. The worker in station 2 working on the first Model 2 car

is able to swim into the previous station to start working on it early. This way, no utility worker is

needed for the second Model 2 car at station 2 because the overload situation is avoided.

95
Figure 19: Worker movement in an optimal solution for the example

In this work, the skip policy with swimming allowed is used in the formulation and the solution

of the mixed model sequencing problem with the objective of minimizing the work overload

situations. The following assumptions are often used in real world assembly lines and are used to

model the problem:

- The mixed model demand and the processing time of all tasks is deterministic.

- Working across the borders of the station is possible (open stations), and we assume that a

worker can swim (drift) to the previous station so he can start working on the work piece

earlier if the worker in that station finished working before the end of the station.

- The processing time for all tasks is always less than the station’s length.

96
- Regular workers will not be required to drift to the next station to complete any task since

utility workers take care of overload situations and tasks are always less than the station

length.

- The length units (LU) and time units (TU) are the same assuming that conveyor velocity is

normalized to 1 LU/TU.

- The operator returns to the next work piece with infinite velocity. This is because the speed

of the conveyor is usually much slower than the human speed.

- Regeneration is not required, i.e. there is no need to reset workers’ position to zero after

the last period.

- Stations may have different lengths and different starting positions for workers.

- The cycle time is fixed and work pieces are launched down the line in equal intervals.

97
4.5 Model formulation

Figure 20: Movement of a worker in a station

In this work, we extend the model provided by Boysen et al. (2011) by adding station to

station interaction via worker swimming. We propose an IP and a heuristic to solve the MMS

problem with swimming and assess their performance using data from Bautista & Cano (2011).

Therefore, we will compare the IP and heuristic performance as opposed to comparing our

heuristic results to Boysen et al. (2011) heuristic results.

The parameters and variables needed for the MMS are shown in Table 20 and are

illustrated in Figure 20. Each station 𝑘 has a length 𝑙𝑘 that is greater than the cycle time 𝑐, and an

early start position 𝑞𝑘 that may be less than 0 to allow the worker to swim to the previous station

so he can start working early on the next work piece in sequence if that is needed. The worker will

not be able to swim to the previous station unless the worker in that station finishes his work early

98
in the previous period (due to the potential for work location overlap). The binary decision variable

𝑥𝑚𝑡 takes the value 1 when car model 𝑚 is assigned to period 𝑡 and 0 otherwise.

If the projected ending time for model 𝑚 at station 𝑘 (𝑝𝑚,𝑘 ) in period 𝑡 is greater than the

station’s length (𝑠𝑘,𝑡 + ∑𝑚 𝑝𝑚𝑘 𝑥𝑚𝑡 ≥ 𝑙𝑘 ), a utility worker is called to help and the regular worker

completely skips the overload causing work piece to start working on the work piece that follows

it (𝑦𝑘,𝑡 = 1). Thus, the starting position of each worker depends on both the finishing position of

the same worker in the previous period and the finishing position of the worker in the previous

station in the previous period. This can be described as:

max(𝑓𝑘,𝑡−1 − 𝑐, 𝑓𝑘−1,𝑡−1 − 𝑙𝑘−1 , 𝑞𝑘 ) 𝑖𝑓 𝑦𝑘,𝑡−1 = 0

𝑠𝑘,𝑡 = { (4.1)
max(𝑠𝑘,𝑡−1 − 𝑐, 𝑓𝑘−1,𝑡−1 − 𝑙𝑘−1 , 𝑞𝑘 ) 𝑖𝑓 𝑦𝑘,𝑡−1 = 1

Where 𝑓𝑘,𝑡 = 𝑠𝑘,𝑡 + ∑𝑚 𝑝𝑚𝑘 𝑥𝑚𝑡

99
4.6 Mathematical model

Table 20: Mixed model sequencing problem’s notations

𝑀 Number of models indexed 𝑚 = 1,2,3, … , 𝑀

𝑇 Number of production cycles indexed 𝑡 = 1,2,3, … , 𝑇
𝐾 Number of stations indexed 𝑘 = 1,2,3, … , 𝐾
𝑐 Cycle time
𝑙𝑘 Length of station 𝑘
𝑝𝑚𝑘 Processing time of model 𝑚 at station 𝑘
𝑑𝑚 Demand for copies of model 𝑚 (𝑑𝑚 = 1 if all cars are unique)
Minimum starting position for the worker at station 𝑘 (𝑞𝑘 = 0 if no
𝑞𝑘
swimming is allowed)
𝑎𝑘 Initial starting position for the worker at station 𝑘
𝜖 A very small number

Decision variables:

1 𝑖𝑓 𝑚𝑜𝑑𝑒𝑙 𝑚 𝑖𝑠 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑 𝑡𝑜 𝑐𝑦𝑐𝑙𝑒 𝑡

𝑥𝑚𝑡 = {
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

1 𝑖𝑓 𝑢𝑡𝑖𝑙𝑖𝑡𝑦 𝑤𝑜𝑟𝑘𝑒𝑟 𝑖𝑠 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑 𝑖𝑛 𝑠𝑡𝑎𝑡𝑖𝑜𝑛 𝑘 𝑐𝑦𝑐𝑙𝑒 𝑡

𝑦𝑘𝑡 = {
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑠𝑘𝑡 = Starting position of worker in station 𝑘 when cycle 𝑡 begins.

The model may be represented with the use of an additional variable 𝑓𝑘𝑡 that represents

the finishing position of worker in station 𝑘 when cycle 𝑡 ends. In the following we use the

expression 𝑠𝑘𝑡 + ∑𝑚 𝑝𝑚𝑘 𝑥𝑚𝑡 instead of introducing an additional decision variable to the integer

program.

Objective:

𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = ∑ ∑ 𝑦𝑘𝑡
𝑡 𝑘

The objective is to minimize the number of overload situations across all production cycles

and stations.

Subject to the following constraints:

100
 Only one model is assigned to each production cycle.

∑ 𝑥𝑚𝑡 = 1 ∀ 𝑡 = 1, … , 𝑇 (C4.1)
𝑚

 Demand for car models should be satisfied during the production horizon.

∑ 𝑥𝑚𝑡 = 𝑑𝑚 ∀ 𝑚 = 1, … , 𝑀 (C4.2)
𝑡

 The variable that indicates an overload situation is flagged using the following constraints.

This variable 𝑦𝑘𝑡 should be equal to 1 if an overload will occur when the finishing position

of the worker 𝑓𝑘𝑡 exceeds the station’s length 𝑙𝑘 . (Recall, 𝑠𝑘𝑡 + ∑𝑚 𝑝𝑚𝑘 𝑥𝑚𝑡 represents the

finishing position of worker in station 𝑘 when cycle 𝑡 ends). It should be noted that a

redundant constraint is used (C4.4) compared to the formulation in Boysen et al. (2011) in

which only one constraint is used. The reason is because when one constraint only is used,

𝑦𝑘𝑡 will be free to take any value in some cases but will be pushed to equal zero since the

minimization of the sum of 𝑦𝑘𝑡 is the objective. The redundant constraint will help setting

the flag variable 𝑦𝑘𝑡 correctly regardless of the objective function and will define the

feasible region more completely.

𝑠𝑘𝑡 + ∑ 𝑝𝑚𝑘 𝑥𝑚𝑡 − 𝑙𝑘 − 𝜖 ∀ 𝑡 = 1, … , 𝑇, 𝑘 = (C4.3)

𝑚
1, … , 𝐾
≥ −𝑙𝑘 (1 − 𝑦𝑘𝑡 )

𝑠𝑘𝑡 + ∑ 𝑝𝑚𝑘 𝑥𝑚𝑡 − 𝑙𝑘 ≤ 𝑙𝑘 𝑦𝑘𝑡 ∀ 𝑡 = 1, … , 𝑇, 𝑘 = (C4.4)

𝑚
1, … , 𝐾

 The starting position of the worker of each station depends on the finishing position of the

same worker in the previous cycle if no utility worker is needed (no overload situation). In

101
 the case of overload situation, a utility worker is called to help and the worker can skip the

current cycle’s car and start working on the next cycle’s car.

𝑠𝑘𝑡 ≥ 𝑠𝑘(𝑡−1) + ∑ 𝑝𝑚𝑘 𝑥𝑚(𝑡−1) − 𝑐 ∀𝑡 = (C4.5)

𝑚
2, … , 𝑇, 𝑘 = 1, … , 𝐾
− 𝑙𝑘 𝑦𝑘(𝑡−1)

𝑠𝑘𝑡 ≥ 𝑠𝑘(𝑡−1) − 𝑐 − 𝑙𝑘 + 𝑙𝑘 𝑦𝑘(𝑡−1) ∀𝑡 = (C4.6)

2, … , 𝑇, 𝑘 = 1, … , 𝐾

 A worker cannot start working on a car unless the worker in the previous station has

finished working on it. In case the previous station had a utility worker in the previous

period, the utility worker will always finish working on the car before the end of the station.

𝑠𝑘𝑡 ≥ 𝑠(𝑘−1)(𝑡−1) + ∑ 𝑝𝑚(𝑘−1) 𝑥𝑚(𝑡−1) − 𝑙𝑘−1 ∀𝑡 = (C4.7)

𝑚
2, … , 𝑇,
− (𝑙𝑘−1 − 𝑐)𝑦(𝑘−1)(𝑡−1)
𝑘 = 2, … , 𝐾

 The starting position of workers should be greater than or equal to the minimum starting

position allowed in his station and less than or equal the difference between the length of

the station and the cycle time.

𝑠𝑘𝑡 ≥ 𝑞𝑘 ∀ 𝑡 = 1, … , 𝑇, 𝑘 = 1, … , 𝐾 (C4.8)

𝑠𝑘𝑡 ≤ 𝑙𝑘 − 𝑐 ∀ 𝑡 = 1, … , 𝑇, 𝑘 = 1, … , 𝐾 (C4.9)

 Initial starting positions for workers at the first cycle.

𝑠𝑘1 ≥ 𝑎𝑘 ∀ 𝑘 = 1, … , 𝐾 (C4.10)

 Variable domains.

𝑥𝑚𝑡 ∈ {0,1} ∀ 𝑡 = 1, … , 𝑇,𝑚 = 1, … , 𝑀 (C4.11)

𝑦𝑘𝑡 ∈ {0,1} ∀ 𝑡 = 1, … , 𝑇, 𝑘 = 1, … , 𝐾 (C4.12)

102
4.7 Solution approach

Since the problem of mixed-model sequencing (MMS) is an NP-hard optimization

problem, we expect that there will be limitation in using the Integer Program (IP) to solve

realistically sized problems in a reasonable amount of time. However, it is useful to have an IP

that is able to solve small and medium sized problems and that can be used as a benchmark to any

other solution methods used. In this section we present a meta-heuristic named the greedy

randomized adaptive search procedures (GRASP) to solve the problem at hand.

4.7.1 Greedy randomized adaptive search procedures (GRASP)

Metaheuristics represent a family of approximate optimization general and high-level

procedures that coordinate simple heuristics or rules to find good quality solution for hard

combinatorial optimization problem in reasonable time.

The GRASP metaheuristic is an iterative greedy heuristic that was proposed by Feo and

Resende (1989) as a heuristic to solve the set covering problem and later was named in the

following paper by Feo and Resende (1995). GRASP is composed of two steps: construction of a

solution and local search. In the construction step, the solution is constructed by iteratively adding

elements to the partial solution in a greedy way such that each added element has a minimal impact

on the solution’s objective value. After a solution is constructed, local search is performed to check

the solution’s neighborhood for a better solution. The greedy algorithm is randomized to be able

to generate several solutions. Since the local search depends on the initial solution, this

metaheuristic will be efficient if the constructive heuristic samples different promising parts of the

search space.

103
4.7.1.1 Constructive greedy randomized algorithm

Consider the combinatorial optimization problem of minimizing 𝑓(𝑆) over all

solutions 𝑆 ∈ 𝑋, which is defined by a set of feasible solutions 𝑋, and by an objective function 𝑓.

The first step in the GRASP metaheuristic is to generate a good feasible solution to the problem.

The heuristic starts with an empty solution and appends an element from a restricted candidate list

RCL to the solution at each iteration. The determination of the next element to be added to the

solution is done in part using a greedy evaluation function that represents the incremental increase

in cost that accompanies adding that element to the partial solution. Using that function a list of

candidate elements that have the lowest increase on the cost of the partial solution is generated.

The candidate set is the set of all possible elements that can be incorporated into the solution.

Furthermore, the RCL is the list of candidate elements that will impact the objective the least. In

order to make use of randomness in generating different solutions with good quality, the element

to be added to the partial solution is selected randomly from the RCL. A pseudo code for the greedy

randomized solution generation is given in Figure 21.

104
 Algorithm: Constructive greedy randomized algorithm.
1. 𝑆 ← ∅;
2. Populate the candidate set 𝐶;
3. Evaluate the incremental cost 𝑐(𝑒) for all 𝑒 ∈ 𝐶;
4. while 𝐶 ≠ ∅
5. Build a restricted candidate list with elements that have an
incremental cost equal to the minimum across the candidate set.
𝑅𝐶𝐿 ← 𝑒 ∀ 𝑐(𝑒) = min(𝑐(𝑒): 𝑒 ∈ 𝐶);
6. Select element 𝑠 ∈ 𝑅𝐶𝐿 at random;
7. Incorporate 𝑠 into the solution: 𝑆 ← 𝑆 ∪ {𝑠};
8. Update the candidate set 𝐶;
9. Reevaluate the incremental cost 𝑐(𝑒) for all 𝑒 ∈ 𝐶;
10. end
11. return 𝑆;
end

Figure 21: Greedy randomized algorithm for a minimization problem

4.7.1.2 Neighborhood function

The definition of the neighborhood function is a required step in implementing any

metaheuristic that is based on a single solution (such as simulated annealing and tabu search) and

not a population of solutions (such as genetic algorithms). This function should be selected

carefully because it plays an important role in the performance of the metaheuristic. A

neighborhood of a solution 𝑆 is defined by the set 𝑁(𝑆) ⊆ 𝑋. The members of the neighborhood

set are generated by the application of a move operator that performs a small perturbation to the

solution 𝑆. It is desirable that a neighborhood solution be local by performing a move operation

that yields small changes in the solution value. There are two general types of operators for a

permutation solution representation. The first is the position-based neighborhood such as the

insertion operator in which one random element is removed from its position to be inserted into a

new random position in the sequence. The insertion operation is illustrated in Figure 22. The

105
second is the order-based neighborhood, with operators like the exchange and inversion operators

illustrated in Figure 23 and Figure 24 respectively. In the exchange operator, two randomly

selected elements are swapped while in the inversion operator, two random elements are selected

and the sequence between these elements is inverted.

Figure 22: Insertion operator

Figure 23: Exchange (swap) operator

Figure 24: Inversion operator

4.7.1.3 Local search

Local search iteratively generates solutions from a predefined neighborhood 𝑁(𝑆) and

moves to a better solution once and if it is found. The exploration of the neighborhood is usually

stochastic because enumerating the whole neighborhood is time-prohibitive. A pseudo code

example of a local search is given in Figure 25.

106
 Algorithm: Local search
1. while the max number of iteration is not reached do
2. Find 𝑆 ′ ∈ 𝑁(𝑆) with 𝑓(𝑆 ′ ) < 𝑓(𝑆);
3. 𝑆 ← 𝑆′;
4. end
5. return 𝑆;
end

Figure 25: Local search algorithm

4.7.1.4 GRASP for the MMS problem:

There are two primary design decisions to implement GRASP for the MMS problem: the

objective function and the neighborhood.

Objective function:

For a given sequence, let 𝑏𝑘,𝑡 denote the current processing time for the model assembled

in station 𝑘 in cycle 𝑡. In order to know the number of overload situations for this given sequence,

the starting 𝑠(𝑘, 𝑡) and finishing 𝑓(𝑘, 𝑡) position of workers at all stations 𝑘 = 1, … , 𝐾 for all

production sequences 𝑡 = 1, … , 𝑇 should be calculated. The overload for each station at each

production cycle 𝑜(𝑘, 𝑡) can be easily calculated using the skip policy detailed earlier. The number

of overload situation for the given sequence is equal to the sum of all overload situation across all

the stations and production cycles. The algorithm for the evaluation of a given sequence is given

in Figure 26.

107
 Algorithm: overload situations for a given sequence
1. Initialize starting positions at t=1
for 𝑘 = 1 𝑡𝑜 𝐾
𝑠(𝑘, 1) = 0;
end
2. Calculate workers’ finish & start positions for all production cycles and record any overload situation
for 𝑘 = 1 𝑡𝑜 𝐾
𝑓(𝑘, 1) = 𝑠(𝑘, 1) + 𝑏(𝑘, 1);
end
for 𝑡 = 2 𝑡𝑜 𝑇
if 𝑓(1, 𝑡 − 1) ≤ 𝑙(1) & 𝑜(1, 𝑡 − 1) == 0
𝑠(1, 𝑡) = max{0, 𝑓(1, 𝑡 − 1) − 𝑐} ;
else
𝑠(1, 𝑡) = max{0, 𝑠(1, 𝑡 − 1) − 𝑐} ;
end
𝑓(1, 𝑡) = 𝑠(1, 𝑡) + 𝑏(1, 𝑡);
if 𝑓(1, 𝑡) ≥ 𝑙(1)
𝑜(1, 𝑡) = 1;
𝑠(1, 𝑡) = 0;
𝑓(1, 𝑡) = 𝑠(1, 𝑡) + 𝑏(1, 𝑡);
end
end
for 𝑘 = 2 𝑡𝑜 𝐾
for 𝑡 = 2 𝑡𝑜 𝑇
if 𝑓(𝑘, 𝑡 − 1) ≤ 𝑙(𝑘) & 𝑜(𝑘, 𝑡 − 1) == 0
𝑠(𝑘, 𝑡) = max{𝑞(𝑘), 𝑓(𝑘, 𝑡 − 1) − 𝑐, 𝑓(𝑘 − 1, 𝑡 − 1) − 𝑙(𝑘 − 1)} ;
else
𝑠(𝑘, 𝑡) = max{𝑞(𝑘), 𝑠(𝑘, 𝑡 − 1) − 𝑐, 𝑓(𝑘 − 1, 𝑡 − 1) − 𝑙(𝑘 − 1)} ;
end
𝑓(𝑘, 𝑡) = 𝑠(𝑘, 𝑡) + 𝑏(𝑘, 𝑡);
if 𝑓(𝑘, 𝑡) ≥ 𝑙(𝑘)
𝑜(𝑘, 𝑡) = 1;
𝑠(𝑘, 𝑡) = 0;
𝑓(𝑘, 𝑡) = 𝑠(𝑘, 𝑡) + 𝑏(𝑘, 𝑡);
end
end
end
5. Calculate the total overload
𝑇𝑜𝑡𝑎𝑙𝑜𝑣𝑒𝑟𝑙𝑜𝑎𝑑 = ∑𝑘 ∑𝑡 𝑜(𝑘, 𝑡);

Figure 26: Computing the objective function by evaluating a sequence

108
4.7.1.5 Neighborhood function for the MMS problem:

The solution to the MMS problem is naturally represented by a permutation since it is the

models’ sequence to be assembled on the assembly line. The adjacency of the models in the

sequence affects the number of overload situations. Thus, the order-based neighborhood is more

suitable for the MMS problem and specifically the exchange (swap) operator. This operator swaps

the position of a predefined number of elements in the sequence. Swapping more than two elements

in the sequence of the MMS problem makes a big perturbation that might never lead to a better

solution. For the GRASP algorithm, the 2- exchange operator is used in which two production

orders in the sequence are swapped randomly.

4.7.1.6 The constructive randomized greedy algorithm:

The first step in the randomized greedy algorithm is to populate the candidate set, 𝐶, which

in our case is the different models to be assembled on the assembly line. The inputs to the algorithm

are the task duration on all stations for all models, cycle time, stations’ length, swimming amount

allowed for each station, and the demand of each model. Initially, the candidate list will have all

models and the solution can start by picking a model randomly. However if swimming is not

allowed, the task with highest total task duration across all stations can be picked first since it is

the most demanding model and the starting position for workers are initialized to zero. The next

step in the greedy algorithm is to update the candidate list, by updating the demand for each model

and removing models with a demand of zero, and calculate the cost (overload situations) of adding

each of the models to the sequence. After that, a restricted candidate list (RCL) is generated from

the candidate list by keeping models that will incur the minimum cost when added to the sequence

and removing all the other candidate models. Afterwards, one of the models in the RCL is selected

randomly and appended to the sequence. The procedure will iterate until all demand is satisfied.

109
4.8 Computational experiments

In our computational experiments, we use instances from the Nissan powertrain plant in

Barcelona that are presented in Bautista & Cano (2011). The plant assembles nine types of engines

that are grouped under three families (4x4, vans and trucks) through an assembly line that is

composed of 21 workstations. The stations are arranged serially and are linked to each other. In

addition, each station has the same number of operators under normal operating conditions. The

processing time for each type of engine on each stations is provided in Table 21. This table is a

result of a line balance presented in Bautista and Pereira (2007) that is done for the Nissan

powertrain plant in Barcelona. The 46 instances are grouped into two blocks: the first has 23

demand plans each for 270 units, which is equivalent to a production schedule for a full working

day with two shifts. The second block has 23 demand plans for 540 units, which is equivalent to a

production schedule for two working days. An effective cycle time (c=175s) is used in the

experiments with a fixed station length for all stations (l=195s) with no swimming allowed (q=0).

The demand plans for the 46 instances are shown in Table 22.

110
Table 21: Processing time for the 9 types of engines

Stations
i
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
p1 104 103 165 166 111 126 97 100 179 178 161 96 99 147 163 163 173 176 162 164 177
p2 100 103 156 175 114 121 96 97 174 172 152 106 101 155 152 185 179 167 150 161 161
p3 97 105 164 172 114 122 96 95 173 172 168 105 102 142 156 183 178 181 152 157 154
p4 92 107 161 167 115 124 93 106 178 177 167 97 101 154 152 178 169 180 152 159 168
p5 100 101 148 168 117 127 96 94 178 178 167 101 99 146 153 169 173 172 160 162 172
p6 94 108 156 167 117 130 89 102 171 177 166 100 101 143 152 173 178 173 151 160 170
p7 103 106 154 168 115 120 94 103 177 175 172 96 96 154 154 172 174 173 155 162 167
p8 109 102 164 156 111 121 101 102 171 173 157 104 102 153 156 182 175 168 148 158 149
p9 101 110 155 173 111 134 92 100 174 175 177 96 99 155 156 171 175 184 167 157 169

111
Table 22: Demand plans for 46 instances

i Block 1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
p1 30 30 10 40 40 50 20 20 70 10 10 24 37 37 24 30 30 60 10 20 60 20 10
p2 30 30 10 40 40 50 20 20 70 10 10 23 37 37 23 30 30 60 10 20 60 20 10
4x4

p3 30 30 10 40 40 50 20 20 70 10 10 23 36 36 23 30 30 60 10 20 60 20 10
P4 30 45 60 15 60 30 75 30 15 105 15 45 35 45 55 35 55 30 90 15 15 90 30
Van

p5 30 45 60 15 60 30 75 30 15 105 15 45 35 45 55 35 55 30 90 15 15 90 30
p6 30 23 30 30 8 15 15 38 8 8 53 28 23 18 23 28 18 8 15 45 15 8 45
p7 30 23 30 30 8 15 15 38 8 8 53 28 23 18 23 28 18 8 15 45 15 8 45
Trucks

p8 30 22 30 30 7 15 15 37 7 7 52 27 22 17 22 27 17 7 15 45 15 7 45
p9 30 22 30 30 7 15 15 37 7 7 52 27 22 17 22 27 17 7 15 45 15 7 45
Total 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270 270
i Block 2
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46
p1 60 60 20 80 80 100 40 40 140 20 20 47 74 74 47 60 60 120 20 40 120 40 20
p2 60 60 20 80 80 100 40 40 140 20 20 47 73 73 47 60 60 120 20 40 120 40 20
4x4

p3 60 60 20 80 80 100 40 40 140 20 20 46 73 73 46 60 60 120 20 40 120 40 20

P4 60 90 120 30 120 60 150 60 30 210 30 90 70 90 110 70 110 60 180 30 30 180 60
Van

p5 60 90 120 30 120 60 150 60 30 210 30 90 70 90 110 70 110 60 180 30 30 180 60

p6 60 45 60 60 15 30 30 75 15 15 105 55 45 35 45 55 35 15 30 90 30 15 90
p7 60 45 60 60 15 30 30 75 15 15 105 55 45 35 45 55 35 15 30 90 30 15 90
Trucks

p8 60 45 60 60 15 30 30 75 15 15 105 55 45 35 45 55 35 15 30 90 30 15 90
p9 60 45 60 60 15 30 30 75 15 15 105 55 45 35 45 55 35 15 30 90 30 15 90
Total 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540 540

112
 The integer program experiments were run on AMPL version 20161220 using two different

versions of the Gurobi solver, namely Gurobi 6.5.0 and Gurobi 7.0.2 solvers. The GRASP

experiments were coded and run on MATLAB R2016a. Block 1 and Block 2 experiment with

50,000 iterations were run on a machine that has an Intel core i7-3770 @3.4 Ghz processor, 32GB

RAM, and Windows 7. Block 2 experiments with 100,000 iterations were executed on a machine

with an Intel core i7-2600 @ 3.4 Ghz processor, 16GB RAM, and Windows 7.

IP results using the two versions of the Gurobi solver are shown in Table 23. The lower

bound (LB) and best solution (number of overload situations) are reported along with the

computation time required to reach the best solution. Computational time might be very small

compared to the 3600 seconds time limit because the branch and bound method used by the IP

solver might get stuck on a solution early on and need over 3600 seconds to find a better solution.

The time column is the time at which the reported solution was found. The newer version of the

solver found better solutions in 17 out of the 46 instances, while it found worse solutions in 14 out

of the 46 instances (best values bolded). Furthermore, for instances in which both versions found

the same objective function value, the newer version was faster in 6 instances and slower in 6

(lower time bolded). Also, the newer version found a better lower bound in 3 instances but had a

worse lower bound in 1. The better result, regardless of which solver found it, is used in the rest

of this chapter.

113
Table 23: IP results for two versions of Gurobi

instance Gurobi ver 6.5.0 (3600 Seconds time limit) Gurobi ver 7.0.2(3600 Seconds time limit)
LB Solution Time* (sec) LB Solution Time* (sec)
1 3 8 2856 3 7** 324
2 11 20 1929 11 22 2135
3 18 24 582 18 24 1611
4 9 13 3558 9 13 2501
5 23 34 3203 24 36 3443
6 11 21 1292 11 21 2858
7 30 38 203 30 38 2823
8 4 4 3111 4 4 1680
9 25 36 1178 25 34 3122
10 51 60 1810 51 58 474
Block 1

11 3 3 281 3 3 360
12 8 17 2921 10 17 1908
13 7 14 23 7 15 223
14 13 26 1129 12 25 2441
15 17 24 1681 17 28 217
16 5 13 168 5 13 479
17 18 29 267 18 28 186
18 18 29 1730 18 28 109
19 40 47 571 40 49 67
20 3 3 3322 3 3 2283
21 18 30 1703 18 27 2140
22 41 49 2174 41 50 951
23 5 5 17 5 5 9

24 6 26 800 6 20 1939
25 23 44 2905 23 42 1006
26 38 50 801 38 53 465
27 20 34 1197 20 33 2004
28 50 84 41 50 77 1125
29 24 49 2602 24 54 2392
30 63 81 2710 63 84 602
31 7 12 2817 7 15 909
32 52 80 3397 52 77 927
33 103 120 1261 103 116 3479
Block 2

34 6 7 1492 6 7 1229
35 21 36 2745 21 46 810
36 16 38 1607 16 36 878
37 27 55 959 27 55 183
38 35 52 1516 35 52 1346
39 11 28 3547 11 27 2309
40 37 53 2521 37 62 1022
41 36 75 185 37 66 1449
42 81 100 3595 81 97 731
43 6 13 2903 6 13 3530
44 38 65 852 38 67 1789
45 84 105 1920 84 110 484
46 10 11 3598 10 11 2350
* The time at which the reported solution was found.
** Bolded is better.

114
 In order to benchmark the GRASP algorithm in solving the MMS problem, the results

obtained using the algorithm are compared against the best IP solution from Table 23. Since

GRASP is based on randomization, a number of replications is done for each instance to better

assess the performance of the algorithm. Each instance is solved 30 times using GRASP and the

following is recorded: minimum solution, maximum solution, solution average, solution standard

deviation, number of times the best solution is reached, and the average computation time. The

GRASP was limited to 100,000 iterations for block 1 instances, which is equivalent to

approximately 10 minutes of computation time. For block 2 instances, the 50,000 iterations is

equivalent to approximately 10 minutes of computation time while the 100,000 iterations takes

approximately 20 minutes to finish.

Results of the IP and GRASP algorithm for Block 1 and Block 2 are shown in Table 24

and Table 25 respectively. Note that the IP was able to find the optimal solution for only 4 instances

of block 1 within the one hour time limit. On the other hand, no optimal solutions were found for

block 2, which is expected due to the size of the problem. The GRASP algorithm managed to find

better or optimal solutions to all instances including the ones in which the IP has failed to reach

the optimal solution. In addition, the worst GRASP solution found among the 30 replications is

still better than the IP solution for all but two of these instances. The average time for the GRASP

to reach the best solution is around 4.4 minutes for block 1 instances, 6.2 minutes for block 2

instances with 50,000 iterations, and 14.7 minutes for block 2 instances with 100,000 iterations.

Looking at Table 24 it is clear that some instances are easier than the others for the IP and GRASP.

Instances 8, 11, 20, 23 in which the IP has found an optimal solution are easier than the other

instances even for the GRASP algorithm. Another way to look at the difficulty of the different

115
instances is by looking at the number of times the best solution is found by the GRASP algorithm

across the 30 replications. This number varies between instances but it is clear that the number of

times the best solution is reached for instances in block 1 is higher than the same number for

instances of block 2. This is expected as block 2 instances are bigger and require more computation

time for the algorithm to be efficient. Furthermore, the average solution is improved when the

number of iterations is doubled as we can see in Table 25. It should be noted that the instances in

which the demand patterns with more truck engines (p6, p7, p8, p9) are easier to solve when

compared to other instances and have lower bounds from the IP for the total number of overload

situations. This is a result of how the line balance distributed the truck engines’ related tasks across

the assembly line stations.

Table 24: IP vs GRASP Block 1 results

Instance IP (3600 seconds time limit) GRASP (600 seconds time limit, 30 replications)
LB Solution Time* (sec) Min Max Avg. SD # of best Avg. Time* (sec)
1 3 7 324 5 6 5.53 0.51 14 321.89
2 11 20 1929 15 17 15.40 0.56 19 294.85
3 18 24 582 21 22 21.27 0.45 22 184.20
4 9 13 2501 12 13 12.50 0.51 15 236.68
5 24 34 3203 29 32 30.37 0.81 4 344.70
6 11 21 1292 15 18 17.10 0.84 1 389.64
7 30 38 203 34 35 34.17 0.38 25 238.29
8** 4 4 1680 4 5 4.07 0.25 28 204.29
9 25 34 3122 30 33 31.60 0.86 3 325.19
10 51 58 474 55 56 55.17 0.38 25 234.60
Block 1

11 3 3 281 3 3 3.00 0.00 30 37.22

12 10 17 1908 12 13 12.53 0.51 14 332.26
13 7 14 23 10 13 11.50 0.63 1 315.72
14 13 25 2441 17 19 18.13 0.78 7 336.19
15 17 24 1681 20 22 20.20 0.48 25 291.06
16 5 13 168 8 10 8.57 0.63 15 322.96
17 18 28 186 21 24 22.33 0.71 3 363.42
18 18 28 109 23 26 24.37 0.67 1 354.98
19 40 47 571 43 45 43.80 0.55 8 221.81
20 3 3 2283 3 5 3.87 0.68 9 233.42
21 18 27 2140 23 25 24.17 0.65 4 307.45
22 41 49 2174 45 46 45.47 0.51 16 186.23
23 5 5 9 5 5 5.00 0.00 30 63.56
* The time at which the reported solution was found.
** Bold indicates that an optimal solution was found.

116
Table 25: IP vs GRASP Block 2 results

Instances IP (3600 seconds time limit) GRASP (50,000 iterations, 30 replications) GRASP (100,000 iterations, 30 replications)
LB Solution Time* Min Max Avg. SD # of Avg. Min Max Avg. SD # of Avg.
(sec) best Time* best Time*
(sec) (sec)
24 6 20 1939 15 19 16.73 1.01 2 440.90 13 17 15.13 1.01 1 840.63
25 23 42 1006 33 39 36.63 1.27 1 410.05 34 37 34.97 0.85 9 965.92
26 38 50 801 43 46 44.90 0.76 1 317.87 43 46 44.33 0.71 3 722.22
27 20 33 2004 27 31 28.77 0.97 2 451.55 26 29 27.63 0.85 3 919.57
28 50 77 1125 65 69 67.27 1.20 4 429.57 64 67 65.97 0.89 1 954.46
29 24 49 2602 38 44 41.17 1.49 1 441.83 37 41 38.67 1.06 5 1055.59
30 63 81 2710 71 74 72.40 0.81 3 415.26 70 73 71.37 0.72 2 861.42
31 7 12 2817 9 13 11.30 0.84 1 353.50 8 12 10.23 1.01 2 957.99
32 52 77 927 69 72 70.23 1.04 8 425.39 66 70 67.93 1.36 6 1080.62
33 103 116 3479 112 115 113.90 0.80 1 261.68 112 114 113.27 0.58 2 668.76
Block 2

34 6 7 1229 7 8 7.10 0.31 27 252.69 7 7 7.00 0.00 30 413.72

35 21 36 2745 28 33 30.63 1.19 1 408.86 27 30 28.83 0.91 1 991.19
36 16 36 878 27 32 29.60 1.28 1 394.73 26 30 27.87 0.97 2 925.01
37 27 55 183 41 45 43.17 1.09 2 394.87 39 45 41.77 1.14 1 1025.49
38 35 52 1346 43 47 44.87 1.11 2 376.74 42 46 43.53 1.04 4 940.22
39 11 27 2309 21 25 22.70 0.99 3 426.92 19 23 20.80 1.00 2 1056.14
40 37 53 2521 47 53 50.77 1.41 1 410.32 48 51 49.23 0.82 4 975.48
41 37 66 1449 52 59 56.27 1.55 1 404.08 51 56 53.77 1.22 2 1056.06
42 81 97 731 90 91 90.73 0.45 8 254.97 89 91 90.27 0.52 1 670.71
43 6 13 2903 9 13 10.73 0.98 3 375.46 8 11 9.97 0.56 1 819.02
44 38 65 852 53 57 54.67 1.35 7 427.03 51 54 52.37 1.03 7 1023.20
45 84 105 1920 93 96 94.83 0.65 1 356.54 92 95 94.03 0.76 2 868.90
46 10 11 2350 11 12 11.67 0.48 10 226.12 11 12 11.20 0.41 24 609.56

117
 In order to assess the benefit of swimming on minimizing the overload situation,

experiments are conducted using GRASP to compare the results when 𝑞𝑘 = 0 and 𝑞𝑘 =

−17 for 𝑘 = 2, . . , . 𝐾. The swimming amount 𝑞𝑘 is chosen to be 10% of the cycle time 𝑐 =

175. The experiments are done with the same random number seed for both cases. Table

26 and Table 27 show the results for blocks 1 and 2 respectively. Results show that by

allowing workers to swim 10% of the cycle time the work overload situations is reduced

by 46.6% and 48% on average for blocks 1 and 2 respectively. Swimming had most impact

on instance 16 of block 1 in which the average of work overload situations is reduced by

%58.8 from 9.7 to 4, and in instance 24 of block 2 in which the average of work overload

situations is reduced by %63.2 from 14.5 to 5.3.

Table 26: Effect of swimming on Block 1 overload situations
Instances Overload situations with no swimming, Overload situations with swimming
𝑞𝑘 = 0 allowed, 𝑞𝑘 = −17
Min Max Avg SD # of Min Max Avg SD # of
best best
1 5 7 6.30 0.65 3 2 3 2.70 0.47 9
2 14 18 15.93 0.94 2 7 9 8.27 0.52 1
3 21 22 21.57 0.50 13 12 12 12.00 0.00 30
4 12 14 12.93 0.58 6 7 8 7.50 0.51 15
5 29 33 31.30 0.84 1 16 18 17.27 0.52 1
6 16 19 18.30 0.65 1 8 10 9.00 0.45 3
7 34 35 34.47 0.51 16 19 20 19.20 0.41 24
8 4 6 4.23 0.50 24 2 2 2.00 0.00 30
9 31 34 32.37 0.72 2 18 19 18.93 0.25 2
10 55 56 55.67 0.48 10 32 33 32.13 0.35 26
Block 1

11 3 3 3.00 0.00 30 2 2 2.00 0.00 30

12 12 15 13.57 0.97 5 7 7 7.00 0.00 30
13 11 13 12.23 0.68 4 5 6 5.97 0.18 1
14 17 20 18.83 0.79 1 9 11 9.47 0.57 17
15 20 22 20.60 0.67 15 11 11 11.00 0.00 30
16 8 11 9.70 0.70 1 4 4 4.00 0.00 30
17 22 24 23.33 0.61 2 12 12 12.00 0.00 30
18 24 27 25.40 0.67 2 13 14 13.40 0.50 18
19 43 45 44.33 0.55 1 25 25 25.00 0.00 30
20 3 5 4.13 0.51 2 2 2 2.00 0.00 30
21 23 26 24.67 0.76 2 13 15 14.07 0.37 1
22 44 47 45.63 0.61 1 26 27 26.50 0.51 15
23 5 5 5.00 0.00 30 3 3 3.00 0.00 30

119
Table 27: Effect of swimming on Block 2 overload situations
Instances Overload situations with no swimming, Overload situations with swimming
𝑞𝑘 = 0 allowed, 𝑞𝑘 = −17
Min Max Avg SD # of Min Max Avg SD # of
best best
24 13 16 14.50 0.94 5 5 6 5.33 0.48 20
25 33 37 35.07 1.23 2 16 18 16.63 0.72 15
26 43 46 44.43 0.82 4 24 25 24.93 0.25 2
27 26 29 27.27 0.83 5 14 16 15.03 0.72 7
28 63 68 65.70 1.24 1 34 36 35.03 0.76 8
29 37 41 39.27 1.01 1 19 21 19.70 0.53 10
30 71 73 71.67 0.61 12 40 42 40.87 0.43 5
31 9 12 10.57 0.82 3 5 5 5.00 0.00 30
32 65 70 67.73 1.17 1 37 40 38.47 0.63 1
33 112 115 113.30 0.65 2 65 67 65.83 0.53 7
34 7 8 7.03 0.18 29 4 4 4.00 0.00 30
Block 2

35 27 31 28.90 0.92 2 14 14 14.00 0.00 30

36 25 29 27.50 1.07 1 12 14 13.10 0.71 6
37 39 44 41.37 1.10 2 19 22 20.73 0.58 1
38 42 46 43.80 0.85 1 23 24 23.03 0.18 29
39 19 23 21.03 1.03 1 8 11 8.83 0.79 11
40 48 51 49.40 0.89 4 24 26 25.43 0.73 4
41 51 55 53.93 1.01 1 27 30 28.37 0.72 3
42 90 91 90.20 0.41 24 52 52 52.00 0.00 30
43 9 12 10.00 0.83 9 5 6 5.10 0.31 27
44 50 54 52.37 1.16 1 28 30 29.27 0.74 5
45 93 95 94.13 0.57 3 54 55 54.17 0.38 25
46 10 12 11.07 0.37 1 6 6 6.00 0.00 30

120
4.9 Conclusion

In this chapter, the MMS problem with open stations and skip policy is studied. The

problem is introduced first and then compared to the car sequencing problem. This is

followed by a literature review on related work. In this work, we adopt the skip policy in

managing work overload in which a utility worker takes over all the work required for a

work piece in a given station whenever an overload is expected. In the case a utility worker

is called, the normal worker skips the current work piece to the next one in sequence. In

addition, interaction between stations is allowed through workers swimming beyond

stations boundaries. The problem description is given and both the skip policy and

swimming is illustrated by examples. Next the mathematical model formulation is given

followed by the details of the proposed GRASP algorithm that is developed to solve the

problem.

To assess the performance of the IP and GRASP, computation experiments are done

on data from the literature. Two different versions of the IP solver are used and the best

result is compared against the GRASP result. Results show that the GRASP algorithm

performed better in a fraction of the time when compared to the IP solution. In addition,

results show that work overload situations can be reduced by up to 63.3% by allowing

workers to swim a distance (LU) equal to 10% of the cycle time (TU).

The MMS problem discussed in this work and in the literature consider the

movement of one worker per station. A possible future research is considering the

movement of multiple workers per station along with the mounting position they are

working on. For example, a worker can still swim to work on a free mounting zone while

121
other workers are still working on other mounting zones. This will open up the possibility

of having more room for swimming. This could also change the way line balancing is done

by trying to schedule tasks on certain mounting zones early to anticipate for workers

swimming from the following station.

122
 CHAPTER FIVE

5 CONCLUSION AND FUTURE WORK

In this dissertation, the model by Becker and Scholl (2009) is extended by considering

the suggestions, in Falkenauer (2005) and Pearce (2015), to build a realistic mixed model

line balancing model with parallel stations, zoning constraints, assignment restrictions, and

ergonomics. The IP formulation is shown followed by a scheduling CP model that is

developed to solve industrial sized assembly line balancing problems. Both models were

compared using data provided from an OEM. Results show that the CP model

outperformed the IP model and is considered a potential tool for solving the mixed model

assembly line balancing problem.

In the following chapter, the OEM data are used for two sensitivity analysis

experiments. The first experiment studies the effect of changing the maximum ergonomic

score limit on the performance of both the IP and CP models. The second experiment

examines the effect of disabling a subset of constraints on the performance of the IP and

CP. Results show that CP computation time was always consistent except in some hard

instances which required more computation time. The problem becomes harder to solve

for some levels of maximum ergonomic score. Results also show that disabling the

assignment constraint increased the time needed for the IP to reach the optimal solution.

Finally, the mixed model sequencing problem is studied. Interaction between

stations are allowed through workers swimming. The skip policy is used to manage the

work overload situations. An IP formulation is presented along with a GRASP algorithm

that is developed to solve the problem. Results show that the GRASP algorithm managed

123
to find better solutions in less computation time than the IP. By allowing workers to swim,

work overload situations can be reduced by up to 63.3%. This managerial insight is a key

contribution of this chapter.

Future work can be based on either a microscopic or macroscopic view of the

assembly systems modeled in this dissertation. A microscopic view explores the details to

increase the realism of the models presented. This would make the models more applicable

but would add to the complexity of the problem. One possible future direction would be to

add constraints to the IP or CP models such as a cognitive constraint. Since each assembly

task requires some skill and knowledge from the worker to perform it, would it be more

realistic to assume that each worker has a limited knowledge of how different assembly

tasks are performed? Moreover, could the demand of switching tasks, as in a station or

work assignment that manages many different models, be cognitively tiring, even if it is

not ergonomically or physically challenging? Another extension to the line balancing

problem would be to consider a better ergonomics model instead of using a simple

constraint with an upper bound. Is it possible to implement a more realistic ergonomics

model in the CP and how would that affect solution time and quality? Furthermore, since

tasks duration and demand values are assumed to be deterministic in this work, what if they

are assumed to be stochastic? In addition, the effect of line balancing on the robustness of

the sequencing problem can be studied. How would the model/task correlation in two

consecutive stations affect the sequencing? Also, since line balancing is done using parallel

workers; how would the mixed model sequencing problem differ if the movement of more

than one worker is considered?

124
 A macroscopic extension looks at the big picture and not the details. This approach

ignores the details for the sake of solving an integrated problem. An interesting research

direction would be to integrate the mixed model line balancing problem with the

sequencing problem. Is it better to solve the line balancing and sequencing problems in one

step? Another research idea would be to develop a model that can solve bigger problems.

Is it possible to come up with a line balance for the whole assembly line? Finally, line

balancing and sequencing problems can be also integrated with the incoming and outgoing

supply chains. How would integrating supply chain affect the line balancing and

sequencing decisions?

125
 REFERENCES

Akpınar, S., & Bayhan, G. M. (2011). A hybrid genetic algorithm for mixed model
assembly line balancing problem with parallel workstations and zoning constraints.
Engineering Applications of Artificial Intelligence, 24(3), 449-457.
AkpıNar, S., Bayhan, G. M., & Baykasoglu, A. (2013). Hybridizing ant colony
optimization via genetic algorithm for mixed-model assembly line balancing problem
with sequence dependent setup times between tasks. Applied Soft Computing, 13(1), 574-
589.
Bautista, J., & Cano, J. (2008). Minimizing work overload in mixed-model assembly
lines. International Journal of Production Economics, 112(1), 177-191.
Bautista, J., & Cano, A. (2011). Solving mixed model sequencing problem in assembly
lines with serial workstations with work overload minimisation and interruption rules.
European Journal of Operational Research, 210(3), 495-513.
Bautista, J., Cano, A., & Alfaro, R. (2012). Modeling and solving a variant of the mixed-
model sequencing problem with work overload minimisation and regularity constraints.
An application in Nissan’s Barcelona Plant. Expert systems with applications, 39(12),
11001-11010.
Bautista, J., Alfaro, R., & Batalla, C. (2015). Modeling and solving the mixed-model
sequencing problem to improve productivity. International Journal of Production
Economics, 161, 83-95.
Bautista, J., & Pereira, J. (2007). Ant algorithms for a time and space constrained
assembly line balancing problem. European journal of operational research, 177(3),
2016-2032.
Baybars, I. (1986). A survey of exact algorithms for the simple assembly line balancing
problem. Management science, 32(8), 909-932.
Becker, C., & Scholl, A. (2006). A survey on problems and methods in generalized
assembly line balancing. European journal of operational research, 168(3), 694-715.
Becker, C., & Scholl, A. (2009). Balancing assembly lines with variable parallel
workplaces: Problem definition and effective solution procedure. European Journal of
Operational Research, 199(2), 359-374.
Bock, S. (2008). Using distributed search methods for balancing mixed-model assembly
lines in the automotive industry. Or Spectrum, 30(3), 551-578.

126
Bockmayr, A., & Pisaruk, N. (2001). Solving assembly line balancing problems by
combining IP and CP. Proceedings of the 6th Annual Workshop of the ERCIM Working
Group on Constraints, Prague, Czech Republic.
Bolat, A., & Yano, C. A. (1992a). Scheduling algorithms to minimize utility work at a
single station on a paced assembly line. Production Planning & Control, 3(4), 393-405.
Bolat, A., & Yano, C. A. (1992b). A surrogate objective for utility work in paced
assembly lines. Production Planning & Control, 3(4), 406-412.
Boysen, N., Fliedner, M., & Scholl, A. (2007). A classification of assembly line balancing
problems. European Journal of Operational Research, 183(2), 674-693.
Boysen, N., & Fliedner, M. (2008). A versatile algorithm for assembly line balancing.
European Journal of Operational Research, 184(1), 39-56.
Boysen, N., Fliedner, M., & Scholl, A. (2008). Assembly line balancing: which model to
use when? International Journal of Production Economics, 111(2), 509-528.
Boysen, N., Fliedner, M., & Scholl, A. (2009). Assembly line balancing: Joint
precedence graphs under high product variety. IIE Transactions, 41(3), 183-193.
Boysen, N., Fliedner, M., & Scholl, A. (2009a). Sequencing mixed-model assembly lines:
Survey, classification and model critique. European Journal of Operational Research,
192(2), 349-373.
Boysen, N., Kiel, M., & Scholl, A. (2011). Sequencing mixed-model assembly lines to
minimise the number of work overload situations. International Journal of Production
Research, 49(16), 4735-4760.
Bukchin, J., Dar-El, E. M., & Rubinovitz, J. (2002). Mixed model assembly line design in
a make-to-order environment. Computers & Industrial Engineering, 41(4), 405-421.
Bukchin, Y., & Rabinowitch, I. (2006). A branch-and-bound based solution approach for
the mixed-model assembly line-balancing problem for minimizing stations and task
duplication costs. European Journal of Operational Research, 174(1), 492-508.
Cano-Belmán, J., Ríos-Mercado, R. Z., & Bautista, J. (2010). A scatter search based
hyper-heuristic for sequencing a mixed-model assembly line. Journal of
Heuristics, 16(6), 749-770.
Cortez, P. M., & Costa, A. M. (2015). Sequencing mixed-model assembly lines operating
with a heterogeneous workforce. International Journal of Production Research, 53(11),
3419-3432.

127
Chutima, P., & Chimklai, P. (2012). Multi-objective two-sided mixed-model assembly
line balancing using particle swarm optimisation with negative knowledge. Computers &
Industrial Engineering, 62(1), 39-55.
Ege, Y., Azizoglu, M., & Ozdemirel, N. E. (2009). Assembly line balancing with station
paralleling. Computers & Industrial Engineering, 57(4), 1218-1225.
Erel, E., & Gokcen, H. (1999). Shortest-route formulation of mixed-model assembly line
balancing problem. European Journal of Operational Research, 116(1), 194-204.
Falkenauer, E. (2005, July). Line balancing in the real world. In Proceedings of the
International Conference on Product Lifecycle Management PLM (Vol. 5, pp. 360-370).
Falkenauer, E. (1998) Genetic Algorithms and Grouping Problems. New York: John
Wiley & Sons.
Feo, T. A., & Resende, M. G. (1989). A probabilistic heuristic for a computationally
difficult set covering problem. Operations research letters, 8(2), 67-71.
Feo, T. A., & Resende, M. G. (1995). Greedy randomized adaptive search procedures.
Journal of global optimization, 6(2), 109-133.
Gökcen, H., & Erel, E. (1997). A goal programming approach to mixed-model assembly
line balancing problem. International Journal of Production Economics, 48(2), 177-185.
Gökcen, H., & Erel, E. (1998). Binary integer formulation for mixed-model assembly line
balancing problem. Computers & Industrial Engineering, 34(2), 451-461.
Golle, U., Rothlauf, F., & Boysen, N. (2014). Car sequencing versus mixed-model
sequencing: A computational study. European Journal of Operational Research, 237(1),
50-61.
Gujjula, R., Werk, S., & Günther, H. O. (2011). A heuristic based on Vogel's
approximation method for sequencing mixed-model assembly lines. International
Journal of Production Research, 49(21), 6451-6468.
Haq, A. N., Rengarajan, K., & Jayaprakash, J. (2006). A hybrid genetic algorithm
approach to mixed-model assembly line balancing. The International Journal of
Advanced Manufacturing Technology, 28(3-4), 337-341.
Hop, N. V. (2006). A heuristic solution for fuzzy mixed-model line balancing problem.
European Journal of Operational Research, 168(3), 798-810.
Jin, M., Luo, Y., & Eksioglu, S. D. (2008). Integration of production sequencing and
outbound logistics in the automotive industry. International Journal of Production
Economics, 113(2), 766-774.

128
Kara, Y., Özgüven, C., Seçme, N. Y., & Chang, C. T. (2011). Multi-objective approaches
to balance mixed-model assembly lines for model mixes having precedence conflicts and
duplicable common tasks. The International Journal of Advanced Manufacturing
Technology, 52(5-8), 725-737.
Mahdavi, I., Javadi, B., Sahebjamnia, N., & Mahdavi-Amiri, N. (2009). A two-phase
linear programming methodology for fuzzy multi-objective mixed-model assembly line
problem. The International Journal of Advanced Manufacturing Technology, 44(9-10),
1010-1023.
Manavizadeh, N., Hosseini, N. S., Rabbani, M., & Jolai, F. (2013). A Simulated
Annealing algorithm for a mixed model assembly U-line balancing type-I problem
considering human efficiency and Just-In-Time approach. Computers & Industrial
Engineering, 64(2), 669-685.
Matanachai, S., & Yano, C. A. (2001). Balancing mixed-model assembly lines to reduce
work overload. IIE Transactions, 33(1), 29-42.
McMullen, P. R. (1998). JIT sequencing for mixed-model assembly lines with setups
using tabu search. Production Planning & Control, 9(5), 504-510.
McMullen, P. R., & Frazier, G. V. (1997). A heuristic for solving mixed-model line
balancing problems with stochastic task durations and parallel stations. International
Journal of Production Economics, 51(3), 177-190.
McMullen, P. R., & Tarasewich, P. (2003). Using ant techniques to solve the assembly
line balancing problem. IIE transactions, 35(7), 605-617.
McMullen, P. R., & Tarasewich, P. (2006). Multi-objective assembly line balancing via a
modified ant colony optimization technique. International Journal of Production
Research, 44(1), 27-42.
Mosadegh, H., Fatemi Ghomi, S. M. T., & Süer, G. A. (2016). Heuristic approaches for
mixed-model sequencing problem with stochastic processing times. International Journal
of Production Research, 1-24.
Otto, A., & Scholl, A. (2011). Incorporating ergonomic risks into assembly line
balancing. European Journal of Operational Research, 212(2), 277-286.
Özcan, U., Toklu, B. (2009). Balancing of mixed-model two-sided assembly lines.
Computers & Industrial Engineering, 57(1), 217-227.
Öztürk, C., Tunalı, S., Hnich, B., Örnek, A. (2013). Balancing and scheduling of flexible
mixed model assembly lines with parallel stations. The International Journal of
Advanced Manufacturing Technology, 67(9-12), 2577-2591.

129
Öztürk, C., Tunalı, S., Hnich, B., & Örnek, M. A. (2013b). Balancing and scheduling of
flexible mixed model assembly lines. Constraints, 18(3), 434-469.
Pastor, R., Ferrer, L., & García, A. (2007). Evaluating optimization models to solve
SALBP. In International Conference on Computational Science and Its Applications (pp.
791-803). Springer Berlin Heidelberg.
Pearce, B. (2015). A study on general assembly line balancing modeling methods and
techniques. (Doctoral dissertation). Retrieved from Clemson University’s Dissertation
database (tigerprints No. 1549).
Pil, F.K., & Holweg, M. (2004). Linking product variety to order-fulfilment strategies.
Interfaces, 34, 394–403.
Sarker, B. R., & Pan, H. (1998). Designing a mixed-model assembly line to minimize the
costs of idle and utility times. Computers & industrial engineering, 34(3), 609-628.
Sawik, T. (2002). Monolithic vs. hierarchical balancing and scheduling of a flexible
assembly line. European Journal of Operational Research, 143(1), 115-124.
Scholl, A., Klein, R., & Domschke, W. (1998). Pattern based vocabulary building for
effectively sequencing mixed-model assembly lines. Journal of Heuristics, 4(4), 359-381.
Scholl, A. (1999). Balancing and sequencing assembly lines (2nd ed.). Physica,
Heidelberg.
Scholl, A., & Becker, C. (2006). State-of-the-art exact and heuristic solution procedures
for simple assembly line balancing. European Journal of Operational Research, 168(3),
666-693.
Sewell, E. C., & Jacobson, S. H. (2012). A branch, bound, and remember algorithm for
the simple assembly line balancing problem. INFORMS Journal on Computing, 24(3),
433-442.
Simaria, A. S., & Vilarinho, P. M. (2004). A genetic algorithm based approach to the
mixed-model assembly line balancing problem of type II. Computers & Industrial
Engineering, 47(4), 391-407.
Simaria, A. S., & Vilarinho, P. M. (2009). 2-ANTBAL: An ant colony optimisation
algorithm for balancing two-sided assembly lines. Computers & Industrial Engineering,
56(2), 489-506.
Sivasankara, P., & Shahabudeen, P. M. (2013). Genetic Algorithm for Concurrent
Balancing of Mixed-Model Assembly Lines with Original Task Times of Models.
Intelligent Information Management, 5(3), 84-92.

130
Sparling, D., & Miltenburg, J. (1998). The mixed-model U-line balancing problem.
International Journal of Production Research, 36(2), 485–501.
Tamura, T., Okumura, T., Singh Dhakar, T., & Ohno, K. (2011). Optimal production
sequencing problem to minimise line stoppage time in a mixed-model assembly line.
International Journal of Production Research, 49(14), 4299-4315.
Tasan, S. O., & Tunali, S. (2008). A review of the current applications of genetic
algorithms in assembly line balancing. Journal of intelligent manufacturing, 19(1), 49-69.
Tonelli, F., Paolucci, M., Anghinolfi, D., & Taticchi, P. (2013). Production planning of
mixed-model assembly lines: a heuristic mixed integer programming based approach.
Production Planning & Control, 24(1), 110-127.
Tsai, L. H. (1995). Mixed-model sequencing to minimize utility work and the risk of
conveyor stoppage. Management Science, 41(3), 485-495.
Tsujimura, Y., Gen, M., & Kubota, E. (1995). Solving fuzzy assembly-line balancing
problem with genetic algorithms. Computers and Industrial Engineering, 29(1-4), 543–
547.
Van Zante-de Fokkert, J. I., & de Kok, T. G. (1997). The mixed and multi model line
balancing problem: a comparison. European Journal of Operational Research, 100(3),
399-412.
Vilà, M., & Pereira, J. (2014). A branch-and-bound algorithm for assembly line worker
assignment and balancing problems. Computers & Operations Research, 44, 105-114.
Vilarinho, P. M., & Simaria, A. S. (2002). A two-stage heuristic method for balancing
mixed-model assembly lines with parallel workstations. International Journal of
Production Research, 40(6), 1405-1420.
Vilarinho, P. M., & Simaria, A. S. (2006). ANTBAL: an ant colony optimization
algorithm for balancing mixed-model assembly lines with parallel workstations.
International journal of production research, 44(2), 291-303.
Wilhelm, W. E., & Gadidov, R. (2004). A branch-and-cut approach for a generic
multiple-product, assembly-system design problem. INFORMS Journal on Computing,
16(1), 39-55.
Yano, C. A., & Rachamadugu, R. (1991). Sequencing to minimize work overload in
assembly lines with product options. Management Science,37(5), 572-586.

131
 APPENDIX

Mixed Model Line Balancing Integer Program (AMPL)

# SETS
set task;
param nbrstations;
param nbrworkplaces;
set station=1..nbrstations;
set workplace=1..nbrworkplaces;
set worker={station,workplace};
set prec within {task,task};
set ss within {task,task};
set sw within {task,task};
set ns within {task,task};
set nw within {task,task};
set tasklimit within task;

# PARAMETERS
param c >= 0; # Cycle time
param LB >= 0; # Lower Bound
param t {h in task} >= 0; # t[h] is the expected duration of task h
param q {h in task} >= 0; # q[h] is the mounting position of task h
param es {h in task} >= 0; # es[h] is the earliest time to start task h
param lf {h in task} >= 0; # lfn[h] is the latest time to finish task h
param EST {h in task} = ceil(es[h]/c);
param LFN {h in task} = ceil(lf[h]/c);
param ER {h in task} >= 0; # ER[h] is the expected ergonomic risk of
task h
param Eli1 {h in tasklimit} >= 0;
param Eli2 {h in tasklimit} >= 0;
param Eli3 {h in tasklimit} >= 0;

param Erg >= 0 ; # Ergonomic risk target

param ST {i in station}; # Earliest starting time of worker (i,j)

# VARIABLES

var X {(i,j) in worker, h in task} binary;

var Y {(i,j) in worker} binary;
var V {h in task, l in task} binary;
var S {h in task} >= 0 integer;

# OBJECTIVE

minimize number_of_workers:
sum {(i,j) in worker} Y[i,j];

# CONSTRAINTS

132
subject to C_1 {h in task}: sum {(i,j) in worker: i>=EST[h] and i<=LFN[h]}
X[i,j,h] = 1;
subject to C_3 {(i,j) in worker}: sum {h in task:i>=EST[h] and i<=LFN[h]}
X[i,j,h]*t[h] <= c*Y[i,j];
subject to C_4 {(h,l) in prec}: S[h]+t[h]<=S[l];
subject to C_5 {(i,j) in worker}: sum {h in task} X[i,j,h]*ER[h] <= Erg;
subject to C_6 {h in task}: S[h] >= sum {i in station,j in workplace:i>=EST[h]
and i<=LFN[h]} ST[i]*X[i,j,h];
subject to C_7 {h in task}: S[h]+t[h]<= sum {i in station,j in
workplace:i>=EST[h] and i<=LFN[h]} (ST[i]+c)*X[i,j,h];
subject to C_8 {h in task,l in task, (i,j) in worker: h<>l and
i>=max(EST[h],EST[l]) and i<=min(LFN[h],LFN[l])}: V[h,l]+V[l,h] >=
X[i,j,h]+X[i,j,l]-1;
subject to C_9 {i in station,h in task,l in task: (q[h]<>0 && q[l]<>0) and
q[h]==q[l] and h<>l and l>h and i>=max(EST[h],EST[l]) and
i<=min(LFN[h],LFN[l])}: V[h,l]+V[l,h] >= sum {j in workplace}
(X[i,j,h]+X[i,j,l])-1;
subject to C_10 {h in task,l in task: h<>l}: S[h]+t[h] <= S[l]+(1-
V[h,l])*(lf[h]-es[l]);
subject to C_11 {i in station,(h,l) in ss: i>=max(EST[h],EST[l]) and
i<=min(LFN[h],LFN[l])}: sum {j in workplace} X[i,j,h]= sum {j in
workplace}X[i,j,l];
subject to C_12 {(i,j) in worker,(h,l) in sw: i>=max(EST[h],EST[l]) and
i<=min(LFN[h],LFN[l])}: X[i,j,h]=X[i,j,l];
subject to C_13 {i in station,(h,l) in ns: i>=max(EST[h],EST[l]) and
i<=min(LFN[h],LFN[l])}: sum {j in workplace} (X[i,j,h]+X[i,j,l]) <= 1;
subject to C_14 {(i,j) in worker,(h,l) in nw: i>=max(EST[h],EST[l]) and
i<=min(LFN[h],LFN[l])}: X[i,j,h]+X[i,j,l] <= 1;
subject to C_15 {(i,j) in worker:j<=card(workplace)-1}: Y[i,j]>=Y[i,j+1];
subject to C_16 {(i,j) in worker, h in tasklimit:i<>Eli1[h] and i<>Eli2[h] and
i<>Eli3[h]}: X[i,j,h] = 0;

133
Example data file:
set task := 1 2 3 4 5 8 9 10 15 16 17 18 19 21 23 27 29 30 31 32 33 34 35 38
41 42 44 48 55 56 58 59 60 62 64 65 67 68 69 71 73 74 75 78 79 80 83 85 86 87;
set prec := (1,3) (1,4) (1,5) (1,8) (1,9) (9,10) (15,16) (15,17) (15,18)
(15,19) (15,21) (21,23) (15,27) (15,29) (1,30) (33,31) (38,32) (1,33) (15,34)
(15,35) (15,38) (1,41) (1,42) (30,44) (29,48) (3,55) (62,58) (62,59) (60,62)
(17,64) (71,67) (71,68) (69,71) (4,73) (30,74) (1,75) (83,78) (83,79) (80,83)
(18,85) (29,86) (15,87) (16,27) (59,58) (68,67) (56,73) (79,78) (65,85);
set ss :=;
set sw := (58,59) (59,60) (60,62) (67,68) (68,69) (69,71) (78,79) (79,80)
(80,83);
set ns :=;
set nw :=;
set tasklimit := 8 9 19 21 55 64 73 85;

param nbrstations :=5;

param nbrworkplaces:=5;
param c := 103000;
param Erg := 500;
param: ST :=
1 0
2 103000
3 206000
4 309000
5 412000
;

param: t ER q es ls:=
1 60 20 1 1 5
2 3300 17.90909091 1 1 5
3 13500 19.8 1 1 5
4 11340 19.57142857 3 1 5
5 4020 23.59701493 1 1 5
8 24840 22.60869565 3 1 5
9 18900 35.94285714 3 1 5
10 7020 44 1 1 5
15 60 20 9 1 5
16 3120 26.94230769 7 1 5
17 13500 19.8 7 1 5
18 11340 19.57142857 9 1 5
19 23940 22.09022556 9 1 5
21 22680 35.34920635 9 1 5
23 10620 44 7 1 5
27 3300 17.90909091 9 1 5
29 9828 18.82 9 1 5
30 10008 19.03934426 3 1 5
31 6216 19.40135135 1 1 5
32 6216 19.40135135 7 1 5
33 4536 10.15 1 1 5
34 5724 6.622641509 7 1 5
35 11592 16.38913043 7 1 5

134
38 5166 9.851219512 7 1 5
41 6264 6.822413793 1 1 5
42 12852 15.60588235 1 1 5
44 10122 16.18423237 3 1 5
48 10626 14.81897233 9 1 5
55 70320 24.14846416 1 1 5
56 1440 24 1 1 5
58 3960 24 0 1 5
59 4320 46 0 1 5
60 5220 12 0 1 5
62 11520 49 0 1 5
64 70320 24.12798635 7 1 5
65 1440 24 7 1 5
67 3960 24 0 1 5
68 4320 46 0 1 5
69 5220 12 0 1 5
71 11520 49 0 1 5
73 67260 26.64942016 3 1 5
74 4200 15.64285714 3 1 5
75 13500 19.04 3 1 5
78 3960 27 0 1 5
79 4320 23 0 1 5
80 5280 12 0 1 5
83 11880 90 0 1 5
85 67260 26.87421945 9 1 5
86 4200 14.91428571 9 1 5
87 13500 19.04 9 1 5
;

param: Eli1 Eli2 Eli3 :=

8 1 0 0
9 1 0 0
19 1 0 0
21 1 0 0
55 3 0 0
64 3 0 0
73 4 0 0
85 4 0 0
;

135
Mixed Model Line Balancing Constraint Program (CPLEX Optimization Studio)
using CP;

// Number of WorkStations
int nbWorkstations= ...;
range Workstations= 1..nbWorkstations;

// Tasks names
{int} TaskN=...;

// Tasks assignment limitation

{int} TaskLimit=...;

// Number of Stations
int nbStations=...;
range Stations= 1..nbStations;

// Cycle time
int Cycle= ...;

// Station start and end times for each Worker

int SStart[Stations]=...;
int SEnd[Stations]=...;

// Earliest time and latest finish for each task

int est[TaskN]=...;
int lfn[TaskN]=...;

// Duration, mounting position, and expected ergonomic rating for each task
int duration[TaskN]= ...;
int mount[TaskN]= ...;
float ER[TaskN]= ...;

// Target total expected ergonomic rating for each worker

int Erg= ...;

// Task assignment eligibility list

int Eli1[TaskLimit]=...;
int Eli2[TaskLimit]=...;
int Eli3[TaskLimit]=...;

// Precedence relation
tuple Precedence {
int pre;
int post;
};
{Precedence} precedences = ...;

// Assignment constraints
tuple AssignConst {

136
 int task1;
int task2;
}

{AssignConst} ss = ...;
{AssignConst} sw = ...;
{AssignConst} ns = ...;
{AssignConst} nw = ...;

// Feasible time window

{int} bounds[i in TaskN][j in Stations] = asSet(est[i]..lfn[i]) inter
asSet(SStart[j]..SEnd[j]);

// Decision variables
dvar int usage in 0..nbStations*nbWorkstations;
dvar interval Task[i in TaskN] in est[i]..lfn[i] size duration[i];
dvar interval TaskStaWork[i in TaskN][j in Stations][k in Workstations]
optional in (card(bounds[i][j])==0 ? 0 :
first(bounds[i][j]))..(card(bounds[i][j])==0 ? 0 : last(bounds[i][j]));;
dvar sequence Seq[j in Stations][k in Workstations] in all(i in TaskN)
TaskStaWork[i][j][k];

// Objective
minimize usage;

// Constraints
subject to
{
forall(i in TaskN) {
alternative(Task[i], all(j in Stations,k in Workstations)
TaskStaWork[i][j][k]);
}
// No overlap on Workers
forall(j in Stations)
forall(k in Workstations)
noOverlap(Seq[j][k]);

// Same mounting position

forall(m in 1..9)
noOverlap( all(i in TaskN: mount[i] == m) Task[i] );

// Precednece relation
forall(i in precedences )
endBeforeStart(Task[i.pre], Task[i.post]);

// Same station constraint

forall(t in ss)
forall (j in Stations)
sum(k in Workstations) presenceOf(TaskStaWork[t.task1][j][k]) ==
sum(k in Workstations) presenceOf(TaskStaWork[t.task2][j][k]) ;
// Not same station constraint
forall(t in ns)

137
 forall (j in Stations)
sum(k in Workstations) (presenceOf(TaskStaWork[t.task1][j][k])+
presenceOf(TaskStaWork[t.task2][j][k]))<=1;
// Same worker constraint
forall(t in sw)
forall(j in Stations)
forall(k in Workstations)
presenceOf(TaskStaWork[t.task1][j][k]) ==
presenceOf(TaskStaWork[t.task2][j][k]);
// Not same worker constraint
forall(t in nw)
forall(j in Stations)
forall(k in Workstations)

presenceOf(TaskStaWork[t.task1][j][k])+presenceOf(TaskStaWork[t.task2][j
][k])<=1;

// Ergonomics
forall(j in Stations)
forall(k in Workstations)
sum(i in TaskN) presenceOf(TaskStaWork[i][j][k])*ER[i]<= Erg;

// Eligibility constraint
forall(i in TaskLimit)
forall(j in Stations: j!=Eli1[i] && j!=Eli2[i] && j!=Eli3[i])
forall(k in Workstations)
presenceOf(TaskStaWork[i][j][k])==0;

// Breaking symmetry
forall(j in Stations, k in 1..nbWorkstations-1)
max(i in TaskN) presenceOf(TaskStaWork[i][j][k+1]) <= max(i in TaskN)
presenceOf(TaskStaWork[i][j][k]);

// The number of Workers used

usage == sum(j in Stations,k in Workstations) max(i in TaskN)
presenceOf(TaskStaWork[i][j][k]);
}

138
Example data file:
nbWorkstations=5 ;
nbStations=5;
Cycle=103000;
Erg=500;
TaskN=
{1,2,3,4,5,8,9,10,15,16,17,18,19,21,23,27,29,30,31,32,33,34,35,38,41,42,44,48,
55,56,58,59,60,62,64,65,67,68,69,71,73,74,75,78,79,80,83,85,86,87};
TaskLimit={8,9,19,21,55,64,73,85};
SStart=[0,103000,206000,309000,412000];
SEnd=[103000,206000,309000,412000,515000];
duration=[60,3300,13500,11340,4020,24840,18900,7020,60,3120,13500,11340,23940,
22680,10620,3300,9828,10008,6216,6216,4536,5724,11592,5166,6264,12852,10122,10
626,70320,1440,3960,4320,5220,11520,70320,1440,3960,4320,5220,11520,67260,4200
,13500,3960,4320,5280,11880,67260,4200,13500];
est=[0,60,60,60,60,60,60,18960,0,60,60,60,60,60,22740,3180,60,60,4596,5226,60,
60,60,60,60,60,10068,9888,13560,0,21060,16740,0,5220,13560,0,21060,16740,0,522
0,11400,10068,60,21480,17160,0,5280,11400,9888,60];
lfn=[431180,515000,444680,447740,515000,515000,507980,515000,431180,511700,444
680,447740,515000,504380,515000,515000,504374,504878,515000,515000,508784,5150
00,515000,508784,515000,515000,515000,515000,515000,447740,515000,511040,49520
0,506720,515000,447740,515000,511040,495200,506720,515000,515000,515000,515000
,511040,494840,506720,515000,515000,515000];
mount=[1,1,1,3,1,3,3,1,9,7,7,9,9,9,7,9,9,3,1,7,1,7,7,7,1,1,3,9,1,1,0,0,0,0,7,7
,0,0,0,0,3,3,3,0,0,0,0,9,9,9];
precedences={<1,3>,<1,4>,<1,5>,<1,8>,<1,9>,<9,10>,<15,16>,<15,17>,<15,18>,<15,
19>,<15,21>,<21,23>,<15,27>,<15,29>,<1,30>,<33,31>,<38,32>,<1,33>,<15,34>,<15,
35>,<15,38>,<1,41>,<1,42>,<30,44>,<29,48>,<3,55>,<62,58>,<62,59>,<60,62>,<17,6
4>,<71,67>,<71,68>,<69,71>,<4,73>,<30,74>,<1,75>,<83,78>,<83,79>,<80,83>,<18,8
5>,<29,86>,<15,87>,<16,27>,<59,58>,<68,67>,<56,73>,<79,78>,<65,85>};
ER=[20,17.91,19.8,19.57,23.6,22.61,35.94,44,20,26.94,19.8,19.57,22.09,35.35,44
,17.91,18.82,19.04,19.4,19.4,10.15,6.62,16.39,9.85,6.82,15.61,16.18,14.82,24.1
5,24,24,46,12,49,24.13,24,24,46,12,49,26.65,15.64,19.04,27,23,12,90,26.87,14.9
1,19.04];
ss={};
ns={};
sw={<58,59>,<59,60>,<60,62>,<67,68>,<68,69>,<69,71>,<78,79>,<79,80>,<80,83>};
nw={};
Eli1=[1,1,1,1,3,3,4,4];
Eli2=[0,0,0,0,0,0,0,0];
Eli3=[0,0,0,0,0,0,0,0];

139
Mixed Model Sequencing Integer Program (AMPL)
# SETS
param nbrstations;
param nbrcycles;
param nbrmodels;
set station=1..nbrstations;
set cycle=1..nbrcycles;
set models=1..nbrmodels;

# PARAMETERS
param c >= 0; # Cycle time
param l {k in station} >= 0; # Station length
param d {m in models} >=0; # Demand for each model
param q; # Min station start
param p {m in models, k in station} >= 0; # Processing time for car m at
station k
param e=0.1;

# VARIABLES
var X {m in models, t in cycle} binary;
var Y {k in station, t in cycle} binary;
var S {k in station, t in cycle};

# OBJECTIVE
minimize utility:
sum {k in station, t in cycle} Y[k,t];

# CONSTRAINTS
subject to C_1 {t in cycle}: sum {m in models} X[m,t]=1;
subject to C_2 {m in models}: sum {t in cycle} X[m,t]=d[m];
subject to C_3 {k in station, t in cycle}: S[k,t]+ sum {m in models}
p[m,k]*X[m,t]- l[k]-e>=-l[k]*(1-Y[k,t]);
subject to C_4 {k in station, t in cycle}: S[k,t]+ sum {m in models}
p[m,k]*X[m,t]-l[k]<=l[k]*Y[k,t];
subject to C_5 {k in station, t in cycle:t>=2}: S[k,t] >= S[k,t-1]+ sum {m in
models} p[m,k]*X[m,t-1]-c-(l[k]*Y[k,t-1]);
subject to C_6 {k in station, t in cycle:t>=2}: S[k,t] >= S[k,t-1]-c-((1-
Y[k,t-1])*l[k]);
subject to C_7 {k in station, t in cycle:k>=2 && t>=2}: S[k,t] >= S[k-1,t-1]+
sum {m in models} p[m,k-1]*X[m,t-1]-l[k-1]-(l[k-1]-c)*Y[k-1,t-1];
subject to C_8 {k in station, t in cycle}: S[k,t] >= q;
subject to C_9 {k in station, t in cycle}: S[k,t] <= l[k]-c;
subject to C_10 {k in station}: S[k,1] = 0;
subject to C_11 {t in cycle}: S[1,t] = 0;

140
Example data file:
param nbrstations := 21;
param nbrcycles :=270;
param nbrmodels :=9;
param c := 175;
param q :=0;
param : l :=
1 195
2 195
3 195
4 195
5 195
6 195
7 195
8 195
9 195
10 195
11 195
12 195
13 195
14 195
15 195
16 195
17 195
18 195
19 195
20 195
21 195
;
param d:=
1 10
2 10
3 10
4 105
5 105
6 8
7 8
8 7
9 7
;
param p: 1 2 3 4 5 6 7 8 9 10 11
12 13 14 15 16 17 18 19 20 21 :=
1 104 103 165 166 111 126 97 100 179 178 161
96 99 147 163 163 173 176 162 164 177
2 100 103 156 175 114 121 96 97 174 172 152
106 101 155 152 185 179 167 150 161 161
3 97 105 164 172 114 122 96 95 173 172 168
105 102 142 156 183 178 181 152 157 154
4 92 107 161 167 115 124 93 106 178 177 167
97 101 154 152 178 169 180 152 159 168
5 100 101 148 168 117 127 96 94 178 178 167
101 99 146 153 169 173 172 160 162 172

141
6 94 108 156 167 117 130 89 102 171 177 166
100 101 143 152 173 178 173 151 160 170
7 103 106 154 168 115 120 94 103 177 175 172
96 96 154 154 172 174 173 155 162 167
8 109 102 164 156 111 121 101 102 171 173 157
104 102 153 156 182 175 168 148 158 149
9 101 110 155 173 111 134 92 100 174 175 177
96 99 155 156 171 175 184 167 157 169

142
Mixed Model Sequencing Heuristic Functions (MATLAB)
% Calculate the overload situations for a given sequence
function [a,s,f,o] = overload(nPeriods,nStations,c,l,q,x,dur)
%c = cycle time
%l = stations length (vector)
%q = station start (vector)
%x = sequence (vector)
%dur = duration
%a = total number of overload situations
%s = start times
%f = finish times
%o = overload situations
s=zeros(nPeriods,nStations);
f=zeros(nPeriods,nStations);
o=zeros(nPeriods,nStations);

%period 1
for station= 1:nStations
f(1,station)=s(1,station)+dur(x(1),station);
end

%station 1
for period= 2:nPeriods
if f(period-1,1)<=l(1) && o(period-1,1)==0
s(period,1)=max([0,f(period-1,1)-c]);
else
s(period,1)=max([0,s(period-1,1)-c]);
end
f(period,1)=s(period,1)+dur(x(period),1);
if f(period,1)>l(1)
o(period,1)=1;
s(period,1)=0;
f(period,1)=s(period,1)+dur(x(period),1);
end

end

%complete
for station= 2:nStations
for period= 2:nPeriods
if f(period-1,station)<=l(station) && o(period-1,station)==0
s(period,station)=max([q(station),f(period-1,station)-
c,f(period-1,station-1)-l(station-1)]);
else
s(period,station)=max([q(station),s(period-1,station)-
c,f(period-1,station-1)-l(station-1)]);
end
f(period,station)=s(period,station)+dur(x(period),station);

if f(period,station)>l(station)
o(period,station)=1;

143
 s(period,station)=0;
f(period,station)=s(period,station)+dur(x(period),station);
end
end
end
a=sum(sum(o));

***********************************************************
*****
function [a,cost] = P_overload(nStations,c,l,q,x,dur)
%c = cycle time
%l = stations length (vector)
%q = station start (vector)
%x = sequence (vector)
%dur = duration
%a = total number of overload situations
%s = start times
%f = finish times
%o = overload situations
nPeriods=length(x);
s=zeros(nPeriods,nStations);
f=zeros(nPeriods,nStations);
o=zeros(nPeriods,nStations);

%period 1
for station= 1:nStations
f(1,station)=s(1,station)+dur(x(1),station);
end

%station 1
for period= 2:nPeriods
if f(period-1,1)<=l(1) && o(period-1,1)==0
s(period,1)=max([0,f(period-1,1)-c]);
else
s(period,1)=max([0,s(period-1,1)-c]);
end
f(period,1)=s(period,1)+dur(x(period),1);
if f(period,1)>l(1)
o(period,1)=1;
s(period,1)=0;
f(period,1)=s(period,1)+dur(x(period),1);
end

end

%complete
for station= 2:nStations
for period= 2:nPeriods
if f(period-1,station)<=l(station) && o(period-1,station)==0
s(period,station)=max([q(station),f(period-1,station)-
c,f(period-1,station-1)-l(station-1)]);

144
 else
s(period,station)=max([q(station),s(period-1,station)-
c,f(period-1,station-1)-l(station-1)]);
end
f(period,station)=s(period,station)+dur(x(period),station);

if f(period,station)>l(station)
o(period,station)=1;
s(period,station)=0;
f(period,station)=s(period,station)+dur(x(period),station);
end
end
end
a=sum(sum(o));
cost=sum(max(f(nPeriods,:)-c,0));

***********************************************************
*****
% Greedy random creator
function [x]=GRC(nStations,c,l,q,dur,demand)
count=0;
avail=1:length(demand);
temp=sum(dur,2);
x=datasample(avail,1);
demand(x)=demand(x)-1;
while ~isempty(avail)
zcost=0;
over=0;
tmp=0;
for i=1:length(avail)
[over(i,1)]= P_overload(nStations,c,l,q,[x,avail(i)],dur);
end
over=[avail',over];
zcost=find(over(:,2)==min(over(:,2)));
xn=datasample(zcost,1);
x=[x,over(xn)];
demand(over(xn))=demand(over(xn))-1;
avail=find(demand~=0);
count=count+1
end

***********************************************************
*****
% Randomly swap n elements in the vector x.
function x=randswap(x,n)
l=length(x);
y = datasample(1:l,2*n,'Replace',false);
y=y(randperm(length(y)));
x(y)=x([y(n+1:2*n),y(1:n)]);
end

145
***********************************************************
*****
% GRASP search
function [best,Timetobest]=GRASP(nPeriods,nStations,c,l,q,dur,demand)
iter=100000;
best=0;
Timetobest=0;
xn=0;
fxn=0;
initime=cputime;
x=GRC(nStations,c,l,q,dur,demand);
f=overload(nPeriods,nStations,c,l,q,x,dur);
best=f;
Timetobest=cputime-initime;
for i=1:iter
xn=randswap(x,1);
fxn=overload(nPeriods,nStations,c,l,q,xn,dur);
if fxn<=f
f=fxn;
x=xn;
if f<best
best=f;
Timetobest=cputime-initime;
end
end
fprintf('iter= %d, the current sol= %d , the best solution so far
is %d.\n',i,f,best);

***********************************************************
*****

146