PEOPLE’S DEMOCRATIC REPUBLIC OF ALGERIA

‫الجـمـهـوريـة الجـزائـريـة الديـمـقـراطـيــة الشـعـبـيــة‬

MINISTRY OF HIGHER EDUCATION ‫وزارة التعليـم العالـي والبحـث العلمـي‬

AND SCIENTIFIC RESEARCH

HIGHER SCHOOL IN APPLIED SCIENCES

‫المـدرسـة العليـا في العلوم التطبيقيــة‬
--T L E M C E N-- -‫تلمسان‬-

Final degree project

For obtaining the Engineering Diploma

Field: Industrial engineering

Major: Industrial Management and Logistics

Presented by :
GAOUAR Rim Djazia
TERKI HASSAINE Mohieddine El-Hadi

Subject

Formulation and mathematical

programming heuristic for Flexible
Layout Design Problem

Publicly defended, on 07/07/2024, in front of the jury composed of :

Mrs. SEKKAL Nourelhouda MCB ESSA. Tlemcen Président

Mr. MALIKI Fouad MCA ESSA. Tlemcen Thesis Director

Mr. BRAHIMI Nadjib Prof Rennes Business Thesis Co-Directeur

School

Mr. SOUIER Mehdi Prof Univ. Tlemcen Examiner 1

Mr. HADDOU BENDERBAL Associate Univ. Aix- Examiner 2

Hichem Professor Marseille

Academic Year : 2023 / 2024

Gratitude

First and foremost, we would like to thank ”Allah,” the Almighty and Merciful, who
has given us the strength, will, and courage to complete this work.

We wish to express our sincere gratitude and deep appreciation to:

Our supervisor and academic leader, Mr. Fouad MALIKI, for nurturing us over
these pleasantly spent three years. We are profoundly grateful for his guidance and
for enriching our knowledge in this field.

Our co-supervisor, Mr. Nadjib BRAHIMI, for affording us this opportunity

and for enhancing our understanding of the subject matter. We value immensely
the invaluable advice and direction he provided throughout the execution of this
modest work.

To the members of our jury Mrs. Nourelhouda SEKKAL, Mr.Mehdi SOUIER,

Mr.Hichem HADDOU BENDERBAL for the esteemed honor they bestow
upon us by agreeing to evaluate this modest effort.

We would also like to thank our dear professor who has been with us through these
past three years, Mr. Mohammed BENNEKROUF, for giving us the oppor-
tunity to explore plentiful information and for his invaluable guidance and support
throughout our journey.

To all our teachers who have contributed significantly to the relevance of our
education.

Finally, we extend our thanks to all who have played a part, whether directly or
indirectly, in the realization of this thesis. Your interest in our work is greatly
appreciated.

i
Acknowledgement

I am deeply grateful to the people who have supported me during my academic

studies.

First and foremost, I am very thankful to my parents, Hayet HADJ SLIMANE

and Kamel GAOUAR, for their unwavering love and support. They have provided
me with everything I needed to become the person I am today, and I am forever
grateful for their sacrifices and encouragement.To my brothers Alaa, Abdessamad.

I also appreciate my extended family, including my aunts, uncles, and cousins, es-
pecially Yasmine, Amira, Rihem, and Ibtissem. They have been a constant source
of support and inspiration for me since we were young.

To my second family Yasmine, Linda, Hadjer, Ghizlene ,Feriel and the family with
whom with whom I studied in the same class last 3 years, we have shared countless
memorable moments and created a bond that I will cherish forever. To all my friends
Nassim ,Nadir, Azeddine, Wadie, Salah, Jami,Youcef.

I would also like to thank my project partner, Mohieddine El-Hadi Terki Hassaine,
for his hard work. Without him, I would not have been able to make it this far in
our work.

Finally, I dedicate this work to the younger version of myself, who once thought
this day would be so far away. I want to let her know that she has made it, and in
the most wonderful circumstances.

ii
Acknowledgement

I am deeply grateful to the people who have supported me during my academic

studies.

First and foremost, I am very thankful to my parents, for their unwavering love
and support. They have provided me with everything I needed to become the per-
son I am today, and I am forever grateful for their sacrifices and encouragement.To
my two sisters, Saliha and Awicha and to my brother Younes. Your presence in my
life has been a source of joy, support, and inspiration. Your love and encouragement
have pushed me to give my best.

I also appreciate my extended family, including my aunts, uncles, and cousins. They
have been a constant source of support and inspiration for me since we were young.

I would like to extend my deepest gratitude to Nassim, Azedine, Nadir, Salah,

Wadie, Youcef and Zinedine, who have been like brothers to me. Your presence
has been a source of strength and inspiration, and I am truly grateful for all the
moments we’ve shared. To all my friends Amina, Hadjer, Yasmine, Linda, Ghizlene,
Feriel and the family with whom I studied in the same class last 3 years, we have
shared countless memorable moments and created a bond that I will cherish forever.

To two wonderful individuals, who have always been there for me, even in the
toughest times. Your genuine friendship and unwavering support have helped me
overcome obstacles and persevere. Thank you for your precious friendship

I would also like to thank my project partner, Rim Djazia GAOUAR, for her per-
severance and hard work. Without her, I would not have been able to make it this
far in our work.

To my friends, my loved ones, and my 2024 Industrial Engineering class.

iii
Contents

Gratitude i

Acknowledgement ii

Acknowledgement iii

List of Abbreviations vii

List of Figures viii

List of Tables ix

General Introduction 1

1 General Overview of Manufacturing Systems 3

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Manufacturing Systems . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Craft Manufacturing . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 DMS(Dedicated Manufacturing System) . . . . . . . . . . . . 6
1.2.4 FMS(Flexible Manufacturing System) . . . . . . . . . . . . . . 6
1.2.5 RMS(Reconfigurable Manufacturing System) . . . . . . . . . . 7
1.2.6 MMS (Matrix-structured Manufacturing System) . . . . . . . 11
1.3 Assembly Line Balancing . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 RMS literature review 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 DMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 FMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.3 RMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.4 MMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.5 ALBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.6 Our reference articles . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

iv
3 Considered problem and proposed methods 26
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 Model Notation . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 Model Formulation . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.3 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 2-Phased Resolution Method . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 Set-Covering Notation . . . . . . . . . . . . . . . . . . . . . . 32
3.3.3 Set-Covering model . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Results and discussions 35

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Results(1200 seconds) . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.1 A description of the result obtained . . . . . . . . . . . . . . . 36
4.2.2 Bowman C=19, 8 tasks . . . . . . . . . . . . . . . . . . . . . . 38
4.2.3 EX C=18, 5 tasks . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.4 EX C=20, 5 tasks . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2.5 Jackson C=08, 11 tasks . . . . . . . . . . . . . . . . . . . . . 41
4.2.6 Jaeschke C=07, 9 tasks . . . . . . . . . . . . . . . . . . . . . . 42
4.2.7 Mansoor C=47, 11 tasks . . . . . . . . . . . . . . . . . . . . . 43
4.2.8 Mertens C=08, 7 tasks . . . . . . . . . . . . . . . . . . . . . . 44
4.2.9 Mitchell C=14, 21 tasks . . . . . . . . . . . . . . . . . . . . . 45
4.2.10 Remaining instance (1200s) . . . . . . . . . . . . . . . . . . . 46
4.3 Results(300 seconds) . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3.1 Bowman C=19, 8 tasks . . . . . . . . . . . . . . . . . . . . . . 47
4.3.2 EX C=18, 5 tasks . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.3 EX C=20, 5 tasks . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.4 Jackson C=08, 11 tasks . . . . . . . . . . . . . . . . . . . . . 50
4.3.5 Jaeschke C=07, 9 tasks . . . . . . . . . . . . . . . . . . . . . . 51
4.3.6 Mansoor C=47, 11 tasks . . . . . . . . . . . . . . . . . . . . . 52
4.3.7 Mertens C=08, 7 tasks . . . . . . . . . . . . . . . . . . . . . . 53
4.3.8 Mitchell C=14, 21 tasks . . . . . . . . . . . . . . . . . . . . . 54
4.3.9 Remaining instances (300s) . . . . . . . . . . . . . . . . . . . 55
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

General Conclusion and Perspectives 57

Bibliography 58

Webography 62

Appendix 63
4.4.1 Bowman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.4.2 EX C=18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.4.3 EX C=20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.4.4 Jackson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

v
4.4.5 Jaeschke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.4.6 Mansoor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4.7 Mertens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.8 Mitchell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

vi
List of Abbreviations

Abbreviation Definition
AGV Automated Guided Vehicules
ALB Assembly-Line Balancing
CEGA Cell Evaluation and Genetic Algorithms
CNC Computer Numerical Control
CS Cuckoo Search
DMS Dedicated Manufacturing System
FJSP Flexible Job-Shop Scheduling Problem
FLDP Flexible Layout Design Problem
FMS Flexible Manufacturing System
GA Genetic Algorithms
GRASP Greedy Randomized Adaptive Search Proce-
dure
LB Lower Bound
MA Memetic Algorithm
MALBP Mixed Assembly Line Balancing Problem
MILP Mixed-Integer Linear Programming
MMS Matrix-Structured Manufacturing System
MS Manufacturing System
NSGA Nondominated Sorting Genetic Algorithm
OEM Automotive-Original Equipment Manufac-
turers
PPS Production Planning and Scheduling
PSO Praticle Swarm Optimization
RMS Reconfigurable Manufacturing System
SA Simulated Annealing
SALB Simple Assembly Line Balancing
SH structure heterogeneity
TTH Task Time Heterogeneity
UB Upper Bound

vii
List of Figures

1.1 Manufacturing System source: sketchbubble . . . . . . . . . . . . . . 5

1.2 Dedicated Manufacturing System . . . . . . . . . . . . . . . . . . . . 6
1.3 Reconfigurable Manufacturing System . . . . . . . . . . . . . . . . . . 8
1.4 Flow line configuration . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Reconfigurable Cellular Manufacturing system . . . . . . . . . . . . . 10
1.6 Rotary Machining System . . . . . . . . . . . . . . . . . . . . . . . . 11
1.7 Comparison of a classic MS and a MMS configuration . . . . . . . . . 12
1.8 Diffrent Assembly Lines . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 An example of a precedence diagram . . . . . . . . . . . . . . . . . . 34

4.1 UB and LB Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2 GAP From LB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Result example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4 The objective variation curves example . . . . . . . . . . . . . . . . . 37

viii
List of Tables

2.1 Production planning & Scheduling . . . . . . . . . . . . . . . . . . . 17

2.2 Layout design summarized articles . . . . . . . . . . . . . . . . . . . . 18
2.3 line balancing/ re-balancing summarized articles . . . . . . . . . . . . 18
2.5 RMS design summarized articles . . . . . . . . . . . . . . . . . . . . . 19

3.1 Index sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Decision variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Model notation for set covering . . . . . . . . . . . . . . . . . . . . . 32

4.1 Bowman GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 38

4.2 EX C=18 GAP from LB for MILP and 2 phased (1200s) . . . . . . . 39
4.3 EX C=20 GAP from LB for MILP and 2 phased (1200s) . . . . . . . 40
4.4 Jackson GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 41
4.5 Jaeschke GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 42
4.6 Mansoor GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 43
4.7 Mertens GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 44
4.8 Mitchell GAP from LB for MILP and 2 phased (1200s) . . . . . . . . 45
4.9 Remaining instances: GAP from LP for MILP and 2-phased (1200s) . 46
4.10 Bowman GAP from LB for MILP and 2 phased (300s) . . . . . . . . 47
4.11 EX C=18 GAP from LB for MILP and 2 phased (300s) . . . . . . . . 48
4.12 EX C=20 GAP from LB for MILP and 2 phased (300s) . . . . . . . . 49
4.13 Jackson GAP from LB for MILP and 2 phased (300s) . . . . . . . . . 50
4.14 Jaeschke GAP from LB for MILP and 2 phased (300s) . . . . . . . . 51
4.15 Mansoor GAP from LB for MILP and 2 phased (300s) . . . . . . . . 52
4.16 Mertens GAP from LB for MILP and 2 phased (300s) . . . . . . . . . 53
4.17 Mitchell GAP from LB for MILP and 2 phased (300s) . . . . . . . . . 54
4.18 Remaining instances : GAP from LB for MILP and 2-phased (300s) . 55
4.19 MILP and 2-Phased Results within 20 minutes : Bowman . . . . . . 63
4.20 MILP and 2-Phased Results within 20 minutes : EX C=18 . . . . . . 64
4.21 MILP and 2-Phased Results within 20 minutes : EX C=20 . . . . . . 65
4.22 MILP and 2-Phased Results within 20 minutes : Jackson . . . . . . . 66
4.23 MILP and 2-Phased Results within 20 minutes : Jaeschke . . . . . . 67
4.24 MILP and 2-Phased Results within 20 minutes : Mansoor . . . . . . 68
4.25 MILP and 2-Phased Results within 20 minutes : Mertens . . . . . . . 69
4.26 MILP and 2-Phased Results within 20 minutes : Mitchell . . . . . . . 70

ix
General Introduction

Reconfigurable Manufacturing System (RMS) has been the center of research since
the 1990s and has been evolving ever since, thanks to its multiple characteristics
that positively influence every production line (products) and meet the demands.
To further access the reconfigurability characteristics of RMS, a new manufactur-
ing system (MS) called the Matrix-structured Manufacturing System (MMS) has
been introduced. The MMS has the flexibility and reconfigurability of RMS manu-
facturing systems, featuring matrix-structured layouts with multiple stations where
flows are often transported by AGVs. Assembly line balancing has also been a focal
point of research since the notion of manufacturing systems and stations existed.
Researchers have been striving to find the perfect balance for every type of manu-
facturing system using all existing methodologies or developing their own solutions,
known as heuristics or hybrid-heuristics.

In this thesis, we will focus on the line balancing of a flexible layout design prob-
lem (FLDP), which involves designing a flexible layout for an assembly segment.
This includes the integrated problems of station formation and station location
while also anticipating the operational AGV flow,particularly on the part where the
MMS layout is introduced to the manufacturing system. Our approach proposes
integrating the principle of assigning resources to each station through this matrix-
structured layout to reflect the reality of manufacturing systems. We will work on
the mathematical model already established by Grunow which takes into consid-
eration a multi objective problem minimzing both the number of opened locations
and the transportation flow between the location secured by the AGV’s, introducing
our own principle of resource allocation. The proposed method involves establishing
a mat-heuristic composed of two phases: the first phase compiles Grunow’s model
without any changes, and the second phase determines the allocation of resources
without any constraints on capacity or utilization.

As part of the professional training for obtaining the state engineer’s diploma
in the Industrial Engineering specialty at the Higher School of Applied
Sciences of Tlemcen, a two-month end-of-study internship was carried out within
the IMT Atlantique Nantes. This internship was an enriching and instructive
experience. It allowed us to gain a clearer understanding of the practical application
of the theoretical knowledge acquired during our training.

For the structure of our thesis, we will focus on the flexible layout design problem.
The work presented will be divided into parts that encompass everything we have
learned or accomplished during our two-month internship and since the beginning
of our journey at the university. This work will be divided into four chapters, each

1
detailing the methodologies, techniques, and information utilized to achieve our
objectives:

In Chapter 1, we will explore the evolution of manufacturing systems, starting

from Craft Manufacturing and progressing to the Dedicated Manufacturing System,
then to the Flexible Manufacturing System, and finally to the Reconfigurable Man-
ufacturing System and the Matrix-structured Manufacturing System. Additionally,
we will define the assembly line balancing problem.

In the next chapter, we will review the existing literature on various manufactur-
ing systems. We will present the reference articles that inspired our work on this
problem.

In the third chapter, we will present our mathematical model, including model
notation, explanations of constraints, and objective functions. We will then intro-
duce the mat-heuristic approach which we will call 2-phased method, used to resolve
the problem, along with its formulation. Finally, we will present the datasets used
to test our approach.

In the last chapter, we will test and thoroughly discuss the results of the Mixed
Integer Linear Programming (MILP) and our 2-phased method. We will conduct
tests on 128 instances with time limits of 20 minutes and 5 minutes.

2
Chapter 1

General Overview of
Manufacturing Systems

1.1 Introduction
The industrial sector has undoubtedly undergone significant transformations over
the years, with each revolution bringing about new technologies and systems that
have revolutionized the way manufacturing processes are carried out. The first in-
dustrial revolution, which took place in the late 18th and early 19th century, saw
the introduction of mechanization, which allowed for the mass production of goods.
This revolution was followed by the second industrial revolution, which was charac-
terized by the introduction of new energy sources, such as electricity and oil, and
the emergence of assembly lines, which made production more efficient.

The third industrial revolution, also known as the digital revolution, was marked
by the rise of computer technology and automation, which enabled manufacturers
to streamline their operations and reduce costs. This revolution paved the way for
the fourth industrial revolution, which is currently underway and is set to transform
the industry in ways that were once thought impossible.

The fourth industrial revolution is unique in that it has introduced the Internet
of Things (IoT), the Industrial Internet of Things (IIoT), and the Internet of People
(IoP) to the mechanisms of an industrial place. This integration of advanced tech-
nologies has resulted in the creation of smart factories, which are highly automated
and intelligent, with machines that communicate with each other and with humans.
The use of AI and machine learning has also made it possible for machines to learn
and improve on their own, leading to increased productivity and efficiency. It is
also characterized by a focus on sustainability and the environment. Manufacturers
are now turning to sustainable practices such as recycling, renewable energy, and
eco-friendly materials. This focus on sustainability not only benefits the environ-
ment but also helps companies save costs in the long run. As we continue to move
forward, it is clear that the industry will continue to evolve, with new innovations
and advancements that will shape the future of manufacturing.

3
1.2 Definitions
1.2.1 Manufacturing Systems
The modern manufacturing industry (Fig 1.1) thrives on the essential contribution
of human workers and their ability to operate a collection of machines and tools to
initiate the processes required to produce goods or services that provide the needs
of people. However, the success of a manufacturing system is determined by vari-
ous characteristics such as efficiency, flexibility, quality, safety, and cost-effectiveness.

Efficiency is a vital factor in a manufacturing system as it ensures optimal uti-

lization of resources and minimizes wastage. A manufacturing system that is flexible
can quickly adapt to changing consumer needs and market demands, thereby en-
hancing customer satisfaction. Quality is another essential aspect as it ensures that
the products or services produced are of high standards and meet the expectations
of the consumers, leading to customer loyalty and repeat purchases. Safety is a
crucial factor in a manufacturing system as it ensures the well-being of the workers
and prevents accidents or injuries. A manufacturing system that prioritizes safety
creates a conducive work environment that fosters productivity and job satisfaction.
Moreover, cost-effectiveness is an essential factor as it ensures that the manufactur-
ing system is profitable and sustainable in the long run.

Over the years, the manufacturing industry has undergone several revolution-
ary changes resulting in the emergence of different manufacturing systems. Each
manufacturing system possesses its unique advantages and challenges. For instance,
the mass production system introduced during the industrial revolution was highly
efficient, but it lacked flexibility and produced standardized products. In contrast,
the lean production system, which emerged in the 1990s, is highly flexible and em-
phasizes quality, but it requires a skilled workforce. The success of a manufacturing
system is determined by its ability to meet the needs of the consumers while ensur-
ing efficiency, flexibility, quality, safety, and cost-effectiveness. As the manufacturing
industry continues to evolve, it is essential to embrace innovative technologies and
best practices to remain competitive and meet the demands of the market.

4
 Figure 1.1: Manufacturing System source: sketchbubble

1.2.2 Craft Manufacturing

Craft manufacturing is a process that has been used for centuries, long before the in-
troduction of automated manufacturing lines and systems in the 20th century. This
method relies on the expertise of highly skilled workers and the use of simple but
flexible tools to produce goods that meet the precise demands of the customer. In
this process, the focus is on quality and attention to detail, with the aim of creating
a unique product that stands out from the rest.

Craft manufacturing is a time-honored tradition that has been passed down from
generation to generation. The skills and techniques used in this process are often
learned through apprenticeships, where young workers are taught by experienced
craftsmen. This process not only ensures that the craft is preserved, but it also
helps to maintain the quality of the product. It is known for its attention to detail
and precision. The craftsmen who work in this field take great care to ensure that
each product is made to the highest standards. This often involves using specialized
tools and techniques that have been refined over many years. The result is a product
that is not only functional but also beautiful and unique.

One of the advantages of craft manufacturing is that it allows for a high degree of
customization. Because each product is made by hand, the customer can work with
the craftsman to create a product that meets their exact needs and specifications.
This level of customization is not possible with automated manufacturing, which
produces products in large quantities that are designed to meet the needs of the
masses.

5
1.2.3 DMS(Dedicated Manufacturing System)
In the 1990s, there were two main types of manufacturing systems that were common
in the industry (Fig 1.2) [GŚ12] . The first one was the Continuous Manufacturing
system, which primarily involved producing goods for stock. This system relied on
forecasting to estimate the likely demand for the products. The production pro-
cess was standardized, and the inventory was managed through the first-in, first-out
(FIFO) method. Additionally, the work carried out in this system was not diverse,
and the workload was balanced. On the other hand, the second type of manufac-
turing system was called the Intermittent Manufacturing system. This system was
designed to satisfy orders placed by customers. The production facilities were flex-
ible enough to handle a wide variety of products and sizes. The storage was done
between operations, and the system could accommodate small quantities of products
that were flexible in nature. However, the workload in this system was unbalanced,
and the production process was not standardized.

The Continuous Manufacturing system was ideal for companies that manufac-
tured products in high volumes and had a stable and predictable demand for their
products. This system allowed companies to achieve economies of scale and optimize
their production processes. It was also ideal for companies that wanted to keep their
inventory levels low and minimize the risk of holding excess stock. On the other
hand, the Intermittent Manufacturing system was ideal for companies that produced
a wide variety of products or had a constantly changing demand for their products.
This system allowed companies to be more agile and responsive to their customers’
needs. It also enabled companies to produce small quantities of products efficiently
and cost-effectively.

Both the Continuous Manufacturing system and the Intermittent Manufacturing

system had their unique strengths and weaknesses. Companies had to choose the
system that was best suited to their production requirements and business goals.
The decision to adopt a particular manufacturing system had a significant impact on
the company’s operational efficiency, profitability, and customer satisfaction. There-
fore, it was crucial for companies to carefully evaluate their options before making
a decision.

Figure 1.2: Dedicated Manufacturing System

1.2.4 FMS(Flexible Manufacturing System)

In the 1960s, the market competition was spiraling, and companies were struggling
to keep up with the ever-changing demands of the consumers. It was during this

6
time that FMS came into existence. It was a revolutionary concept that provided a
fast and flexible response to unexpected changes in the market. FMS is a group of
numerically controlled machinery that allows for the production of a large variety of
small quantities of products. The system is designed to load and unload tools and
workpieces automatically, which significantly reduces the need for human interven-
tion. This means that the system can operate virtually unattended for long periods,
making it incredibly efficient and cost-effective.

One of the most significant advantages of FMS is its flexibility. The system can
quickly adjust to changes in demand, allowing companies to produce a wide variety
of products without having to reconfigure their production line. This is particularly
beneficial for companies that produce a range of products or have a fluctuating de-
mand for their products.

FMS has become increasingly popular in recent years, and many companies have
adopted this technology to improve their productivity and efficiency. The system
has also helped companies to reduce their manufacturing costs, as it eliminates
the need for manual labor and reduces the risk of errors in production. This MS
has revolutionized the manufacturing industry by providing a fast, flexible, and
cost-effective solution to the production of small quantities of products. With the
growing demand for customized products and the need for quick response times,
FMS is quickly becoming an essential tool for companies looking to stay ahead of
the competition.

1.2.5 RMS(Reconfigurable Manufacturing System)

An RMS is a production system that is designed to be flexible and adaptable. It
allows us to add, modify, delete, and exchange modules and machines, depending on
the production needs and changes. This means that RMS can easily accommodate
changes in production processes, and it can quickly adjust to new market demands.
The primary focus of RMS is to produce part families. Part families are groups of
parts that have similar characteristics, such as size, shape, or function. By grouping
parts into families, RMS can optimize production processes and reduce the time and
cost of manufacturing.

One of the key advantages of RMS is its ability to reconfigure itself quickly. This
means that if a company needs to change its production processes, it can do so
without having to invest in new equipment or machinery. Instead, it can simply re-
configure its existing RMS to meet the new requirements. Another benefit of RMS
is that it can improve the quality of the products produced. By using advanced
technology and automation, RMS can reduce the risk of errors and defects, which
can lead to higher customer satisfaction and loyalty.

RMS is a flexible and adaptable production system (Fig 1.3) [And17] that can
easily accommodate changes in production processes. Its focus on part families al-
lows for optimization of production processes, and its ability to reconfigure quickly
can save time and cost. Additionally, it can improve the quality of products pro-
duced, leading to higher customer satisfaction and loyalty.

7
 Figure 1.3: Reconfigurable Manufacturing System

In modern manufacturing, the need to remain competitive and adapt to rapidly

changing markets has led to the development of advanced production systems. One
such system is the RMS. This system has become increasingly popular in recent
years due to its many advantages, including scalability, convertibility, customiza-
tion, modularity, diagnosability, and integrability.

• Scalability : it’s a key feature of an RMS, as it allows manufacturers to ad-

just production capacity according to the situation. This can be achieved by
adding or removing machines, changing production lines, or reconfiguring ex-
isting equipment. By doing so, manufacturers can quickly respond to changes
in demand, reduce lead times, and improve overall efficiency.

• Convertibility:is another important characteristic of an RMS. This refers to

the ability to transform the functionality of the system to satisfy specific re-
quirements. For example, an RMS may be configured to produce one product
today and a completely different product tomorrow. This flexibility allows
manufacturers to quickly adapt to changes in market demands, without the
need to invest in new equipment or systems.

• Customization:is a feature of an RMS that is limited to part families. This

means that the system can be tailored to produce a variety of products within a
specific category, such as automotive parts or medical devices. This flexibility
allows manufacturers to produce a wide range of products while maintaining
their competitive edge.

• Modularity: is another key feature of an RMS. This refers to the ability

to change parts of the machinery in order to respond to production changes.
For example, an RMS may be designed with interchangeable tooling, allowing

8
 manufacturers to quickly switch between different product lines. This modu-
larity also makes it easier to maintain and upgrade the system over time.

• Diagnosability:is an important characteristic of an RMS, as it allows for

real-time diagnosing of product quality. By monitoring the production pro-
cess and analyzing data, manufacturers can quickly identify and address any
quality issues, reducing waste and improving overall efficiency.

• Integrability:is the final characteristic of an RMS, and refers to the ability

to rapidly integrate modules by hardware and software interfaces. This al-
lows manufacturers to quickly add new equipment or processes to the system,
without the need for extensive reconfiguration or downtime. This flexibility
is essential for maintaining competitiveness in today’s rapidly changing man-
ufacturing environment.

RMS types
RMS is a flexible and adaptable approach to manufacturing that allows for multiple
types and configurations to meet the specific needs of a given production line. When
first considering the implementation of an RMS, there are various criteria to con-
sider, such as the type of product being manufactured, the volume of production,
and the level of automation required.

• Reconfigurable Flow Lines (RFL): These production lines consist of a

series of workstations, each equipped with reconfigurable machines that can
perform a variety of tasks ( Fig 1.4) [YCGBD21]. This type of RMS is partic-
ularly effective in high-volume production environments where efficiency and
speed are crucial. The flexibility of the RFL allows for quick changes in pro-
duction processes, making it an ideal choice for companies that require a high
level of adaptability.

Figure 1.4: Flow line configuration

9
• Reconfigurable Cellular Manufacturing System (RCMS): This type of
production system is based on the concept of group technology, where a group
of reconfigurable machines is organized into cells that share similar produc-
tion tasks (Fig 1.5) [YCGBD21]. The RCMS is particularly effective in low
to medium volume production environments where product customization is
important. The ability to reconfigure the production line according to product
demands makes the RCMS a highly versatile manufacturing system.

Figure 1.5: Reconfigurable Cellular Manufacturing system

• The Dynamic Cellular Manufacturing System (DCMS): is another

type of RMS that applies the same principles as the RCMS. The only difference
is that the DCMS is composed of movable machines instead of reconfigurable
machines. This type of system is particularly effective in environments where
space is limited, and production requirements are constantly changing.

• the Rotary Machining System: is a type of RMS that utilizes a ro-

tary table to move the product through different modular machines (Fig 1.6)
[YCGBD21]. This type of system is particularly effective in high-precision ma-
chining applications where accuracy and consistency are crucial. The modular
design of the Rotary Machining System allows for easy reconfiguration and
modification of the production line, making it an ideal choice for companies
that require a high level of flexibility.

10
 Figure 1.6: Rotary Machining System

1.2.6 MMS (Matrix-structured Manufacturing System)

The matrix manufacturing system, as the name suggests, has a matrix layout where
all workstations are interconnected. The two main elements of this system are the
products and the workstations. It has some unique principles that distinguish it
from traditional manufacturing systems. For instance, each workstation has its own
pace and cycle time, which helps prevent starvation and blocking. This system can
produce multiple products using routing flexibility with Automated Guided Vehicles
(AGVs) to transfer the product flow between workstations. This layout provides a
great deal of flexibility for products and task assignment, within certain constraints.

As seen in the figure (Fig 1.7) [SHGT15], the difference between the classic
manufacturing system (MS) and the matrix manufacturing system (MMS) is clear.
For example, in the classic MS, product 1 must be completed before product 2 to
minimize changeover time and maximize workstation utilization. However, in the
MMS, a workstation can perform multiple tasks for multiple products, allowing a
reduction in the number of workstations or an increase in the number of product
types while maintaining high utilization.

11
 Figure 1.7: Comparison of a classic MS and a MMS configuration

1.3 Assembly Line Balancing

In the manufacturing industry, line balancing (Fig 1.8) [KM13] is an essential pro-
cess that ensures optimal productivity and efficiency in assembly line operations.
The goal of line balancing is to distribute workloads evenly across the production
line, minimizing idle time and maximizing throughput. To achieve this, manufac-
turers must assign the appropriate number of employees or automated machines to
each section of the assembly line. This process involves analyzing the cycle time of
each station and determining the optimal number of workers or machines required
to complete the task within that time frame.

Streamlining workflow is another crucial aspect of line balancing. This involves

coordinating workstations and tasks to minimize unnecessary movement and im-
prove the overall flow of the production line. Manufacturers must also continuously
evaluate and improve their assembly line processes to identify and eliminate any
bottlenecks or inefficiencies that may arise.

There are two primary types of line balancing methods: SALB-1 and SALB-2.
SALB-1 focuses on minimizing the number of stations based on cycle time. This
involves grouping similar tasks together and eliminating any redundant or unnec-
essary stations. By reducing the number of stations, manufacturers can minimize
setup time and reduce idle time between tasks. On the other hand, SALB-2 aims
to reduce cycle time by adjusting the number of stations. This method involves
adding or removing stations to balance the workload across the production line. By
optimizing the number of stations, manufacturers can achieve faster cycle times and
improve overall productivity.

There are two main types of assembly line balancing problems - single-model and
multi-model. In both types, there are four subcategories each based on whether the

12
problem is deterministic or probabilistic and whether the assembly line is straight-
type or U-type.

The first subcategory, SMDS or Single-Model Deterministic Straight-type, refers

to a scenario where there is only one product being manufactured, and the produc-
tion process is deterministic, meaning that the time required for each task is fixed
and known in advance. The assembly line in this case is straight, meaning that the
flow of work is linear. SMDU or Single-Model Deterministic U-type is similar to
SMDS, but the assembly line is in a U-shape. This is often the case when there are
constraints on the floor space available for the production process.

The third subcategory, SMPS or Single-Model Probabilistic Straight-type, is where

the production process is not deterministic, and there is some variation in the time
required for each task. This could be due to factors such as worker variability or
machine breakdowns. The assembly line is still straight in this case. SMPU or
Single-Model Probabilistic U-type is the same as SMPS, but the assembly line is in
a U-shape.

Moving on to the multi-model subcategories, MMDS or Multi-model Determinis-

tic Straight-type is where there are multiple products being manufactured, but the
production process is still deterministic. The assembly line is straight, and each
product follows the same sequence of tasks. MMDU or Multi-model Deterministic
U-type is similar to MMDS, but the assembly line is in a U-shape.

The seventh subcategory, MMPS or Multi-model Probabilistic Straight-type, is

where there are multiple products being manufactured, and the production process
is not deterministic. The assembly line is still straight, and there is some variability
in the time required for each task. Finally, MMPU or Multi-model Probabilistic
Straight-type is the same as MMPS, but the assembly line is in a U-shape.

Figure 1.8: Diffrent Assembly Lines

1.4 Conclusion
The manufacturing world has undergone rapid changes over the years, transitioning
from the mechanical industry to the energy sector, then to the digital industry, and
currently to Industry 4.0 (the integration of artificial intelligence with machines).

13
There are several types of industrial systems. Craft Manufacturing relies on the
expertise of highly skilled workers, allowing for a high degree of customization, but
it is not feasible to produce large quantities.

Dedicated Manufacturing Systems, along with Continuous Manufacturing systems

and Intermittent Manufacturing systems, are used for large production volumes with
less variety. The Flexible Manufacturing System enables the production of a wide
variety of products without the need to reconfigure the production line, known for
its flexibility, but it is limited in capacity.

Reconfigurable Manufacturing Systems combine the advantages of both Dedicated

Manufacturing Systems and Flexible Manufacturing Systems, allowing for the pro-
duction of high quantities with a wide variety. The Matrix-structured Manufactur-
ing System is a type of RMS where all workstations are interconnected in a matrix
layout. This system can produce multiple products using routing flexibility while
maintaining high utilization. Assembly Line Balancing is an essential process that
ensures optimal productivity and efficiency in assembly line operations.

14
Chapter 2

RMS literature review

2.1 Introduction
In this chapter, we will discuss the existing literature on the various manufactur-
ing systems we previously mentioned, including Dedicated Manufacturing Systems,
Flexible Manufacturing Systems, and Reconfigurable Manufacturing Systems. Our
focus will be primarily on Modular Manufacturing Systems (MMS) and Assembly
Line Balancing Problems (ALBP).

2.2 Literature review

2.2.1 DMS
This paper [LGZ09] introduces a combined approach of cell evaluation and genetic
algorithms (CEGA) to optimize the dedicated remanufacturing system through sim-
ulation. The paper begins by highlighting the unique features and challenges of the
dedicated remanufacturing process. It then presents a simulation model that incor-
porates a prioritized stochastic batch arrival mechanism, accounting for factors that
influence the total profit. Using this simulation model as a foundation, the CEGA
algorithm is developed to optimize the production planning and control strategies
for the dedicated remanufacturing facility. The paper further provides a case study
based on a remanufacturing facility located in Austin, USA.

2.2.2 FMS
In this study [BD13], a cuckoo search (CS)-based strategy has been developed to
optimize the scheduling of a flexible manufacturing system. The goal is to minimize
the penalty cost associated with manufacturing delays and maximize machine uti-
lization time. To showcase the application of the CS-based approach, the researchers
have modified the Levy flight operator to accommodate the discrete nature of the
solution, which was tested on a standard FMS scheduling problem consisting of
43 jobs and 16 machines. The CS scheme was implemented using Matlab, and its
performance was compared to other soft computing-based optimization techniques,
such as genetic algorithm (GA) and particle swarm optimization, found in the litera-
ture. The results demonstrate that the CS-based approach outperforms the existing

15
heuristic algorithms, including GA, for the given problem.

Extensive research has been conducted on the design and operational challenges
of automated manufacturing systems, yet many issues remain unresolved. Notably,
the scheduling task and the control problem during operation are crucial due to
the dynamic nature of flexible manufacturing systems (FMS), such as the flexibility
of parts, tools, and automated guided vehicle (AGV) routes. Various traditional
optimization techniques have been employed to tackle the FMS scheduling problem.
While these methods can provide optimal solutions for small-scale problems, they
often prove inefficient when applied to larger-scale issues. In this work [JAPS05],
diverse scheduling mechanisms have been developed to generate optimal schedules,
including non-traditional approaches like genetic algorithm (GA), simulated anneal-
ing (SA) algorithm, memetic algorithm (MA), and particle swarm algorithm (PSA).
These methods consider multiple objectives, aiming to minimize both machine idle
time and total penalty costs for missed deadlines. The memetic algorithm presented
here combines elements of genetic algorithms and simulated annealing. The results
of these different optimization algorithms (memetic, genetic, simulated annealing,
and particle swarm) are compared, and conclusions are drawn.

Organizing schedules for versatile job workshops is crucial in both production

management and combinatorial optimization domains. Yet, attaining the ideal solu-
tion through conventional optimization methods remains challenging due to the im-
mense computational complexity involved. Combining various optimization criteria
further escalates the intricacy, leading to new challenges. Particle swarm optimiza-
tion, an evolutionary computational technique, mimics the behavior of flying birds
and their information-sharing mechanisms. It harmonizes local search (based on per-
sonal experience) and global search (based on neighboring experience), demonstrat-
ing high search efficiency. Simulated annealing, a local search algorithm, employs
specific probabilities to avoid becoming trapped in local optima, proving effective in
diverse situations, including scheduling and sequencing. By thoughtfully integrating
these two methodologies in this article [XW05], researchers have developed an eas-
ily implemented hybrid approach for the multi-objective flexible job-shop scheduling
problem (FJSP). The computational study results have shown that the proposed al-
gorithm is a viable and effective solution, particularly for large-scale FJSP problems.

2.2.3 RMS
This case of study of the LEGO group [WBH12] they tried to make a conceptual
model to analyze the most suitable areas for RMS implementations, where they de-
vided their process into four steps: Molding,Decoration,Assembly and Packaging.In
each of these steps they narrowed down caracteristics: Responsiveness, Capacity,
Functionality and Cost.They put their own scale which goes from 0 to 1 , 0 meaning
no existance of the RMS caracteristic in the said step and 1 the full existence. They
analyzed the percentage of RMS cacarteristics already found in their process and
where they can implement the rest to make their process even more suitable for
RMS scale.

16
RMS optimization problems are grouped into four different categories: RMS de-
sign, Production planning and scheduling (PPS), Layout design and Line balancing
and re-balancing. This conference paper [BGDJLD21] integrates process planning,
scheduling, and layout design problems into a model formulation. The choice of inte-
grating these three optimization problems was made because when treating process
planning and scheduling separately, they have conflicting objectives, and layout re-
configuration directly impacts system configuration and material handling distances
which affects process planning and scheduling decisions. In the end, the three prob-
lems are interconnected by a shared objective: resource allocation to improve system
performance.This study aims to minimize total costs: production costs+ machine
reconfiguration cost + layout reconfiguration cost + material handling cost and the
tardiness cost.

[YCGBD21] The aim of this document is to review research publications focus-

ing on RMS optimization challenges and their resolution techniques. This involves
describing the types of RMS and their constituent parts. Furthermore, the relevant
objective functions and performance metrics of RMS are discussed. Additionally, a
summary of the most commonly employed solution methods and a categorization
of optimization problems are provided. We will see a summarizing tables of the
search publications in the [YCGBD21]: Production planning & Scheduling (table
2.1), layout design (table 2.2), line balancing/ re-balancing (table 2.3), RMS design
(2.5).

• Production planning & Scheduling :

Objectives Count
min makespan 2
min total cost 2
scheduling min total tardiness 2
min mean flow time 1
min mean tardiness 1
max profit 1
planning min energy consumption 1
max throughput 1
Table 2.1: Production planning & Scheduling

17
• Layout design:

Objectives Count
min material handling cost
min reconfiguration cost 1
min constraint satisfaction 1
penalty 1

Table 2.2: Layout design summarized articles

• Line balancing/ re-balancing

Objectives Count
min the number of CNC
1
machines
Table 2.3: line balancing/ re-balancing summarized arti-
cles

• RMS design:

Objectives Count
min total cost
7
min completion time
4
min total cost
3
min completion time
2
min makespan
1
max system flexibility
Process 1
max system modularity
planning 1
max throughput
1
min energy consumption
1
max machine precision
1
max smoothness
1
min machines exploitation time
1
Rotary
max total cost 2
machining
min cycle time 1
system design

18
 min capital cost 5
min reconfiguration cost 4
min total cost 2
max operational capability 2
max machine reconfigurability 2
min investment cost 2
min production cost 1
min life cycle cost 1
min inventory cost 1
Flow line
min operating cost 1
configuration
min maintenance cost 1
selection
min capital cost 1
max machine utilization 1
max configuration convertibility 1
min reconfiguration cost 1
min total tardiness 1
min reconfiguration cost 1
min tardiness 1
max workload balance 1
min system availability 1
min Inter-cellular movements
2
min Machine related costs
1
min total (Re)-Manufacturing costs
1
RCMS/DCMS min Intra-cellular void
1
Design min Total cost
1
min Total changes in auxiliary
1
modules
1
min Machine workload balancing
Table 2.5: RMS design summarized articles

2.2.4 MMS
In this piece [SWS22], the writer revisits their previous work to examine the effects
of flexibility on MMS systems. A cost-focused perspective is applied to assess the
long-term planning challenges of MMS design. Specifically, three types of flexibility
are explored: Material or Equipment flexibility, Material handling flexibility, and
Operation flexibility. The first two relate to the products that need to be produced.
Material handling flexibility concerns the loading, unloading, and transportation of
materials between stations in the MMS system. Operation flexibility pertains to
the ability to produce products using different processes, resulting in the same end
product. The author conducts simulations for each flexibility, using the same data
and materials as their previous article. The results show that enabling all flexibili-
ties yields the same outcome as disabling operation flexibility.This system remains
unaffected by the degree of flexibility in operations.

19
2.2.5 ALBP
Organizing an assembly line efficiently is a crucial challenge faced by manufacturers.
The task involves allocating the total work required to produce a single unit among
the various workstations along the line. This basic version of the problem, known
as the simple assembly line balancing problem (SALBP), has been extensively stud-
ied by researchers and professionals in the field of operations research for nearly 50
years. In this article [BS06], the authors provide a thorough and up-to-date review
of the SALBP research, with a particular focus on the latest significant and influ-
ential contributions that have shaped the field.

For centuries, from the pioneering days of Henry Ford to the cutting-edge era of
Industry 4.0, flow-based assembly processes have been a cornerstone of mass pro-
duction across diverse industries. At the heart of these systems lies the assembly
line balancing problem – a fundamental optimization challenge that determines the
efficient division of labor among the various stations along the production line. This
paper [BSS22]presents a comprehensive review of the scientific literature on assem-
bly line balancing, covering the advancements published since the last major reviews
in 2006 and 2007. It delves into the essential stages of the decision-making process,
exploring novel methods for effectively gathering the necessary (precedence graph)
data, examining emerging problem variants and models, and highlighting the most
significant algorithmic developments. Moreover, it outlines a prospective research
agenda for the next fifteen years, charting the future direction of this crucial field.

Mass production through assembly lines has been a longstanding and appealing
approach. Since the pioneering work of Henry Ford, the assembly line concept
has evolved from rigid, single-model production to more flexible systems. These
advancements include parallel workstations, customer-oriented mixed-model and
multi-model lines, U-shaped configurations, and unpaced lines with intermediate
buffers.Configuring an assembly line requires addressing the critical assembly line
balancing problem. This involves distributing the total workload for manufacturing a
product across the various workstations along the line. While early research focused
on the simple assembly line balancing problem (SALBP) with restrictive assump-
tions, more recent work has aimed to describe and solve more realistic generalized
problems (GALBP). This paper[BS06] provides an overview of the developments in
GALBP research.

Assembly lines are specialized manufacturing systems that play a crucial role in
the large-scale production of standardized goods. Interestingly, these assembly lines
have also become important in the low-volume production of customized products,
a process known as mass-customization. Given the significant financial investment
required to set up or modify an assembly line, careful planning of its configuration
is of paramount importance for businesses. This challenge has attracted the atten-
tion of numerous researchers, who have developed optimization models to support
real-world configuration planning, known as assembly line balancing problems. De-

20
spite the substantial academic effort in this field, there remains a considerable gap
between the requirements of actual configuration problems and the current state
of research. To bridge this communication gap between researchers and practition-
ers, this article [BFS07]presents a classification scheme for assembly line balancing,
which is a valuable step in identifying the remaining research challenges that can
contribute to closing the existing gap.

In this article [KHBB20] The writers set out with minimizing three main goals:total
completion time ,total production cost , and total energy consumption. They de-
cided not to directly equate energy consumption with expenses due to the complex-
ities of how energy prices fluctuate over time(TOU:Time of use). After researching
different ways to make their processes more efficient and comparing various opti-
mization techniques, they settled on using ϵ-constraint approach. By choosing this
method, they aim to strike a good balance between getting things done faster and
being more environmentally friendly. They hope that by optimizing their processes
in this way, they can not only save time and energy but also contribute to a more
sustainable and cost-effective operation in the long run.

This study [BDNS+ 23] addresses work task and resource allocations to work sta-
tion with buffer capacity to maximize throughput, minimize total buffer capacity
while considering the stochastic behavior of the system. To achieve this, they use a
two-step method called sim-opt. In the first step, they employ FACTS Analyzer to
simulate different scenarios, giving them a picture of how the system behaves under
various conditions. This helps them understand where potential issues might arise
and where improvements can be made. In the second step, they turn to NSGA-2,
a well-known optimization technique, to find the best solutions. NSGA-2 is partic-
ularly useful for handling multiple objectives simultaneously, making it a good fit
for this task. By using NSGA-2, they aim to strike a balance between maximizing
productivity and minimizing the need for buffer space, considering the uncertainties
in the system. Through this combined approach of simulation and optimization,
they’re able to tackle the complexities of task allocation and resource management
in dynamic environments effectively. This not only improves productivity but also
ensures that the system remains adaptable and resilient to unforeseen changes.

Many studies on balancing mixed-model assembly lines assume that a task com-
mon to multiple models must be performed at a single station. However, in this
research[BR06], the authors chose to remove this limitation, enabling a shared task
to be assigned to different stations for each model. Their goal was to minimize the
combined costs of the stations and any necessary task duplication. An optimal solu-
tion method was developed using a backtracking branch-and-bound algorithm, and
its effectiveness was evaluated through extensive experimentation. Additionally, a
heuristic branch-and-bound-based approach was created to tackle larger-scale prob-
lems. The heuristic solutions were compared to a lower bound, and the results
demonstrated that this approach provided significantly better outcomes than tradi-
tional techniques.

This piece [DK19] examines how mixed-model manufacturing systems address

the needs of modern manufacturing, bringing together the benefits of mass produc-

21
tion and the ability to create personalized products to meet individual customer
requirements. The article explores various metaheuristics that can aid in task as-
signment, ultimately concluding that simulated annealing is the most effective op-
tion. Other methods discussed include Genetic Algorithm, Imperial Competitive
Algorithm, and Particle Swarm Optimization. To compare these metaheuristics,
the article looks at how they describe problem elements and evaluate potential so-
lutions. The authors conclude that population-based metaheuristics rely on initial
population allocation and progression direction, while classic algorithms may not be
the best fit for complex multi-objective problems.

[BY21] delves into the type-2 mixed-model assembly line balancing, which involves
determining the ideal cycle time for a specific number of workstations. To achieve
this, the author recommends the use of a Genetic Algorithm as a greedy randomized
adaptive search procedure to minimize cycle time by finding the optimal task as-
signment. To build the initial population, the author employs the GRASP method
and uses RPW to ensure the feasibility of the solution. Local search is also utilized
to enhance the constructed solution. Finally, the author compares the results of
the executions with the final solutions obtained using the hybrid GRASP-GA. A
mathematical formulation is presented, which is akin to the SALB-2 problem. A
numerical example is also provided to support the author’s proposed method.

The article [JW03] discusses the popularity of the goal chasing method for the
mixed-model assembly line balancing problem. However, it has a flaw in that it uses
up the ”good” parts early on, leaving fewer options for the later stages. The author
takes it upon himself to identify these ”good” parts and sequences and analyze their
relationship with the optimal solution. His solution is the ”variance algorithm,”
which he supports with a numerical example. To begin, he provides the necessary
mathematical formulation. He then details the goal chasing method and its short-
comings. He suggests three methods to address the problem: symmetry, horizon,
and rate-preserving. The author discusses the issue of faulty units and suggests
methods to prevent their occurrence. To support his argument, he presents a math-
ematical demonstration that utilizing the ”variance method” can lead to faster and
higher quality results.

The subject of this piece [ESI+ 11]is the integration of Tabu Search metaheuristic into
the MMPAL-2 mixed model. The parallel assembly is utilized to distribute tasks to
the various workstations. A heuristic algorithm generates an initial solution which
is then passed to the TS to ascertain the minimum cycle time. The author provides
a series of test problems to demonstrate the efficacy of this approach. The system
in question differs from those previously discussed, as evidenced in the accompa-
nying figure which depicts multiple parallel assembly lines featuring mixed-model
products. The author then proceeds to outline the mathematical formulation. It
can be concluded that the proposal has had a positive impact on the issue at hand,
as evidenced by the fact that out of 87 test problems, 77 showed improvement when
the proposed procedure was implemented.

The authors of [SV04] utilized a combination of a mathematical programming model

and an iterative genetic algorithm-based approach to tackle the mixed-model as-

22
sembly line problem (MALBP) with parallel workstations. Their objective was to
maximize the production rate of the assembly line with a set number of operators.
To accomplish this, they developed a 3-step process. The first stage involved using
a constructive heuristic to solve the MALBP-2 and create an initial solution for the
genetic algorithm. In the second stage, they applied the genetic algorithm procedure
GA-1 to optimize the initial solution. Finally, in the third stage, they introduced
genetic algorithm procedure GA-2 to fine-tune the solutions obtained from GA-1 by
balancing the workloads within the workstations to ensure each operator performs
roughly equal amounts of work for each model being assembled.

This study [YYO20] examines a single-model system, considering the hierarchy of

work assignments, positional constraints, and the option to parallelize workstations,
with task assignment restrictions. The primary goal is to minimize the costs as-
sociated with opening workstations and hiring workers, while finding the optimal
solution for assigning tasks to stations and workers to each station and task. To ad-
dress this challenge, the researchers developed an integer programming model with a
single objective and 14 constraints. Recognizing the NP-Hard nature of the problem,
they shifted to a more appropriate approach - Simulated Annealing. To generate
an initial solution, they utilized a modified version of the Ranked Positional Weight
heuristic. Additionally, they implemented a local search procedure to enhance the
solution obtained from the previous meta-heuristic.

[RG17] delves into the intricacies of multi-worker assembly lines, where individu-
als collaborate on the same product within the same workstation. The challenge
lies in determining the optimal number of items to produce, utilizing only the ex-
isting workstations without adding new ones. To address this, the authors propose
a mixed-integer mathematical programming model that aims to minimize the cycle
time and the number of workers required. Given the complexity of this problem,
classified as NP-Hard, the researchers employ a metaheuristic approach. They ex-
plore two distinct strategies based on the simulated annealing algorithm: one that
solves the problem directly (DSA) and another that tackles it indirectly (ISA). The
study concludes that the DSA approach outperforms the ISA method in terms of
both solution quality and computational efficiency.

In this article [BR14], the authors address the challenge of worker assignment and
task balancing in an assembly line setting, where individual worker capabilities play
a crucial role. To tackle this problem, they present a Mixed-Integer Programming
model that aims to maximize production rates by optimizing the allocation of work-
ers to stations and tasks. Given the complexity of the problem, which is classified as
NP-Hard, the researchers propose a heuristic approach based on beam search and a
task-oriented branch-and-bound procedure. Through extensive numerical analysis
and comparison, the authors demonstrate that their proposed heuristic outperforms
some existing methods. However, they also acknowledge that the branch-and-bound
method has limitations, as it can only efficiently handle instances with a relatively
small number of tasks. Recognizing the strengths of both approaches, the authors
conclude that the best strategy is to combine their proposed solutions with the
method developed by Vila And Pereira (BBR), effectively leveraging the advantages
of multiple techniques to address this challenging optimization problem.

23
In this research [Per18], the scientists tackle a widespread issue faced by various
industries - the need to balance an assembly line for multiple products while account-
ing for resource demands. The goal is to minimize the total cost of the workforce
and resources required. The problem is inherently complex, classified as NP-Hard,
so the researchers propose a heuristic approach that combines the Hoffman heuristic
and Estimation of Distributions Algorithm. This hybrid method delivers excel-
lent results, generating high-quality solutions while providing a clear framework to
address the challenges of heterogeneity and multiple resources in assembly line bal-
ancing problems.

This research focuses [RE15] on the intricate challenge of managing a mixed-model

assembly line, which involves the simultaneous production of various products on the
same line. The study addresses the problem of assigning workers to workstations,
taking into account their skills and the associated operating costs. The researchers
have developed a model that aims to achieve two primary objectives: minimizing
the overall cycle time and minimizing the operating costs related to the workforce.
Given the high complexity of this problem, the researchers have proposed an alterna-
tive approach - the Imperialist Competitive Algorithm (ICA). This algorithm starts
with a random initial population and then generates supervised-random solutions
in an effort to reach more optimal results. Through a numerical example, the study
has demonstrated the efficiency of the proposed ICA method, showcasing its ability
to achieve near-optimal outcomes.

In this article [TLZ+ 16], the focus is on a unique manufacturing system - the mixed-
model assembly line with sequence-dependent tasks. The problem is modeled with
two key objectives: reducing cycle time and minimizing workload variation. First,
the authors simplify the problem by combining the sequence-dependent connections
and precedence relations into a single precedence graph, effectively transforming it
into a single-model assembly line balancing problem. To address the complexity of
this revised problem, the authors propose using a Genetic Algorithm approach. For
their numerical example and discussion, they rely on an initial solution provided
as input to the GA algorithm, selecting three key factors to base their study on:
processing times, the number of immediate successors, and the number of updated
tasks for each task.

2.2.6 Our reference articles

Modular manufacturing systems (MMS) aspire to attain exceptional operational
agility by incorporating a flexible product flow between stations using automated
guided vehicles, and by furnishing redundant resources for each task, thereby elimi-
nating consistent cycle times and the sequential arrangement of stations. This study
[SWS21]explores the design of MMS with an economic goal in mind. They construct
a mixed-integer program to design MMS. By presenting a numerical case, the au-
thors showcase the efficacy of their approach and identify future research prospects.

24
This study [HG19] examines the initial configuration of such systems. The flexi-
ble layout design problem (FLDP) involves designing a flexible layout for a segment
of the assembly of diverse vehicles. It combines station formation and station loca-
tion issues. Additionally, the FLDP anticipates the operational flow allocation of
the automated guided vehicles. The researchers formulate the FLDP as a mixed-
integer linear program and develop a decomposition-based solution approach that
can optimally solve small to medium-sized instances. Furthermore, they transform
this solution approach into a matheuristic that generates high-quality solutions in
a reasonable time for large-sized instances. They compare the efficiency of flexible
layouts to mixed-model assembly lines and quantify the benefits of flexible layouts,
which increase with vehicle heterogeneity.

2.3 Conclusion
In this chapter, we reviewed research papers related to the Flexible Layout Design
Problem (FLDP).We examined the state-of-the-art in simple line balancing and dis-
cussed the Dedicated Manufacturing Systems (DMS) and Flexible Manufacturing
Systems (FMS) problems. We presented the Reconfigurable Manufacturing Systems
(RMS) problems and summarized them in tables.

Finally, we reviewed articles on Matrix-structured Manufacturing Systems (MMS)

problems and assembly line balancing issues.
These articles led us to study the FLDP problem integrated with resource allocation
problems.

25
Chapter 3

Considered problem and proposed

methods

3.1 Introduction
In this chapter, we dive into the adjustments and implementation of the model origi-
nally developed by [HG19]. Hottenrott’s work solely focused on the task assignment
problem and the minimization of cycle time, which is the duration needed to finish
a set of tasks or a production cycle. Building on this foundational model, we were
inspired by the research conducted by [SWS21] on the economic aspects of resource
distribution. We recognized the need to expand the original model to include the
dynamic allocation of resources, as this is a crucial element in the real-world imple-
mentation of such production systems.

By integrating resource allocation into the problem, we aim to create a more compre-
hensive and realistic model that can better address the challenges faced by modern
manufacturing and service organizations. The distribution of resources, such as
labor, equipment, and materials, significantly impacts the overall efficiency, produc-
tivity, and cost-effectiveness of the production process.

3.2 Model
In this section, we detail the mixed-integer linear program for the FLDP where they
investigate a segment of the final assembly at an automotive OEM in which highly
variant, manual tasks are performed. The purpose is to derive layouts that allow for
the efficient assembly of a given model mix.At first we have the set of model m ∈ M
that needs to be processed, t ∈ Tm the set of tasks for each model that has a starting
task T S and an ending task T E , Vm,t the set of successor tasks that shows wich task
needs to be done for the next one to begin for each model m, the set of routes R that
represents all the possible routes which the AGV can take,the layout is represented
by l ∈ L locations at which stations could be opened as resumed in table 3.1. We
decide whether a station is opened at a location (variables Xl ) and which tasks are
assigned (variables Yt,l ). Also, anticipate the models’ flow allocations along the
positions i ∈ Ir of the routes (variables Zm,r,t,i ).The objectives are to minimize the

26
number of opened stations as well as to minimize the flow intensity. Common entry
and exit points are required, the exact locations of the entry and exit points are
determined in the FLDP. The layout is designed for a given model mix and volume
that should be produced in a given production time.
We use the notation as summarized in Tables 3.1, 3.2, 3.3.

3.2.1 Model Notation

Index sets
m∈M Models
t∈T Tasks
[T S , T E ] ⊂ T Dummy start and end tasks
T \[TS , T E ] Real tasks
Tm ⊂ T Tasks for model m
t2 ∈ V m, t Successor tasks of task t for
. model m(precedence
. relations): t to be finished
. before t2 starts
l∈L Locations
r∈R Routes: potential AGV flow
. paths
i ∈ Ir Position index on route
r:i=1......|r|

Table 3.1: Index sets

Parameters
wr Distance on route r
dm Demand for model m
qm,t Task time of task t for model
. m
τ Production time
bt Maximum number of
. duplicates of task t
. (bT S =bT E =1)
el Level index of location l
fl Row index of location l
pr,l Position index of location l on
route r

Table 3.2: Parameters

27
 Decision Variables
Xl 1 if station at location l is
. opened , else 0
Yt,l 1 if task t is assigned to
. location l,else 0
Zm,r,t,i Units of model m that receive
task t at i th location on route
r

Table 3.3: Decision variables

3.2.2 Model Formulation

X
M in Z1 = Xl (3.1)
l∈L
XX
M in Z2 = wr ∗ Zm,r,T S ,1 (3.2)
r∈R m∈M
X
Yt,l ≤ bt ∀t∈T (3.3)
l∈L
X
Yt,l ≤ |T | ∗ Xl ∀ l∈L (3.4)
t∈T
X X X
qm,t ∗ Zm,r,t,pr,l ≤τ ∀l∈L (3.5)
m∈M t∈Tm r∈R|l∈L
X
Zm,r,0,0 = dm ∀m∈M (3.6)
r∈R
X
Zm,r,t,i = Zm,r,0,0 ∀ m ∈ M, r ∈ R, t ∈ Tm \T S (3.7)
i∈Ir
X
Zm,r,t2 ,i ≤ Zm,r,t,j ∀ m ∈ M, r ∈ R, t ∈ Tm , t2 ∈ Vm,t , i ∈ Ir (3.8)
j∈Ir |j ≤ i
X X X
Zm,r,t,pr,l ≤ Yt,l ∗ dm ∀ t ∈ T, l ∈ L (3.9)
m∈M |t∈Tm r∈R|l∈r m∈M
X
Yt2 ,l2 ≤ |L| ∗ (1 − Yt,l ) ∀ m ∈ M, t ∈ Tm , t2 ∈ Vm,t , l ∈ L (3.10)
l2 ∈L|el2 < el
X X
Xl ≤ |L| ∗ Xl ∀ e2 ∈ 1, . . . ., max el (3.11)
l∈L
l∈L|fl =e2 l∈L|el >e2

Xl ∈ {0, 1} ∀ l ∈ L (3.12)
Yt,l ∈ {0, 1} ∀ t ∈ T, l ∈ L (3.13)
Zm,r,t,i ≥ 0 ∀ m ∈ M, r ∈ R, t ∈ Tm , i ∈ Ir (3.14)

We employ a lexicographic multi-objective formulation as shown in Eqs. 3.1 and

3.2.
The objectives are ranked in lexicographic optimization based on their importance.

28
The optimization process starts with the most important objective, optimizing it
without considering the values of the subsequent objectives. Once the first objective
is optimized, its optimal value becomes a constraint for the next objective, and so
on.
The primary objective (3.1) is to minimize the number of opened stations. Min-
imizing the number of opened stations is equivalent to maximizing the efficiency of
the layout because demand and production time are already set. The efficiency is
defined as:
P P
Total workload m∈M dm · qm,t
Efficiency = = P t∈Tm
Installed capacity τ · l∈L Xl
As the subordinate objective (3.2), we minimize the flow intensity for the mini-
mum number of stations. Flow intensity represents the transportation effort and is
calculated as the sum product of route distance and the number of vehicles assem-
bled along a route.
Constraints (3.3) limit the maximum number of task duplicates, except for the
dummy tasks TS and TE which exist only once. Constraints (3.4) guarantee that
tasks can only be assigned to stations that have been opened. Constraints (3.5)
ensure the workload allocated to a station does not exceed the production time.
Constraints (3.6) guarantee the fulfillment of demand and constraints (3.7) maintain
the balance of flow. Constraints (3.8) satisfy the precedence relations by ensuring
that the number of vehicles receiving the successor task at a location cannot be
higher than the number of vehicles receiving the predecessor task at all preceding
locations along a specific route. Constraints (3.9) link the binary assignment vari-
ables Yt,l to the continuous flow variables Zm,r,t,i allowing positive flow only when
the corresponding task is assigned to the corresponding location. Constraints (3.10)
enforce full routing flexibility by forbidding any duplicate of a task t2 that is a
successor of task t for any model m to be assigned to the left of any duplicate of
task t. Essentially, they guarantee that all duplicates of the successor task t2 are
reachable from any duplicate of the predecessor task t by the AGVs, which cannot
travel backward. This enhanced flexibility is especially beneficial in the event of
disruptions. We therefore refer to constraints (3.10) as robustness constraints. Con-
straints (3.11) ensure the layout is longer than its width. These constraints permit
the use of locations only if their row index is less than the level index of the last used
location on the right. Finally, in constraints (3.12)–(3.14), we restrict the domains
of the decision variables.

3.2.3 Proposed Method

In our optimization model, we have customized the initial formulation by including
the concept of resource allocation to each station and task. This represents a cru-
cial extension to the basic model, as it allows us to capture the practical reality of
resource constraints and their impact on the overall system performance.

The key addition to the model is the introduction of a new binary decision vari-
able, denoted as Wrs,l . This variable takes on a value of 1 if a specific resource rs
from the set of available resources Rs is assigned to a particular location l, and 0
otherwise. By incorporating this decision variable, we can now model the available

29
resources, their assignment to various stations, and the execution of tasks at those
stations. We also have the set of parameters Ars which represents the set of tasks
that the ressource rs can execute.

By including the Wrs,l variable, the model can now determine the most efficient
allocation of resources to different stations, taking into account factors such as the
specific requirements of each task, and the overall system objectives,in our case we
consider that the resources are reconfigurable, meaning that we can use them for
doing multiple tasks and for multiple stations, if the same resource is needed for
two diffrent stations then it is considered bought two times , each resource for each
station .This level of detail allows the optimization process to generate more realistic
and practical solutions that can be effectively implemented in the actual operational
environment.

The incorporation of resource allocation into the model is a significant step forward,
as it enables a more comprehensive and realistic representation of the underlying
system. This, in turn, leads to improved decision-making capabilities and the devel-
opment of more robust and effective optimization strategies for the problem at hand.

First we change the second objective (3.2) where we add minimizing the cost of
allocated resources (Cr) and the cost of flow intensity (Ct) as shown in Eq (3.15):
XX X X
M in Z2 = wr ∗ Zm,r,T S ,1 ∗ Ct + Cr ∗ Wrs,l (3.15)
r∈R m∈M rs∈Rs l∈L

And we add the constraint needed as well:

X
Yt,l ≤ Wrs,l ∀ t ∈ T, l ∈ L (3.16)
rs∈Rs|t∈Ars

Wrs,l ∈ {0, 1} ∀ rs ∈ Rs, l ∈ L (3.17)

Constraints (3.16) link the binary variables Yt,l to Wrs,l . This ensures that a task is
allocated to a location that has at least one resource to perform the task. Constraints
3.17 define the nature of the variables Wrs,l .

3.3 2-Phased Resolution Method

In this section, we present a unique resolution approach that effectively addresses the
challenges of resource assignment without directly integrating the resource exten-
sion into the underlying mathematical model. This method, named the ”2-phased
resolution method” consists of two distinct yet complementary phases that work
together to provide a comprehensive solution.It is also considered as a matheuris-
tic approach which is a hybrid optimization approach that combines elements of
mathematical programming with heuristic methods to solve complex optimization
problems efficiently.

The first phase involves developing the fundamental mathematical model. This
model serves as the foundation, without capturing the core constraints and objec-
tives of the resource assignment problem. By focusing on this essential model, we

30
can ensure that the core problem is addressed in a thorough and structured manner,
laying the groundwork for the subsequent phase.

The second phase of our 2-phased resolution method introduces a heuristic ap-
proach for assigning resources to the various locations. This phase utilizes the well-
established optimization technique of set covering to efficiently allocate resources
based on the insights gained from the basic mathematical model. By separating the
resource assignment process from the core model, we can explore a wider range of
heuristic strategies, potentially leading to more effective and flexible resource allo-
cation solutions.

The mathematical foundation serves as the solid basis, while the intricate resource
extension needs are also addressed. This divide between the primary model and
supplementary requirements allows for more flexibility in the resource allocation
process, enabling the investigation of diverse heuristic techniques without adding
undue complexity to the core model.

3.3.1 Motivation
We are going to outline our own proposed approach in two distinct phases, where
we allocate resources to the stations and tasks using our own method, rather than
integrating it into the MILP. This approach provides us with greater flexibility and
control over the resource allocation process, allowing for more customized and opti-
mized solutions.

The initial phase involves running the MILP without any resources, essentially exe-
cuting the Grunow model as-is. This step serves as a crucial precursor to our second
phase, as it provides us with a baseline understanding of the problem and the un-
derlying task structure. By executing the Grunow model without any resources,
we can gain valuable insights into the inherent complexities and constraints of the
problem, which will inform our subsequent resource allocation strategy.

Once we have the results from the initial phase stored in an Excel sheet, we can move
on to the subsequent phase, which entails implementing a set-covering approach to
assign the necessary resources to the identified stations and their respective tasks.
This approach allows us to carefully analyze the resource requirements for each sta-
tion and task, and then devise an optimal way to allocate the available resources
to meet these requirements. By using a set-covering approach, we can ensure that
all the necessary tasks are covered by the assigned resources, while also considering
factors such as resource availability, task priorities, and cost-effectiveness.

The set-covering approach involves several steps, including identifying the resource
requirements for each station and task, determining the available resources and their
capabilities, and then developing an optimization model to assign the resources in
the most efficient manner.

31
3.3.2 Set-Covering Notation
For the set-covering problem, we require mathematical program notation. We define
the set of resources rs ∈ Rs ,for the location l ∈ L and for the tasks t ∈ T . As
for the parameters used are Ars representing the group of tasks each resource rs
can perform, Yl representing the tasks allocated to each station l which is the result
from phase 1. As for this problem it has one decision variable Xr,l which equals 1 if
resource rs is allocated to station l otherwise 0 as resumed in Table 3.4.

Index set
rs ∈ Rs Resources
l∈L Locations
t∈T Tasks

Parameters
Ars Set of tasks which resource rs
. can do
Yl Which tasks t are in location l

Decision Variables
Xrs,l 1 if resource rs is allocated to
station l, 0 else

Table 3.4: Model notation for set covering

3.3.3 Set-Covering model

X X
M in X = Xrs,l (3.18)
rs∈Rs l∈L
X
Xrs,l ≥ 1 ∀l ∈ Y |t ∈ Yl (3.19)
rs∈Rs|t∈Ars

By adopting this two-phase approach, we can leverage the strengths of both the
MILP and the set-covering approach, ultimately leading to a more comprehensive
and effective resource allocation strategy. The initial phase provides us with a solid
foundation for understanding the problem, while the subsequent phase allows us to
fine-tune the resource allocation to better suit our specific needs and preferences.
The objective function Eq 3.18 aims to minimize the number of stations used.
Constraint 3.19 to ensure availability of resource for task t.

3.4 Data Sets

To assess our techniques and determine the more efficient one, we have selected a
data set to walk through our numerical example and discussion. We have used the
same data set that [HG19] employed for their approach and numerical example for
data related to models (task time, successors...). For location-related data (routes,

32
distances...), we relied on [BGK23].

While all our cases share the same number of locations (9), the same set of re-
sources R (20 in total), and the same set of routes (62 possible for the demand to
traverse our mms layout), they also have the same number of models (5 per case).In
these cases, Grunow has incorporated his own understanding, varying the percent-
age change from one case to the next. For Cr , we will select the value of 40000 as
referenced in [SWS21], and for Ct , we will choose a cost of 1.

To generate instances with diverse vehicles, we consider four levels of structural

heterogeneity (sh): 0%, 10%, 25%, and 50%. Structural heterogeneity refers to the
degree of dissimilarity in the precedence graph structures of the models. It is defined
as the average percentage difference between the minimum supergraph’s task count
and the precedence graphs of all models.

In the base case, where model-task assignments remain unchanged (sh = 0%), all
models must handle every task. Consequently, the precedence graphs of the models
are identical and match the data set from Scholl (1993). However, for the case where
25% of the assignments are randomly removed (sh = 25%), we ensure that each task
is still required by at least one model and that each model needs a minimum of two
real tasks. We also consider four levels of task time heterogeneity (tth), where the
task times for each model can deviate from the demand-weighted mean task time
by 0%, ±10%, ±25%, and ±50%. In the base case (tth = 0%), all models have
the same task times, but for tth = 25%, the task times are chosen to be within the
range of 0.75 and 1.25 times the demand-weighted mean task time, maximizing the
heterogeneity. Importantly, the demand for each model and the overall workload for
each task remain constant across all structure and task time heterogeneity levels to
ensure comparable instances.

We have used matplotlib python’s library to draw a precedence diagram to vi-

sualize the precedence relationships between tasks for a given model, as shown in
figure 3.1.

33
 Figure 3.1: An example of a precedence diagram

3.5 Conclusion
Throughout this chapter, we have explored the FLDP with its two objectives: mini-
mizing the number of opened stations and optimizing flow transportation. We have
identified that the new FLDP integrated with resources presents an NP-hard prob-
lem. To tackle this challenge, we have introduced a 2-phased heuristic aimed at
reducing computational time for solving this problem.

In the next chapter, we will discuss the results of the integrated model and the
2-phased approach.
m Repeter ce qu’on a fait dans ce chapitre (une petite intro sur le model, le travail
proposé et la discussion sur le resultat assumptions

34
Chapter 4

Results and discussions

4.1 Introduction
In this chapter, we will discuss the results obtained for all the datasets mentioned
in Chapter 3. Each dataset has a specific number of tasks: Bowman has 8 tasks
with a cycle time of 19, EX has 5 tasks with cycle times of 18 and 20, Jackson
has 11 tasks with a cycle time of 8, Jaeschke has 9 tasks with a cycle time of 7,
Mansoor has 11 tasks with a cycle time of 47, Mertens has 7 tasks with a cycle
time of 8, and Mitchell has 21 tasks with a cycle time of 14. Each dataset also has
its own configurations of ov and tth, we note that in our data set ov=sh and ttv=tth.

Our solution approaches and the instances are implemented in a program appli-
cation written in Python 3 and interfaced with Gurobi 11. All experiments are run
on a computer using an AMD Ryzen 7 5700U with Radeon Graphics processor with
1.8 GHz and 16 GBRAM.

4.2 Results(1200 seconds)

In this section, we are going to talk about the results obtained from simulating our
two approaches for 20 minutes. The reason for this choice of time is that we noticed
the UB has converged rather quickly for some instances, and so that will be our
focus in the next results and discussions.

To substantiate our claim, we selected a random instance ”MITCHELL c=14

ov=25 ttv=25 s=00” as depicted in Figure 4.1. The figure illustrates that our
primary objective (UB) begins to stabilize near the 20-minute mark, with marginal
further improvement that may be insignificant to our analysis. Additionally, it
demonstrates that the lower bound (LB) remains stable for a considerable duration
well before the 20-minute threshold, after which it steadily progresses closer to the
UB, thereby reducing the gap between the two results.

35
 Figure 4.1: UB and LB Convergence

The fact that we also chose to compare the two approaches based on the CPU
time and the GAP from LB (figure 4.2) , which is calculated by:
U B − LB
GAP f rom LB = (4.1)
UB

Figure 4.2: GAP From LB

4.2.1 A description of the result obtained

We will use the Matplotlib library in Python to draw and visualize the solutions
without taking into account resources, and the variation of the objectives. We take
this instance as an example: JACKSON c=08 ov=00 ttv=10 s=00 to explain the
form of the results (Figures 4.3, 4.4).

36
 Figure 4.3: Result example

Locations in green are the opened locations. For this example, we have 6 opened
locations. In each location, we have the set of tasks for this model (for example,
model 0 at location 0 includes tasks: {T0, T1, T2, T3, T5}). The flow of products
is represented by arrows, with each color indicating a different route. The quantity
allocated to each route is shown in the picture legend.

Figure 4.4: The objective variation curves example

As seen in Figure 4.4, the optimization begins with the first objective without

37
considering the value of the second objective. The latter undergoes oscillations be-
fore the first objective completes optimization (around 350 seconds), after which it
begins its minimization process.

To access all the 128 instances results, check this drive 1.

4.2.2 Bowman C=19, 8 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 34,20 1200 39,42 392
10 14,70 1200 21,41 203
00
25 24,80 1200 30,76 314
50 24,00 1200 30,57 383
Average 24,43 1200 30,54 323
00 33,00 1200 38,24 95
10 22,50 1200 28,66 284
10
25 33,70 1200 38,89 99
50 33,90 1200 39,57 173
Average 30,78 1200 36,34 162,5
00 31,20 1200 37,50 113
10 26,00 1200 32,77 94
25
25 29,30 1200 35,80 235
50 17,70 1200 25,24 206
Average 26,05 1200 32,83 162
00 5,23 1200 30,25 139
10 4,05 1199 28,28 162
50
25 15,50 1200 24,48 182
50 9,18 1200 18,84 53
Average 8,49 1199,75 25,47 134

Table 4.1: Bowman GAP from LB for MILP and 2 phased (1200s)

38
4.2.3 EX C=18, 5 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 00 118 15,87 30
10 00 142 18,46 68
00
25 00 133 13,09 31
50 00 143 18,46 71
Average 00 134 16,47 50
00 00 122 18,46 51
10 00 159 00 69
10
25 00 122 18,46 51
50 00 145 15,87 31
Average 00 137 13,19 50,5
00 00 164 18,46 52
10 00 174 8,41 33
25
25 00 150 00 51
50 00 161 18,46 52
Average 00 162,25 11,33 47
00 00 174 00 71
10 00 173 18,46 52
50
25 00 162 00 52
50 00 151 18,46 52
Average 00 165 9,23 56,75

Table 4.2: EX C=18 GAP from LB for MILP and 2 phased (1200s)

39
4.2.4 EX C=20, 5 tasks

OV TTV MILP 2-Phased

Gap from CPU(s) Gap from CPU(s)
LB(%) LB(%)
00 00 99 19,70 40
10 00 98 19,70 38
00
25 00 111 19,70 39
50 00 100 20,08 25
Average 00 102 19,79 35,5
00 00 121 19,70 41
10 00 140 19,70 57
10
25 00 132 1,38 26
50 00 140 19,70 41
Average 00 133,25 15,12 41,25
00 00 123 2,60 26
10 00 125 00 42
25
25 00 134 19,70 43
50 00 125 19,70 42
Average 00 126,75 10,5 38,25
00 00 123 19,70 42
10 00 123 1,91 26
50
25 00 129 19,70 42
50 00 122 0,95 26
Average 00 124,25 10,56 34

Table 4.3: EX C=20 GAP from LB for MILP and 2 phased (1200s)

40
4.2.5 Jackson C=08, 11 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 1200 — 1200
10 43,98 1200 53,25 1022
00
25 42,51 1200 54,36 1187
50 39,63 1200 53,08 1094
Average 42,04 1200 53,56 1101
00* — 1200 — 1168
10 39,95 1194 54,82 879
10
25 40,23 1200 54,39 666
50 49,33 1200 50,57 925
823,33 43,17 1198 53,26 1200
00 42,04 1200 52,03 578
10 40,24 1199 55,16 826
25
25 43,05 1200 54,29 561
50 42,03 1200 53,86 470
Average 41,84 1199,75 53,83 608,75
00 41,85 1278 56,83 329
10 45,90 1200 56,56 415
50
25 42,09 1200 57,42 369
50 46,76 1200 57,02 330
Average 44,15 1219,5 56,95 360,75

Table 4.4: Jackson GAP from LB for MILP and 2 phased (1200s)

41
4.2.6 Jaeschke C=07, 9 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 1200 — 1200
10 30,77 1200 47,75 506
00
25 26,17 1200 49,61 488
50 20,85 1200 45,61 631
Average 25,93 1200 47,65 541,66
00 26,83 1200 53,12 1035
10 25,09 1195 47,58 659
10
25 19,67 1182 40,63 722
50 24,82 1200 47,51 583
Average 24,10 1194,25 47,21 749,75
00 26,97 1200 51,11 561
10 23,32 1199 48,67 319
25
25 26,80 1200 49,81 264
50 23,49 1200 50,84 480
Average 25,14 1199,75 50,10 406
00 7,56 1200 41,15 374
10 6,21 1200 36,34 342
50
25 5,60 1200 35,94 159
50 28,26 1200 53,42 382
Average 11,90 1200 41,71 314,25

Table 4.5: Jaeschke GAP from LB for MILP and 2 phased (1200s)

42
4.2.7 Mansoor C=47, 11 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 26,77 1200 50,86 642
10 26,80 1200 46,48 405
00
25 26,96 1200 50,99 489
50 26,40 1200 46,19 567
Average 26,73 1200 48,63 525,75
00 30,40 1200 49,11 556
10 30,25 1200 49 512
10
25 30,48 1200 49,17 529
50 30,27 1200 53,21 569
Average 30,35 1200 50,12 541,5
00 34,29 1200 41,47 212
10 34,12 1200 51,83 229
25
25 34,02 1200 51,76 360
50 34,06 1200 51,79 187
Average 34,12 1200 49,21 247
00 48,28 1200 62,19 62
10 38,21 1199 58,54 65
50
25 49,10 1200 66,85 57
50 47,07 1200 61,30 230
Average 45,66 1200 62,22 103,5

Table 4.6: Mansoor GAP from LB for MILP and 2 phased (1200s)

43
4.2.8 Mertens C=08, 7 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 00 630 27,43 236
10 00 586 20,13 166
00
25 00 1007 20,13 222
50 00 925 27,37 209
Average 00 787 23,76 208,25
00 00 575 20,13 175
10 00 560 27,43 85
10
25 00 610 27,43 174
50 00 658 20,13 191
Average 00 600,75 23,78 156,25
00 00 762 27,43 153
10 00 595 20,13 136
25
25 00 605 11,19 132
50 00 628 27,43 124
Average 00 647,50 21,54 136,25
00 00 470 27,43 117
10 00 617 11,19 117
50
25 00 470 11,19 121
50 00 662 20,13 108
Average 00 554,75 17,48 115,75

Table 4.7: Mertens GAP from LB for MILP and 2 phased (1200s)

44
4.2.9 Mitchell C=14, 21 tasks

OV TTV MILP 2-Phased

Gap from CPU(s) Gap from CPU(s)
LB(%) LB(%)
00* — 1200 — 1200
10* — 1200 — 1200
00
25* — 1200 — 1200
50* — 1200 — 1005
00* — 1200 — 1200
10* — 1200 — 1069
10
25* — 1200 — 1156
50 73,10 1200 61,73 1011
Average 73,10 1200 61,73 1011
00* — 1200 — 1201
10 56,61 1177 66,74 885
25
25 62,91 1200 65,26 810
50 57,45 1200 64,63 821
Average 58,99 1192,33 65,54 838,66
00 63,05 1196 69,95 625
10 61,76 1200 70,56 1101
50
25 59,59 1200 69,59 443
50 57,98 1200 69,61 739
Average 60,59 1199 69,92 652

Table 4.8: Mitchell GAP from LB for MILP and 2 phased (1200s)

Our data set was limited to 128/528 instances due to the restricted number of loca-
tions and routes available. We will now dive into a deep discussion between the 2
methods with the 128 instances.

From Bowman C=19 instances (Table 4.1), we can see that as the structure
heterogeneity (sh) increases, the optimization converges more rapidly. The differ-
ence between the 2-phased GAP and the MILP GAP values is (6.11%, 5.56%, 6.78%,
16.98%) for sh=00, 10, 25, 50 respectively, with an average difference GAP of 8.86%.
We notice that the less similar the structure of the models, the better the MILP
performs. The 2-PHASED method gives us satisfactory results for small variations
in the structure of the models (sh). Task time heterogeneity (tth) hardly affects
the GAP. The solution time for the 2-phased method is almost 1/4 of the MILP
time. For instances with 5 tasks, without a doubt, the MILP is better than the
2-phased method, with CPU time less than 130 seconds. As we can see in Tables
4.2 (Tc=18) and 4.3 (Tc=20), the 2-phased method did not give us good results.
This is because in phase 1, we focused only on flow transport. From Jackson’s Table
4.4 (Tc=8), we can see that changing SH and TTH does not influence the GAP.
For TTH=50%, the 2-phased method solved within less than 10 minutes (≤ 600s).In
Jaeschke’s Table 4.5, the 2-phased method is not suitable for 9 tasks and Tc=7. The
difference between the two GAPs is greater than 20%. Mansoor’s Table 4.6 (Tc=47)

45
showed that as SH increases, the GAP also increases. This means with 9 tasks and
a high cycle time, the GAP highly depends on SH values. The MILP in Mertens’
Table 4.7 is better than the 2-phased method. However, the 2-phased method gives
us near-optimal solutions in less time (600s). In Mitchell’s Table 4.8 (Tc=14), the
GAP was elevated for both methods. For sh=00 and sh=10, the MILP method did
not have enough time to optimize the second objective within 1200 seconds. The
2-phased method is better than MILP with low changes in structure heterogeneity.
For example, with 5 types of vehicles, if the models are nearly similar for all vehicles,
the 2-phased method can solve the problem effectively. The 2-phased method gives
us satisfactory results for sh=25 and sh=50, with a difference in GAP of less than
10%.
To summarize, the MILP method remains better to the 2-phased approach for a
small number of tasks (≤ 8 tasks). The 2-phased method provides satisfactory re-
sults for more than 8 tasks, particularly when SH=00 or SH=10. The 2-phased
method has yielded satisfactory results with 21 tasks and outperformed in certain
instances. The MILP method did not achieve optimization of the second objective
for 11 out of 128 instances.

4.2.10 Remaining instance (1200s)

Now we are going to tackle the instances that gave us no optimal solution for the
second objective and so we didn’t put them in our first tables, and let them for after
to discuss even better the results that we obtained for these particular cases.

Instance OV TTV MILP 2-Phased

OB1 OB2 CPU (s) OB1 OB2 CPU (s)
Jackson 00 00 6 7.106 1200 6 7.105 1200
6 4
10 00 6 7.10 1200 6 5.10 1168
Jaeschke 00 00 6 7.106 1200 6 4.105 1200
6 5
00 8 7.10 1200 8 9.10 1200
6 6
10 8 7.10 1200 8 1.10 1200
00
Mitchell 25 8 7.106 1200 8 1.106 1200
6 5
50 8 7.10 1200 8 9.10 1005
00 8 7.106 1200 8 9.105 1200
6 5
10 10 8 7.10 1200 8 8.10 1069
6 5
25 8 7.10 1200 8 8.10 1156
25 00 8 7.106 1200 8 9.105 1201

Table 4.9: Remaining instances: GAP from LP for MILP and 2-phased (1200s)

The instances presented in Table 4.9 were stopped during the optimization of the
first objective using the MILP method. However, the 2-phased method yielded satis-
factory results. This is due to the MILP requiring additional time to initiate efficient
optimization for specific instances and parameters in our problem. Consequently,
the 2- phased method proves notably superior in this context.

46
4.3 Results(300 seconds)
We will now test our approach for 5 minutes for several reasons:
First, imagine you are in a meeting and would like to explore different perspectives
by adjusting inputs. Having a fast solution means you don’t have to wait 20 minutes
to see the changes.
Secondly, as observed in the tables of Section 4.2, the 2-phased method found solu-
tions in less than 400 seconds for some instances. Comparing these solutions with
those obtained within the same timeframe, we found that the 2-phased method per-
formed better. In this section, we will test our methods with a time limit of 300
seconds.

4.3.1 Bowman C=19, 8 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 300 — 273
10 34,26 300 39 215
00
25 91,51 300 89,36 229
50 37,18 300 41,85 215
Average 54,31 300 56,73 219,66
00 44,90 303 44,90 113
10 38,71 300 43,55 274
10
25 53,55 314 54,79 53
50 43,62 312 43,63 170
Average 45,19 307,25 46,71 152,5
00 41,89 300 45,59 108
10 34,32 338 40,31 95
25
25 41,38 299 45,42 144
50 34,53 300 40,56 200
Average 38,03 309,25 42,89 136,75
00 25,43 300 45,07 161
10 21,26 300 41,15 104
50
25 49,24 300 54,64 177
50 30,81 300 38,18 52
Average 31,68 300 44,76 123,5

Table 4.10: Bowman GAP from LB for MILP and 2 phased (300s)

47
4.3.2 EX C=18, 5 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 00 71 15,87 29
10 00 85 18,46 66
00
25 00 72 13,09 29
50 00 85 18,46 66
Average 00 78,25 16,47 47,5
00 00 74 18,46 49
10 00 87 00 67
10
25 00 74 18,46 49
50 00 87 15,87 30
Average 00 80,5 13,19 48,75
00 00 91 18,46 51
10 00 104 8,41 30
25
25 00 91 00 50
50 00 90 18,46 50
Average 00 94 11,33 45,25
00 00 104 00 68
10 00 104 18,46 50
50
25 00 90 00 50
50 00 90 18,46 50
Average 00 97 9,23 54,5

Table 4.11: EX C=18 GAP from LB for MILP and 2 phased (300s)

48
4.3.3 EX C=20, 5 tasks

OV TTV MILP 2-Phased

Gap from CPU(s) Gap from CPU(s)
LB(%) LB(%)
00 00 58 19,70 39
10 00 58 19,70 37
00
25 00 59 19,70 38
50 00 59 20,08 24
Average 00 58,50 19,79 34,5
00 00 72 19,70 41
10 00 85 19,70 57
10
25 00 73 1,38 25
50 00 84 19,70 41
Average 00 78,50 15,12 41
00 00 73 2,60 26
10 00 73 00 41
25
25 00 74 19,70 41
50 00 73 19,70 40
Average 00 73,25 10,5 37
00 00 72 19,70 40
10 00 73 1,91 25
50
25 00 73 19,70 41
50 00 73 0,95 25
Average 00 72,75 10,56 32,75

Table 4.12: EX C=20 GAP from LB for MILP and 2 phased (300s)

49
4.3.4 Jackson C=08, 11 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 300 — 300
10* — 300 — 300
00
25* — 300 — 300
50* — 300 — 300
Average — 300 — 300
00* — 300 — 300
10* — 300 — 300
10
25* — 300 — 300
50* — 300 — 300
Average — 300 — 300
00* — 300 — 300
10 49,83 300 63,08 96
25
25 44,12 300 62,39 29
50 46,02 299 58,27 31
Average 46,65 299,66 61,24 52
00 99,45 301 92,93 26
10 46,94 300 60,18 28
50
25 46,74 300 78 21
50 50,01 301 60,49 297
Average 60,78 300,5 72,90 93

Table 4.13: Jackson GAP from LB for MILP and 2 phased (300s)

50
4.3.5 Jaeschke C=07, 9 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 300 — 300
10 34,71 300 53,87 245
00
25* — 300 — 300
50* — 300 — 300
Average 34,71 300 53,87 295
00* — 300 — 300
10 32,98 301 53,90 300
10
25 28,33 300 46,89 319
50 36,76 300 54,06 300
Average 32,69 300,33 51,61 306,33
00* — 300 — 300
10 26,99 300 53,19 215
25
25 35,69 300 50,65 271
50 30,64 300 55,43 275
Average 31,10 300 53,09 253,66
00 21,27 300 49,88 235
10 22,75 300 51,18 350
50
25 21,37 300 46,35 194
50 47,44 304 60,47 347
Average 28,20 300 51,97 281,5

Table 4.14: Jaeschke GAP from LB for MILP and 2 phased (300s)

51
4.3.6 Mansoor C=47, 11 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00* — 300 — 300
10 26,96 300 55,82 300
00
25 26,96 300 50,99 289
50 26,96 299 58,89 300
Average 26,96 299,66 55,23 296,33
00 30,40 300 57,83 300
10 30,42 300 54,62 300
10
25 30,60 299 50,78 300
50 30,48 300 58,07 300
Average 30,47 299,75 55,32 300
00 34,29 300 41,47 208
10 34,29 300 51,96 231
25
25 34,02 300 51,76 249
50 34,06 300 51,79 176
Average 34,16 300 49,24 216
00 52,77 322 65,42 60
10 47,36 303 64,65 62
50
25 53,60 300 68,87 58
50 49,36 300 62,98 226
Average 50,77 306,25 65,48 101,5

Table 4.15: Mansoor GAP from LB for MILP and 2 phased (300s)

52
4.3.7 Mertens C=08, 7 tasks

OV TTV MILP 2-Phased

Gap from LB CPU (s) Gap from LB CPU (s)
(%) (%)
00 16,33 300 39,28 235
10 00 290 20,13 170
00
25 20,24 300 36,30 220
50 20,98 300 42,61 213
Average 14,38 297,5 34,58 209,5
00 8,63 300 27,37 178
10 00 296 27,43 83
10
25 1,56 300 35,47 146
50 13,98 300 31,30 148
Average 6,04 299 30,39 138,5
00 8,20 300 33,79 150
10 4,37 300 24,40 140
25
25 00 293 11,19 131
50 12,94 300 36,82 125
Average 6,37 298,25 26,55 136,5
00 00 286 27,43 115
10 8,06 300 18,63 117
50
25 4,66 300 16,58 120
50 00 288 20,13 105
Average 3,18 293,5 20,69 114,25

Table 4.16: Mertens GAP from LB for MILP and 2 phased (300s)

53
4.3.8 Mitchell C=14, 21 tasks

OV TTV MILP 2-Phased

Gap from CPU(s) Gap from CPU(s)
LB(%) LB(%)
00 — 300 — 300
10 — 300 — 300
00
25 — 300 — 300
50 — 300 — 300
00 — 300 — 300
10 — 300 — 300
10
25 — 300 — 300
50 — 300 — 300
00 — 300 — 300
10 — 300 — 300
25
25 — 300 — 300
50 — 300 — 300
00 — 300 — 300
10 64,95 300 73,47 301
50
25 63,01 286 70,25 300
50 59,70 300 72,03 300
Average 62,55 295,33 71,91 300,33

Table 4.17: Mitchell GAP from LB for MILP and 2 phased (300s)

For Bowman’s instances (Table 4.10), the results are comparable for lower values of
SH. No significant changes were observed for EX C=18 and EX C=20 (Tables 4.11,
4.12) as the number of tasks is only 5. Jackson’s instances (Table 4.3.4) did not
reach the optimization of the second objective for SH=10. The 2-phased method
showed a substantial difference in GAP compared to the first method for Jaeschke
(Table 4.3.5), Mansoor (Table 4.15), and Mertens (Table 4.16). Mitchell’s instances
(Table 4.17) did not achieve optimization of the second objective except for SH=50.

54
4.3.9 Remaining instances (300s)

Instance OV TTV MILP 2-Phased

OB1 OB2 CPU (s) OB1 OB2 CPU (s)
Bowman 00 00 4 6.106 300 4 3.105 300
6 5
00 00 8 7.10 300 6 6.10 300
6 5
10 8 7.10 300 6 6.10 299
Jackson 25 8 7.106 300 6 7.105 300
6 5
50 8 7.10 300 6 6.10 300
6 5
00 8 7.10 300 6 6.10 300
10 8 7.106 300 6 6.105 299
10 6 5
25 8 7.10 300 6 7.10 300
50 8 7.106 300 6 6.105 300
6 5
25 00 6 7.10 300 6 7.10 300
6 5
00 8 7.10 300 6 7.10 300
25 8 6.106 300 6 6.105 298
00 6 5
Jaeschke 50 8 7.10 300 6 6.10 300
6 5
10 00 8 7.10 300 6 6.10 300
25 00 8 6.106 300 6 5.105 311
6 5
Mansoor 00 00 6 7.10 300 4 7.10 300
6 5
00 8 7.10 287 9 9.10 300
10 8 7.106 300 8 9.105 300
00 6 5
Mitchell 25 8 7.10 299 8 9.10 300
50 8 7.106 300 8 1.106 300
6 5
00 8 7.10 300 8 9.10 300
6 5
10 8 7.10 300 8 9.10 300
10
25 8 7.106 300 8 9.105 297
6 5
50 8 7.10 300 8 8.10 300
00 8 7.106 284 8 9.105 286
6 5
10 8 7.10 300 8 9.10 300
25 6 5
25 8 7.10 300 8 9.10 300
50 8 7.106 300 8 9.105 300
6 6
50 00 8 7.10 299 8 1.10 296

Table 4.18: Remaining instances : GAP from LB for MILP and 2-phased (300s)

For this table, we are going to discuss the instances that didn’t give us a solution for
the second objective. We can see that the 2-phased method is privileged for these
instances. This may be because the MILP needs more time to start optimizing the
second objective efficiently, especially for these multiple instances, as opposed to the
previous results we got for the majority of our dataset.

As we can see, for the instances with a smaller number of tasks (8-11), the 2-phased
method is better than the MILP even in the opened locations. This indicates that
the 2-phased method started off with a good solution and then had the time to opti-
mize it further within the time limit. The initial phase provided a solid foundation,
allowing the method to refine and improve the solution effectively. This advantage

55
highlights the efficiency of the 2-phased method in handling smaller task sets, where
it quickly reaches a viable solution and utilizes the remaining time to enhance it
further.In our case, the same thing goes for the instances with a bigger set of tasks.
As we can see for Mitchell, which has the greatest number of tasks, it is optimized
better with the 2-phased method than MILP.

4.4 Conclusion
Within this chapter, we have explored and observed the distinction between our
methods, employing a 20-minute simulation since the majority of the dataset has
converged towards a feasible and consistent solution. Subsequently, we assessed the
performance of our methods under a more constrained time frame of 5 minutes.

56
General Conclusion and
Perspectives

Within our study, we have addressed the problem using the MILP method and the
2-phased method. We tested the problem with a 20-minute simulation since the
majority of the dataset converged towards a feasible and consistent solution. We
also assessed the performance of our methods under a more constrained time frame
of 5 minutes. Thus, we can draw the following conclusions:

- Our approaches can handle up to 21 tasks due to restricted number of locations

- When the number of tasks is less than 9, the MILP method remains better
than 2-phased approach

- 2-phased method performs better when sh=0 or sh=10 (low values of structure
heterogeneity)

- 2-phased method provides good results in a short time when the number of
tasks is ≥ 11

- 2-phased method approach yields better solutions within a short time restric-
tion.

We suggest some future research directions as follows:

- explore the problem with other configurations of locations and routes

- explore more instances with more tasks

- add that if we purchase multiple units of the same resource, for example 2 or
more, the single cost will be reduced.

- impose restrictions on resources such as availability of resources

57
Bibliography

[And17] Ann-Louise Andersen. Development of changeable and reconfigurable

manufacturing systems: supporting context-specific design of change-
ability. 2017.

[BD13] Shashikant Burnwal and Sankha Deb. Scheduling optimization of

flexible manufacturing system using cuckoo search-based approach.
The International Journal of Advanced Manufacturing Technology,
64:951–959, 2013.

[BDNS+ 23] Carlos Alberto Barrera-Diaz, Amir Nourmohammadi, Henrik Smed-

berg, Tehseen Aslam, and Amos H. C. Ng. An Enhanced Simulation-
Based Multi-Objective Optimization Approach with Knowledge Dis-
covery for Reconfigurable Manufacturing Systems. Mathematics,
11(6):1527, March 2023.

[BFS07] Nils Boysen, Malte Fliedner, and Armin Scholl. A classification of

assembly line balancing problems. European journal of operational
research, 183(2):674–693, 2007.

[BGDJLD21] Isabel Barros Garcia, Joanna Daaboul, Antoine Jouglet, and Julien
Le Duigou. An approach to jointly optimize the process plan, schedul-
ing, and layout design in reconfigurable manufacturing systems. In
International Workshop on Service Orientation in Holonic and Multi-
agent Manufacturing, pages 403–415. Springer, 2021.

[BGK23] Osman Baturhan Bayraktar, Martin Grunow, and Rainer Kolisch.

Instance set and implementation codes of dynamic reconfigurations
of matrix assembly layouts, 2023.

[BR06] Yossi Bukchin and Ithai Rabinowitch. A branch-and-bound based so-

lution 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, 2006.

[BR14] Leonardo Borba and Marcus Ritt. A heuristic and a branch-and-

bound algorithm for the assembly line worker assignment and balanc-
ing problem. Computers & Operations Research, 45:87–96, 2014.

[BS06] Christian Becker and Armin Scholl. A survey on problems and meth-
ods in generalized assembly line balancing. European journal of oper-
ational research, 168(3):694–715, 2006.

58
[BSS22] Nils Boysen, Philipp Schulze, and Armin Scholl. Assembly line bal-
ancing: What happened in the last fifteen years? European Journal
of Operational Research, 301(3):797–814, 2022.
[BY21] LAKHDAR Belkharroubi and KHADIDJA Yahyaoui. A hybrid grasp-
genetic algorithm for mixed-model assembly line balancing problem
type 2. International Journal of Computing, 20(3):424–432, 2021.
[DK19] K Dziki and D Krenczyk. Mixed-model assembly line balancing prob-
lem with tasks assignment. In IOP Conference Series: Materials
Science and Engineering, volume 591, page 012013. IOP Publishing,
2019.
[ESI+ 11] Gholam Reza Esmaeilian, Shamsuddin Sulaiman, Napsiah Ismail,
Maryam Hamedi, and Megat Mohammad Hamdan Megat Ahmad. A
tabu search approach for mixed-model parallel assembly line balanc-
ing problem (type ii). International Journal of Industrial and Systems
Engineering, 8(4):407–431, 2011.
[GŚ12] Arkadiusz Gola and Antoni Świć. Directions of manufacturing sys-
tems’ evolution from the flexibility level point of view. Innovations
in Management and Production Engineering, Oficyna Wyd. Polskiego
Towarzystwa Zarzadzania Produkcja, Opole, pages 226–238, 2012.
[HG19] Andreas Hottenrott and Martin Grunow. Flexible layouts for the
mixed-model assembly of heterogeneous vehicles. OR Spectrum,
41(4):943–979, 2019.
[JAPS05] J Jerald, Prabaharan Asokan, G Prabaharan, and R Saravanan.
Scheduling optimisation of flexible manufacturing systems using par-
ticle swarm optimisation algorithm. The international journal of ad-
vanced manufacturing technology, 25:964–971, 2005.
[JW03] Mingzhou Jin and S David Wu. A new heuristic method for mixed
model assembly line balancing problem. Computers & Industrial En-
gineering, 44(1):159–169, 2003.
[KHBB20] A. Khezri, H. Haddou Benderbal, and Lyes Benyoucef. Sustainable
Multi-objective Process Plan Generation in RMS Through Modelling
Energy Consumption. In Lyes Benyoucef, editor, Reconfigurable Man-
ufacturing Systems: From Design to Implementation, pages 161–177.
Springer International Publishing, Cham, 2020. Series Title: Springer
Series in Advanced Manufacturing.
[KM13] Naveen Kumar and Dalgobind Mahto. Assembly line balancing: a re-
view of developments and trends in approach to industrial application.
Global Journal of Researches in Engineering Industrial Engineering,
13(2):29–50, 2013.
[LGZ09] Jianzhi Li, Miguel González, and Yun Zhu. A hybrid simulation opti-
mization method for production planning of dedicated remanufactur-
ing. International Journal of Production Economics, 117(2):286–301,
2009.

59
[Per18] Jordi Pereira. Modelling and solving a cost-oriented resource-
constrained multi-model assembly line balancing problem. Interna-
tional Journal of Production Research, 56(11):3994–4016, 2018.

[RE15] Reza Ramezanian and Abdullah Ezzatpanah. Modeling and solving

multi-objective mixed-model assembly line balancing and worker as-
signment problem. Computers & Industrial Engineering, 87:74–80,
2015.

[RG17] Abdolreza Roshani and Davide Giglio. Simulated annealing algo-

rithms for the multi-manned assembly line balancing problem: min-
imising cycle time. International Journal of Production Research,
55(10):2731–2751, 2017.

[SHGT15] Malte Schönemann, Christoph Herrmann, Peter Greschke, and Sebas-

tian Thiede. Simulation of matrix-structured manufacturing systems.
Journal of Manufacturing Systems, 37:104–112, 2015.

[SV04] Ana Sofia Simaria and Pedro M Vilarinho. A genetic algorithm based
approach to the mixed-model assembly line balancing problem of type
ii. Computers & Industrial Engineering, 47(4):391–407, 2004.

[SWS21] Patrick Schumacher, Christian Weckenborg, and Thomas S Spengler.

Economic design of matrix-structured manufacturing systems. In Ad-
vances in Production Management Systems. Artificial Intelligence for
Sustainable and Resilient Production Systems: IFIP WG 5.7 Inter-
national Conference, APMS 2021, Nantes, France, September 5–9,
2021, Proceedings, Part II, pages 516–524. Springer, 2021.

[SWS22] Patrick Schumacher, Christian Weckenborg, and Thomas S Spengler.

The impact of operation, equipment, and material handling flexibility
on the design of matrix-structured manufacturing systems. IFAC-
PapersOnLine, 55(2):481–486, 2022.

[TLZ+ 16] Qiuhua Tang, Yanli Liang, Liping Zhang, Christodoulos A Floudas,
and Xiaojun Cao. Balancing mixed-model assembly lines with
sequence-dependent tasks via hybrid genetic algorithm. Journal of
Global Optimization, 65(1):83–107, 2016.

[WBH12] Ying Wang, Arne Bilberg, and Ronen Hadar. Implementation of re-
configurable manufacturing systems, the case of the lego group. In
4th World Conference on Production & Operations Management/19th
International Annual EurOMA Conference Amsterdam, Amsterdam,
Netherlands,(Submitted manuscript), pages 1–9, 2012.

[XW05] Weijun Xia and Zhiming Wu. An effective hybrid optimization

approach for multi-objective flexible job-shop scheduling problems.
Computers & industrial engineering, 48(2):409–425, 2005.

[YCGBD21] Abdelkrim R Yelles-Chaouche, Evgeny Gurevsky, Nadjib Brahimi,

and Alexandre Dolgui. Reconfigurable manufacturing systems from an

60
 optimisation perspective: a focused review of literature. International
Journal of Production Research, 59(21):6400–6418, 2021.

[YYO20] Seyda Topaloglu Yildiz, Gokalp Yildiz, and Raziye Okyay. Assembly
line balancing problem with hierarchical worker assignment, positional
constraints, task assignment restrictions and parallel workstations.
International Journal of Industrial Engineering, 27(3), 2020.

61
Webography
[1] All results of the MILP and 2-phased methods for 20 minutes and 5 minutes :
https://drive.google.com/results
Appendix

4.4.1 Bowman

Method OV TTV OB 1 OB 2 GAP

00 4 346165 34.23
10 4 346139 14.67
00
25 4 346107 24.81
50 4 343395 24
00 4 346161 32,49
10 4 346135 22,54
10
25 4 346105 33,65
50 4 343395 33,86
MILP
00 4 341609 31,24
10 4 341531 25,87
25
25 4 341401 29,32
50 4 341197 17,66
00 4 335789 5,23
10 4 335789 4,05
50
25 4 335789 15,5
50 4 335789 9,18
00 4 375789 39,42
10 4 375789 21,41
00
25 4 375789 30,76
50 4 375789 30,76
00 4 375789 38,24
10 4 375789 28,66
2-Phased 10
25 4 375789 38,89
50 4 375789 39,57
00 4 375789 37,50
10 4 375789 32,77
25
25 4 375789 35,80
50 4 375789 25,24
00 4 456251 30,25
10 4 449253 28,28
50
25 4 375789 24,48
50 4 375789 18,84

Table 4.19: MILP and 2-Phased Results within 20 minutes : Bowman

63
4.4.2 EX C=18

Method OV TTV OB 1 OB 2 GAP

00 2 176665 0
10 2 176665 0
00
25 2 176665 0
50 2 176665 0
00 2 176665 0
10 2 176665 0
10
25 2 176665 0
50 2 176665 0
MILP
00 2 176665 0
10 2 176665 0
25
25 2 176665 0
50 2 176665 0
00 2 176665 0
10 2 176665 0
50
25 2 176665 0
50 2 176665 0
00 2 209995 15,87
10 2 216665 18,46
00
25 2 203275 13,09
50 2 216665 18,46
00 2 216665 18,46
10 2 176665 00
2-Phased 10
25 2 216665 18,46
50 2 209995 15,87
00 2 216665 18,46
10 2 192881 8,41
25
25 2 176665 00
50 2 216665 18,46
00 2 176665 00
10 2 216665 18,46
50
25 2 176665 00
50 2 216665 18,46

Table 4.20: MILP and 2-Phased Results within 20 minutes : EX C=18

4.4.3 EX C=20

Method OV TTV OB 1 OB 2 GAP

00 2 163000 0
10 2 163000 0
00
25 2 163000 0
50 2 163000 0
00 2 163000 0
10 2 163000 0
10
25 2 163000 0
50 2 163000 0
MILP
00 2 163000 0
10 2 163000 0
25
25 2 163000 0
50 2 163000 0
00 2 163000 0
10 2 163000 0
50
25 2 163000 0
50 2 163000 0
00 2 203000 19,70
10 2 203000 19,70
00
25 2 203000 19,70
50 2 203952 20,08
00 2 203000 19,70
10 2 203000 19,70
2-Phased 10
25 2 165280 1,38
50 2 203000 19,70
00 2 167350 2,60
10 2 163000 00
25
25 2 203000 19,70
50 2 203000 19,70
00 2 203000 19,70
10 2 166166 1,91
50
25 2 203000 19,70
50 2 164566 0,95

Table 4.21: MILP and 2-Phased Results within 20 minutes : EX C=20

4.4.4 Jackson

Method OV TTV OB 1 OB 2 GAP

00 6 7364830 33,33
10 6 505960 43,98
00
25 6 488364 42,51
50 6 471438 39,63
00 6 7347684 42,9
10 6 455594 39,95
10
25 6 462500 40,23
50 6 551868 49,33
MILP
00 6 450586 42,04
10 6 424548 40,24
25
25 6 455606 43,05
50 6 450076 42,03
00 6 420178 41,85
10 6 454158 45,90
50
25 6 416016 42,09
50 6 456418 46,76
00 6 686892 100
10 6 606266 53,25
00
25 6 615104 54,36
50 6 606568 53,08
00 6 610736 100
10 6 605605 54,82
2-Phased 10
25 6 606080 54,39
50 6 5656688 50,57
00 6 544328 52,03
10 6 565812 55,16
25
25 6 567548 54,29
50 6 565430 53,86
00 6 565934 56,83
10 6 565528 56,56
50
25 6 565732 57,42
50 6 565430 57,02

Table 4.22: MILP and 2-Phased Results within 20 minutes : Jackson

4.4.5 Jaeschke

Method OV TTV OB 1 OB 2 GAP

00 6 7018433 33,33
10 6 431425 30,77
00
25 6 391425 26,17
50 6 391425 20,85
00 6 391425 26,83
10 6 427123 25,09
10
25 6 391425 19,67
50 6 425321 24,82
MILP
00 6 383673 26,97
10 6 383673 23,32
25
25 6 391425 26,80
50 6 391425 23,49
00 6 388777 7.56
10 6 388777 6.21
50
25 6 388777 5,60
50 6 396367 28.26
00 6 706735 100
10 6 571523 47,75
00
25 6 573413 49,61
50 6 569559 45,61
00 6 610935 53,12
10 6 610327 47,58
2-Phased 10
25 6 529567 40,63
50 6 609149 47,51
00 6 573083 51,11
10 6 573167 48,67
25
25 6 570823 49,81
50 6 609155 50,84
00 6 510629 41,15
10 6 572043 36,34
50
25 6 569711 35,94
50 6 610467 53,42

Table 4.23: MILP and 2-Phased Results within 20 minutes : Jaeschke

4.4.6 Mansoor

Method OV TTV OB 1 OB 2 GAP

00 4 326381 26,77
10 4 326381 26,80
00
25 4 326381 27
50 4 326381 26,40
00 4 326381 30,40
10 4 326381 30,25
10
25 4 326381 30,48
50 4 326381 30,27
MILP
00 4 326381 34,29
10 4 326381 34,12
25
25 4 326381 34,02
50 4 326381 34,06
00 4 326381 48.28
10 4 326381 38.20
50
25 4 326381 49.10
50 4 326381 47.10
00 4 486381 50,86
10 4 446381 46,48
00
25 4 486381 50,99
50 4 446381 46,19
00 4 446381 49,11
10 4 446381 49
2-Phased 10
25 4 446381 49,17
50 4 486381 53,21
00 4 366381 41,47
10 4 446381 51,83
25
25 4 446381 51,76
50 4 446381 51,79
00 4 446381 62,19
10 4 486381 58,54
50
25 4 486381 65,85
50 4 446381 61,30

Table 4.24: MILP and 2-Phased Results within 20 minutes : Mansoor

4.4.7 Mertens

Method OV TTV OB 1 OB 2 GAP

00 4 317500 3,88
10 4 317500 2,93
00
25 4 317500 2,98
50 4 317500 2,39
00 4 317500 4,61
10 4 317500 4,90
10
25 4 317500 1,77
50 4 317500 2,69
MILP
00 4 317500 4,10
10 4 317500 3,70
25
25 4 317500 3,62
50 4 317500 3,78
00 4 317500 9.05
10 4 317500 4.61
50
25 4 317500 2.70
50 4 317500 2.89
00 4 437500 27,43
10 4 397500 20,13
00
25 4 397500 20,13
50 4 437124 27,37
00 4 397500 20,13
10 4 437500 27,43
2-Phased 10
25 4 437500 27,43
50 4 397500 20,13
00 4 437500 27,43
10 4 397500 20,13
25
25 4 357500 11,19
50 4 437500 27,43
00 4 437500 27,43
10 4 357500 11,19
50
25 4 357500 11,19
50 4 397500 20,13

Table 4.25: MILP and 2-Phased Results within 20 minutes : Mertens

4.4.8 Mitchell

Method OV TTV OB 1 OB 2 GAP

00 8 7639386 62,50
10 8 7643066 37,50
00
25 8 7622974 62,50
50 8 7624702 25
00 8 7625222 62,50
10 8 7625626 25
10
25 8 7632854 62,50
50 8 1192536 73,10
MILP
00 8 7626212 37,50
10 8 699408 56,61
25
25 8 819136 62,91
50 8 694334 57,45
00 8 708588 63,05
10 8 701094 61,76
50
25 8 656916 59,59
50 8 629732 57,98
00 8 911648 100
10 8 1033862 100
00
25 8 1036854 100
50 8 807140 100
00 8 954686 100
10 8 839586 100
2-Phased 10
25 8 875820 100
50 8 838280 61,73
00 8 918578 100
10 8 912440 66,74
25
25 8 874378 65,26
50 8 835206 64,63
00 8 871260 69,95
10 8 910476 70,56
50
25 8 872922 69,59
50 8 870650 69,61

Table 4.26: MILP and 2-Phased Results within 20 minutes : Mitchell

Abstract
This thesis investigates the Flexible Layout Design Problem (FLDP), which involves
a set of stations arranged in a matrix-structured manufacturing system (MMS). Ini-
tially, the main objectives are to minimize the number of stations used and to
minimize the flow of transport between stations. Subsequently, we implement a
Mixed-Integer Linear Programming (MILP) model which integrate resource alloca-
tions into the FLDP and propose a matheuristic method to address the problem.
We conduct tests using a chosen dataset for 20 minutes and then for 5 minutes.
Finally, we discuss the results obtained and conclude with perspectives for future
research.
Key words: Reconfigurable Manufacturing Systems, Matrix-structured manufac-
turing system, Assembly line balancing, Flexible layout design

Résumé
Ce projet de fin d’études étudie le Problème de Conception de Layout Flexible
(FLDP), qui implique un ensemble de stations disposées dans un système de fab-
rication à structure matricielle (MMS). Initialement, les principaux objectifs sont
de minimiser le nombre de stations utilisées et de réduire le flux de transport entre
les stations. Par la suite, nous mettons en œuvre un modèle de Programmation
Linéaire Mixte en Nombre Entier (MILP) qui intègre les allocations de ressources
dans le FLDP, et proposons une méthode matheuristique pour résoudre le problème.
Nous réalisons des tests sur un jeu de données choisi pendant 20 minutes puis pour
5 minutes. Enfin, nous discutons des résultats obtenus et concluons en évoquant les
perspectives pour les recherches futures.
Mots clés: Systèmes de Fabrication Reconfigurables, Systèmes de Fabrication à
Structure Matricielle, Équilibrage de Lignes d’Assemblage, Conception de Layout
Flexible

‘jÊÓ
áҒJK úæË@  , (FLDP) éKQÖÏ @ É¿AJêË@ ÕæҒ éʾ  ‚Ó  Èñk h. QjJË@ ¨ðQå„Ó ¨ñ“ñÓ PðYK
¬@Yë B@ , éK @YJ.Ë@ ú¯ . (MMS) é¯ñ®’Ó ɾJë ð X ©J’ ÐA¢ ú¯ éÒ  ¢JÖÏ @ HA¢jÖ  ×
Ï @ áÓ é«ñÒm .

Ðñ®K ,½Ë X YªK. . HA¢jÖ      
Ï @ áK. É®JË@ ‡¯YK ÉJÊ®Kð éÓYjJ‚ÖÏ @ HA¢jÖ    
Ï @ XY« ÉJÊ®K ùë éJ‚KQË@ 
ú¯ XP@ñÖÏ @ HA’J  ’m' l×. YK ø YË@ (MILP) iJj’Ë@ XYªË@ð é¢Ê  JjÖÏ @ éJ ¢mÌ '@ ém ×QÊË h XñÖß YJ®JK
. . . .
       
20 èYÖÏ èPAJm× HA  KAJK. é«ñÒm.× úΫ H@  P AJ.Jk@ ø Qm.' . éʾ‚Ö Ï @ ÉmÌ AJ“AK P AK. ñʃ @ h Q®Kð , (FLDP)

éJ ÊJ.® J‚ÖÏ @ IjJ
 . Ë@ †A ¯ @ é‚ ¯A JÖß. ÕæJm' ð éʒjÖ
 Ï @ l' AJJË@  ¯A JK , ÐAJmÌ '@ ú¯ . ‡ KA¯X
.
 5 èYÖÏ Õç' é®J ¯X

,ú¯ñ®’ÖÏ @ ɾJêË@ ð X ©J’JË@ ÐA¢ , áK ñºJË@ èXA«B éÊK  A¯ ©J’JË@ éÒ
.
 ¢ @ : éJ kAJ®ÖÏ @ HAÒʾË@

éKQÖÏ @ É¿ AJêË@ ÕæҒ , h. AJKB @ ¡k à P@ñK