UNIVERSITY OF SOUTHAMPTON

FACULTY OF BUSINESS, LAW AND ART

Southampton Business School

Cutting Patterns for Efficient Production of Irregular-shaped Pieces

by

Ranga Prasad Abeysooriya

Thesis for the degree of Doctor of Philosophy

August 2017
 UNIVERSITY OF SOUTHAMPTON

ABSTRACT

FACULTY OF BUSINESS, LAW AND ART

Southampton Business School

Doctor of Philosophy in Operational Research

CUTTING PATTERNS FOR EFFICIENT PRODUCTION OF

IRREGULAR-SHAPED PIECES

by Ranga Prasad Abeysooriya

The research presented in this thesis belongs to the subject area of operations research.
The study investigates and utilises the solution methodologies known as heuristics and
local search for three practical problems related to cutting and packing using irregular
shapes and multiple bins. From an application point of view, the problems domains
remain in manufacturing, specifically where minimising the resources is required to meet
a particular outcome. Many manufacturing processes begin with cutting desired items
from a stock sheet of material, hence this study focuses on generating efficient cutting
patterns, which is applicable in the manufacture of furniture, shoes, tools, ships, and
garments.

First, we consider designing an efficient solution procedure for solving two-dimensional

irregular shape single bin size bin packing problem and two-dimensional irregular shape
multiple bin size bin packing problem. Our intention is to consider alternative strategies
such as placement policies, hole-filling and handling rotation of pieces; particularly with
unrestricted rotations. Despite the fact that both problems are widely applicable in
sheet cutting, their consideration in the literature is limited. To our knowledge, only
a few authors have attempted to incorporate the first problem with the unrestricted
rotation of pieces while the second problem with unrestricted rotation has not been
considered at all. Being applicable in the real world, both the problems require powerful
algorithms to determine the arrangement of irregular pieces on stock sheets in order
to minimise the material waste. In this thesis, our focus is on developing algorithms
to solve each problem efficiently. These algorithms draw on concepts in computational
geometry, computer science as well as operations research.

We investigate a set of newly proposed single-pass constructive algorithms that builds a

feasible solution by adding pieces sequentially to a packing area defined by a set of bins
which can either be homogeneous or heterogeneous. Each problem has a large solution
space due to the different combinations of bins and arrangements of irregular pieces.
We adopt the optimisation power of local search methods and metaheuristics to find
iv

good solutions. As one of the useful heuristic procedures, we use the Jostle heuristic
(JS) to solve irregular shape single bin size bin packing problem and irregular shape
multiple bin size bin packing problem due to its promising performances in handling
both allocation and placement decisions of the pieces together. The Jostle was used in
earlier studies for solving irregular strip packing problems and in this study we adopt
it first time to solve irregular bin packing problems. Also, our implementation of Jostle
handles identifying promising orientation angles of the pieces using the newly proposed
angled tuning mechanism when placing pieces. Experimental results reveal that the
proposed algorithms can manage different variants of the problem and find solutions
with good utilisation of material.

For the third problem of this study, we consider multi-period irregular bin packing prob-
lem with use of residuals. This allows using leftovers of a certain period which are usable
as input stock material for the next periods. Here, we expand the previous work on ir-
regular bin packing algorithms for heterogeneous stock sheets to consider the inventory
and production process of sheet cutting. We propose two models to test the impact of
a variety of operational policies around the retention and reuse of residual materials in
the sheet cutting process. It also examines the cost sensitivity of using residuals with
respect to nine practical scenarios within those operational policies. The computational
results demonstrate that the proposed multi-period approach with residuals derives bet-
ter results than solving each order individually for a selected set of operational scenarios
and disclose which policy would be more advantageous to operate in each scenario. The
results facilitate developing a tool to guide the manufacturers to take effective decisions
based on the scenarios applicable to their sheet cutting process.

Key words: Irregular shape bin packing, Jostle, Usable leftovers

Contents

Declaration of Authorship xiii

Acknowledgements xv

Abbreviation xvii

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Focused Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Aim and Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Research Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Literature Review 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Overview of Cutting and Packing Problems . . . . . . . . . . . . . . . . . 10
2.3 Overview of 2D Irregular Bin Packing Problems/Cutting Stock problems . 14
2.4 Overview of Problems considering Usable Leftovers . . . . . . . . . . . . . 16
2.5 Computational Method for Irregular shape Strip Packing Problems . . . . 22
2.5.1 Constructive heuristics for irregular shape packing problems . . . . 22
2.5.1.1 Placement rules . . . . . . . . . . . . . . . . . . . . . . . 22
2.5.1.2 Placement sequences: . . . . . . . . . . . . . . . . . . . . 26
2.5.2 Work with complete solutions . . . . . . . . . . . . . . . . . . . . . 27
2.5.2.1 Searching over with feasible solution . . . . . . . . . . . . 28
2.5.2.2 Search over the layouts allowing infeasible solutions . . . 29
2.5.3 Exact methods for irregular shape packing problems . . . . . . . . 31
2.6 Computational Method for Irregular shape Bin Packing Problems . . . . . 32
2.6.1 Constructive methods . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.2 Search methods for irregular bin packing problems . . . . . . . . . 40
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Methodology 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Computational Geometric Techniques . . . . . . . . . . . . . . . . . . . . 47
3.2.1 Representation of pieces . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.2 Overlap calculation and finding feasible placement positions . . . . 49
3.3 Heuristics-based Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.1 Constructive heuristics . . . . . . . . . . . . . . . . . . . . . . . . . 55

v
vi CONTENTS

3.3.2 Local search methods . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.3.3 Metaheuristic Algorithms . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.3.1 Simulated Annealing (SA) . . . . . . . . . . . . . . . . . 58
3.3.3.2 Tabu Search (TS) . . . . . . . . . . . . . . . . . . . . . . 58
3.3.3.3 Iterated Local Search (ILS) . . . . . . . . . . . . . . . . . 59
3.3.3.4 Variable Neighbourhood Search (VNS) . . . . . . . . . . 60
3.3.3.5 Genetic Algorithm (GA) . . . . . . . . . . . . . . . . . . 60
3.3.3.6 Greedy Randomised Adaptive Search Procedure (GRASP) 62
3.3.3.7 Parameter tuning . . . . . . . . . . . . . . . . . . . . . . 63
3.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.1 Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4 Homogeneous Bin Packing Problems with Irregular Shapes 69

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Packing Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4.1 Construction algorithm . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.1.1 Generating candidate placement positions . . . . . . . . . 74
4.4.1.2 Placement orientation of a piece . . . . . . . . . . . . . . 75
4.4.2 Placement policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5 Solution Improvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.5.1 Jostle heuristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.5.2 Iterated Jostle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
1-piece insert move: . . . . . . . . . . . . . . . . . . . . . . . 82
Bin-swapping move: . . . . . . . . . . . . . . . . . . . . . . 83
Iterated Jostle with join and release of pieces . . . . . . . . 83
4.6 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.6.1 Data instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.6.2 Performance of different solution strategies . . . . . . . . . . . . . 86
4.7 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.7.1 Test 01: Jostle search vs. Random order of pieces . . . . . . . . . 87
4.7.2 Test 02: IJS vs JS . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.7.3 Test 03: Comparing with benchmark results in literature . . . . . 90
4.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5 Irregular shape Bin Packing in Heterogeneous Bins 97

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3.1 Evaluation function of a solution . . . . . . . . . . . . . . . . . . . 100
5.4 Packing Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.4.1 Constructive Algorithm 1 (CA1) . . . . . . . . . . . . . . . . . . . 101
5.4.2 Solution improvement by Jostle heuristic . . . . . . . . . . . . . . 102
5.4.3 Hybrid approach of GA and Jostle . . . . . . . . . . . . . . . . . . 103
5.4.3.1 Outline of HGAJ . . . . . . . . . . . . . . . . . . . . . . . 104
5.4.3.2 Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
CONTENTS vii

5.4.3.3 Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.4.3.4 Implementation of Jostle within the GA framework . . . 106
5.4.3.5 Selection for the next population . . . . . . . . . . . . . . 107
5.4.3.6 Addressing some issues of GA Jostle approach . . . . . . 109
5.4.4 Iterated Jostle with Random Assignment of Bins (IJRAB) . . . . . 109
5.4.5 Sequential Packing with a Bin Centric Heuristic (SPBCH) . . . . . 111
5.4.5.1 Constructive Algorithm 2 (CA2) . . . . . . . . . . . . . . 111
5.4.5.2 Solution improvement . . . . . . . . . . . . . . . . . . . . 113
5.5 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.5.1 Data instances and implementation . . . . . . . . . . . . . . . . . . 114
5.5.2 Performance of two constructive procedures (CA1 vs. CA2) . . . . 115
5.5.2.1 Best performing GA Jostle approach: . . . . . . . . . . . 116
5.5.2.2 Performances of SPBCH approaches: . . . . . . . . . . . 117
5.5.2.3 Best performing IJRAB approach: . . . . . . . . . . . . . 118
5.5.3 Comparison among three computational methods . . . . . . . . . . 120
5.5.4 Performance comparison: 2D IMBSBPP vs. 2D ISBSBPP . . . . . 121
5.5.5 Best solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

6 Use and Reuse of Materials in Cutting Industries 129

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1.1 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2 Overview of Using and Reusing of Material in Sheet Cutting Process . . . 131
6.2.1 Identifying operational policies: . . . . . . . . . . . . . . . . . . . . 133
6.3 Model for Processing Successive Cut Orders . . . . . . . . . . . . . . . . . 133
6.3.1 Finding candidate usable leftovers . . . . . . . . . . . . . . . . . . 134
6.3.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.3.3 Problem to be solved at each order: . . . . . . . . . . . . . . . . . 137
6.3.4 Modifications proposed for IJRAB to solve 2D IRBPP . . . . . . . 138
6.4 Model 01: Minimizing the Trim Loss . . . . . . . . . . . . . . . . . . . . . 139
6.4.1 Computational experiments . . . . . . . . . . . . . . . . . . . . . . 139
6.4.1.1 Simulation framework . . . . . . . . . . . . . . . . . . . . 140
6.4.1.2 Multi-period run of orders: Period wise investigation . . 141
6.4.1.3 Assessing the results . . . . . . . . . . . . . . . . . . . . . 142
6.4.1.4 Multi-period run of orders: Investigation for whole time
horizon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.4.1.5 Assessing the results for the planned time horizon . . . . 144
6.4.2 Findings and discussion . . . . . . . . . . . . . . . . . . . . . . . . 147
6.5 Model 02: Minimizing the Cost . . . . . . . . . . . . . . . . . . . . . . . . 148
6.5.1 Computational experiments . . . . . . . . . . . . . . . . . . . . . . 150
6.5.2 Findings and discussion . . . . . . . . . . . . . . . . . . . . . . . . 158
6.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

7 Conclusion and Future Work 163

7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
viii CONTENTS

A Period wise Results for Model 01 169

References 173
List of Figures

1.1 Research flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Seven intermediate problem types under output maximization - extracted

from Wäscher et al. (2007) . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Seven intermediate problem types under input minimization - extracted
from Wäscher et al. (2007) . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Ways of generating usable leftovers by Andrade et al. (2014) . . . . . . . . 20
2.4 Placement rules (extracted from Oliveira et al. (2000)) . . . . . . . . . . . 23
2.5 Bottom left and Bottom left with Hole-filling (Dowsland et al., 2002) . . . 24
2.6 Feasible placement region for the next inserting piece (image is from Ben-
nell and Oliveira (2009)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Pieces merge in TOPOS and missing the enclosed gaps (extracted from
Bennell and Song (2010)) . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Piece moves (extracted from Bennell and Oliveira (2009)) . . . . . . . . . 28
2.9 One step of Jostle Algorithm (extracted from Bennell and Oliveira (2009)) 29
2.10 Constructive approach (Terashima-Marı́n et al., 2010) . . . . . . . . . . . 35
2.11 Slide along the edges (extracted from Han et al. (2013)) . . . . . . . . . . 36
2.12 Search forest (extracted from Han et al. (2013)) . . . . . . . . . . . . . . . 37
2.13 Containment Polygon (extracted from Martinez-Sykora et al. (2015)) . . . 41

3.1 Discrete representation methods (extracted from Bennell and Oliveira

(2008)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 NFP (A,B): The reference point on B traces around A (Bennell et al.,
2001)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Different cases of generating NFPs (Bennell and Oliveira, 2008): a) B fits
inside the concavity space of A, b) B can slide into the concavity of A in
one direction, c) B fits only in one point inside the concavity of B . . . . . 51
3.4 Inner-fit Polygon (extract from Bennell and Song (2010)) . . . . . . . . . 51
3.5 Shape representation of a piece . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6 Iterative Improvement Local search (Decent) behavior (extracted from
Talbi (2009)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7 Working principle of iterated local search (extracted from Talbi (2009)) . 60
3.8 Variable neighbourhood search in two neighbourhood (extracted from
Talbi (2009)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.1 Strip to Bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.2 Partial solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 Hole-filling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4 Feasible placement for a new piece . . . . . . . . . . . . . . . . . . . . . . 76

ix
x LIST OF FIGURES

4.5 Tuning the orientation angle . . . . . . . . . . . . . . . . . . . . . . . . . . 77

4.6 Angle tuning at the edges of bins . . . . . . . . . . . . . . . . . . . . . . . 78
4.7 BL and ML placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.8 Placement of piece on different vertices along the feasible placement segment 80
4.9 Possibilities when pieces are paired by the Join and Release mechanism . 85
4.10 JS and IJS1 iterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.1 Strip to Bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5.2 Representation of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.3 Crossover operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6.1 Defining usable leftovers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

6.2 Variation of Ut with periods . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.3 Model 01: Use of leftovers with different Jmax values . . . . . . . . . . . . 146
6.4 Model 01: Use of leftovers with different Jmax values . . . . . . . . . . . . 147
6.5 Cost Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.6 Model 02: Total cost graphs by running simulation for 10 periods for each
policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.7 Model 02: Total cost graphs by running simulation for 30 periods for each
policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.8 Supportive tool for policy selection . . . . . . . . . . . . . . . . . . . . . . 159
List of Tables

2.1 Base characteristics of Dyckhoff’s typology Dyckhoff (1990) . . . . . . . . 12

2.2 New characteristics proposed by Wäscher (2012) . . . . . . . . . . . . . . 13
2.3 Solution methods attempted to solve C&P problems with usable leftovers 21

3.1 Nesting Instances (extracted from Martinez-Sykora et al. (2017)) . . . . . 66

3.2 Characteristics of the JP1 instances (extracted from Lopez-Camacho et al.
(2013b)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.3 Characteristics of the JP2 instances (extracted from Lopez-Camacho et al.
(2013b)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.1 Notation of Algorithms: placement rule and rotation . . . . . . . . . . . . 86

4.2 Notation of Algorithms: algorithm and kick type . . . . . . . . . . . . . . 86
4.3 Performance of construction algorithm for different rotation criteria . . . 88
4.4 Performance results JS, IJS1, IJS2 and IJS3 for nesting data instances . . 90
4.5 Performance results of IJS1, IJS2 and IJS3 vs JS for JP1 and JP2 instances 90
4.6 Results of the IJS1, IJS2 and IJS3 algorithm applied for JP1 instances . . 92
4.7 Results of the IJS1, IJS2 and IJS3 algorithms applied for JP2 instances . 93
4.8 Results of the IJS1 algorithm for each nesting instances with 2md Large
bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.9 Results of the IJS1 algorithm for each nesting instances with 1.5md Medium
bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.10 Results of the IJS1 algorithm for each nesting instances with 1.1md small
bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5.1 Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.2 Bin Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.3 Performance of construction algorithms . . . . . . . . . . . . . . . . . . . 115
5.4 Different parameter settings to tune GA, HGA v1 and HGA v2 . . . . . . 117
5.5 Performance of GA and HGA approaches for 800 seconds . . . . . . . . . 117
5.6 Performance of LS CA2 and ILS CA2 approaches for 800 seconds . . . . . 118
5.7 Performance of the IJRAB algorithms for different bin configurations
(with Restricted Rotation of pieces) . . . . . . . . . . . . . . . . . . . . . 120
5.8 Performance of the IJRAB algorithms for different bin configurations
(with Unrestricted Rotation of pieces) . . . . . . . . . . . . . . . . . . . . 120
5.9 Performance comparison of IJRAB, HGA and SPBCH methods . . . . . . 120
5.10 Performance comparison: 2D IMBSBPP vs.2D ISBSBPP (with restric-
tions of piece rotation) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.11 Performance comparison: 2D IMBSBPP vs.2D ISBSBPP (No restriction
of piece rotation) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

xi
xii LIST OF TABLES

5.12 Best solutions received in 10 trials by running algorithm for 800 seconds
-(with Mix set of bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.13 Best solutions received in 10 trials by running algorithm for 800 seconds
-(with LB set of bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

6.1 Material Saving Scheme: Use of leftovers with different Jmax values for
10 periods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.2 Material Saving Scheme: Use of leftovers with different Jmax values for
30 periods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.3 Characteristics of scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.4 Model 02: NonUL-Mix vs. UL-Mix results for with different α and h
values (simulation results for 10 periods) . . . . . . . . . . . . . . . . . . . 151
6.5 Model 02: NonUL-Single L vs. UL-Single L results for with different α
and h values (simulation results for 10 periods) . . . . . . . . . . . . . . . 152
6.6 Model 02: NonUL-Mix vs. UL-Mix results for with different α and h
values (simulation results for 30 periods) . . . . . . . . . . . . . . . . . . 152
6.7 Model 02: NonUL-Single L vs. UL-Single L results for with different α
and h values (simulation results for 30 periods) . . . . . . . . . . . . . . . 153

A.1 Multi period cutting with residual use (starting with Single standard
2.25m bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
A.2 Multi period cutting with residual use (starting with Single standard
2.00m bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
A.3 Multi period cutting with residual use (starting with Single standard
1.75m bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
A.4 Multi period cutting with residual use (starting with Single standard
1.50m bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
A.5 Multi period cutting with residual use (starting with Single standard
1.25m bins) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
A.6 Multi period cutting with residual use (starting with Mix of standard sizes)172
 Declaration of Authorship

I, Ranga Prasad Abeysooriya , declare that the thesis entitled Cutting Patterns for
Efficient Production of Irregular-shaped Pieces and the work presented in the thesis are
both my own, and have been generated by me as the result of my own original research.
I confirm that:

• this work was done wholly or mainly while in candidature for a research degree at
this University;

• where any part of this thesis has previously been submitted for a degree or any
other qualification at this University or any other institution, this has been clearly
stated;

• where I have consulted the published work of others, this is always clearly at-
tributed;

• where I have quoted from the work of others, the source is always given. With the
exception of such quotations, this thesis is entirely my own work;

• I have acknowledged all main sources of help;

• where the thesis is based on work done by myself jointly with others, I have made
clear exactly what was done by others and what I have contributed myself;

• none of this work has been published before submission

Signed:.........R. Abeysooriya.........................................................................................

Date:............19.07.2017.................................................................................................

xiii
 Acknowledgements

Undertaking this PhD has been a life-changing experience for me which would not have
been possible without the support and guidance from numerous people.

I would first like to express my heartfelt gratitude to my supervisors Professor Julia

Bennell and Dr Antonio Martinez-Sykora, for bearing all my concerns and guiding me
throughout the study. I have been lucky to have such enthusiastic supervisors who
have been a tremendous support for me in many ways throughout this massive learning
experience.

Next, my sincere gratitude goes to the Head of School of Business and Law of Univer-
sity of Southampton and the Director of the postgraduate research programme and all
the members of staff at the University of Southampton for their enormous support in
several aspects throughout this PhD journey. In particular, I would like to thank the
examiners of MPhil to PhD upgrade and anonymous reviewers of research articles, for
their invaluable reviews and suggestions on my work.

I am indebted to the Faculty of Business, Law and Arts of University of Southampton

for funding my PhD with the aid of Faculty scholarship. The faculty funds provided me
with the opportunity to attend conferences which provided me the opportunity to meet
my research community all over the world, including experts in my research area. Also,
I appreciate the financial support by the Association of European Operational Research
Societies (EURO) when I was attending some internationally recognised workshops in
operations research.

Completing this PhD would have been much more difficult without the support and
friendship provided by the other colleagues of the Business School and our research
group CORMSIS in Southampton. A very special gratitude goes to my friend Carlos
Lamas-Fernandez for his invaluable feedback and advises especially during the hard
times.

I am also grateful to all the staff of University of Moratuwa, Sri Lanka for allowing me
to work through my PhD by releasing me from all my duties and offering me a study
leave to complete it without any trouble.

Finally, but by no means least, I express my gratitude to my beloved wife Tolusha, for
her continued support, encouragement and patience. Throughout this journey, I was
so lucky to have invaluable blessings of my dearest family and in-laws who have always
have faith in me and motivating me to pursue my dreams.

Southampton, United Kingdom, May 2017

xv
Abbreviation

1D : One-dimensional
2D : Two-dimensional
2D IBPPUL : Two Dimensional Irregular shape Bin Packing Problem with Usable Leftovers
2D IMBSBPP : Two Dimensional Irregular shape Multiple Bin Size Bin Packing Problem
2D ISBSBPP : Two Dimensional Irregular shape Single Bin Size Bin Packing Problem
BFD : Best Fit Decreasing
BL : Bottom Left (placement rule)
BPP : Bin Packing Problem
C&P : Cutting and Packing
CA : Constructive Approach
CAA : Constructive Approach with minimum Area
CAD : Constructive Approach with maximum Adjacency
CC : Circle Covering
CSP : Cutting Stock Problem
DJD : Djang & Finch
FF : First Fit
FFD : First Fit Decreasing
FL : Filler
GA : Genetic Algorithm
GD : Greedy Ddecisions
GRASP : Greedy Randomised Adaptive Procedure
HGA : Hybrid Genetic Algorithm
HGAIJ : Hybrid approach of Genetic Algorithm and Iterated Jostling
HGAJ : Hybrid approach of Genetic Algorithm and Jostling
IBPP : Irregular Bin Packing Problem
IFP : Inner-Fit Polygon
IJRAB : Iterated Jostling with Random Assignment of Bins
IJS : Iterated Jostling
ILS : Iterated Local Search
IP : Integer Programming
JP : Jigsaw Puzzle
JS : Jostling

xvii
xviii ABBREVIATION

LB : Large Bins
MB : Medium Bins
MBSBPP : Multiple Bin Size Bin Packing Problem
MIP : Mixed Integer Programming
ML : Minimum Length (placement rule)
MSSCSP : Multiple Stock Size Cutting Stock Problem
MU : Maximum Utilisation (placement rule)
NF : Next Fit
NFD : Next Fit Decreasing
NFP : No-Fit Polygon
ODP : Open Dimension Problem
PBP : Partial Packing Problem
RBPP : Residual Bin Packing Problem
RCSP : Residual Cutting Stock Problem
RR : Restricted Rotation
SA : Simulated Annealing
SB : Small Bins
SBSBPP : Single Bin Size Bin Packing Problem
SPBCH : Sequential Packing with a Bin Centric Heuristic
SSSCSP : Single Stock Size Cutting Stock Problem
TS : Tabu Search
UL : Usable Leftovers
UR : Unrestricted Rotation
VNS : Variable Neighbourhood Search
Chapter 1

Introduction

Cutting and packing is a very common scenario in day-to-day life. For example, when
packing bags to go on a vacation or wrapping gifts for Christmas, we tend to find the
most efficient way of utilising the available space. Also, an efficient cutting and packing
is regarded to a significant degree when planning the layout of a magazine, arranging
articles on a web page, or tiling walls of an office space. Above all, manufacturing
industries such as garment, furniture, component and tools use cutting and packing to
make efficient use of material. Examples include; a metal casing of a personal computer
requires assembly of several parts to cut from a sheet metal, a decorative sofa requires
different cut shapes from a set of cushion foam sheets, a garment requires to cut panels
from a roll of fabric and an assembly of a ship requires cut metal parts from sheets. In
each of these cases, manufacturers are either aiming at minimising the material waste or
maximising the value of cut pieces as much as possible under essential constraints such
as overlapping and containment of pieces.

Out of numerous application areas, this research is based on the applications arise in
business and manufacturing. The main focus of this study is on applications related to
sheet cutting industries where irregular shaped items are cut from sheets of material.
Therefore the targeted problems of the study are mainly laid under the category of
two-dimensional (2D) irregular shape problems with multiple stock sheets (or bins). In
general, the packing problems with irregular pieces is defined as arranging a given set
of irregular pieces inside a multiple number of bins so that the overall utilisation of bin
spaces is maximised. This problem can also be denoted as a waste minimization problem
considering the minimization of trim loss which occurs when irregular items are cut out
of the stock sheet material. Alternatively, some practical problems related to irregular
shape bin packing consider objectives such as reducing the number of bins, or reducing
the cost of bins.

1
2 Chapter 1 Introduction

1.1 Background

In general, the Cutting & Packing (C&P) is a category of optimisation problems that
is focused on finding a good arrangement of multiple items inside large containers. The
problem category of this thesis is irregular bin packing problems (IBPP) and they are
widely applicable in a diverse range of application fields. As a result, a number of re-
searchers have attempted to solve the problems related to this category even though
there is limited existing research. The literature review on Chapter 2 reviews those at-
tempts in detail. As a related problem, the irregular shape packing in a single strip; also
known as the nesting problem, is one of the mature areas of C&P research. While the
problem definition differs, there are many common elements in the solution approaches
available for irregular strip packing problems and irregular bin packing problems. Simi-
larly, some important features of regular bin packing problems are also applicable when
solving IBPP. As in general C&P problems, all the solution methods applicable for IBPP
should reach to a feasible solution and finding the best packing layout during a reason-
able computation time is the best to be achieved. However, in reality, this is still a
challenging task with irregular shapes due to the high computational time it takes to
find local optimal solutions. Moreover, finding the global optimal solution with irregular
shapes is still far away since the existing research is limited to work out the problem
with few pieces.

The irregularity of the pieces, range of orientations of the pieces and different placement
orders of the pieces cause IBPPs to have a higher number of solutions. For the problems
with such complexities, researchers use heuristic-based techniques combined with meta-
heuristics approaches since they provide good solutions at reasonable computational
cost.

We have two main motivations for studying IBPPs. First, 2D irregular shape sheet
cutting problems have numerous applications. The problem occurs in many industries as
metal, wood, paper and plastic, where the application of automatic packing algorithms
can yield considerable savings in bin space and number of bins used. Usually, the
cost of sheet material is significant in a manufacturing process and a saving in sheet
material would lead to significant reduction of cost. According to Guide (2000) reducing
waste, on the other hand, is much appreciated in terms of economic and environmental
implications. Therefore an efficient algorithm which facilitates a better utilisation of
material while reducing trim loss is definitely advantageous for manufacturers. Even
though there are good computational approaches to solve a similar packing problem for
placing irregular shapes into a single container or to a strip with fixed width and variable
length, computational methods developed for packing irregular shapes into the multiple
stock sheets are rare in the literature as well.

As an emerging concept in manufacturing, reuse of waste is the second motivation of the

study. In most scenarios, sheet cutting processes generate off-cuts which can be reused.
Chapter 1 Introduction 3

Having the policy to reuse off-cuts than discarding them can lead to a reduction of
material waste. However, due to the additional handling and storing it brings, the effort
may have a trade-off effect on the overall cost of production. Therefore, we are motivated
to investigate the economic impact of production decisions and policies for sheet cutting
processes. Such effort will include managerial perspectives when solving the related
operations research problems. This will provide useful insights for the managers to take
production decision effectively. To our knowledge, so far, no such tool is available related
to the irregular shapes sheet cutting process.

1.2 Focused Problems

This thesis investigates and provides solution methods for three problems in the category
of irregular shape bin packing problems.

In investigating each problem, it is necessary to define each problem clearly with their
objective function and constraints. As a common objective, all three problems cover
the objective of input minimization which is common to any irregular shape bin packing
problem. As a common set of constraints, all three problems cover the generic con-
straints of C&P problems, which are non-overlapping constraint among the pieces and
containment constraint ensuring that pieces are placed within the boundaries of the bin.

When defining problems, we use the term ”irregular pieces” or ”irregular shapes” to
describe a simple polygon which can either be convex or non-convex. The term ”bin”
is referred to a rectangular-shaped container with a given width and length. Therefore,
in a general sense, in all three problems, a given set of simple polygons is arranged in
multiple bins without overlapping each other to minimise an objective. As another two
variants, for each problem, we discuss the effect of the restricted rotation of pieces and
unrestricted rotation of pieces.

Problem 1: Two-dimensional (2D) irregular shape single size multiple bin packing
problem (2D ISBSBPP)

The main objective is to minimise the number of bins used in the solution. It is also
assumed that the input number of bins is a very large number.

Problem 2: Two-dimensional irregular shape multiple bin size bin packing problem
(2D IMBSBPP)

The main objective is to minimise the sum of bin costs. Out of a finite number of bin
types available (i.e. Heterogeneous rectangular bins), the aim is to find a solution with
the minimum sum of bin cost where the area of the bin is proportional to its cost. It
is also assumed that the input number of bins from each available type is a very large
number.
4 Chapter 1 Introduction

Problem 3: Two-dimensional irregular shape bin packing problem with residual reuse

With reference to this problem, we discuss two aspects. First, we attempt to solve
the problem with the aim of minimising the trim loss. A sequence of cut orders with
different quantities of pieces are considered as one set of inputs. The other inputs include
standard sheets purchased, which can be either one single standard size or multiple
standard sizes. The problem is a multi-period irregular shape bin packing problem, in
which a set of demand quantities in consecutive periods are considered. The problem
considers using leftovers of a certain period which are usable as input stock material for
the next periods. Corresponding to each time period, an irregular shape bin packing
problem is solved to satisfy the demand quantity of pieces required in that period. The
usable leftovers (ULs) gather from one period may be used in future periods and this
makes two forms of input bins; standard bins and ULs, to be considered when solving the
irregular bin packing problem at each period. We solve this problem, assuming that the
future demand quantity of pieces is unknown when the current order is being processed.
Also, the demand quantity of a certain period must be fully met within that period.

1.3 Aim and Scope

The purpose of this study is to provide a contribution to the C&P field by solving a set
of practical problems arise in sheet cutting industries. Our objective is to explore the
different forms of sheet cutting problems and develop computational methods for each
of them with a focus on cutting irregular shaped pieces.

Exploration of the literature reveals two categories of developing computational methods;

exact methods which find the globally optimal solution or heuristics and metaheuristic
approaches which do not guarantee to find the globally optimal solution but provide good
solutions within reasonable computation time. In current studies, the exact methods
are limited in solving irregular shapes packing problem only with a small number of
pieces (Toledo et al., 2013; Alvarez-Valdes et al., 2013; Cherri et al., 2016). Therefore,
we exclude exact approaches and use heuristics and metaheuristics based local search
methods to solve the aforementioned three problems. Our focus is on developing fast
heuristic methods to generate feasible solutions. Our conjecture is that each problem
may contain many local optima and it will require the involvement of search mechanisms
to avoid becoming trapped in local optima.

In our investigation, we discuss solutions methods for ISBSBPP, IMBSBPP and IBPP
with ULs and expand the investigation of irregular shape bin packing problems. The
optimisation routine followed in solving each of the problems relies on the use of geometry
as they have to primarily deal with placement of irregular pieces with no overlaps.
Accuracy of solutions is one of our primary concerns and we intend to use efficient and
Chapter 1 Introduction 5

accurate computational geometric techniques within the proposed solution generation

procedures.

The contribution of this study can be briefly outlined as follows.

The computational methods developed in this study are able to identify the promising
rotation of pieces when they are being placed in the bins rather than restricting the
piece rotation to a predefined finite set of orientation angles. Only a few papers related
to irregular shape packing approaches (e.g. Han et al. (2013), Rocha et al. (2013),
Stoyan et al. (2015a), Martinez-Sykora et al. (2017)) have considered this factor, yet
their procedures take a substantial amount of computational time which causes the
overall computation to be inefficient. This is a realistic requirement for many sheet
cutting industries such as sheet metal, foams, artificial leather, plastics which have
sheets with homogeneous surfaces. We design a new method to find such promising
rotations efficiently. In essence, this allows pieces to rotate within the layout to any
angle.

We design a new computational method to solve 2D ISBSBPP and 2D IMBSBPP fol-

lowing the Jostle procedure which was implemented by Dowsland et al. (1998) for an
irregular shape strip packing problem with the finite rotation of pieces and hole-filling.
Our implementation of Jostle in bin packing problems is new and includes additional
features such as unrestricted rotation of pieces which was not available in the previous
implementations of the Jostle algorithm. We introduce this method as a fast heuristic
for solving irregular shape bin packing problems.

Having algorithms for the 2D ISBSBPP and 2D IMBSBPP allow us to tackle another
variant of IBPP by considering usable leftovers. To our knowledge, no literature is
available in considering IBPP with ULs. While introducing a new method to solve
this problem, we examine the potential opportunities to implement reusing practices in
material waste arise in sheet cutting processes by considering ULs in the problem. The
existing literature has not discussed well a solution method for this problem where the
scope is limited only to 1D cutting stock problems and 2D rectangular item packing
problems using ULs. We further evaluate outcomes of such reuse practices and their
effect to the material costs, setup costs and handling costs of the sheet cutting process.
As one of the major contributions of this study, we present how a set of well-known
operational policies is applied within nine different operational scenarios, with regards
to saving material and direct costs incur in the sheet cutting processes.

1.4 Research Strategy

This thesis converts three types of real problems in sheet cutting industries to optimisa-
tion models and finds solution methods to solve those modelled optimisation problems.
6 Chapter 1 Introduction

In order to solve each problem, we design different algorithms and then evaluate and
validate those to test whether the solution methods can satisfy the objectives associ-
ated with each problem. We conduct evaluation and validation through computational
experiments based on the quantitative analysis of different features such as improve-
ment of bin utilisation, minimization of trim loss and minimization of cost. We design
computational experiments to assess the efficiency of the solution methods and draw con-
clusions to provide useful insights for the process of sheet cutting, by analysing different
geometric features in the derived cut layouts.

The initial phase of this study explores the current practices of C&P research. This
includes recognising the existing methods for different C&P problems. In this phase
we study and review what has been considered, to identify the research gaps. This fact
leads us to solve three different optimisation problems; 2D ISBSBPP, 2D IMBSBPP and
2D IBPP with ULs, by taking into account the degree of rotation of pieces, variety of
bin sizes and use of produced leftovers. For each problem, we follow a modelling phase,
a solution design & development phase and experiments & validation phase, to achieve
the aims of the study. In Figure 1.1, we demonstrate the key steps of each phase to
provide a general overview of the flow of this study. In the next chapters we describe
each step in detail, justifying the accepted methodology.
 Chapter 1 Introduction
7

Figure 1.1: Research flow

8 Chapter 1 Introduction

1.5 Outline

The thesis is arranged in seven chapters including this chapter. Chapter 2 reviews
literature related to the three problem areas investigated in this thesis. This includes
a review of different versions of irregular shape cutting and packing problems and their
solution methods.

Chapter 3 presents the methodologies followed that are common to all the problems. The
methodology related to formulating problems, developing models, designing algorithms
and implementations are discussed separately in the following chapters for each problem
addressed.

In Chapter 4, we present our solution method to solve 2D ISBSBPP with free rotation
of pieces and discuss the constructive and improvement heuristic approaches in solving
the problem.

Chapter 5 presents the details of three solution methods we propose to solve 2D IMBS-
BPP in which we deal with heterogeneous bins.

In Chapter 6, we provide an overview of the problem two-dimensional irregular shape

cutting in multiple periods, where we discuss how the residuals can be used to minimise
the overall usage of input material. Then we extend this discussion towards the economic
use of resources for a selected set of production and procurement policies, which applies
in several practical scenarios in sheet cutting industries.

Finally, Chapter 7 concludes the overall study with a discussion of the advantages and
limitations of the proposed computational methods.
Chapter 2

Literature Review

2.1 Introduction

This chapter presents the literature that relates to this thesis. The thesis focuses mainly
on solving three types of two dimensional cutting and packing problems, particularly
with irregular shapes packing into rectangular containers. The chapter presents the
sources of literature which make a significant contribution to an understanding of the
topic while interpreting the major issues, research gaps and the potential areas of future
research related to the irregular shape bin packing problems and its variants.

The scope of this review addresses two main areas. First, it provides a general overview
of cutting and packing problems and their variants in Section 2.2. Our aim of this is
to review and highlight the areas which have been applied in different industries even
though they have been considered only a little in existing literature. By introducing
these problem types, in Section 2.3 and 2.4, we focus on reviewing the specific problems
relevant to the scope of this thesis.

The second interest area is reviewing the existing computational approaches of packing
irregular pieces. In this case we start reviewing the methods available for irregular shape
strip packing problem which is widely discussed in literature than methods related for
IBPP. The main difference of the irregular strip packing problem compared to IBPP is
that instead of using multiple bins, the packing uses only one large strip with fixed width
and variable length as the stock object to pack the pieces. Although our problems are
on IBPP, we review those approaches of strip packing since they provide useful insights
and understanding to work with IBPPs.

In Section 2.5 we review the published approaches of irregular shape strip packing prob-
lem, covering a wide range of techniques available. The majority of those methods are
established on the basis of heuristics and metaheuristics while there is a few noticeable

9
10 Chapter 2 Literature Review

exceptions based on the exact methods. Then we review the computational approaches
specific to the irregular shape bin packing problems in Section 2.6.

2.2 Overview of Cutting and Packing Problems

In a manufacturing organization, the quality of cutting and packing procedures greatly

affects the utilization of resources; i.e. minimizing input material or maximizing the
output. The automation of cutting and packing is highly motivated by the context of
mass scale production due to its efficiency in getting quality solutions in quick time. The
research in C&P spread over a wide range of problem types occur in various industries.
With the improvement of computer technology, research in C&P have been mostly sup-
ported by the benefits gained through powerful computer facilities, and become more
dynamic area of research. Our intention of this section is to review the different types of
problems addressed by the previous authors. The aim of this review is to gain insights
into the problem types and identify which problem types been considered rarely by the
previous authors. As a guide we follow systematic classifications used to differentiate
each problem type separately by their characteristics.

The C&P problems have over 75 years of long history starting from 1940 (Brooks et al.,
1940). During this time period, many problems have been investigated and many solu-
tion methods have been developed to produce quality layouts. Sweeney and Paternoster
(1992) arranged 359 published research articles in chronological order from 1940 to 1990.
This is evidence that C&P is not a new stream of research and it has reached to a cer-
tain maturity. In such context, it is interesting to investigate which areas of the stream
has developed and which areas being addressed less and rarely. The majority of the
publications listed by Sweeney and Paternoster (1992) have addressed one dimensional
cutting stock problems and two dimensional regular shape packing problems while the
publications related to the two dimensional irregular shape packing problems were rarely
addressed.

Publications in early years denoted the focused problems in various names (e.g. Cut-
ting stock problem, trim loss problem, bin packing problem, nesting problem, knapsack
problem) rather than following a consistent, systematic approach to denote each prob-
lem type. This leads to an issue of identifying and concentrating on the problems with
similar categories and therefore further studying of problems become harder. Also, a
categorization enhances the theoretical background of the research area, which is always
supportive of a new person who is interested in C&P research. With the aim of clar-
ifying those problems and standardising their notations, Dyckhoff (1990) introduced a
typology based on a logical structure of C&P problems. He described four important
characteristics of C&P problems; dimensionality, assignment type of large objects or
items, assortment of large objects, and assortment of small items (see in Table 2.1), to
Chapter 2 Literature Review 11

differentiate each category of C&P problem. Considering the new problem types after
1990, Wäscher et al. (2007) presented an improved typology for C&P. In one sense this
is an extension of the earlier typology, since a part of the Wäscher et al. (2007)’s typol-
ogy was based on the classification criteria presented in Dyckhoff (1990). Also, the new
typology introduces new categories of problems.

According to Wäscher et al. (2007), the new typology addressees six basic types of C&P
problems based on the two characteristics; kind of assignment and assortment of small
items. The typology extends these six basic problems into a set of intermediate problems
based on the characteristic of assortment of large objects. The characteristic; kind of as-
signment, categories problems as output maximization problems and input minimization
problems. The output maximization problems focus on cutting small items as much as
possible within a (given) limited number of large objects, whereas the input minimiza-
tion emphasizes on minimizing the number of large objects or the space occupied by
the large objects required to pack the items. Based on the characteristics of small item,
which Wäscher et al. (2007) denoted as degree of assortment of small items, classifies
items as identical items, weakly heterogeneous items and strongly heterogeneous items.
The large objects are also classified based on the assortment, depending on the number
of large objects used; one or many. In the category of output maximization problems, the
problem is classified as a single object (for one), identical (for many) and heterogeneous
(for many). Similarly, for the category of input minimization problem, the classification
was made as identical, weakly heterogeneous and strongly heterogeneous. Figures 2.1
and 2.2 illustrate 14 intermediate problems described in Wäscher et al. (2007)’s paper.
The classification further refines the problem types at its tertiary level (named as Refined
Problem Types) by adding the properties of dimensionality and shape of small items as
adjectives to the intermediate problem types. Accordingly, the problems were denoted
as 1-dimensional, 2-dimensional or 3-dimensional and regular or irregular shape prob-
lems using those adjective phrases. Wäscher et al. (2007)’s typology can be used to
identify unexplored or rarely explored areas of cutting and packing problem types. As
a systematic typology, it provides some guidance to find new problem types which are
necessary to evolve the research stream. Wäscher et al. (2007)’s typology has been cited
in many papers since several authors adapt this typology to denote problem they con-
sider in their studies. In one sense, this standardizes the current and future researchers’
effort to a common framework so that identifying and referring someone’s work become
easier.

A common complication of these typologies is, they cannot explicitly differentiate the
different variants of the problem. As an example, Trkman and Gradisar (2007) proposed
an extension to the scope of Dyckhoff (1990)’s typology. They proposed to include the
variant multi-period when a certain problem’s scope extends to multiple time periods,
since the typology only denoted the cutting and packing problems discussed for a certain
time period. On the other hand, the available typologies are limited in incorporating the
12 Chapter 2 Literature Review

Characteristic Description
Dimensionality Either one-dimensional, two-dimensional, three-dimensional or N-dimensional where N > 3
Assignment type All objects and a selection of items (needs to fill all the objects available)
A selection of objects and all items (need to pack all the items)
Assortment of large objects One object
Identical objects
Heterogeneous objects
Assortment of small objects Few items of different figures
Many items of many different figures
Many items of relatively few different figures
Congruent figures

Table 2.1: Base characteristics of Dyckhoff’s typology Dyckhoff (1990)

Figure 2.1: Seven intermediate problem types under output maximization -

extracted from Wäscher et al. (2007)

Figure 2.2: Seven intermediate problem types under input minimization - ex-
tracted from Wäscher et al. (2007)
Chapter 2 Literature Review 13

practically relevant constraints arise in production processes applying C&P problems.

As an example, Morabito and Belluzzo (2007) presented a hardboard cutting problem
which is categorized as an extended two-dimensional multiple stock size cutting stock
problem, since it includes the additional aspects such as the cutting pattern sequencing.
Considering such extensions and variants, recently, Wäscher (2012) discussed a new set of
characteristics to describe C&P problems specifically following the different constraints
attached to the problems. This includes input-related constraints, output-related con-
straints, assignment constraints of small items, orientation constraints of small items,
and technological constraints, as depicted in Table 2.2. The Wäscher (2012)’s presen-
tation also discusses some variants for C&P problems based on some factors directly
related to the problem. These variants can be categorized based the demand pattern
of small items, regularity (i.e. Regular or Irregular) and quality of large objects and
placement of items. Exploring such variants helps to understand the characteristics of
a particular problem in depth.
Characteristic Description
Input-related Constraints with limited availability of large objects of a particular type
with an upper limit on the number of large objects types
Output-related Constraints Output small items of a particular type exactly to the right quantity
A lower limit on the quantity of small items of a particular type
An upper limit on the quantity of small items of a particular type
Lower limit and upper limit on the quantity of small items of a particular type
Assignment constraints Upper limit on the number of items in a large object
(limited assignment of items Upper limit on the number of different item types
to large objects) Upper limit on the number of different item classes; e.g. colours in a large object
Specific item types must not appear in the same large object
Orientation constraints Fixed orientation, no rotation is allowed
Finite orientation, rotation of small items restricted to a finite number of angles
Finite orientation, however a small deviations are allowed
Unrestricted orientation, free rotation of items
Technological constraints Guillotine unrestricted
(based on the type of Guillotine restricted
permitted cuts)

Table 2.2: New characteristics proposed by Wäscher (2012)

Wäscher et al. (2007) reviewed the 413 C&P papers from 1995 to 2004 and categorized
the problems according to their typology. Out of the statistics for considering each
problem type, it is interesting to investigate the rarely addressed problem types which
have considerable practical value. According to their study, the input minimization
problem of packing 2D irregular with multiple stock sheets is one of the areas being
rarely discussed in the literature 1995 to 2004.

Within the scope of this chapter, one of our attention is on such rarely discussed input
minimization problems related for 2D irregular shapes. Here onward we discuss about
those rarely discussed input minimization problems related for 2D irregular shapes,
following the terminology proposed in Wäscher et al. (2007)’s typology.

In comparison, the number papers on 2D irregular shapes with open dimension problems
are higher than 2D irregular SSSCSP, 2D irregular SBSBPP, 2D irregular MSSCSP, 2D
14 Chapter 2 Literature Review

irregular MBSBPP, 2D irregular RCSP or 2D irregular RBPP from 1995 to 2004. The
paper reveals only two papers considered 2D irregular SSSCSP and one paper considered
the 2D irregular SBSBPP, compared to the 49 papers addressed 2D irregular ODP. Re-
cently Martinez-Sykora et al. (2015) and Xu (2016) comment that number of researches
in 2D irregular CSP are lower than 2D irregular strip packing problem. On the other
hand, the irregular shape packing with multiple stock sheets has greater applicability in
several industries such as ship building, shoe making, garments (Xu, 2016), glass cutting
(Han et al., 2013), leather (Baldacci et al., 2014), furniture (Baldacci et al., 2014) and ce-
ramic plate cutting (Martinez-Sykora et al., 2017) industries. Considering this research
gap and the practical importance, this thesis addresses on 2D ISBSBPP, 2D IMBSBPP
and 2D IMBSBPP with usable leftovers. The three problems are addressed considering
the appropriate problem variants gathered from industry related requirements. As an
example, unrestricted rotation of pieces is one scheme that can be applied in cutting
sheet metal, ceramic plates, foams, and artificial leather, to get a better utilization of
material (Rocha et al., 2013; Martinez-Sykora et al., 2017).

In the next section, a review of the related literature focusing particularly on the prob-
lems considered in this thesis.

2.3 Overview of 2D Irregular Bin Packing Problems/Cut-

ting Stock problems

Only a limited number of publications considered the bin packing problem with irregular
pieces. As a result, the computational methods for irregular bin packing problems are
limited in the literature. In this section, we only present the previous work on the
irregular bin packing problems and their solution approaches in brief. Details of these
solution approaches and their computation methods are reviewed in Section 2.6 of this
chapter.

According to Lopez-Camacho et al. (2013a), a solution method for irregular bin packing
problems focuses on two types of heuristics; 1) bin selection heuristic to pack the next
piece and 2) the placement heuristic which decides exactly where to locate the piece
within the selected bin. Lopez-Camacho et al. (2013a) is one of the studies considered
on irregular bin packing problems with homogeneous bins. The problem is modelled in a
way that the pieces are restricted to rotate only for a pre-defined set of angles. They used
several heuristic methods to solve the problem. Lopez-Camacho et al. (2013a), Lopez-
Camacho et al. (2013b), Lopez-Camacho et al. (2014), and Terashima-Marı́n et al. (2010)
all described a hyper-heuristic approach based on a genetic algorithm for the 2D BPP
for regular and irregular shapes. They describe several selection heuristics to solve the
allocation of pieces to bins including variants of First fit, Best fit, Filler and Djang &
Finch (DJD). They determined the placement position of the pieces inside the allocated
Chapter 2 Literature Review 15

bin using the placement rule; Bottom-Left (BL) algorithm developed by Jakobs (1996)
or an improved bottom left version proposed by Liu and Teng (1999) or a variant of the
constructive approach by Hifi and M’Hallah (2003). We further review the details of
these approaches in Section 2.5.1.

Song and Bennell (2014) proposed a column generation procedure to solve a related
problem called irregular shape cutting stock problem. However the cutting stock problem
has many copies of each piece leading to cut multiple copies of a particular layout to
meet demand. Song and Bennell (2014) investigated three solution approaches; column
generation, adapted column generation that permits fewer patterns, and a sequential
heuristic procedure to compare the quality of packing solutions. Song and Bennell
(2014)’s approach determines required number of copies of each layout by solving the
master problem and the sub-problem is solved heuristically by using the beam search
approach to generate a bin layout.

While above-mentioned studies were limited to restricted rotation of pieces, Han et al.
(2013) and Martinez-Sykora et al. (2015) considered placing pieces with unrestricted
rotation into multiple identical stock sheets, particularly targeting the application of
glass cutting. The considered pieces were limited to convex and adapted the requirement
of guillotine cuts while placing the pieces. The mirror images of the pieces were also
considered in these cases. These two studies are good examples of different variants
attached to irregular bin packing problems as they covered several variants discussed in
Wäscher (2012).

Martinez-Sykora et al. (2017) presented a study of 2D ISBSBPP and considered both

convex and non-convex polygons with unrestricted rotation of pieces. They presented
a constructive procedure to pack pieces, allowing both free orientation and finite set
of rotations. This paper discussed 5 different strategies of assigning pieces to the bins
and compared the performance of them. In order to satisfy a feasible placement; while
considering the orientation aspects, they proposed a Mixed Integer Programming (MIP)
model. Two local search methods were introduced to search over the different assign-
ments of pieces among the bins.

All the above-mentioned approaches are focused on homogeneous bins where the sizes of
the rectangular bins are identical. As an exception, Babu and Babu (2001) presented an
idea to pack irregular pieces into the irregular shape surfaces. This is irregular shapes
packing in multiple irregular stock sheets. The solution method considers a sequence
of surfaces; i.e. irregular stock sheets, placement order of pieces and rotation angle of
pieces, where the assignment of pieces to the bins is decided according to this sequence.
The pieces are placed into the bins using the bottom-left heuristic following the piece
order and they propose a genetic algorithm to search over these sequences. Babu and
Babu (2001) packed irregular pieces with restricted rotation angles given by an integer
degree from 0 to 89. The mirror images of pieces are also considered in this case. Unlike
16 Chapter 2 Literature Review

the previous approaches, Babu and Babu (2001) use discrete placement positions when
placing the pieces. Baldacci et al. (2014) introduced a heuristic approach to solve the
irregular multiple stock-size cutting stock problem with the objective of minimizing
the total cost of the used stock sheets, focusing on the leather and garment industry.
Baldacci et al. (2014) omit the defect area on stock surfaces and packed pieces into the
irregular shaped bins. Similar to Babu and Babu (2001), Baldacci et al. (2014) employs
the Discrete/raster representation of pieces and irregular shaped bin. One major issue
of this Discrete/raster representation is the accuracy of representing the exact shape of
the irregular pieces. The technique approximates the shape profile of pieces and makes
mislead the overlapping tests in some cases either by showing pieces are overlapped
when they are actually not or by showing pieces are touched when they are actually
not touched. In Chapter 3 we discuss characteristics of this representation technique in
detail.

In summary, it is clear that there are different forms of irregular bin packing problems
and only a few of them were considered in the literature. Out of all the approaches, the
irregular MBSBPP is one area that hasn’t been considered with unrestricted rotation
of pieces. In terms of practical applications, industries such as sheet metal, foams,
furniture requires the cutting of irregular shapes from sheets having different standard
sheet sizes. The placement-orientation are not restricted in most cases since the surfaces
of sheets are homogeneous. On the other hand, we believe that the existing literature
available for solving 2D ISBSBPP can be further improved by applying efficient heuristic
techniques which can provide reasonably good solutions in quick time so that they can
be implemented easily in production and procurement environments.

2.4 Overview of Problems considering Usable Leftovers

Another important trend in manufacturing scenarios is checking for the reusing possi-
bilities of material during the manufacturing process. This is one of the well known
philosophies of manufacturing more biased toward the material saving and environmen-
tal friendliness. In this section, we review the literature related to C&P problems with
off-cuts. In such problems, the off-cuts which smaller than a standard bin size used in
the C&P problem are considered to be reused in future orders. To our knowledge, no
literature is available for two-dimensional irregular shapes with usable leftovers. There-
fore, our discussion includes other studies related to one-dimensional cutting problems
and two-dimensional regular shape cutting problems, using leftovers. Reviewing the
solution approaches of those problems are important as they contain useful concepts
that can be applicable to the solution methods for irregular shape cutting problems.
Leftovers are reusable when their sizes are adequate to reuse as input material for the
next orders. Cherri et al. (2009) categorized leftovers either as small” which considered
Chapter 2 Literature Review 17

as scrap or sufficiently large” which is considered as retail (usable leftovers) to return to

the warehouse.

There are very few investigations that consider usable leftovers in cutting and pack-
ing problems and most of them are found with one-dimensional cutting stock problems
(1DCSP) (Kantorovich, 1960; Brown, 1971; Roodman, 1986). Scheithauer (1991) inves-
tigated this problem by satisfying customer demand so that cost is minimized. They
calculated the cost of material as the cost of stock length minus the cost of resid-
ual lengths. In their methodology, the residual lengths are considered using a linear
programming model and they claim that the same approach can be used to solve gen-
eral two-dimensional guillotine cutting problem with rectangular pieces. In their study,
Sinuany-Stern and Weiner (1994) considered two objectives when formulating 1DCSP
with usable leftovers. Out of those two, the first priority is assigned for trim-loss min-
imization and then on maximizing the quantity of leftovers accumulated for one bar
(Sinuany-Stern and Weiner, 1994). In order to solve this problem Sinuany-Stern and
Weiner (1994) used a heuristic approach.

Since 1999, the research work for one-dimensional cutting stock problem with usable
leftovers had increased significantly. The major studies include Gradišar et al. (1999b),
Trkman and Gradisar (2007), Cherri et al. (2009), Cui and Yang (2010) and Cherri
et al. (2013), with the objective of minimizing the trim loss to satisfy the exact demand.
Gradisar and Trkman (2005) discussed two cases, one of which is focused on minimizing
trim loss related to utilized stock objects where an order can be fulfilled with enough
material availability. If trim loss of a utilized stock object is higher than a defined
upper bound, such stocks are returned to the inventory for reuse. In addition to trim
loss, Arbib and Marinelli (2005) considered scheduling and cutting preparation costs
when developing the cut plans with the possibility of using residuals in subsequent
optimizations. Trkman and Gradisar (2007) emphasized that allowing leftovers can
cause accumulation of partly used stock lengths in the warehouse. This is a tough
situation for a manufacturer since handling and storing become problematic with large
number of partly used quantities which are heterogeneous in size. As a solution, Trkman
and Gradisar (2007) suggested to allocate handling and storage cost for leftovers so that
the objective function is based on the cost of trim loss and the stock return cost of
leftovers. As one of the application focused studies, Koch et al. (2009) proposed a
decision support application based on an integer programming model for a company in
the wood industry, in which the same form of the problem applied.

Reviewing all these investigations, we summarize that leftovers are mainly used to satisfy
the primary objective of minimizing the cost of material. The same objective is denoted
in different ways; i.e. minimizing the trim loss, in some of the studies. However, re-
cently the problem is attached with several secondary objectives and constraints due to
the practical issues that arise when implementing the residual reuse task in a manufac-
turing environment. Authors are concerned more with demand and supply constraints,
18 Chapter 2 Literature Review

inventory constraints and handling of material when determining the number of left-
overs and their allowable sizes. Since C&P problems in practice are not independently
progressed, they are always linked to the different kind of costs; i.e. trim loss, machine
setup, inventory and handling. As examples, Henn and Wäscher (2013) included setup
cost in the objective function aiming to minimize the total cost. Combining several
objectives, Erjavec et al. (2012) and Wang and Liu (2014) studied 1DCSP and propose
a new decision model to reduce trim loss and inventory in the paper industry. Cui
et al. (2013) solved two-dimensional guillotine cutting stock problem having the main
objective as input-minimization and the auxiliary objective as pattern reduction so that
the material and setup cost can be minimized. Usually, the number of machine setups
occur during a cutting process has a significant effect on the total cost. Whenever man-
ufacturer has to process a cut layout, there is a need to adjust the knives and some
other settings in the cutting machine, which occupies time. Since excessive times of
such setups are unproductive, minimization of the number of cutting layouts (i.e. the
stock sheets with placed pieces) is usually considered as the second goal when generating
cutting plans (Foerster and Wäscher, 2000).

While some studies apply exact methods to solve the problem with usable leftovers,
particularly in one dimensional cutting problem, some researchers use heuristic methods
to find solutions. Cherri et al. (2014a) presented an interesting discussion on challenges
and issues arising when dealing with usable leftovers (ULs). Even though their study
is solely focused on 1DCSP, there are some important findings to consider when ad-
dressing usable leftovers with two-dimensional cutting and packing problems as well.
Therefore, we summarize some interesting facts of their study by providing some related
examples from the literature. When dealing with this problem, a clear definition of the
context of the application scenario is necessary. Cherri et al. (2014a) identified three
situations where the context can be different; 1) cases where the demands of the pieces
in subsequent periods are completely unknown, 2) cases where the demands of the pieces
in subsequent periods are unknown, but there is a set of positive demands and 3) de-
mands of the subsequent periods are unknown, however their probability distributions
are known.

In the first option, manufacture has no other option other than keeping usable leftovers
in the warehouse and waiting for a future demand of the items to use them. According
to the second case, some items will be in future demand, hence manufacturers can either
overproduce these items and store them in the warehouse to fulfill the future demand
or generate usable leftovers. In the third approach, only the probability distribution
is known so that the manufacturer has to estimate the demand. Similarly, Trkman
and Gradisar (2007) discussed about considering consecutive orders operating within a
selected time-span to deal with usable leftovers (ULs) by presenting the following three
categories; 1) optimization solution of the C&P problem for a given data at one time-
period. They refer to this as instantaneous, 2) optimization for two or more consecutive
Chapter 2 Literature Review 19

time-periods when the demand and supply of material for all time-periods are known
in advance. Therefore, the demands are deterministic for consecutive time periods, and
3) optimization for two or more consecutive time-periods when only the demand for the
first period is known.

Another challenge is how to determine the size of the usable leftovers. According to
Cherri et al. (2014b), the following categories are available in the literature. 1) Determine
the size of the leftover as the length of the largest or the smallest item, 2) Determine
the size as a pre-established value, 3) define a ’good’ size which is cost effective, by
considering possible future demands.

It is also important to find solutions for C&P problems with leftovers within a reason-
able computation time. As stated in Cherri et al. (2014b), most of the approaches use
heuristic methods which solves the problem in a reasonable time, even though the opti-
mal solution is not guaranteed. On the other hand, unlike a single cut order, the context
of this problem has to solve a sequence of cut orders so that it is important to use a
solution method which provide solutions for all those cut orders within a reasonable
computation time. Controlling the number of different UL types also has a significant
effect on the computational complexity (Trkman and Gradisar, 2007; Cui et al., 2016).
A higher number of heterogeneous UL types makes the problem complex, hence a mech-
anism to control UL generation is recommended in Trkman and Gradisar (2007) and
Cui et al. (2016).

As an additional issue, Cherri et al. (2014a) considered about the number of UL types
generated during the cutting process and its significance in managing the ULs effectively.
Usually researchers focus on the ULs generated from standard large objects, so that
chances of generating larger UL become higher. However, when dealing with these
problems, Cherri et al. (2014a) emphasize the importance of evaluating other factors
such as handling off-cuts.

In most cases, the existing models available, either focusing on simple variations of the
traditional C&P problems, or broader problem focused on a specific industry (e.g. Tex-
tile industry (Gradišar et al., 1997), agricultural light aircraft manufacturing (Abuabara
and Morabito, 2009), and wood-processing industry (Koch et al., 2009)). Any of these
approaches have not been able to guarantee the investigation of all the possible scenar-
ios of the problem considering varying costs of material, setups and handling off-cuts.
Cherri et al. (2014a) stated that a company should have an economic analysis mecha-
nism to compare the cost of handling leftovers. They believed in keeping usable leftovers
in the warehouse is economical for 1DCSP when the cost of the material is high, since
the savings can compensate for the costs of handling, transportation and warehouse
(Trkman and Gradisar, 2007; Cherri et al., 2014a).
20 Chapter 2 Literature Review

Figure 2.3: Ways of generating usable leftovers by Andrade et al. (2014)

Only a handful of research has published in two-dimensional rectangular shape cutting

problems with usable leftovers. Andrade et al. (2014) and Andrade et al. (2016) pre-
sented two-dimensional residual bin-packing problems with heterogeneous rectangular
items considering non-guillotine and guillotine cutting cases respectively. Andrade et al.
(2014)’s method arbitrarily assume that two guillotine leftovers (vertical and horizon-
tal) can be produced from each used object that each leftover is generated by a single
guillotine cut. The proposed models consist of cutting the item quantity using a set of
stock-sheets at the minimum cost. If there are any ties in the solutions of minimum
cost, the solution which maximizes the value of the usable leftovers are selected. If there
are any ties in the solution with minimum cost and maximum value of leftovers, the
solution with the minimum number of usable leftovers is selected. This study was con-
tinued by Andrade et al. (2016) and presented two mixed integer programming models
for the non-exact two-stage two-dimensional guillotine cutting/packing problem with
usable stock-sheet remainders.

In Table 2.3 we summarize a set of solution methods used to solve the C&P problems
with usable leftovers. Due to the complexity of the problem, the majority of the existing
solution approaches is limited to using heuristic approaches for developing models and
solving them.
 Table 2.3: Solution methods attempted to solve C&P problems with usable leftovers

Problem type Solution method Source

1DCSP Residual lengths are considered using a linear programming model Scheithauer (1991)
Chapter 2 Literature Review

1DCSP Algorithm which provides an optimal solution for small problems. Sinuany-Stern and Weiner (1994)
A heuristic approach is suggested for the larger problem
1DCSP A sequential heuristic procedure to solve the problem on a large scale aiming Gradišar et al. (1997)
to minimize trim loss in clothing industry
1DCSP Sequential Heuristic Procedure (SHP) applied to heterogeneous stock lengths. Gradišar et al. (1999a)
Based on this algorithm the computer program called CUT is developed.
1DCSP A hybrid approach for optimizing 1DCSP combining two methods: the pattern-oriented Gradišar et al. (1999b)
LP-based method, and the item-oriented sequential heuristic procedure.
1DCSP A solution combining a linear programming based method and a sequential Gradisar and Trkman (2005)
A heuristic procedure so that the method leads to almost optimal solutions
1DCSP A review two frequency used heuristic method; COLA (Gradišar et al., 1997) Cherri et al. (2009)
and CUT Gradišar et al. (1999a), and modifies existing heuristics to solve the problem
2D Regular Introduced Mixed Integer Programming models for 2D non-guillotine problems with ULs Andrade et al. (2014)
2D Regular Two bi-level mathematical programming models to represent two-stage two-dimensional residual Andrade et al. (2016)
bin-packing problem. Based on the of special characteristics of these bi-level models, the models
are reformulated as one-level mixed integer programming models.
21
22 Chapter 2 Literature Review

2.5 Computational Method for Irregular shape Strip Pack-

ing Problems

In this section we review key computational methods available for irregular shape strip
packing problems. Out of different approaches, the majority are based on the category
of heuristics and metaheuristics while there is few methods based on the exact methods.
We review both these categories referring the up-to-date research articles published
related to this problem.

The section divides into three subsections. First, Section 2.5.1 reviews the methods
used to build up a final packing layout. In C&P literature, this is generally denoted
as constructive heuristics (Bennell and Oliveira, 2009). As the name implies, these
methods determine the construction of a complete solution (i.e. packing layout) piece
by piece. Next, in Section 2.5.2 we review those methods that work with complete
solutions where changes are made to find an improved solution. In this case, we review
the metaheuristic approaches methods used when it is required to find a good solution
by searching over feasible set of solutions or required to find a good solution by searching
over the physical layout with the allowance of infeasible solutions. Third, our focus is
on the exact methods which find global optimal solutions. A review of those methods
are presented in Section 2.5.3

2.5.1 Constructive heuristics for irregular shape packing problems

The main task of constructive heuristics is to grow the partial packing layouts. Bennell
and Oliveira (2009) referred to these growing partial layouts as Partial solutions at each
insertion of a piece, till it reaches to a Complete packing solution. In this section, the re-
view is specific to the methods described as single pass constructive methods where pieces
and placed in sequence following a certain placement strategy. Constructive methods
follow legal placement of pieces and do not violate the piece overlapping constraint. The
method can be used with several repeated efforts using different orderings or different
placement strategies to choose the best solution.

In this section, the review is based on different constructive methods, ordering pieces
and placement rules discussed for irregular shapes packing problems.

2.5.1.1 Placement rules

The Placement rule finds an acceptable placement position for the next candidate piece
to be placed on the packing area according to a certain rule. By applying a placement
rule, the authors should ensure that placing position of the piece is contained within the
stock sheet without overlapping with other pieces. According to Bennell and Oliveira
Chapter 2 Literature Review 23

(2009), left most placement is one of the most popular placement strategies used in
irregular packing problems. The rule follows placing pieces toward the left of the stock
sheet as much as possible. The rule can be applied with being biased toward the top
or bottom of the stock sheet. Art (1966) selected bottom side in his study and justified
it as bottom left placement policy. Qu and Sanders (1987) described the bottom left
placement in an alternative way so that the pieces are placed along the bottom of the
stock area. Dowsland et al. (2002) attempted bottom left placement (BL) to locate the
next piece horizontally farthest to left and vertically to the lowest within the packing
space. The objective is to minimize the total length required when placing each piece
into a strip. This placement rule has been used later in several research studies (e.g.
Jakobs (1996), Dowsland et al. (2002), Gomes and Oliveira (2002)).

Dagli and Tatoglu (1987) proposed pairwise matching of the sides of the pieces. In or-
der to support this, the placement position yields the minimum rectangular enclosure is
ensured when placing the pieces. Oliveira et al. (2000) described three placement rules.
The first one is Minimum area placement in which the placement point is selected based
on the area of the enclosing rectangle. Pieces are placed in a position where the area
of enclosing rectangle is minimum (Figure 2.4a). Accordingly the heuristic controls the
build of the partial solution to be as rectangular as possible. Second is Minimum length
placement where the placement point is selected based on the length of enclosing rect-
angle by placing pieces in a position where the length of enclosing rectangle is minimum
(Figure 2.4b). The third rule: maximum overlap of rectangular enclosures selects the
placement point based on the overlap between rectangular enclosure of the placing piece.
Pieces are placed in a position where the overlap between rectangular enclosure of the
placing piece is maximum, while keeping no overlap between the pieces(Figure 2.4c).

Figure 2.4: Placement rules (extracted from Oliveira et al. (2000))

Blażewicz et al. (1993) paid attention on the gaps created by the packed pieces (i.e Holes)
and highlighted the importance of filling holes to obtain a better solution. Dowsland
et al. (2002) and Gomes and Oliveira (2002) described holes as the spaces either sur-
rounded by the packed pieces or spaces generated between the edges of stock-sheet and
the packed pieces.

While Segenreich and Braga (1986) allowed hole-filling relying completely on grid repre-
sentation, Burke et al. (2006) used a strategy in which the horizontal movement of the
24 Chapter 2 Literature Review

Figure 2.5: Bottom left and Bottom left with Hole-filling (Dowsland et al., 2002)

pieces is discrete and the vertical movement of the piece is continuous when filling holes.
Since grid/raster methods are approximation methods, it is interesting to review a hole
filling mechanism without involving grid/raster method to ensure high accuracy.

Similar to previous studies, the C&P researchers who believed in constructive heuristic
approaches tested their constructive methods using different placement rules with the
objective of generating reasonably better solution in a single attempt with less time. One
issue of constructive methods is finding the feasible placement positions accurately for
each piece in an efficient manner. In order to find feasible placement positions accurately,
considering a partial solution with previously placed pieces in fixed locations, one can
compute the set of NFPs between the candidate piece and each of those placed pieces.
Any part of edge/edges of a certain NFP which are not surrounded by any of other NFPs
results the placement region for the next piece to place without overlapping with other
pieces. In other words, the edges or parts of the edges of a NFP inside another NFP, do
not denote placement positions. When this region is bounded by the inner fit polygon
between the candidate piece and the stock sheet object, the corresponding intersecting
edges denote the feasible region for the next piece. Gomes and Oliveira (2002) used this
method to find the feasible placement region for the next piece. If this feasible region
is found by means of set of vertices, then these vertices are the vertices of NFPs and
the points created at the intersection of NFP edges which are bounded by the IF Pk as
illustrated in Figure 2.6. Accordingly, the the final placement position of the next piece
can be found following a placement rule such as bottom-left. This method can be also
used to find the feasible placement positions within the holes in the partial solution.

Oliveira et al. (2000) proposed an alternative approach called TOPOS to construct

layout. They generate the No-fit polygon between the candidate piece and the whole
partial solution which is represented in one polygon. At the insertion of each piece,
the candidate piece is placed and merged into the existing packing layout and updates
the partial solution. Therefore the whole partial solution is considered as one merged
polygon of packed pieces. The reference point of the partial solution is not fixed in
the packing area and consists a floating origin, so that the packed pieces have fixed
positions among themselves but float within the strip together. However the width
Chapter 2 Literature Review 25

Figure 2.6: Feasible placement region for the next inserting piece (image is from
Bennell and Oliveira (2009))

of partial solution is always maintained to be less than or equal to the stock sheet
width. The study tested several placement strategies such as minimum area rectangular
enclosure of the new partial solution, minimum length rectangular enclosure of the new
partial solution and maximum overlap between the rectangular enclosure of the pieces
placed without overlapping pieces. Since the existing partial solution is not fixed, the
partial solution can adjust their position within the boundary of packing space (Oliveira
et al., 2000). This allows a new piece to fit into the partial solution along any side of it.

The original TOPOS cannot detect the holes surrounded by the pieces and this is con-
sidered to be a major limitation of TOPOS.

Figure 2.7: Pieces merge in TOPOS and missing the enclosed gaps (extracted
from Bennell and Song (2010))

Bennell and Song (2010) suggested a revised approach for TOPOS. Instead of merging
pieces, at every insertion, Bennell and Song (2010) proposed to merge the NFPs between
candidate piece and the pieces already in the partial layout. These NFPs are used to
find the feasible positions for the new piece. The method allowed automatic generation
of inner-fit polygons from the enclosed gaps and accordingly established hole-filling.
26 Chapter 2 Literature Review

2.5.1.2 Placement sequences:

The placement sequence of pieces describes the sequence of positioning each piece in the
packing area. Similar to the placement rule, the placement sequence of irregular pieces
also makes a significant impact to the final complete layout. Bohme and Graham (1979)
used a constructive method by throwing pieces on to the stock sheet randomly. They
used at least 2000 random orders of pieces to obtain a reasonably acceptable solution.
In order to find the best solution, they believed on improvement given by small random
perturbations. These random sequences of pieces produced highly variable solutions and
the issue of this approach was to consider a higher number of multiple runs to find a
better solution.

Qu and Sanders (1987) proposed two methods to construct solutions. The first method
sorts pieces by decreasing length and then places them along two adjacent edges of the
stock sheet, whereas the second method sorts pieces by decreasing heights. The second
method was later improved by the authors to have an advantage of possible interlocks
between the pieces. In this case, the tallest piece is selected first as the starting piece.
The other pieces are then dynamically sorted so that the best fitting piece can be found
in subsequence placements. In order to calculate the fit between the partial solution
and each of the remaining pieces, the area ratio between the piece and the free space
remaining from along the partial solution to the right border of the piece was calculated.
Dagli and Tatoglu (1987) conducted an experimental analysis of different initial order-
ing of pieces such as sorted by area, irregularity and number of of sides. Oliveira et al.
(2000) described ordering pieces using the following criteria; decreasing length, decreas-
ing area, decreasing concavity, increasing rectangularity and total area (quantity ×area).
As similar approaches, the authors also test the piece orders; decreasing width, decreas-
ing perimeter, decreasing aspect ratio, and decreasing area utilization, in their study.
Dowsland et al. (2002) used eight static orderings; Decreasing area of enclosing rectan-
gle, Decreasing length of enclosing rectangle,Decreasing width of enclosing rectangle, De-
creasing perimeter of enclosing rectangle, Decreasing aspect ratio of enclosing rectangle,
Decreasing area of polygon, Decreasing perimeter of enclosing polygon and Decreasing
utilisation ratio of enclosing rectangle. The strategy of each placement sequence was
designed to place the difficult-to-place pieces first. All these methods discussed in this
paragraph represent the placement sequences based on a fixed sequence. In comparison
to the random selection, the heuristics are slightly expensive in terms computation to
generate a single solution. However, there are some rules, making reasonably promising
results (e.g. sorted by area) for some sets of instances (Oliveira et al., 2000).

Albano and Sapuppo (1980) used a construction method using a placement rule and
found that focusing only a fixed ordering leads to an inefficient result. As a solution,
they proposed a search tree where each node represents the partial solution at each
stage of solution construction. The approach permits all piece types to be selected
Chapter 2 Literature Review 27

as the next piece to be placed. Bennell and Oliveira (2009) denote this as dynamic
selection approach, where a greedy selection mechanism decides which piece to select
for placement at the current state (i.e search node), based on the characteristics of
the previous state. Using this principle in, Albano and Sapuppo (1980), Bennell and
Song (2010) have modelled the progression of partial solution generation as a search
tree. The lowest level of the tree represented the final solution. In some cases, authors
pruned some branches of the search tree depending on the evaluation results of each
partial solution. In comparison with other rules of ordering, the dynamic selection
investigates several combinations of order sequence which causes higher computation
efforts. However, exploring several combinations in a search tree leads to finding an
improved solution if the criterion for evaluating partial solutions is effective and directs
the search towards finding a better complete solution when it reaches to the end of the
search tree.

According to Bennell and Oliveira (2009); Gomes (2013), working with partial solutions
is one of the two main computational approaches for solving irregular shaped C&P
problem. Generally, the constructive procedures are developed to find a feasible packing
layout by placing pieces one after another. The other approach works with all the pieces
as a whole where all pieces are placed in the layout and tries to derive the final solution
by moving, swapping, or rotating the pieces with a focus of improving the complete
solution. According to Bennell and Oliveira (2009), the constructive procedures are
capable of generating feasible better solutions if it is exposed to an efficient strategy for
finding a better order of pieces and placement strategy. In order to find high quality
solution, it is necessary to work through multiple solutions. According to Bennell and
Oliveira (2009), researchers achieved good solutions using constructive methods through
the approaches like beam search based on a dynamic selection of pieces (Bennell and
Song, 2010) or by iteratively improving the complete solutions based on sequence of
pieces. The constructive heuristics typically support to construct feasible initial solutions
and requires an improvement stage which implement a search mechanism to search over
the solution space to find a good solution.

2.5.2 Work with complete solutions

This section reviews how a search is conducted to find a better solution by making
changes iteratively to the complete solution. Usually these approaches use metaheuris-
tics to search for better solutions. E.g. Simulated Annealing (SA) by Heckmann and
Lengauer (1995) and Oliveira and Ferreira (1993), Tabu Search (TS) by Bennell and
Dowsland (1999) and Burke et al. (2006), Genetic Algorithm (GA) by Jakobs (1996)
and Babu and Babu (2001). The articles Dowsland and Dowsland (1995) and Bennell
and Oliveira (2009) present useful surveys and categorization of using search methods
28 Chapter 2 Literature Review

have been applied to irregular shape packing problems. Bennell and Oliveira (2009) re-
viewed two categories of search algorithms for irregular shape packing problems. While
a certain set of approaches is working directly with the complete solution by moving
the pieces around in the physical layout, another set of approaches is searching over a
sequence of pieces and explore different solutions relying on a certain placement rule.

2.5.2.1 Searching over with feasible solution

These search methods are working similar to the search methods applied for sequencing
problems in operational research. In sequence searching methods, researchers apply
neighbourhood search strategies such as swap moves (e.g. Gomes and Oliveira (2002)),
insert moves (e.g. Takahara et al. (2003)) and change orientation move (e.g. Jakobs
(1996)) to the insertion sequence of pieces. We briefly describe the methodology of each
of these metaheuristics and their features in Chapter 3.

The methods work with solutions which are decoded into sequences of pieces. The
solutions are always feasible and the requirement is to search for a sequence of pieces
to have a better layout using a given placement rule. Figure 2.8 illustrates how each of
these move performs to change the permutation of the pieces.

Figure 2.8: Piece moves (extracted from Bennell and Oliveira (2009))

Jakobs (1996) used GA to search over the piece orders. In this approach the initial
offsprings are generated through a randomly ordered sequence of pieces and progressed
the search through crossover and mutation operators. For mutation, they introduced
an orientation-move by rotating the pieces in 90 degrees. Babu and Babu (2001) dealt
with packing irregular pieces in multiple heterogeneous stock sheets using GA. Their
chromosome coding included three parts; sequence of stock sheet types, sequence of
pieces and sequence of orientation of the pieces. The crossover operator of this algorithm
Chapter 2 Literature Review 29

randomly selects two crossover points, one from the sequence of stock sheets and the
other from the sequence of pieces and changes the sequences by swapping the genes of
two chromosomes. They used mutation operators to implement the swap moves and
orientation moves.

Dowsland et al. (1998) proposed the Jostle algorithm, which packs the pieces to the left
end of the stock-sheet and then to the right end and vice versa. The placement order of
the pieces is determined by the x-coordinate of each piece in the previous iteration. The
heuristic starts with an arbitrary or defined ordering of the pieces and packs pieces from
left to right following the leftmost placement policy. Next, the pieces are re-ordered in
the decreasing order of the x-coordinates of their rightmost points and packed following
the rightmost placement policy (see Figure 2.9). Then the pieces are re-ordered in the
increasing order of their leftmost points and pack again using the leftmost policy. This
will continue for a fixed number of iterations, allowing pieces to oscillate from left to
right and right to left.

Figure 2.9: One step of Jostle Algorithm (extracted from Bennell and Oliveira
(2009))

Takahara et al. (2003) proposed a neighbourhood search based only on insert moves.
The pieces are moved according to a probability which is dependent on their relative
weight. The probability of move is increased by increasing the weight by one unit, if it
makes any improvement to the solution.

Burke et al. (2006) applied the hill climbing local search method and TS to find rea-
sonably better solutions. In this study, the neighbourhood size was limited to randomly
selected five solutions by applying a random moves called; insert, pairwise swap, three-
way swap and n-way swap.

2.5.2.2 Search over the layouts allowing infeasible solutions

According to Bennell and Oliveira (2009), allowing infeasible solutions during the search
space is one of the main characteristics of this category. Accepting solutions with over-
laps and penalizing the degree of overlap in the objective function open ups the search
space. However, this form of search, sometimes struggles to reach a local optimum with
feasible placement of pieces. Bennell and Oliveira (2009) also highlights that the issues
arising in designing a mechanism is mainly due to the size of the neighbourhood, which
30 Chapter 2 Literature Review

is infinite since placement and orientation of the pieces can be infinite in a stock sheet
which is defined on a continuous scale.

Many local search approaches have been used to improve the physical layout. SA and TS
are two initial local search techniques used in solving irregular shape packing problems.
In these search methods, the search process starts with a starting solution and explores
one or more neighbour solution as a replacement for the current solution. As described
in Dréo (2006), the entire search method focuses on the solution space, neighbourhood
structure and the cost function.

Oliveira and Ferreira (1993) used SA, which places pieces randomly on the packing area
and performs neighbourhood movement one piece at a time. They move a piece in the
grid layout by one grid unit at each move. As states in (Bennell and Dowsland, 2001),
SA was most favoured in the studies of the 90s than TS since the fundamental approach
of TS does not favour the infinite neighbourhoods in a continuous stock sheet. However,
Blażewicz et al. (1993) used TS in a discretized packing area which allows the conversion
infinite piece placement positions into a finite and discrete set.

Egeblad et al. (2007) developed an efficient search method which translates pieces in a
specified direction and finds a promising position for a piece that minimizes its overlap
area. They used Guided Local Search to escape local minima.

Bennell and Oliveira (2009) review two major frameworks to change the solution and de-
fine the neighbourhood of a solution. They are; 1) moving a single piece or a set of pieces
either using inset-piece move, swap move, changing orientations or horizontal transla-
tion, vertical translation, or flipping of pieces and 2) Use compaction and separation
procedure, to change positions of pieces simultaneously. Bennell and Dowsland (2001)
first introduced compaction using linear programming. Gomes and Oliveira (2006) used
a hybrid methodology of Simulated Annealing (SA) and linear programming. Once an
initial layout is constructed using the bottom-left placement rule, SA was used to guide
the search over the solution space. The linear programming was used to apply the com-
paction algorithm at each neighbourhood structure. The objective of compaction was
to slide the pieces in a given layout to achieve better layout. A separation algorithm
was used to remove the overlapping while moving the pieces.

Imamichi et al. (2009) proposed a new separation algorithm based on unconstrained Non-
linear programming. They also designed an algorithm which swaps two pieces placed in
the layout and this was combined into an iterated local search algorithm to minimize
the overlap by those swap moves. Leung et al. (2012) proposed an extended local search
algorithm based on Non-linear programming. In this case, two neighbourhoods were
used to change piece positions during the local search where as the feasibility of the
solution is determined by a set of non-linear programming separation and compaction
models. Elkeran (2013) proposed a methodology which uses a heuristic to cluster the
pieces in pairs and then use guided cuckoo search to pack the pieces into the packing
Chapter 2 Literature Review 31

area. This study achieved the best results in the literature for the irregular strip packing
problem.

According to Bennell and Oliveira (2009), searching with feasible solutions is compu-
tationally more expensive than searching for a piece move within the layout since the
sequence change requires generating the whole solution to evaluate each change. The
quality of the solution depends on the placement rule used in the constructive algorithm.
The approach guarantees evaluation of feasible solutions at every time it generates a com-
plete solution. In contrast, the moving pieces within layout can create overlaps among
the pieces while searching for better positions. In some situations, it takes longer to find
a feasible solution.

2.5.3 Exact methods for irregular shape packing problems

Though heuristic and metaheuristic algorithms are popular among the majority of the
studies and recorded the best performance (e.g Imamichi et al. (2009), Bennell and
Song (2010), Elkeran (2013), Lopez-Camacho et al. (2013a)), they are not able find
optimal solution or determine how far the resulting solution is in the distance to the
optimal. This research gap is trying to be filled by a few researchers using exact solution
approaches which are based mathematical programming models.

As an initial approach, Ribeiro et al. (1999) developed an exact method to solve the strip
packing problem with convex pieces based on constraint programming. The constraints
of the mathematical programming model ensured the overlapping and containment con-
straints. They defined the allowable placement points for each piece by imposing non-
overlap constraints for each pair of pieces. As another initial work, Carravilla et al.
(2003) used the exact computational approach for solving nesting problem with non-
convex pieces and was able to solve the problem to optimality with maximum of 7
pieces. In both studies, the pieces reference point were associated with a discretized
placement within the stock sheet.

Fischetti and Luzzi (2008) developed a mixed integer programming model and a branch-
ing strategy which is used with CPLEX. The non-overlapping constraints of this model
uses tighter coefficients instead of big-M constants (Martinez-Sykora, 2013). Kallrath
(2008) introduced an exact nesting algorithm for convex polygons using the idea of sep-
arating hyperplane. The convex polygons are considered pairwise and a hyperplane that
separates the two polygons is searched to satisfy the non-overlap constraint. The ap-
proach can also handle non-convex polygons by decomposing them into convex polygons.
The main issue with this approach is the difficulty of finding optimal solutions for the
problems with more than two polygons, especially with non-convex polygons.

For strip packing problems, Toledo et al. (2013) proposed the Dotted-Board Model where
the placement points on the stock sheet (they referred this as a board) are allowed in
32 Chapter 2 Literature Review

the integer positions. This approach was efficient and able to solve to the optimality
for 21 pieces in 7 different instances. In some cases it solved to the optimality for
56 pieces only with two different instances. In a bin packing problem, Baldacci et al.
(2014) discretized both bins and the pieces into a bit-matrix, and imposed constraints
to cover each position of the matrix. Due to this reason, the exact model does not
allow placement position in a continuous scale. The issue of this model is the accuracy
which is highly determined by the resolution of discretization. However, the discretized
methods simplify the formulation and solution procedure.

The studies related to continuous positioning models using mixed-integer programming

have used no-fit polygons (NFPs) so that the outside area of the NFPs is used to set up
the non overlapping constraints of polygons relative to each other (Gomes and Oliveira,
2006; Alvarez-Valdes et al., 2013). When compared to the discretized methods, the ex-
acts models based on continuous placement are able to solve to optimality for a smaller
number of piece types (Cherri et al., 2016). As a better performance, Alvarez-Valdes
et al. (2013)’s continuous exact approach, the model solves to optimality instances up to
16 pieces within 5 hours of computation time. Alvarez-Valdes et al. (2013) developed an
exact Branch & Bound Algorithm and explored different branching strategies. Cherri
et al. (2016) produced mathematical programming models based on two directions. The
first direction they used direct trigonometry to derive non-overlapping constraints with-
out using NFPs. The second direction decomposes the irregular pieces into convex
shapes and generated NFPs for the convex shapes. The intension of these approaches
was to reduce the complexity of the exact model by reducing the complexity of of geom-
etry components attached to it. The models addressed irregular strip packing problems
with polygonal shape geometries without any approximation. The major drawback of
this approach was the size of the instances possible to solve to the optimality. The
approach NFP-CM discussed in Cherri et al. (2016) finds an optimal solution for most
of the instances with up to 12 pieces whereas the method cannot find a feasible solution
with more than 30 pieces. In summary, the exact method approaches developed so far
haven’t been able to deal with the large instances (>30 pieces).

2.6 Computational Method for Irregular shape Bin Pack-

ing Problems

2.6.1 Constructive methods

In this section, the review scope is limited only for irregular shape bin packing prob-
lems and discusses the constructive methods applied for irregular bin packing problems.
Only a limited number of publications consider the bin packing problem with irregular
pieces where most of them are only limited to pack irregular shapes in single size bins.
Chapter 2 Literature Review 33

Therefore, the computational methods developed for irregular bin packing problems are
not plenty in the literature.

In general, the existing studies focus on two decisions; 1) selecting both the next piece
and the bin, and 2) finding the exact placement locating of the piece within the selected
bin, specific to the 2D IBPP (Lopez-Camacho et al., 2013a, 2014). Lopez-Camacho et al.
(2014) claimed that for non-trivial irregular bin packing problems the exhaustive search
is impractical, and highlighted the necessity of heuristic methods.

Terashima-Marı́n et al. (2010) and Lopez-Camacho et al. (2013a) discussed following

bin selection heuristics; First Fit Decreasing (FFD), Best fit Decreasing (BFD), Filler
(FL) and Djang and Finch (DJD) for the first decision. According to Lopez-Camacho
et al. (2013a), the above heuristics generated reasonably better solutions compared to
the other heuristics methods investigated. In this study, Lopez-Camacho et al. (2013a)
investigated 11 selection heuristics, including a few of worst performing heuristics such
as Worst fit (WF), First fit Increasing (FFI). In general, the first three heuristics; FFD,
BFD and FL are used for the cases of 1D and 2D bin packing problems in the literature.
In all three heuristics, the pieces to be packed are arranged in an a certain sequence and
one piece is packed at a time.

The FF heuristic opens a bin if the piece cannot fit in any of the opened bins. As
a subversion, the First-Fit Decreasing (FFD) heuristic considers the pieces sorted in
decreasing order of the area and places each piece inside the first bin where it fits
(Lopez-Camacho et al., 2013b, 2014). The NF heuristic always selects the current bin
to place the next piece. If the new piece fits in the current bin, it is placed there,
otherwise open a new bin and place the piece in the new bin (Lopez-Camacho et al.,
2013b, 2014). The BF heuristic tries to place the piece in the opened bins where it best
fits. A new bin is opened only when there is no space to allocate the next piece in the
opened bins. A subversion of this called Best Fit Decreasing (BFD) sorts the unplaced
pieces by decreasing area and places each piece in the opened bins where it fits best
(Lopez-Camacho et al., 2013b, 2014).

The key idea of Filler (FL) is to pack pieces as much as possible for an opened bin
without considering the order of the bins. The heuristic starts with sorting the pieces in
order of decreasing area and then packs as many pieces as possible into the opened bin.
In a situation where no single piece can be placed in the open bin, a new bin is opened
to pack the remaining unpacked pieces from largest to smallest (Lopez-Camacho et al.,
2013b, 2014).

Lopez-Camacho et al. (2013a) denoted Djang and Finch(DJD) as a complete single-

pass constructive heuristic for 1D problems. However, for 2D problems it is only a
bin selection heuristic (Lopez-Camacho et al., 2013a). The heuristic works in one open
bin at a time and does not consider the other open bins. Pieces should have an order
which is usually decreasing area. The heuristic works with an assigned value of fullness,
34 Chapter 2 Literature Review

usually taken as 1/3, 1/4, 1/2, for example, if 1/3 of fullness is allowed (i.e. then the
algorithm is called as DJD 1/3), then the heuristic fills the bin with the sorted pieces
until one-third of it becomes full. As the next step, it initializes a variable waste = 0.
The heuristic then attempts to find either one, two, or three pieces that completely
fill the bin leaving a free area up to waste. If there is no combination matches with
the requirement and waste is less than the bin free area, then the value of waste is
increased a certain amount (usually a fraction of the bin capacity), otherwise a new bin
is opened. Lopez-Camacho et al. (2013a) empirically found that, increasing this waste
as one-twentieth of the total bin area, provides a good balance to generate efficient
solutions. Lopez-Camacho et al. (2013a)’s experiments demonstrate that the DJD 1/3
and DJD 1/4 produce better results than the other selection heuristics. However, DJD
heuristics are time consuming since it tries several combinations before placing the 3
pieces.

Lopez-Camacho et al. (2013a) considered four placement heuristics to work with the
selection heuristics. They considered bottom left and three versions of constructive ap-
proach placement rule which is originally presented by Hifi and M’Hallah (2003). They
used BL heuristic in a way that piece starts moving the right hand corner of the bin and
slides down and left with a set of movements until no further movement is possible. The
piece does not overlap with the bin edges and it is not possible to pass over the already
placed pieces while sliding the candidate piece. The constructive approach heuristic
starts by placing the first piece at the bottom-left of bin. Each candidate piece starts
its placement from each of the nine placement positions. These nine points include the
four corners of the bin and coordinate positions; (0, ȳ), (x, ȳ), (x̄, ȳ), (x̄, y), (x̄, 0) where
x̄, x, ȳ, and y are the maximum and minimum coordinates in x and y of the previously
placed piece as illustrated in Figure 2.10.

From each of these starting points, the feasible movement positions are recorded when
the next piece slides vertically and horizontally following down and left movements as
illustrated in Figure 2.10. Out of the feasible points reordered, the most bottom-left
position is selected as the placement point for the next piece.

Lopez-Camacho et al. (2013a) modified the CA rule to select the best placement position
based on the position that yields the minimum bounding rectangle (the minimum area)
which contains all placed pieces and the new piece. This was called as CA with Minimum
Area (CAA). Another CA version was tested in the same study where the placement po-
sition is evaluated based on the largest adjacency. The adjacency describes the common
boundary between piece perimeter and the placed pieces and the bin edges (Terashima-
Marı́n et al., 2010). As described in CA rule, the candidate piece is sliding down and
left, to place in the position where the largest adjacency is found. Lopez-Camacho et al.
(2013a) denoted this heuristic as CA with maximum Adjacency.(CAD).
Chapter 2 Literature Review 35

Figure 2.10: Constructive approach (Terashima-Marı́n et al., 2010)

When placing the pieces, orientation of the piece (angle of rotation) impacts significantly
for the final placement. Terashima-Marı́n et al. (2010) tested two schemes of rotating
pieces; the first one rotates each piece by multiples of 90 degrees (0, 90, 180 and 270
degrees) and the second rotates each piece in multiples of 5 degrees. They found that
there is no significant different in both approaches. We believe this is mainly due to the
jigsaw type pieces in which the optimal solution can only be found with the multiples
of 90 degrees (i.e. 0, 90, 180, 270) rotation angles.

The placement rules suggested in Lopez-Camacho et al. (2013a) is time consuming due
to their sliding operations which performs several times till the piece reaches its final
placement position. In addition, CA, CAA and CAD heuristics are more time consuming
than BL due to the exploration of the movements start from several starting points as
described in the CA placement rule. The other issue of Lopez-Camacho et al. (2013a)’s
constructive procedures is they are only functioning with a predefined discrete set of an-
gles. The study mostly used jigsaw type pieces to test each of the constructive methods.
Usually such pieces are aligned to a better solution with a selected set angle. However,
in practical situations, most pieces are not jigsaw type and requires to consider a higher
number of orientation angles efficiently by the constructive method use.

Han et al. (2013) proposed two constructive heuristics for bin packing problems with
irregular pieces. The problem only considered convex pieces arises from glass industry
where the cutting process forces use of guillotine cuts. Han et al. (2013) discussed a
one-stage solution approach as a constructive heuristic which begins with generating
efficient clusters of polygons. These clusters were used to construct a complete solution.
36 Chapter 2 Literature Review

A part of this clustering process possesses the matching of two polygons. This includes
placing each polygon relative to the other polygon based on the objective of minimizing
the waste. When two polygons are placed (touching with each other) relative to each
other, Han et al. (2013) calculated the utilization gained by the pieces over the area
of the convex hull of these two polygons, and the utilization gained by the pieces over
the area of the rectangle enclosure of those (see Figure 2.11). In order to evaluate the
best matching position, these two measurements were used following a weighted sum
approach. Always pieces are touched by its edges and along the edges it finds the best
relative placement position.

Figure 2.11: Slide along the edges (extracted from Han et al. (2013))

Han et al. (2013) defined a search forest to represent the constructive stages of the
solution. A node denotes the matching of a set of polygons if the utilization ratio of
those polygons achieves or exceeds an acceptance threshold. The forest is continued till
there are no further matches for the blocks. Han et al. (2013) defined blocks as a set of
polygons grouped together by a feasible transformation. Figure 2.12 illustrates a forest
made with composition of blocks through the levels of the search. The top level contains
all pieces individually. The second level contains pairs of pieces match to a higher
utilisation. The last level (level six in this case) contains all six pieces transformed to a
block.

The next step of this approach is to pack the blocks into the bins. In this step Han
et al. (2013) placed those blocks in bins efficiently. They defined a utilization threshold
to check whether a bin packing pattern is acceptable. Han et al. (2013) also proposed
a two-stage algorithm that groups pairs of pieces into rectangles and then packs those
rectangles using the guillotine bin packing algorithm proposed by Charalambous and
Fleszar (2011). The approach works well for convex pieces and presents quality results
within a reasonable time. However, this approach includes guillotine constraints and
edge matching which addresses significantly a different problem than the problems con-
sidered in this thesis. Also, the proposed constructive algorithms are only capable of
dealing with convex pieces.

As a similar problem, Martinez-Sykora et al. (2015) presented a constructive algorithm

to build irregular shape bin packing solutions with guillotine constrains by adding pieces
sequentially to the current bin. They use a mixed integer programming (MIP) model
Chapter 2 Literature Review 37

Figure 2.12: Search forest (extracted from Han et al. (2013))

to determine the position of the pieces in the bin and the corresponding guillotine con-
straints. In their constructive method, the relative position of the piece with respect to
the other pieces is fixed, however the absolute position of each piece can be changed at
each insertion of pieces when the constructive algorithm is being run. The constructive
algorithm uses the set of angles of rotation and reflection for each piece sequentially
place to the bin. For a given initial permutation of the pieces, within an open bin,
the candidate piece generates the rotation angles using an algorithm called GR which
produces the rotation angles of the piece by matching the edges of the piece with the
edges of the other pieces. For each rotation the authors solve the MIP model and try to
find the best feasible placement position. If such a feasible solution is found, the algo-
rithm adds the candidate piece into the bin and updates the orientation angle; otherwise
they leave the piece unpacked, and move on to the next and try to place it in the bin.
When no other piece can feasibly fit into the bin, the algorithm closes that bin, opens
a new bin and start packing from the first unpacked piece in the list. After packing
all pieces, the authors evaluate the occupied bin with the lowest utilization using two
evaluation functions and rebuilt the bin with the aim of packing the pieces as tightly
38 Chapter 2 Literature Review

as possible. Martinez-Sykora et al. (2015) denotes these two objective functions as FO1
and FO2, where FO1 is used to minimize the width when there is tie in length used
and the FO2 used is to minimize the length, when there is tie in width used. At the
last step of the constructive algorithm, the rebuilding of the lowest utilized bin involves
applying a horizontal or vertical cut respectively to differentiate the occupied area of the
bin to find the fractional number of bins. At each insertion of a piece, Martinez-Sykora
et al. (2015)’s method calls the MIP model several times. This requires a significant
computational effort. In Martinez-Sykora et al. (2015)’s approach, computational time
to run the constructive algorithm highly depends on the number of pieces in a problem
instance.

The same study proposed another constructive algorithm having a change to the pre-
viously discussed constructive algorithm which identifies the guillotine cuts after each
piece insertion and then sets as constraints to the MIP. In the previous constructive
method, the next piece is forced to satisfy all the guillotine cuts previously defined be-
fore finding the best insertion. Martinez-Sykora et al. (2015) proposed that this can over
constrain the solutions space. Therefore, they propose a change to their first method so
that the next piece is inserted ignoring the guillotine cut constraints. After finding the
best insertion, they set the algorithm to perform a procedure to identify a new guillotine
cut structure, if one exists. They denoted this procedure as two phase constructive al-
gorithm. According to this new modification, they also changed the constraints in their
MIP model to work with this second constructive procedure as well. Martinez-Sykora
et al. (2015) experimented three alternatives sorting criteria for the pieces; random,
Non-increasing area and by shape.

Martinez-Sykora et al. (2017) described a constructive algorithm considering the two

aspects of the irregular bin packing problem including free rotation of pieces. As the
first aspect, the constructive method considers the assignment of pieces to bins where
as the second aspect considers the arrangement of assigned pieces in the bin. Martinez-
Sykora et al. (2017) implemented five different strategies for assigning pieces to the bins;
1) Greedy Decisions, 2) First Fit Algorithm, 3) Partial Bin Packing, 4) Two Phases
Strategy and 5) a simple construction heuristic. In the constructive algorithm, one of
these was implemented as the first phase.

The first method greedy decisions (GD) was based on the idea of retaining the bin
that has a feasible arrangement with some of the assigned pieces. Then the focus is
to add pieces from the other bins without violating the feasibility and improve the
bin utilization. This strategy assigns pieces to bins by solving an IP model used in a
1D bin packing problem. Since it solves an IP formulation, the first strategy requires a
considerable time to generate a complete solution. The second strategy uses the popular
first fit heuristic (FF). According to Martinez-Sykora et al. (2017), this algorithm is
very fast and the quality of the solutions depends on the initial ordering of the pieces.
The third strategy Partial Bin Packing (PBP) uses an objective function that favours
Chapter 2 Literature Review 39

assigning larger pieces. This assignment strategy follows a knapsack formulation as in

1D BPP, so that the objective is to maximize the value of bin by packing some of the
available pieces. If packing procedure (This is separately explained in the next paragraph
briefly, more details of it are available in Martinez-Sykora et al. (2017)) fails to place a
given piece j feasibly, then they solve the knapsack formulation again for the remaining
capacity by removing j from the consideration. Therefore, with the new assignment,
the packing procedure attempts to place the newly assigned piece feasibly from the
available pieces. These tasks; feasible packing (performed by the packing procedure)
and reassignment, continues until it finds a bin with feasible placements of pieces. Any
piece(s) j which are removed during the process are returned back to the available set of
pieces to pack. As the next step, the algorithm then solves the knapsack formulation to
generate a new bin. The fourth strategy is a two phase one, which executes the GD first.
In this case, after solving the assignment model of the GD, it provides the minimum
number of bins. Martinez-Sykora et al. (2017) solve another IP model to reassign pieces
across bins. This model forces the assignment of large pieces into the earlier bins so
that it can minimize the total number of bins. The fifth strategy was called simple
construction heuristic which packs pieces into a bin following a given sorted order of
piece. The method uses Next-Fit Decreasing (NFD) strategy, in which the next piece i
is packed feasibly into the opened bin by solving the IP model described in Martinez-
Sykora et al. (2017). In this case, the algorithm considered packing the next piece (i + 1)
to the current bin. If piece i cannot be placed feasibly, close that bin and open a new
bin.

In the second phase, the subset of pieces assigned to each bin is sequentially placed
inside the bin using the packing procedure. This includes determining a set of promis-
ing orientations for the piece and for each orientation angle determining whether the
piece fits into the bin according to the MIP model. The packing procedure allowed
all pieces to move within the bin when placing the next piece. When assigned pieces
are failing to pack feasibly, reassignment takes place as explained in the previous para-
graph. Comparing each version of the constructive algorithm, out of these five options,
Martinez-Sykora et al. (2017) found the best results for the PBP strategy. The packing
procedure described in this article allows placement of pieces with free rotation. This is
also the first study to deal with free rotation of pieces for both convex and non convex
pieces in irregular bin packing. Since the problem definitions are the same in this ar-
ticle and the first problem addressed in this thesis, we use the results of this study to
compare the performance of our new computational method. As the key improvement
areas we focus on improving the computational time, since the packing procedure has
taken considerable computational effort to solve the MIP due to the higher number of
constraints attached to it.
40 Chapter 2 Literature Review

2.6.2 Search methods for irregular bin packing problems

Terashima-Marı́n et al. (2010) presented a GA-based method to produce general hyper-

heuristics for the irregular shape bin-packing problem. In the hyper-heuristics approach,
they attempted to explore the efficient combination of heuristics to apply for a given
situation of the problem. Hyper-heuristics design generic methods to produce solutions
with acceptable quality(Burke et al., 2013). The performances are based on the set of
simple low-level heuristics where as the hyper-heuristic acts as a high-level guidance
methodology (Burke et al., 2013). After evaluating the individual performance of each
heuristic, Terashima-Marı́n et al. (2010) used GA to evolve combinations of rules to
produce the suitable hyper-heuristic. Terashima-Marı́n et al. (2010) observed a slightly
better performance when the hyper-heuristics are tested with more irregular instances.
They also found that, when compared to the 12 hours in solving 541 instances with 40
single heuristics, the hyper-heuristic approach takes 11 hours to run the GA to generate a
hyper-heuristic. Once a suitable hyper-heuristic is found, solving each instance in the set
with the produced hyper-heuristic taken only a few minutes. Accordingly, Terashima-
Marı́n et al. (2010) believed that hyper-heuristic approach is faster than solving the
instances with each of the 40 single heuristics. However, one issue with this approach is
the performance of low level heuristics when it investigates a higher number of rotation
angles. All 40 instances investigated by Terashima-Marı́n et al. (2010) worked with a
finite set of angles (< 4 in most cases). In Terashima-Marı́n et al. (2010)’s study the
performance of the hyper heuristic approach is based on the range of low level heuristics
and the selection mechanism of heuristics. However, as reviewed above, the low level
heuristics of this particular study shows its efficiency only with a finite set of rotation
angles. In addition, no constructive method is described to consider unrestricted rotation
of pieces and only provision there is to consider the higher number of finite rotation
angles which cause higher computational time ultimately.

Martinez-Sykora et al. (2015) discussed the solution method for irregular bin packing
problem considering guillotine placement of convex pieces. In their solution method,
they developed an improved procedure embedded into the construction algorithm so
that utilization of a bin can be improved before that bin is closed. Then the improve-
ment phase is executed within the constructive algorithm before opening a new bin. The
authors applied this to the bins only with utilization below a given threshold. According
to this method, the improvement procedure removes the piece with the worst utiliza-
tion and rebuilds the bin without changing the placement and orientation of the other
pieces in the bin. In this case, the utilization is calculated as the ratio between the area
of the piece to the area of the containing polygon. As illustrated in Figure 2.13, the
guillotine cuts define the containment polygon around each piece. As the next step the
algorithm tries to insert pieces in the unpacked list. In this phase, the removed piece is
also considered at least to insert to the same bin if it is possible to place considering ten
Chapter 2 Literature Review 41

Figure 2.13: Containment Polygon (extracted from Martinez-Sykora et al.

(2015))

alternative rotations at each reflection option. As the acceptance criteria, the bin solu-
tion is accepted and identify the next piece with the lowest utilization if the utilization
of the repacked bin is higher, otherwise the procedure continues until no improvement
can be found.

Martinez-Sykora et al. (2017) presented two local search strategies based on hill climbing
first found improvement. The search method was simple, did not include any diversifi-
cation scheme and terminated when a local optimum solution is found. They considered
the coefficient F proposed by Lopez-Camacho et al. (2013a) to evaluate solutions. If a
move makes any reduction in the number of occupied bins, as a result F increased and
such move was accepted during the search process. Also, if a move makes any improve-
ment in utilization in one bin maintaining the same number of bins, F was increased,
and accepted the move. During the search process, both properties are important to
find good solutions as it differentiates solutions with the same number of bins.

Out of two search methods, the first method starts with a solution of occupied bins
sorted by non-decreasing utilization. Then the local search procedure considers all pairs
of bins, say i, j and tries to move pieces one after the other, from one bin (i) into another
(j). The move accepted if all pieces from both bins fit into only one bin or if utilization
of j is increased and utilisation of i is decreased. The objective of this search method
is to reduce the usage of the least utilised bins with the aim of emptying those bins.
The second method proposed by Martinez-Sykora et al. (2017) was an extended version
of the first search method, In this case, in order to empty bin i, a set occupied bins
were proposed instead using a single bin j. This increases the chances of packing all
pieces of bin i into a set of bins. As the set of j bins, the bins with higher utilisation
than bin i and lower than 0.99, are selected for new placement. In the first local search,
Martinez-Sykora et al. (2017) rebuild bin i when a new bin j is built. However, in
the second search method, they consider more than one bin of js before rebuilding
the bin i. A fail to pack all the pieces from bin i, or in other words, a fail to reduce
42 Chapter 2 Literature Review

at least one bin from the previous solution, fails the whole movement and replaces to
the previous solution. According to Martinez-Sykora et al. (2017) both local search
procedures obtained significant improvement and the second LS2 performed better than
first, even though it consumed almost similar computational time.

Both Martinez-Sykora et al. (2015) and Martinez-Sykora et al. (2017) approaches present
the best result in each of their contexts. The approaches established high accuracy due
to the use of the geometric representation and they allow free rotation of the pieces
which is rare in the literature. One main issue arises is the high computational time
they take to find a reasonably good solution. Reviewing all the solution approaches for
irregular shape bin packing problems, we find that it is strongly significant to reduce
the computational time of the constructive algorithm especially when pieces orientation
is unrestricted. Meanwhile, improvements in search method are also critical and require
experimenting different search mechanisms to improve the solutions.

2.7 Summary

Wäscher et al. (2007) and Wäscher (2012) provide substantial information about the
entire scope of cutting and packing problems and their variants in existing literature.
Within this context, our review of literature mainly concentrated on irregular shape
strip packing and irregular shape bin packing problems. Even though the objective of
the thesis is on irregular bin packing problems, we reviewed the solution methods used
for irregular strip packing problem since some of the concepts are useful in developing
solution methods for the irregular bin packing problems.

Out of the different solution approaches for irregular shape packing problems, the liter-
ature contains two main categories of solution approaches; exact methods and heuristic
based approaches. Within the scope of this thesis, we intend to develop algorithms for
the industries running mass scale production, which is dealing with a higher number of
pieces and a higher number of stock sheets. Out of two main computational approaches;
heuristic based approaches are favoured more as they can provide good solutions within
reasonable computation time even for large problems. On the other hand, the use of
exact method demonstrated its power in finding the global optimum solution only for a
limited number of pieces.

Reviewing heuristic based approaches, we noted that there are two main approaches
for dealing with the problem. One approach always works with the feasible solutions
and attempts to find better solution using a search method. The other approach accepts
overlaps (i.e. infeasible solutions) and continues searching over the layout until it reaches
to feasible, good solution. The chapter reviews the advantages and disadvantages of both
approaches based on the outcomes of different studies.
Chapter 2 Literature Review 43

We also review the literature specific to the IBPP and note that the computational
time for solution construction is a significant issue when the packing orientations of
shapes are not restricted. Even-through the number of studies related to this area is
limited, so far, a few researchers attempted to solve this problem using metaheuristics
and math-heuristics approaches.
Chapter 3

Methodology

3.1 Introduction

This chapter presents a discussion on the key methodologies adopted in solving the three
problems mentioned in Chapter 1. According to Bertrand and Fransoo (2002), the re-
search discussed in this thesis belongs to the category of axiomatic research where the
primary concern is to produce solutions within a defined model while ensuring that the
solutions generated provide insights to the real problem. The problems are described
based on the objective models we developed during the study, which explain the be-
haviour of real-life operational processes with 2D cutting and packing functions. With
respect to each problem, the models investigated in this study are determined by various
methods and techniques in mathematics, computer science and management science.

As explained in most of the axiomatic researches, the study discussed in this thesis is
interested in developing strategies, tools and techniques to find good solutions for a
newly defined problem or to improve the results for a problem already considered in the
literature. In general, each problem of the study has been passed through three phases.
In the first phase, the model of the problem is designed, which includes conceptual-
ization, reviewing similar types of problems, making decisions about variables, defining
the scope of the problem and formulating the model. During the second phase, the
formulated problem is solved using the principles of mathematics, geometry, computer
science and operational research. In the third phase, an analysis is conducted using the
implementation of results, so that useful insights can be obtained about the problem
and its relatives.

The problem modelling is started with the intention of identifying the characteristics
of the problems within the operational process we aim at. A comprehensive literature
review is conducted using peer reviewed articles specific to the problem type and its
variants. By focusing on the research gaps of existing literature the problem types that

45
46 Chapter 3 Methodology

have been rarely considered are identified. Chapter 2 reviews those gaps and presents the
respective problem types and their features. We then focus on developing the models for
each problem in a way that these can be read through a computer program and solvable
with a designed algorithm. The concept of geometry used in this study is a key research
topic within irregular shape C&P problems and many techniques have been published
describing how to handle the irregularities of pieces, especially in the overlapping and
containment calculations.

Literature review in Chapter 2 reveals two main categories of solution methods; exact
methods and heuristic methods. Exact methods have been applied in order to find
the global optimal solution to the given problem. However, the existing studies reveal
that exact methods are not able to prove optimality for large-size” instances which
causes a high level of complexity to the C&P problems. As reviewed in Chapter 2,
so far, the exact methods find an optimal solution for most of the instances with up
to 12 pieces and the methods even cannot find a feasible solution with more than 30
pieces for most instances (Cherri et al., 2016). Heuristic techniques, on the other hand,
have been developed to provide a good feasible solution in such cases. Though the
heuristic approaches are generally fast, they do not guarantee to solve the problem
towards optimality. For the improvement of the solutions from the heuristic approaches,
the local search techniques have been used. The purpose of local search techniques is to
reach a local optimum of the specific area it explored over the solution space. Usually,
the local search methods cannot perform an intelligent search as they do not access
other areas of the solution space. To avoid this limitation of being trapped in a local
optimum, researchers use metaheuristic algorithms. Metaheuristics are a set of high-
level algorithms which apply a set of guiding strategies in designing problem-specific
heuristics and local search techniques. Meta-heuristics conduct an intense search over
the solution space starting with one or more input solutions and find a good solution.
Unlike exact methods, the heuristic methods are more popular in solving irregular shape
packing problems. Even though this approach is not able to prove optimality or to
provide any measure of how far it is away from the optimal, still they are capable in
finding a reasonably good solution when they are applied in the cases that arise in
practical scenarios where large-size instances are involved. Therefore, in this study, we
use heuristic methods when solving the three versions of IBPP.

The remainder of this chapter is organised as follows. Section 3.2 describes the mostly
used computational techniques involve in calculating overlaps and containments of ir-
regular shaped pieces when they are placing in the containers. In this section, we also
describe the specific techniques we use in this study by reviewing the pros and cons of
each technique. Section 3.3 presents a variety of basic heuristics and some examples of
popular metaheuristic frameworks used in irregular shape packing problems. The pur-
pose of this presentation is to identify the features of each approach to involve them in
solving the problems wherever they are necessary. In Section 3.4 we provide an overview
Chapter 3 Methodology 47

of implementation details of our algorithm in general. This includes the specification

of the instances and specification of computer tools and hardware used to conduct our
experiment.

3.2 Computational Geometric Techniques

The main purpose of involving geometric techniques is the calculation of overlaps and
containments of pieces within the objects. The overlap calculation evaluates whether
two pieces are overlapped, touched or separated. The containment evaluation checks
that the piece is placed within the object (i.e. bin). Several methods have been adopted
by researchers to perform this task, where the performance and accuracy of them are
different.

3.2.1 Representation of pieces

Some representation techniques approximate the geometric shape of the irregular piece
while some retain the shape as a simple polygon. The accuracy of piece placements is not
100% accurate in approximation methods since it cannot represent the exact shape of
the piece. However, the approximation can reduce the complexity of shapes and makes
the execution of operations simple and efficient. Therefore, there is always a trade-off
between speed and accuracy that both affect the quality of the solutions. Usually the
overlap and containment calculation is based on the technique adapted for representing
the pieces and objects.

As one of the approximation techniques; Segenreich and Braga (1986) and Oliveira
and Ferreira (1993) discretized the whole packing area into several discrete areas (i.e.
a grid). In this case, identifying non-overlapping placement positions is easy since it
can be easily determined by the grid intervals. Dagli and Hajakbari (1990) placed
a grid inside the rectangular enclosure of the piece. This technique has been widely
used in different papers (Oliveira and Ferreira, 1993; Babu and Babu, 2001; Xu et al.,
2013; Guo et al., 2015). The major aim of this technique in common is identifying non-
overlapping placement positions as it can be easily determined by the grid intervals. The
same concept is also used as a pixel / raster method which divides the packing space
into the pixels and each pixel has been assigned a value to indicate if it is occupied
by a piece (Babu and Babu, 2001). In this case, the geometric information of pieces
and stock sheet are transformed to a data matrix following a certain codification, to
represent the irregular pieces, holes and defective areas within the stock sheet. The
intuition of this technique is to convert the geometric information to a coded data matrix.
Bennell and Oliveira (2008) and Bennell and Oliveira (2009) described several features
of those discretized methods. According to their review, discrete/raster methods are
48 Chapter 3 Methodology

simple to code and they are efficient in checking the geometric feasibility. However, the
representation of the shape is not accurate and that sometimes misleads the overlapping
tests by showing pieces as overlapped when they are not actually overlapping. On the
other hand, the technique is memory intensive which may require computers with higher
memory capacity to realise computational efficiency. In order to increase the accuracy,
raster approaches increase the resolution of pixels which leads more towards expensive
computation. Figure 3.1 illustrates a few uses of discretized methods with different
codification schemes as examples.

Figure 3.1: Discrete representation methods (extracted from Bennell and

Oliveira (2008))

Some researchers approximate the irregular shapes into rectangles or convex polygons
(Hopper and Turton, 2001). Heckmann and Lengauer (1995) represented irregular pieces
by their polygonal enclosure. The method has only provided an approximation rather
than representing the exact shape of the piece. When dealing with C&P problems,
accurate representation of pieces directly affects the accuracy of overlap calculation.
Since a given piece is constructed by means of segments, for a complex shape, it requires
a higher number of segments to represent the shape accurately. Encoding an arbitrary
closed curve to a minimal enclosed rectangle (Freeman and Shapira, 1975), packing
several arbitrary polygons and curved pieces into rectangles (Martin and Stephenson,
1988), algorithms to find the convex enclosure with the smallest area for two concave
polygons (Grinde and Cavalier, 1995) are a few attempts for approximating irregular
shapes to a finite number of geometric segments. However, these techniques seem to be
outdated as they have not been used in recent methods for solving irregular shape C&P
problems.

Rocha et al. (2013) introduced circle covering (CC) for overlap computation in nesting
problem allowing free rotations of pieces. The technique represents a piece as a set of
overlapping circles whose sum of areas approximates the piece area. This approximation
technique aims to minimise the necessary number of circles so that the complexity of
Chapter 3 Methodology 49

related operations can be reduced. As the paper states, the main difficulty of CC ap-
proximation is to achieve a lower approximation error with the minimum of a number of
circles. Rocha et al. (2014) discussed the trade-off between piece representation and the
corresponding number of circles required. They demonstrated that the low-resolution
approximation achieves better computational efficiency if the accuracy of the approxi-
mation is compromised. In this case, the low-resolution circle coverings contain a less
number of circles than the high resolution covering which leads to a lower quality of
representation. Details of this technique are available in Rocha (2014).

To take advantage of accurate representation of pieces, several researchers use the exact
shape of an irregular piece as a simple polygon (e.g. Gomes and Oliveira (2006); Bennell
and Song (2008)). However, in order to take the benefit of accuracy, it is necessary to
use an accurate overlap calculation technique (Bennell, 1998).

3.2.2 Overlap calculation and finding feasible placement positions

D-function is used as an efficient joint approach to giving the relative position of a

certain point with respect to an oriented edge (Konopasek, 1981). Mahadevan (1984)
used the D-functions to find the relative position of two oriented edges. When two pieces
are represented accurately by a closed series of vertices and edges, the relative positions
of the two polygons can be identified using this approach (Bennell and Oliveira, 2008).
For example, Mahadevan (1984) and Ghosh (1991b) used trigonometry and D-function
to detect the intersection of edges in the overlapping pieces. According to Bennell
and Oliveira (2008), the approach can further identify whether the points of a piece
are located within another piece. This method maintains a better accuracy in overlap
calculation which is governed by a set of hierarchical tests to evaluate if two polygons
are overlapping or not. A detail description of these hierarchical tests is available in
Bennell and Oliveira (2008). An advantage of using trigonometry with D-function is the
accuracy they provide in overlaps calculations since the polygons are represented in their
exact shape. Bennell and Oliveira (2008) reports that the techniques are not efficient in
two terms; First, they review that the computational effort put on these calculations is
higher with the number of floating points used. Second, they believe it is not efficient
to apply such calculations in an iterative search process since these calculations have to
be repeated from scratch at every step. Due to these reasons, it is rare to find using
D-functions in recently published papers of C&P steam.

Adamowicz and Albano (1976) exploited the use of No-fit polygons to find the minimum
enclosure of the number of nested irregular pieces. Later, the concept was widely devel-
oped and used by many researchers in C&P stream (e.g. Mahadevan (1984), Milenkovic
et al. (1991), Li and Milenkovic (1995), Bennell et al. (2001), Gomes and Oliveira (2006),
Huyao et al. (2007), Burke et al. (2007), Bennell and Song (2008), Imamichi et al. (2009)
,Bennell and Song (2010), Burke et al. (2010), Leung et al. (2012), Junior et al. (2013),
50 Chapter 3 Methodology

Hu et al. (2015) Martinez-Sykora et al. (2015), Leao et al. (2016), Martinez-Sykora et al.
(2017)). These examples reveal that no-fit polygon (NFP) is the most popular technique
used in literature recently. Bennell et al. (2001) and Bennell and Oliveira (2008) denote
NFP of two polygons A and B as N F PAB the resulting polygon from a sliding operation
between A and B. Polygon orientations are fixed. When polygon A is in a fixed position,
then polygon B is considered to be the tracing polygon that moves along the perimeter
of A so that B slides around A. Then N F PAB is described as the path of touching
points of B with A, marking the locus of a reference point on B as it traces around the
boundary of A (Bennell et al., 2001; Bennell and Oliveira, 2008). Figure 3.2 illustrates
how the locus of the reference point on polygon B creates the N F PAB as polygon B
traces around polygon A.

Figure 3.2: NFP (A,B): The reference point on B traces around A (Bennell
et al., 2001))

When the position of A is known and B is positioned so that its reference point is
inside N F PAB , then it overlaps with polygon A. If the reference point of B is on the
boundary of NFP, then B touches A. When the reference point is outside of NFP, it
shows that A and B neither overlap nor touch (Junior et al., 2013). As the locus of
points where the reference point of polygon B should be placed in contact with polygon
A, its exterior, represents the set of feasible coordinates for the reference point of polygon
B so that it doesn’t overlap A (Leao et al., 2016). Since this is a promising technique
for overlap calculation and placement decision of irregular pieces, it is vital to find an
efficient technique to derive NFPs. Usually, the calculation of NFP for purely convex
polygons is simple (Cuninghame-Green, 1989). This is the most simple case in terms
of complexity of polygonal shapes. However, Bennell and Oliveira (2008) illustrates
some cases (illustrated in Figure 3.3) which require a considerable effort to calculate the
correct NFP. Bennell and Song (2008) referred those as the degenerate cases.

A related concept called inner-fit-rectangle was discussed in Gomes and Oliveira (2001).
They denote the rectangle it creates when an irregular piece slides along the internal
contour of the rectangle object as the inner-fit-rectangle. Extending this idea, Liu and
Chapter 3 Methodology 51

Figure 3.3: Different cases of generating NFPs (Bennell and Oliveira, 2008): a)
B fits inside the concavity space of A, b) B can slide into the concavity of A in
one direction, c) B fits only in one point inside the concavity of B

Figure 3.4: Inner-fit Polygon (extract from Bennell and Song (2010))

He (2006) described Inner-fit NFP as the locus of the reference point of the piece when
the piece slides around the inside boundary of an irregular container (as illustrated in
3.4). In Bennell and Song (2010)’s paper uses the same concept and called it the Inner-
fit polygon (IFP). Since IFP finds the feasible placement path of a polygon, Bennell and
Song (2010) states that IFP can be useful for packing pieces inside an irregular shape
(e.g. cutting irregular shapes natural leather hides).

In order to generate NFPs, Mahadevan (1984) used a sliding algorithm to find the
movement path of the tracing polygon around the fixed polygon. However, this only
finds the outer boundary of the NFP. The approach does not work with holes and does
not work for the cases illustrated in Figure 3.3.

Milenkovic et al. (1991) and Bennell (1998) use the Minkowski Sum concept to generate
NFPs. They describe the approach as Minkowski difference. Using Ghosh (1991a)’s
Boundary Addition Theorems, Bennell (1998) described how NFP can be generated for
the case where one polygon has concavity and the other polygon is convex. In order to
deal with the case where neither polygon is convex, Bennell et al. (2001) proposed an
approach that first replaces B by its convex hull and solves the simple case of deriving
the NFP of a convex polygon and a polygon with the concavity, as mentioned in Ghosh
52 Chapter 3 Methodology

(1991a). Then the resulted NFP (this is considered as an intermediary NFP) is repaired
to obtain the real NFP. Details of this can be found in Bennell et al. (2001) and Bennell
and Oliveira (2008).

Burke et al. (2007) introduced a different method called robust orbital method to generate
NFPs using the Mahadevan (1984)’s sliding algorithm. This resolved the limitations of
Mahadevan (1984)’s approach related with the degenerate cases. Later, Bennell and
Song (2008) presented a better approach based on the Minkowski Sum to generate true
boundary of the NFP, including holes, slits and exact fits, without generating the convex
hull of one polygon and then repairing the resulting intermediate NFP as in Bennell et al.
(2001). The authors presented an empirical analysis to demonstrate that this method is
reasonably efficient. This method is also capable of generating NFP for the degenerate
cases and the polygons containing spiralling concavities.

Based on the quick method to find the NFP of two convex polygons proposed by
Cuninghame-Green (1989), Watson and Tobias (1999) and Agarwal et al. (2002) de-
composed simple polygons into convex sub-polygons. The first step of the approach is
to decompose each given polygon into convex sub-polygons and generate the NFP of
each pair of sub-polygons. Next, they combine these NFPs of the sub-polygons and
generate the final NFP of the given two polygons. In comparison with other methods
of generating NFP, the method possesses higher complexity in the decomposition and
combining operations.

Stoyan et al. (2001, 2004) developed and applied Phi-functions to formulate the mutual
position of simple geometric objects. A specific research group in C&P stream involves
using Phi-functions to describe the placement and interactions of geometric objects.
The phi-function measures the relative position of objects. The phi-function normalizes
its value as the Euclidean distance between the two phi-objects (i.e shapes) (Bennell
et al., 2010; Chernov et al., 2012). At early stages, Stoyan et al. (2004) derived the
phi-function for primary objects such as rectangles, circles, regular polygons, convex
polygons as well as the compliment of these shapes, based on trigonometric theories.
Bennell et al. (2010) pointed out the effectiveness of Phi-functions in terms of measur-
ing the degree of overlap or distance between shapes. Even though Phi-functions can
be used directly in the mathematical modelling part of the problem, it is necessary to
derive the corresponding function for the objects beforehand. Chernov et al. (2012)
have derived ready-to-use basic phi-functions to construct radical free phi-functions for
arbitrary shaped 2D objects. The major disadvantage of this method is the limitation
of constructing phi-functions for a selected set of pieces; e.g. ellipses. Extending the
concept of phi-functions, a new concept; quasi-phi-functions is introduced by Stoyan
et al. (2015b). Quasi-phi-functions are simpler than phi-functions. For example, it is
possible to develop Quasi-phi-functions even for the objects that construction of phi
functions are limited. (Stoyan et al., 2015b). Stoyan et al. (2015b) highlight that the
new method would lead to a larger set of parameters and extra variables due to the
Chapter 3 Methodology 53

derivation of new functions. The Stoyan et al. (2015b)’s quasi-phi-functions have also
been successfully implemented with various types of objects (including ellipses) that can
be continuous rotated. Stoyan et al. (2015a) used phi-functions and quasi-phi-functions
to solve placement problem of irregular shapes using a non-linear programming model.
Both Phi-functions and Quasi-phi-functions approaches deal with an accurate repre-
sentation of objects with no approximation. However, the approach becomes inefficient
when highly irregular objects where the compositions of those objects are complex, those
objects are compiled from basic objects (Stoyan et al., 2015b,a).

In summary, our methodology of addressing the geometric aspect involved mainly the
steps in representing pieces, identifying the data structures that the algorithm is imple-
mented and selecting a proper overlap calculation technique to find feasible placement
positions for pieces.

As described in Section 3.2.1, there are three ways of modelling the positioning of pieces
in a stock sheet and accordingly the representing method is decided. The first approach
uses the continuous positioning of the pieces along the area of the stock sheet, while
the second approach considers discretized positioning. The third approach is based on
positioning with semi-discrete geometric representation where one axis of the placement
coordinate system along the stock sheet is discrete and the other axis is continuous. Out
of the three, we used the continuous positioning which is based on geometric representa-
tion when modelling the problems. This approach addresses one of the objectives of our
study by guaranteeing accurate placement of pieces without any approximation. Also,
it was evident that existing research that used geometric representation for irregular
packing problems have generated better results than the other two approaches (Egeblad
et al., 2007; Imamichi et al., 2009; Umetani et al., 2009; Bennell and Song, 2010; Burke
et al., 2010; Elkeran, 2013).

Irregular pieces and rectangular bins were the two main elements of the model of our
problem. An irregular piece was represented as a closed series of vertices and edges
oriented in anti-clockwise direction. Hence, the model allowed irregular shapes to be
defined as simple polygons. Any other geometric cases such as partial solutions and
holes were also defined as simple polygons. Accordingly, the data structure of a piece
is described by a list of vertices in the polygon, an orientation angle which defines the
rotation of the piece with reference to the initial orientation and a reference point which
defines as the bottom left point of the enclosing rectangle of the piece at its current
orientation.
54 Chapter 3 Methodology

Figure 3.5: Shape representation of a piece

The geometric representation enabled close nesting of pieces, hence brings the higher
accuracy of the solution. However, the higher number of vertices consume longer com-
putational time; since the calculation of no-fit polygon need more time when there is a
higher number of edges in the corresponding polygons. Due to this reason, we eliminated
the very narrow concavities of some polygons (i.e. especially in the merged polygon of
irregular shapes inside the container) without affecting to the accurate placement of
pieces.

Similar to pieces, a bin was described by a list of four vertices (i.e. rectangular bin) and
a reference point (bottom left point of the bin).

The third step; the overlap calculation was performed involving the concept, no-fit poly-
gons (NFPs). As we reviewed in Section 3.2.2, the boundary of NFP represents the path
of feasible placement positions of a polygon with respect to the other neighbour poly-
gon. Therefore by finding this path, we can find a feasible placement positions of a
polygon with respect to the other so that those two polygons are neither overlapped
nor separated. As the most famous tool in the recent literature, the recently developed
methods for generating NFPs work efficiently with various types of polygons, without
compromising the benefit of accuracy (Burke et al., 2007; Bennell and Song, 2008; Ben-
nell and Oliveira, 2008; Burke et al., 2010). We used the NFP generation method based
on the Minkowski Sum approach by Bennell and Song (2008), to calculate NFPs in our
constructive algorithms, since it was able to facilitate the expected requirements of our
algorithms such as dealing with jigsaw-type shapes and interlocking polygons.
Chapter 3 Methodology 55

3.3 Heuristics-based Approaches

In this section, we describe some design principles of heuristics and metaheuristics ap-
proaches and briefly introduce some of the key facts to be understood when implementing
those approaches.

Heuristics based approaches find good solutions to large-size problems at acceptable

computational costs. They are simple in understanding, easy in implementation and less
costly in computational time than the exact algorithms. Generally, the heuristics based
methods are distinguished between constructive heuristics and local search methods.

3.3.1 Constructive heuristics

Constructive heuristics always generate a solution from scratch by adding the compo-
nents which require to build up a solution. Initially, it starts with an empty partial
solution and then it is gradually extended until the solution is complete. As reviewed
in Chapter 2, Sections 2.5.1 and 2.6.1, C&P researchers have used constructive methods
to generate solutions by adding pieces to the partial solution, one after the other. In
this case, the solution of a constructive heuristic contains a sequence of partial solutions.
The heuristic criteria of the algorithm decide the extension required at each insertion
of components. The extension made at one insertion affects the next extension of the
partial solution.

The design structure of constructive heuristics is straight forward. Some methods insert
each piece following a given order of pieces. Also, there is another method which chooses
the next insert piece using dynamic selection (see. Chapter 2, Section 2.5.1.2). In this
case, the constructive methods follow a selection procedure to select the piece to be
positioned next, by evaluating the best partial solution among other partial solutions
correspond to the pieces to be placed next. For C&P problems, the decision related to
constructive heuristics are attached with the insertion order of pieces and the placement
rule used in the algorithm to insert the new piece. Chapter 2 reviews a set of well-known
ordering and placement strategies used in irregular shape packing problems.

3.3.2 Local search methods

As described in Talbi (2009), the local search is operated on the appropriately defined
neighbourhood of the selected feasible solution s. The neighbourhood is a set of solutions
that are close to s. They are derived from s so that they share some amount of structure
of solution s. The constructive heuristic can be used to find neighbour solutions, by
making small modifications; called moves.
56 Chapter 3 Methodology

As a basic local search algorithm, iterative improvement tries to explore the neigh-
bourhood as follows. Starting with an initial solution the algorithm moves from one
neighbour to another to find a better solution until a termination criterion is met. If a
neighbour has an improved objective value than the current solution, then the neighbour
becomes the new current solution. The search process is iterated until no such accessible
improved neighbour is found (see Figure 3.6). The solution found by the local search
method may not be the best solution to the problem, but it finds a good solution.

Figure 3.6: Iterative Improvement Local search (Decent) behavior (extracted

from Talbi (2009))

Even for a basic local search method, it requires deciding the way of generating the
initial solution and the procedure for producing neighbours. Usually, the initial solution
is produced by the constructive heuristic following a random approach (e.g. a random
order of pieces) or a predefined strategy (e.g. a predefined rule of ordering pieces). The
selection of an initial solution may cause a significant effect in a simple local search
heuristic. However, more powerful improved local search techniques with broad search
space will provide a solution which is almost independent from the initial solution.

In addition to the iterative improvement, Talbi (2009) discusses some other variations of
basic local search algorithms. In the above case, it accepts the solution corresponds to
the first improving move. This is referred as the first-improvement way of choosing the
improved neighbour. Instead of selecting the first improving solution, there is another
way called best-improvement, in which the whole neighbourhood is searched before a
solution is accepted as a better one. As Gonzalez (2007) states, the local search methods
based on the first-improvement is comparatively faster in finding individual improvement
steps since the full neighbourhood will only be used when the current solution is the
local optimal.

The local search algorithms are intuitively understandable and easier to implement than
exact algorithms. Most importantly, they are very useful when solving problems with
large instances. However, a local search algorithm only finds a local optimal solution.
Chapter 3 Methodology 57

The local optimal solution is the best solution that can be found considering a certain
neighbouring configuration that the local search being applied. This situation in existing
literature has been referred as being stuck in local optima which researchers consider
as a limitation of local search. In order to overcome this, researchers use repeated local
search; known as Restarts, or more complex algorithm schemes based on iterations such
as Iterated Local Search, Tabu Search and Simulated Annealing.

3.3.3 Metaheuristic Algorithms

Metaheuristics are problem-independent heuristic frameworks which guide the heuris-

tics which are problem-specific. Meta-heuristics find good solutions in an acceptable
computational time. They provide reasonable solutions to the problem when the search
being stuck in local optima. Meta-heuristics resolve this by allowing worsening moves or
generating new starting solutions for the local search in an intelligent manner than the
random initial solutions in multi-start local approaches. A key feature of metaheuristic
frameworks to improve the solution search process is the implementation of random-
ization while controlling intensification and diversification of the search process. The
intensification is used when there is a solution which needs to be further explored by
its neighbours. This is referred as the exploitation of the best solutions found. This is
usually implemented with the Local Search techniques so that the selected promising
regions are explored thoroughly to find better solutions.

In the diversification phase, instead of conducting a search in a particular area which is

not further appealing, the search is guided towards a new area. Here the non-explored
regions are explored to ensure that most regions of the search space are explored. This
is usually implemented with random search in the search space.

Different metaheuristic algorithms have different techniques to implement these intensi-

fication and diversification phases. Before applying a metaheuristic, it is also important
to know the size of input instances the algorithm is supposed to solve and the level
of complexity of the problem. There are some hard problems which are solvable using
the exact approach if they are with small instances. Also, the time it takes to search
and obtain a reasonably acceptable solution is an important concern in the selection of
metaheuristic.

Next, we briefly describe the main features of metaheuristic approaches focusing partic-
ularly on those which are related to the irregular shape C&P problems and the solution
methods present in this thesis.
58 Chapter 3 Methodology

3.3.3.1 Simulated Annealing (SA)

According to Gendreau and Potvin (2010), the principle behind SA is derived from
physics. The idea is to simulate the annealing process of solid, where the solid changes
its structure as the temperature cools, which was proposed by Metropolis et al. (1953).
Kirkpatrick et al. (1983) introduced Simulated Annealing in solving optimisation prob-
lems. In combinatorial optimisation scenario, the elements of problems can be mapped
to the elements of the physical cooling process.

The objective of SA is to escape from local optima thereby exploring the opportunity
of searching a broader search space. The general procedure of SA is described in Talbi
(2009) as follows. Starting from an initial solution which becomes the current best
0
solution (x) at the first step, a tentative neighbour solution (x ) is generated at each
iteration with respect to x. If the neighbour solution improves the value of the objective
0
function (i.e. f (x ) < f (x)), then it is accepted. Otherwise, the non-improving solution
is accepted based on a given probability which depends on the amount of improvement
0
of the objective function f (x) − f (x ) and the control parameter (T). At higher values
of T, the probability of acceptance of non-improving solutions are high, hence almost
all the solutions are accepted. When the algorithm finds a local optimum at a certain
T value, T is decreased according to the cooling schedule, and continue the local search.
When T is gradually decreased, the probability of acceptance decreases, hence less non-
improving solutions is accepted. Accordingly, at the end of the procedure, the best
solution is found.

In comparison to the basic local search which is getting stuck at local optima. SA solves
this issue by allowing more time for less quality moves during the search process. In
other words, it allows some high-cost solutions so that the search can escape from local
optima. Also, SA selects a random move from the neighbourhood while basic local
search selects the best move from all those available at higher temperatures the search is
analogous to random search. Then the temperature is decreasing though out the search.
When it is zero, then only the better moves will be accepted similar to the basic local
search.

3.3.3.2 Tabu Search (TS)

According to Gendreau and Potvin (2010), Tabu Search was introduced as an optimisa-
tion approach in Glover (1986). The general idea of TS is to search a space by choosing
a point, and then going to its best neighbour which is not in the tabu list. The tabu
list records forbidden moves, which are referred to as Tabu moves. As key features, the
method uses memory to conduct a local search. As described in Dréo (2006) and Gen-
dreau and Potvin (2010), the Tabu search prevents revisiting and maintains increased
diversity in exploration which helps to escape from local optima.
 Chapter 3 Methodology 59

Tabu search uses a search approach which is similar to the best-improvement local
search. However, it escapes from local optima by accepting non-improving solutions
when there is no improvement in neighbours. Unlike in SA, TS usually explores the
whole neighbourhood in a deterministic way.

A significant difference between simulated annealing and tabu search is their level of use
of memory. Simulated annealing has no memory, whereas Tabu Search has the ability
to refer information stored about previous moves. As described in Talbi (2009), there
are three purposes of using memory in TS. First, it prevents the search from revisiting
previously visited solutions. Second, it helps diversification by searching the unexplored
areas of the search space. Third, TS uses it memory structure to store best-found
solutions and explore them further (i.e. application of intensification).

3.3.3.3 Iterated Local Search (ILS)

As described in Gendreau and Potvin (2010), the general idea of Iterated Local Search
is to apply iteratively a local search until a local optimum is reached and then apply
a perturbation (also known as a kick), in order to escape from the local optimum and
explore further solution space. In this case local search only moves towards better
solutions. The main role of the kicks is to change the current solution enough while
keeping the good properties of the current local optimum. According to Lourenço et al.
(2010), the idea of ILS has been proposed in many forms in its early research. One early
implementation of ILS was stated as iterated decent in Baum (1986) and Baum (1987).

Basic procedure:

Algorithm 1: ILS
1 Initialize;
2 Generate initial solution, x ∈ S;
3 x∗ ← LocalSeach(x);
4 Repeat
0
5 x ← P erturbation(x∗ , history);
0 0
6 x∗ ← LocalSearch(x );
0
7 x∗ ← AcceptanceCriterion(x∗ , x∗ , history);
8 Until termination condition met

The entire procedure is repeated until a stopping criterion; generally, a limit on compu-
tation time or the total number of iterations is satisfied.

As described in Talbi (2009), the ILS procedure allows using single solution-based meta-
heuristics (e.g. simple iterative improvement local search, TS, SA). However, it is vital
to control the strength of perturbation when implementing ILS as described in Luke
(2013). If perturbation is too strong, then it is close to random restart behaviour. If
60 Chapter 3 Methodology

Figure 3.7: Working principle of iterated local search (extracted from Talbi
(2009))

perturbation is too weak, then the algorithm will not perform as expected since the
Local Search may undo the effect of the perturbation.

3.3.3.4 Variable Neighbourhood Search (VNS)

Variable Neighbourhood Search is proposed in Mladenović and Hansen (1997) as a search

method which explores the neighbourhood to find a better solution. As described in
Figure 3.8, VNS explores a set of neighbourhoods and visits different local optima. If
a search method can explore all possible neighbourhood structures, then finding global
optimum is possible. However, finding all possible neighbourhood structures is hard for a
complex problem. In this context, VNS deals with different neighbourhood structures by
changing the neighbourhood during the search process. When implementing VNS, the
algorithm generates different neighbourhood structures such as Cross, Swap, Exchange,
and Flip. The movement between neighbourhoods is either deterministic, random or
both.

3.3.3.5 Genetic Algorithm (GA)

GA is a population-based search metaheuristic which is inspired by the evolution theory

of the natural world which follows the principle; survival of the fittest. GA was formally
recognised by Holland in 1975 (Holland, 1992) (the book was originally published in
1975). A solution in search space is referred as individuals (or chromosomes) that genes
of a chromosome represent parts of the solution or solution construction. The numerical
evaluation of an individual is known as its fitness which is an evaluation value of the
quality of each individual.
Chapter 3 Methodology 61

Figure 3.8: Variable neighbourhood search in two neighbourhood (extracted

from Talbi (2009))
With respect to neighbourhood 1, the first local optimum is derived. Then the second
local optimum in neighbourhood 2 is derived from the first local optimum.

When searching occurs, the population is being replaced by new individuals (chromo-
somes). The parents are selected for mating based on the fitness. The aim is to propagate
the good properties of a chromosome into new individuals so that the population can
be improved when generations progress. The new chromosomes are made by applying
the below two types of genetic operators that mimic the evolution of natural systems;

1) Crossover operator: This considers two chromosomes (i.e. parents), and combine their
features by changing their genes to produce new chromosomes (i.e. children). Usually,
two parents make two children. The crossover probability (Pc ) parameter describes the
rate of crossover happens. If Pc = 0% then all offsprings become exact copies of parents.
If Pc = 100% then all offsprings are generated by crossover.

2) Mutation operator. This introduces diversity and allows enter new genetic features
to the population. After crossover, the chromosomes are subject to mutation. Mutation
supports exploration during a search while crossover supports exploitation. There are
different forms of mutation operators; such as mutation operators for binary or integer
representation (e.g. bit flipping) and mutation operators for permutation representation
(e.g. swap mutation, inversion mutation, scramble mutation) as described in Eiben and
Smith (2013). The main difference between those two methods is that the mutation in
binary/integer representation mutates each gene individually while the other consider
mutation of the whole chromosome rather focusing each gene individually. Similar to
(Pc ), the mutation probability (Pm ) describes the how often of the part in the chromo-
some is mutated.

In order to handle a GA search, we have to consider following specification of the oper-

ators and parameters.

• Population Size: Number of chromosomes in the population

 62 Chapter 3 Methodology

• Fitness function: The function which evaluates the quality of chromosomes

• Crossover Operators: The mechanism of crossover to produce children

• Crossover Probability: Chance of crossover; generally 0.6-1.0

• Mutation Operators: The mechanism of mutating the features within chromosomes

• Mutation Probability: Chance that a part of a child is changed randomly

• Elitism: The proportion of best chromosomes be ensured a position in the next

generation

The general structure of a simple GA is represented as follows

Algorithm 2: Genetic algorithm

1 Produce an initial population of individuals;
2 Evaluate the fitness of all individuals;
3 while termination condition not met do
4 select fitter individuals for reproduction;
5 recombine between individuals;
6 mutate individuals;
7 evaluate the fitness of the modified individuals;
8 generate a new population;
9 end

According to Talbi (2009), GA requires handling two main search strategies. First,
the selection strategy determines which parents are selected for reproduction. Usually,
the chromosomes having better fitness has more chance of being parents. However, GA
usually allows some worst-fit individuals to be selected as well. Second, the reproduction
strategy determines which type of mating operators are used. It is possible to use one
or more recombination operators in GA, following the rules below. The combination
operators must inherit the same features from each parent when making children and the
outcome of it must produce valid chromosomes according to the solution representation
technique used. Both strategies support GA search to keep a balance between the
diversity and convergence speed.

3.3.3.6 Greedy Randomised Adaptive Search Procedure (GRASP)

GRASP is a multi-start metaheuristic that combines constructive heuristics and local

search, which was first proposed in Feo and Resende (1989). As explained in Feo and
Resende (1989), Resende and Ribeiro (2010) and Resende and Ribeiro (2016), GRASP
consists two phases; solution construction and solution improvement. These two phases
are performed iteratively and the best solution is always updated and stored in the
memory till the algorithm terminates.
Chapter 3 Methodology 63

The solution construction method is characterised by a greedy constructive heuristic

and randomization. The constructive heuristic constructs the solution by adding each
element of the solution to the current partial solution, in which the choice of the next
element to be added to the partial solution is picked by randomly from a candidate list
of elements. This list of candidates is denoted as the Restricted Candidate List (RCL).
At each construction iteration, the RCL is formed with all the elements which can be
added to the current partial solution in the order of their short-term benefit which
is measured by a greedy function. At each iteration of the construction phase, each
candidate element is updated with a score (i.e. myopic benefit of adding that element
to the partial solution) measured by the greedy function.

3.3.3.7 Parameter tuning

When applying metaheuristics, it requires tuning many parameters used within the
algorithm. The performance of the algorithm heavily depends on its parameters hence
it shows a significant difference in performance improvement by a well-tuned algorithm.
According to Peng-Yeng (2012), the optimal tuning values of the parameters are mainly
based on the problem, the instance and the search time assigned by the user.

As mentioned in Peng-Yeng (2012), the parameter tuning can be done in two different
ways; on-line parameter tuning and off-line parameter tuning. In the on-line tuning, the
parameter values are dynamically manipulated and updated within the execution of the
metaheuristic. The on-line tuning techniques are typically used as a machine learning
technique, which deals with a large number of instances. In the off-line parameter tuning,
the parameter values assigned to different parameters are pre-defined before the running
of the metaheuristic.

As mentioned in Peng-Yeng (2012), the earlier use of off-line parameter tuning considered
only the one-by-one parameter tuning (each parameter in tuned independently) which
doesn’t guarantee the optimal combination of parameter values to maximise the overall
performance of the algorithm. To resolve this, researchers use the design of experiments,
which uses different levels of potential values for each parameter in the algorithm. With
a full design of each level in each parameter (including the combinations) best levels
are determined. This is reasonably a manageable method if the designer has some
understanding about possible levels of each parameter even though it takes a significant
time to conduct the parameter tuning.

As other methods, Talbi (2009) formulates the tuning parameters as the optimisation
problem and try to find the parameter values. Hutter et al. (2009) propose automatic
algorithm configuration framework called PramILS, as a procedure of automatic tuning
of parameters. Recently, many papers have been published on parameter tuning and
Peng-Yeng (2012) presents most of those approaches in his publications.
64 Chapter 3 Methodology

Another promising approach for parameter tuning is Irace: Iterated Race for Automatic
Algorithm Configuration. It is used as a package to support the configuration of algo-
rithm parameters automatically. It finds the most suitable settings for a given set of
instances when solving an optimisation problem. The package was developed on the
F-race package proposed by Balaprakash et al. (2007) and later improved by Birattari
et al. (2010). In fact, Irace is a generalization of the iterated F-race procedure which
runs in R. As a useful tool for parameter tuning, López-Ibáñez et al. (2016) states Irace
as a user-friendly, easy and simple tool to tune parameters even for users who are not
familiar with R. In terms of limitations, López-Ibáñez et al. (2016) highlight the impor-
tance of enough assignment of computational effort for the configuration tasks in order
to result in a better configuration than a random configuration. In general, such auto-
matic configuration methods are hard to use for parameter tuning if problem instances
are highly computationally expensive (López-Ibáñez et al., 2016).

According to Birattari et al. (2002) and Birattari (2009), in the majority of the cases,
the metaheuristics are tuned using the trial-and-error procedure guided by some rules
of thumb. Despite most of the research papers adopted this method, trial-and-error ap-
proach presents following drawbacks. It may need the involvement of a skilled designer,
usually the programmer who implemented the algorithm to make necessary changes
when tuning is required. Also, this method is error-prone and tedious.

Out of these approaches, in this study, we use constructive heuristics to build the solution
by placing one piece at a time, investigating the applicability of different placement
strategies for pieces and sequencing strategies for bins and pieces. Both the local search
and metaheuristic approaches are used to find an improved solution by working over the
search. As described in this chapter, this involves the key tasks of identifying the solution
space, setting up a mechanism to identify neighbourhoods, generating techniques to move
from solution to solution and maintaining the dynamic balance between diversification
and intensification of the search process.

3.4 Implementation

All the constructive and improvement algorithms were implemented by coding in Vi-
sual C++ 2012 and boost C++ 1.63 environments. For each problem, we experimented
with different types of constructive algorithm by testing the different placement and se-
quence strategies. As an example; we implemented six different construction algorithms
for solving the 2D ISBSBBP and compared the performance of them with each other.
Similarly, the search algorithms were also implemented and tested to find better results.

Each algorithm was run on one 2.6 GHz CPU (Intel Sandy Bridge) with 4GB memory
in the Southampton University Iridis 4 computer environment. No algorithm contained
any part of parallel programming.
Chapter 3 Methodology 65

When conducting the experiments, we ran the experiments for multiple trials and average
values were reported if any algorithm contained random components. Once algorithms
were implemented, we conducted parameter tuning for each problem. Details of these
parameters for each case are included in proceeding chapters.

3.4.1 Instances

The different variants of the algorithm were tested on the instances with irregular shapes
available in the literature. In this case, we used the instances published by EURO Special
Interest Group on Cutting and Packing (ESICUP) web site (http://paginas.fe.up.pt/ es-
icup/datasets).

In Table 3.1 we summarise the characteristics of each nesting instance used in solving
2D ISBSBPP in Chapter 4. In order to compare with other published results, we also
use two more sets of instances with Jigsaw type pieces from ESICUP website. The
first set; denoted as JP1, includes a collection of 540 instances where all the pieces of
them are convex. The specifications of those instances are presented in Table 3.2. The
540 instances are divided into 18 classes. The second set; denoted as JP2, includes
480 instances and they are divided into 16 in which some pieces have concavities. The
specifications of those 480 instances are presented in Table 3.3.

For the experiments in solving 2D IMBSBPP in Chapter 5, we also involve the nesting
instances. Since some of the instances contained less number of pieces to be applied in
a bin packing problem (i.e. with the combination of large and small size bins), we make
sure that we consider enough number of pieces to have more than one bin in a solution.
In order to achieve this, we multiply the number the pieces in those instances so that it
can have a sufficient number of pieces to be considered in a bin packing problem with
multiple bin sizes. Table 5.1 presents the details of these instances. A similar approach
is used when deciding the instances used for the experiments discussed in Chapter 6.
The details of these instances and their uses are described in Section 6.4.1.
Chapter 3 Methodology

Table 3.1: Nesting Instances (extracted from Martinez-Sykora et al. (2017))

Instance Types of Pieces Number of Pieces Average of Vertices Allowed Rotation Angles Problem Type Authors
dighe2 10 10 4.7 0 Jigsaw Puzzle Dighe and Jakiela (1995)
fu 12 12 3.58 0-90-180-270 Artificial, convex Fujita et al. (1993)
poly1a 15 15 4.6 0-90-180-270 Artificial Hopper (2000)
dighe1 16 16 3.87 0 Jigsaw Puzzle Dighe and Jakiela (1995)
mao 9 20 9.22 0-90-180-270 Garment Bounsaythip and Maouche (1997)
albano 8 24 7.25 0-180 Garment Albano and Sapuppo (1980)
jakobs1 25 25 5.6 0-90-180-270 Artificial Jakobs (1996)
jakobs2 25 25 5.36 0-90-180-270 Artificial Jakobs (1996)
shapes2 7 28 6.29 0-180 Artificial Oliveira and Ferreira (1993)
poly2a 15 30 4.6 0-90-180-270 Artificial Hopper (2000)
poly2b 30 30 4.53 0-90-180-270 Artificial Hopper (2000)
shapes1 4 43 8.75 0-180 Artificial Oliveira and Ferreira (1993)
shapes0 4 43 8.75 0 Artificial Oliveira and Ferreira (1993)
poly3a 15 45 4.6 0-90-180-270 Artificial Hopper (2000)
poly3b 45 45 4.6 0-90-180-270 Artificial Hopper (2000)
swim 10 48 21.9 0-180 Garment Oliveira and Ferreira (1993)
poly4a 15 60 4.6 0-90-180-270 Artificial Hopper (2000)
poly4b 60 60 4.6 0-90-180-270 Artificial Hopper (2000)
trousers 17 64 5.06 0-180 Garment Oliveira and Ferreira (1993)
poly5a 15 75 4.6 0-90-180-270 Artificial Hopper (2000)
poly5b 75 75 4.57 0-90-180-270 Artificial Hopper (2000)
shirts 8 99 6.63 0-180 Garment Dowsland et al. (1998)
66
 Table 3.2: Characteristics of the JP1 instances (extracted from Lopez-Camacho et al. (2013b))
Piece
Percentage Percentage Average Average Optimal
Average Size
Average of of of of Ratio (number
Problem Type Number of Instances Number of Pieces Piece Standard
Rectangularity Right Orthogonal Concavity area/convex of
Size Devia-
Angles Sides Degree hull objects)
Chapter 3 Methodology

tion
Convex 2D problem instances
Minimum 0.033 0.014 0.35 11 34 1 1 2
Average 0.154 0.100 0.68 42 65 1 1 5.94
Maximum 0.354 0.280 1 100 100 1 1 15
Average of instances per type
Conv A 30 30 0.100 0.069 0.70 42 68 1 1 3
Conv B 30 30 0.333 0.162 0.87 67 84 1 1 10
Conv C 30 36 0.167 0.124 0.68 36 63 1 1 6
Conv D 30 60 0.050 0.036 0.57 23 51 1 1 3
Conv E 30 60 0.050 0.035 0.41 12 38 1 1 3
Conv F 30 30 0.067 0.050 0.59 29 57 1 1 2
Conv G 30 36 0.332 0.156 0.87 67 83 1 1 Unknown
Conv H 30 36 0.333 0.158 0.86 67 84 1 1 12
Conv I 30 60 0.053 0.017 1 100 100 1 1 3
Conv J 30 60 0.067 0.034 0.83 68 83 1 1 4
Conv K 30 54 0.154 0.150 0.63 34 60 1 1 6
Conv L 30 30 0.100 0.075 0.51 23 50 1 1 3
Conv M 30 40 0.125 0.102 0.55 28 55 1 1 5
Conv N 30 60 0.033 0.024 0.62 32 60 1 1 2
Conv O 30 28 0.250 0.223 0.57 27 58 1 1 7
Conv P 30 56 0.143 0.173 0.49 17 43 1 1 8
Conv Q 30 60 0.250 0.053 0.89 51 76 1 1 15
Conv R 30 54 0.167 0.153 0.63 36 62 1 1 9
67
Chapter 3 Methodology

Table 3.3: Characteristics of the JP2 instances (extracted from Lopez-Camacho et al. (2013b))
Piece
Percentage Percentage Average Average Optimal
Average Size
Average of of of of Ratio (number
Problem Type Number of Instances Number of Pieces Piece Standard
Rectangularity Right Orthogonal Concavity area/convex of
Size Devia-
Angles Sides Degree hull objects)
tion
Non-Convex 2D problem instances
Minimum 0.044 0.036 0.38 6 27 1.004 0.834 2
Average 0.160 0.135 0.59 26 50 1.130 0.930 5.9
Maximum 0.333 0.314 0.84 60 74 1.560 0.987 12
Average of instances per type
Nconv A 30 35-50 0.074 0.062 0.60 28 52 1.12 0.935 3
Nconv B 30 40-52 0.214 0.158 0.69 38 58 1.22 0.923 10
Nconv C 30 42-60 0.123 0.111 0.59 25 49 1.11 0.939 6
Nconv F 30 35-45 0.051 0.045 0.53 20 46 1.10 0.940 2
Nconv H 30 42-60 0.245 0.163 0.73 46 64 1.15 0.944 12
Nconv L 30 35-45 0.076 0.065 0.47 16 41 1.10 0.941 3
Nconv M 30 45-58 0.099 0.092 0.50 20 46 1.07 0.956 5
Nconv O 30 33-43 0.186 0.190 0.51 19 46 1.10 0.940 7
Nconv S 30 17-20 0.106 0.097 0.45 10 33 1.16 0.918 2
Nconv T 30 30-40 0.293 0.239 0.60 26 51 1.24 0.916 10
Nconv U 30 20-33 0.197 0.161 0.55 17 44 1.19 0.888 5
Nconv V 30 15-18 0.306 0.236 0.62 27 54 1.09 0.936 5
Nconv W 30 24-28 0.155 0.097 0.78 53 69 1.12 0.931 4
Nconv X 30 25-39 0.097 0.072 0.66 32 53 1.17 0.895 3
Nconv Y 30 40-50 0.135 0.129 0.61 25 51 1.09 0.943 6
Nconv Z 30 60 0.200 0.234 0.54 19 45 1.09 0.940 12
68
Chapter 4

Homogeneous Bin Packing

Problems with Irregular Shapes

4.1 Introduction

This chapter is devoted to addressing the Two-Dimensional Irregular Bin Packing Prob-
lem with homogeneous bins. Following Wäscher et al. (2007)’s typology, the problem
can be also denoted as 2D Irregular Single Bin Size Bin packing Problem (2D ISBSBPP).
The problem arises in many industries where there is a requirement for irregular shape
pieces to be cut from multiple fixed dimension stock sheets such as metal, wood, paper
and plastic. Despite the many real-life applications, the problem has received relatively
little attention in the literature when compared to the rectangle bin packing problem
or the irregular shape strip packing problem. In this chapter, we describe a heuristic
solution method for the 2D ISBSBPP where the stock sheets are homogeneous and a
feasible solution is found when all the pieces are packed in the bins with no overlap
between pieces. The objective is to minimise the total number of bins needed to cut all
of the pieces.

Our approach draws on the literature for irregular strip packing and applies these con-
cepts to bin packing. We describe new features that significantly enhance the perfor-
mance of the algorithm and make it more generally applicable.

The main components of our approach include:

• A single-pass constructive algorithm that constructs the bins

The constructive algorithm (CA) is based on TOPOS originally proposed by
Oliveira et al. (2000) and later improved by Bennell and Song (2010). Both these
implementations were for strip packing problems. In our approach, we propose a
method to convert the strip packing problem into the bin packing problem.
69
70 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

• Placement decisions that allow either finite or free rotation of the pieces
A key component of the placement decision is the orientation of the pieces. Most
data instances restrict the orientation of the pieces to one, two or four angles of
orientation. However, many types of material are homogeneous and pieces can be
cut in any orientation. In this chapter, we develop placement rules that permit
pieces to rotate either by a pre-defined finite set of orientation (restricted rotation
of pieces) or any angle in the range of 0 to 360 degrees.

• An improvement heuristic to search over the solutions

The improvement heuristic uses the principles of the Jostle approach proposed
by Dowsland et al. (1998) for strip packing. The combination of Jostle and the
CA means that the approach solves both the placement and allocation problems
together rather than considering them as two independent problems. Finally,
further improvements are made to the Jostle heuristic to enhance its ability to
search the solution space.

Out of the papers reviewed in Section 2.3, the majority of the existing literature on
irregular packing, has achieved efficient solutions with restricted rotation of the irregular
pieces e.g. Burke et al. (2007), Bennell and Song (2010), Lopez-Camacho et al. (2013a),
while only a few papers (Martinez-Sykora et al., 2015; Rocha et al., 2014; Han et al.,
2013; Stoyan et al., 2015a; Martinez-Sykora et al., 2017) consider unrestricted rotation of
irregular pieces. To our knowledge, Han et al. (2013), Martinez-Sykora et al. (2015) and
Martinez-Sykora et al. (2017) are the only papers that consider unrestricted rotation
with multiple identical stock sheets. Out of those the first two studies consider an
application of cutting glass where pieces are convex and require guillotine cuts. The
guillotine cuts may be at any angle relative to the edge of the stock sheet. This constraint
has a dominating influence over the algorithm design and strongly influence the rotation
angle of the pieces.

4.2 Contribution

In this section, we briefly outline the contribution of this chapter.

• A new method for identifying promising rotations for the pieces, in the case when
any angle is permitted.

• Development of an efficient heuristic adapted from the Jostle approach that is

competitive across all nesting benchmark data sets. In addition, we introduce
a diversification mechanism from local search to improve the performance of the
Jostle approach.
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 71

• This study is one of the very early studies to tackle the irregular bin packing prob-
lem allowing free rotation of the pieces, which is more general and more realistic
for many industries.

4.3 Problem Description

Let P = {p1 , . . . , pn } be a set of n polygons to be packed into identical rectangular

bins. We denote the bins as rectangular objects with fixed width W and fixed length L.
A feasible solution of the 2D ISBSBPP is described by a set of bins B = {b1 , . . . , bN }
where N is the number of bins required to pack the given n irregular polygons. We
denote Pj ⊆ P as the set of packed pieces in each bin bj and nj denotes the number
of placed pieces in bin bj . Each packed piece pi , has an orientation angle oi ∈ [0, 360o )
and a position in its allocated bin relative to the piece origin of (xi , yi ). We consider the
reference point of each piece as the bottom-left corner point of the enclosing rectangle
for a given angle of rotation and the bottom left corner of the first bin is the origin.

The objective is to minimise N ; the number of bins utilised. However, this measure is
not very sensitive and many competing solutions will have the same value of N for a
given problem instance. As a result, we use the following two measurements to evaluate
the bin packing solutions.

1. Fractional number of bins K = N − 1 + R∗ where R∗ is the proportion of the least

utilised bin used to pack pieces after applying either a vertical or horizontal cut.
This assumes that offcuts can be kept and used later. See Han et al. (2013).
PN
j=1 Uj2
2. F = N where Uj is the utilization of each bin bj , proposed by Lopez-Camacho
et al. (2013a). The utilization of each bin Uj is defined as the total area of pieces
inside the bin bj divided by the are of the bin (L × W ).

Both measures reward solutions that highly utilise bins while trying to empty one weak-
est bin, which supports the search in reducing the number of bins.

The measurement F is helpful in terms of differentiating solutions during the search

process while the algorithm runs its iterations and evaluates its solutions. When there
are solutions with an equal number of used bins, the F value favours solutions that have
tightly packed bins and a tail of low utilisation bins over those solutions which evenly
distributed with pieces among the bins. In this case, encouraging some bins to be poorly
utilised makes it easier to reduce the total number of bins in the solution, since only a
few items need to be relocated into other bins.

The measurement K measures the fractional number of bins considering the partial bin
R∗ generated after applying either a vertical or a horizontal cut to the least utilised bin.
72 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

This measurement is useful to differentiate solutions with an equal number of used bins
and it also takes account of the reuse of residuals of the least utilised bin. Usually, it
is advantageous to select the solution with lower K value if number of bins used are
equal in the solutions since the lower R∗ represents less fractional use of bins and higher
saving of residuals which can be reusable.

In terms of using each measurement, K reflects the preference of the manufacturer in

minimising material use, while F is better at guiding the search.

4.4 Packing Procedure

A solution method of 2D IBPP addresses two main decision problems: assigning pieces to
bins and placing those pieces in the bin. The majority of cutting and packing literature
that considers 2D irregular pieces focuses on strip packing where there is a single stock
sheet. In this case, the main decision is where to place the piece on the strip with
the aim of minimising length. Our approach follows the strip packing approach while
imposing the division between bins, therefore making the bin allocation implicit within
the placement decisions. The proposed method always leads to a feasible placement of
the next piece by addressing the allocation and placement decisions together.

Our approach seeks to simulate the jostling of pieces. Dowsland et al. (1998) used this
principle to reduce the packing length of the layout for the irregular shape strip packing
problem by shaking the pieces from one end to another. They called the algorithm Jostle
and suggested it was inspired by the natural human behaviour to shake a container to
encourage items to pack together as tightly as possible. The basic idea is to begin with
an initial random ordering of the pieces and apply a constructive heuristic following
a placement policy of left most position packing from the outside edges of the strip
towards the centre. Once all the pieces are packed, Jostle would proceed to take the
pieces in order of largest to smallest x-coordinate and pack at the opposite end the
stock sheet i.e. a right most policy, filling holes where possible. The next iteration
would take the packing order from smallest to largest x-coordinate and again pack with
a left most policy, and so on simulating the jostling of pieces from one end of the stock
sheet to the other. Note that packing from left to right and then right to left is called
a Jostle cycle. At the time of publishing, the approach generated the best results on
the benchmark data sets. Since this work, researchers have developed more effective
methods for handling the geometry and utilise more powerful computers. In this study,
we apply the principles of Jostle to the homogeneous multiple bin packing problem.

In order to adapt Jostle to the bin packing problem, we first define a very long packing
strip of width W and length n × L. This strip represents n bins joined together. We
then define bin spaces by assigning a set of barrier lines at fixed intervals of length L
along the strip. The bottom left corner of bin i ∈ {1, . . . , n} is placed at (L × (i − 1), 0)
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 73

(see Figure 4.1). We can directly apply Jostle to the strip with the additional constraint
that pieces may not be placed across a barrier. Along with the usual containment and
overlap constraints, this ensures pieces are always placed entirely within the bins.

Figure 4.1: Strip to Bins

Dowsland et al. (1998) imposed a grid of feasible placement positions over the strip,
which formed a search grid for finding the position of the next piece to be packed.
Restricting the placement positions by a grid over constrains the solution space and
good solutions can be lost. More recent work in irregular packing has developed effective
placement policies based on the no-fit polygon (e.g. Bennell and Song (2010)). As a
result, we design several new placement policies that are incorporated within Jostle,
these are described in Section 4.4.1. In Section 4.5 we describe the Jostle approach in
more detail including some modifications to improve the effectiveness of the search.

4.4.1 Construction algorithm

In this section, we introduce the single-pass construction algorithm, which is made up of

the packing order of pieces; and the placement policy. The packing order is predefined for
the first layout construction and subsequently decided by Jostle. The placement policy
sets the criteria for comparing the candidate placement positions and orientations of the
next piece to be placed, where any candidate position must be feasible with respect to
overlap and containment in a bin.

Any solution approach to a packing problem involving irregular shapes requires compu-
tational tools for handling the geometry. Specifically, search heuristics need an efficient
test for identifying feasible placement positions of pieces. Our tool set is based on the no
fit polygon (NFP) and inner fit polygon (IFP) obtained using the approach of Bennell
and Song (2008). Since pieces are permitted to be placed in any orientation, there are
infinitely many possible NFPs, hence it is necessary to generate these on-line as they
are needed. In order to place a piece, the algorithm must generate multiple NFPs, one
for each of the placed piece with the next piece to be placed. Our investigation shows
that it is significantly more efficient to sequentially merge all the placed pieces into a
single polygon and then generate a single, albeit large and complex, NFP. This is the
74 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

approach of Oliveira et al. (2000), called TOPOS, which was later improved by Bennell
and Song (2010).

The proposed constructive algorithm represents the existing partial solution as an array
of polygons mj where each mj is the polygon generated by merging the set of pieces
packed within the region bj (equivalent to a bin). Unlike Oliveira et al. (2000), we fix
the position of the partial solution within the bin. As in Bennell and Song (2010), the
algorithm retains any enclosed gaps between the pieces when a piece is merged with
existing partial solution. These gaps form holes in the merged polygon, which are made
available for placement of subsequent pieces, provided the area of the hole is greater than
the area of the smallest unpacked piece. Let Hk : h1 , .., hk be the set of holes generated
after placing the kth piece. Figure 4.2, provides an example of three regions, which
represent bins and contain a number of packed pieces. The groups of packed pieces are
merged into three polygons mj , which contain usable holes. hk . The spaces outside of
mj are denoted by Sj .

4.4.1.1 Generating candidate placement positions

In this section, we describe our approach for identifying the feasible candidate positions
for the next piece, pi , including its orientation. The placement policy criteria, described
in Section 4.4.2, determines which candidate is chosen. pi may be placed within a hole,
in a space or in a new bin region at the end of the layout. Below we describe the
procedure for determining whether pi will fit in a particular hole or space in a partially
packed bin and the range of orientations.

Each hole hl is represented as a simple polygon. If the area of hl is less than the area of
pi then pi will not fit into hl . Otherwise, the procedure generates the IFP between hl and
pi (IFPhl pi ). The interior and boundary of the IFP represent all the relative positions of
the two component polygons so that the second subscript polygon, pi , fits inside the first
subscript polygon, hl . See Figure 4.3. The boundary of the IFP contains all positions
where the boundaries of the polygons touch and this forms the set of feasible placement
positions within the hole. The exact placement position is determined according to the
placement rule described in Section 4.4.2. If the IFP does not exist then pi will not fit
into hl .

In order to find the feasible placement positions within a bin region that contains polygon
mj , we generate the NFP between the external boundary of mj and pi , NFPmj pi and the
IFP between the bin region bj and pi , IFPbj pi . The exterior and boundary of the NFP
represents all relative positions of the two component where they do not overlap. The
intersection of the IFPbj pi and the complement of NFPmj pi provides all the positions of
pi such that it is contained within bj and not overlapping mj , assuming bj and mj have
a fixed position. We only consider the boundary segment of the NFPmj pi that intersects
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 75

Figure 4.2: Partial solution

with IFPbj pi as the feasible set, which is all the feasible touching positions between mj
and pi . See Figure 4.4 for an example of this evaluation.

Finally, if we place a piece in a new empty bin, then the position and the orientation is
determined by the placement policy (see Section 4.4.2).

4.4.1.2 Placement orientation of a piece

The above discussion assumes pi is in a given orientation. In this chapter, we consider

two scenarios with respect to rotating pieces. First, we allow pieces to rotate by a
pre-assigned set of angles, which we define as the finite rotation approach. This is the
most common case in the literature where the rotation angles of pieces are restricted to
oi = 0o , 180o or α = 0o , 90o , 180o , 270o . Second, we consider rotating pieces freely.
76 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Figure 4.3: Hole-filling

Figure 4.4: Feasible placement for a new piece

In the case of finite rotation angles, we simply consider pi in each of the four orientations.
For each orientation we generate the set of feasible placement positions, as described in
4.4.1.1, and select the best by applying the placement policy, described in Section 4.4.2.
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 77

We then compare the best placement for each orientation, using the same placement
policy criteria, selecting the best as the final position and orientation of pi .

For free rotation, there are an infinite set of orientations, which must be reduced to a
finite set of candidates. Hence we need a mechanism to identify a subset of promising
angles. We first describe the approach assuming pi can be placed in a region that
contains a merged polygon mj .

Our approach begins by placing pi on the packing layout in a temporary position and
orientation using the finite rotation placement approach, where the best orientation is
selected. As discussed in the previous section, the position will be touching the boundary
of mj . Each touching point or edge defines two new angles of orientation. These are
illustrated in figure 4.5.

Figure 4.5: Tuning the orientation angle

Figure 4.5(a) shows an edge-vertex combination, Figure 4.5(b) shows a vertex-vertex

combination and figure 4.5(c) shows an edge-edge combination. In each case, piece p is
rotated in a counter-clockwise direction by angle θ and in a clockwise direction by angle
φ. The second example in figure 4.5(a), shows two separate points of contact and as a
result there are four new orientations. These form the new candidate orientations and
following the placement procedure, we find the best placement position for the pieces in
these orientations. These are compared with the temporary position and orientation, if
none of these new angles provides a better placement position then the algorithm takes
the temporary position and orientation as the placement of the piece. If the next piece
touches the boundary of a bin, then we also consider the angles created by the edges of
the bin (see Figure 4.6).
78 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Figure 4.6: Angle tuning at the edges of bins

4.4.2 Placement policy

The above procedures provide a set of feasible placement positions and orientations for
piece pi . The placement policy sets the criteria that determine which position is chosen.
Note that for some of the placement rules, we can often avoid the computational cost
of assessing all positions by controlling the order in which they are assessed. In the
following, we describe the placement rules in the context of packing from the left-hand
end of the strip extending towards the right. However, Jostle also packs from right to
left, in which case simply swap left and right in the following explanation.

We evaluate three well-known placement rules; bottom left placement, minimum length
placement and maximum utilisation placement, to determine the position of each piece.
However, these rules are dominated by first trying to fill holes in the layout.

The first step when placing pi is to sort all the holes in Hk in order of leftmost x-
coordinate. These holes are evaluated in that order using the hole filling procedure
described in Section 4.4.1.1. If pi fits in a given hole, it is placed and no further holes
are considered. If pi does not fit within any hole, the algorithm attempts to find a
feasible placement position in each of the bin regions bj starting with the left most.
Using the procedure described in Section 4.4.1.1 for finding placement position in a
partially packed bin, the candidate positions are found for each bin region bj in order.
Depending on the placement policy, once a feasible position is found, the search stops
after identifying the best orientation for that position.

The placement policies are as follows:

• Bottom-left (BL) prioritises placing the reference point of the piece in the leftmost
feasible position on the stock sheet breaking ties by selecting the bottom-most of
the leftmost positions. This is a well-known single pass placement rule applied in
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 79

Figure 4.7: BL and ML placement

many cutting and packing papers (Albano and Sapuppo, 1980; Blażewicz et al.,
1993; Dowsland et al., 1998). By applying this rule when hole filling fails, once a
feasible position is found in the region bj the search can stop. If more than one
orientation of pi gives the same BL position, select the orientation that minimises
the width of the bounding box of the piece. If there is still a tie, select the
orientation randomly. Apply the same tie breaker rules for placing pi in a new bin.

• Minimum Length (ML) placement rule selects the position that minimizes the
length from the origin of the bin to the right-most x-coordinate of pi . The piece is
orientated to minimise the right-most x-coordinate. In the case of ties, we select
the bottom-left placement position.

• Maximum Utilisation (MU) placement rule selects the position that provides the
maximum area utilisation in the earliest bin. We assume the occupied area of a
region bj is the area of the convex hull of mj . Let the area of the convex hull of mj
be CHarea(mj ), the area of pi be area(pi ) and the area of the convex hull of the
newly merged polygon after pi is placed in a certain position be CHarea (mj + pi ).
Then the utilisation of the placement position (U tp) is calculated as follows.

CHarea(mj ) + area(pi )
U tp =
CHarea(mj + pi )

Given the set of feasible placement positions on the boundary of the NFP, we
select each feasible vertex of the NFP as a candidate placement position. This is
80 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

illustrated in Figure 4.8. Placement positions are evaluated bin by bin starting at
the left-most bin. If a feasible placement is found in a bin then no further bins are
considered. Ties are broken by selecting the bottom left position and minimum
length orientation.

Figure 4.8: Placement of piece on different vertices along the feasible placement
segment

Algorithm 3 denotes the constructive algorithm discuss in this chapter.

 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 81

Algorithm 3: Constructive Algorithm

Input : P0 , Placement rule
Output: P ∗ : A feasible packing of all pieces in N Bins
1 Set N = 1;
2 for i = 1 : n do
3 Set l = 1, placed = f alse, j = 1;
4 while l ≤ q and placed = f alse do
5 if pi fits in hl with any rotation in {0, π2 , π, 3π
2 }) then
6 Place pi with the best angle derived from the Placement rule used ;
7 Update Hi−1 → Hi and Update q;
8 Sort all the holes in Hi in order of left most x-coordinate;
9 placed = true;
10 else
11 l = l + 1;
12 end
13 end
14 while placed = f alse do
15 if pi fits in bj with any rotation in {0, π2 , π, 3π
2 }) then
16 Compute best angle α in {0, π2 , π, 3π
2 } according to the Placement rule used;
17 Compute new angles by angle tuning mechanism starting with α denoted as θ and
φ; Place pi in bj with the best angle between α, θ and φ according to the
Placement rule used;
18 Merge pi to mj ;
19 Capture any usable hole created (area of the hole ≥ area of the smallest piece);
20 Update Hi−1 → Hi and Update q;
21 Sort holes Hi by left most x-coordinate;
22 If (j > N ) N = j ;
23 placed = true ;
24 else
25 j = j + 1;
26 end
27 end
28 end

4.5 Solution Improvement

4.5.1 Jostle heuristic

In this section, we discuss the design of the Jostle heuristic to address the 2D ISBSBPP.
The Jostle heuristic is used to determine the piece sequence that feeds into each iteration
of the constructive algorithm. The first application of the constructive algorithm uses
an arbitrary order of pieces and packs the pieces from left to right along the strip. After
generating a complete solution, all the pieces are re-ordered according to the piece with
the right-most x-coordinates continuing to the left-most. Following this order, pieces
are packed starting at the right-most position of the strip building the packing layout
from right to left. Jostle continues to oscillate from packing left to right then right to
left until the termination criterion is met.
82 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Jostle is analogous to a local search where the search occurs over the sequence of the
pieces and each Jostle iteration generates a neighbour sequence. The constructive algo-
rithm takes a sequence of pieces and creates a layout. The layout is converted back to
a sequence according to the x-coordinate order. Hence, the new layout defines a new
sequence of pieces, this is a neighbour solution based on the arrangement of pieces of the
previous packing layout. Each iteration generates only one neighbour, which is always
accepted. The optimisation arises from the neighbour generation. Instead of applying
a fixed rule to the sequence i.e. swap or insert, the constructive algorithm changes the
sequence by optimising the packing arising from the current piece sequence. Each com-
plete solution is evaluated according to K and F defined in Section 4.3. At the end of
the search, the best solution according to these measures is kept.

Since the constructive algorithm includes filling of holes and multiple alternative orienta-
tions, the piece sequence and placement positions change between iterations. However,
our experiments show that the search can converge and get stuck in local optima. In
order to overcome this, we incorporate the diversification ideas from iterated local search
to escape from local optima. The new variant is called Iterated Jostle.

4.5.2 Iterated Jostle

The Iterated Jostle algorithm combines the basic Jostle procedure and a kick intended
to move the solution to a new area of the solution space and escape local optima. This
feature is motivated by our observation that the Jostle heuristic becomes trapped in
local optima after running several Jostle cycles. This is more commonly found for data
sets that are small and have similar size pieces. We perform a kick from the best local
optimum found so far. We consider that we are at a local optima after cmax Jostle cycles
with no improvement. We evaluate alternative values for cmax in the results section. We
investigate two kick strategies, described below. Note that in cutting and packing, a
small change in the permutations can lead to a significant change in the layout of the
pieces once it has been decoded by the constructive algorithm. In the case of their first
kick, we check that the move does not produce an identical permutation, which can
occur when there are multiple pieces of the same type,

1-piece insert move: We perform a single insert move as follows:

1. Select two random positions along the permutation of pieces,
2. Select the piece corresponds to the highest position selected in step 1,
3. Remove the selected piece from the permutation and insert it in front of the other
position selected in step 1,
4. Order the remaining pieces following the existing permutation.
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 83

Bin-swapping move: This strategy changes the order of pieces by effectively swap-
ping two bins as follows:
1. Select the bin with the lowest utilisation and let sub-sequence A be the pieces in
that bin in x-co-ordinate order,
2. Select another bin randomly and let sub-sequence B be the pieces in that bin in
x-coordinate order,
3. Swap sub-sequence A and sub-sequence B in the permutation,
4. Obtain the new permutation of pieces.

Algorithm 4 describes the iterated Jostle procedure; IJS. Let t be the Jostle cycle, PtL
be the permutation of pieces to be packed from left to right in cycle t, PtR be the
permutation of pieces to be packed from right to left in cycle t and P0 be the initial
random permutation. P ∗ stores the permutation of pieces of the solution with the
smallest objective function value found so far, K ∗ . The variable c counts the number of
consecutive Jostle iterations with no improvement in solution quality (K). We set cmax
as the maximum value of c. The termination criterion is the maximum total number of
Jostle cycles that the algorithm will run, Tmax .

Iterated Jostle with join and release of pieces This final version of Iterated
Jostle draws on the ideas of Adamowicz and Albano (1972), which was more recently
applied by Elkeran (2013). They combine pieces that fit together well reducing the
overall number of pieces. In our approach we allow two pieces to be combined into one
merged piece for some of the Jostle iterations. The algorithm scans the arrangement of
pieces in each bin and attempts to join adjacent pieces in their current position. If the
combined piece is convex with no holes then they are merged for the following Jostle
iterations. Each scan only permits pairwise merging. However, scans in later iterations
may join a piece to a merged piece resulting in three or more original pieces joining
together. Pieces are joined after each Jostle iteration until a kick is performed, at which
time they are separated again. The type of kick alternates between 1-piece insert and
bin-swapping.

The following steps outline the IJS with join and release procedure.

1. From a given initial permutation, apply Jostle until the kick criterion is met.

2. Apply a kick alternating between bin-swap and 1-piece insert,

3. Jostle for one iteration

4. Execute pairwise joining of pieces for the existing layout. In this case the algorithm
scans the whole layout from left to right and searches for the possible pairs of pieces
to join as described in Figure 4.9.

5. If stop criteria met then stop else

 84 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Algorithm 4: IJS1
Input : P0 , Placement rule, kick type cmax , Tmax , D∗ = Lef t(L) or Right(R)
Output: P ∗ , K ∗ ,
1 Set t = 1;
2 P ∗ = PtL = P0 ;
3 Set D∗ = L;
4 Set K ∗ = a suitably large value;
5 Initialize c = 0;
6 while t ≤ Tmax do
7 Generate solution layout from PtL ;
8 Evaluate the solution and find KtL ;
9 Derive PtR from solution;
10 Generate solution layout from PtR ;
11 Evaluate the solution and find KtR ;
12 if KtL < KtR then
13 if KtL < K ∗ then
14 Set K ∗ = KtL , P ∗ = PtL , D∗ = L;
15 Reset c = 0;
16 else
17 c = c + 1;
18 end
19 else
20 if KtR < K ∗ then
21 Set K ∗ = KtR , P ∗ = PtR , D∗ = R;
22 Reset c = 0;
23 else
24 c = c + 1;
25 end
26 end
27 if c > cmax then
28 Apply kick to P ∗ and obtain PtD∗ ;
29 Generate solution layout from PtD∗ ;
30 Derive PtL from solution;
31 end
32 t = t + 1;
33 end
34 return P ∗ , K ∗ , D∗ ;

6. If kick criteria met, release all joined pieces and go to step 2, else

7. go to step 3
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 85

Figure 4.9: Possibilities when pieces are paired by the Join and Release mech-
anism

4.6 Experimental Design

All the algorithms were coded in C++11 and utilised version 1.58.0 of the ’boost com-
putational geometry library’ to implement some of the geometric operations. The tests
were performed by 2.6 GHz processor and 4 GB of RAM to execute the program.

4.6.1 Data instances

Our experiments utilise two sets of benchmark instances: the nesting instances used
by a wide variety of authors for the irregular strip packing problem and instances pro-
posed by Lopez-Camacho et al. (2013a) that are jigsaw puzzle instances as they have
known optimal solutions of 100% utilisation of the stock sheet. Both are available
from the ESICUP (EURO Special Interest Group on Cutting and Packing) website
http://paginas.fe.up.pt/~esicup/datasets.

The nesting instances contain 23 irregular strip packing data sets and contain both
convex and non-convex pieces. As these are strip packing instances we need to define
bin sizes for each data set. For each instance, we define three different bin sizes based
on the maximum dimension of the enclosing rectangles of each piece in their original
orientation denoted as md . The bin sizes are: small bin (1.1md for length and width),
86 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

medium bin (1.5md for length and width) and large bin (2md for length and width).
These bin sizes were proposed by Martinez-Sykora et al. (2017).

The jigsaw puzzle (JP) instances contain two sets, which we denote as JP1 and JP2.
JP1 has a collection of 540 instances in which all the pieces are convex. These instances
are divided into 18 classes with 30 problems and each class carries a different number
of pieces. The second set, JP2, has 480 instances (16 classes with 30 problems in each
class) in which the pieces have concavities.

4.6.2 Performance of different solution strategies

Our solution strategy includes a number of alternative approaches in terms of the place-
ment policy in the constructive heuristic, the permitted orientations and the kick strate-
gies in Jostle. Table 4.1 and Table 4.2 provide a summary and the notation for the
alternative approaches. Any variant that includes a random input is repeated for 10
trials and we report the average result. We conduct the following set of computational
experiments to compare the performance of each approach.

Table 4.1: Notation of Algorithms: placement rule and rotation

Notation Placement Rule Piece orientation

BL4 Bottom-Left Restricted (0, π2 , π, 3π
2 )
BL∞ Bottom-Left Unrestricted
ML4 Minimum-Length Restricted (0, π2 , π, 3π
2 )
ML∞ Minimum-Length Unrestricted
MU4 Maximum Utilization Restricted (0, π2 , π, 3π
2 )
MU∞ Maximum Utilization Unrestricted

Table 4.2: Notation of Algorithms: algorithm and kick type

Notation Description Kick-move strategy

JS Jostle -
IJS1 Iterated Jostle 1-piece insert move
IJS2 Iterated Jostle Bin swapping move
IJS3 Iterated Jostle 1-piece insert move , Bin swapping move , Join and Release of pieces

Test No. 1: Evaluation of the constructive algorithm and the proposed Jostle heuris-
tic.
These experiments compare the performance of the two rotation criteria in terms of
quality of packing and computational time and the alternative placement policies. A
trial is made up of 500 random permutations of pieces, each packed using a constructive
heuristic. The best layout is recorded from each trial. For each constructive heuristic,
there are ten trials and we report the average of the ten trials in Table 4.3. We com-
pare constructive algorithm variants BL4, BL∞, M L4, M L∞, M U 4, and M U ∞. In
addition, we evaluate the Jostle heuristic and compare with packing random permuta-
tions of pieces using the constructive algorithm. These experiments intend to investigate
whether the Jostle heuristic provides a good search mechanism that improves on ran-
dom trials. We report the average of ten random permutations initialise ten trails of the
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 87

basic Jostle heuristic where each trial is run for 250 Jostle cycles (which generates 500
layouts). Table 4.3 details the results and considers measure K and F . Note we wish to
minimise K and maximise F .

This is a preliminary test to validate that Jostle is more effective than the traditional
multi-start approach. We test it using the nesting instances which are generally used
as the standard set of instances in 2D irregular shape packing problems, rather being
biased more on the specific category of instances like jig-saw puzzle. Table 4.3 contains
results from instances with a substantial variety of pieces representing convex pieces,
non-convex pieces, pieces with a higher number of edges and jig-saw puzzle type pieces.

Test No. 2: Comparative performance analysis of IJS and JS

We report the results for four strategies; i.e. JS, IJS1, IJS2 and IJS3. Ten random
permutations are used as initial permutations for each variant and each trial runs for
500 Jostle cycles. The average is reported in Table 4.2. These experiments evaluate the
impact of introducing a kick and compare the types of kick. After initial experiments
with a range of values for cmax , we set cmax = 4, 6, 6 Jostle cycles for IJS1, IJS2 and
IJS3 respectively.

Test No. 3: Comparative analysis of the performance of IJS algorithms with literature
best.

In this case we compare our results with the work of Lopez-Camacho et al. (2013a) and
Martinez-Sykora et al. (2017). Since these algorithms are run on different computing
platforms, it is not possible to run for the same amount of time. Hence we retain our
termination criteria of 500 Jostle cycles and report the average result and computational
time for the ten random trials.

4.7 Experimental Results

4.7.1 Test 01: Jostle search vs. Random order of pieces

A key component of the algorithms is the level of freedom to rotate the pieces. We
evaluated the benefits of different rotation criteria; 1) finite angles of rotation and 2) free
rotation, for building packing solution. The experimental results in Table 4.3 show that
allowing free rotation of pieces can significantly improve the solution quality. Considering
a pairwise comparison of four angles of rotation and free rotation for each placement
rule (BL, ML, MU), 16 of 18 comparisons produced better solutions using free rotation
for both K and F , and the remaining two comparisons showed free rotation is better
for measure K but not for F .

Considering the performance of Jostle against the randomly generated permutations,

the experimental results are shown in Table 4.3 reveal that the Jostle heuristic performs
88 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Table 4.3: Performance of construction algorithm for different rotation criteria

Random order Jostle

Small bins Medium bins Large bins Small bins Medium bins Large bins
K F K F K F K F K F K F
BL4 9.91 0.379 4.792 0.389 2.601 0.395 9.61 0.387 4.712 0.392 2.602 0.398
BL∞ 9.73 0.383 4.632 0.394 2.555 0.392 9.48 0.389 4.612 0.397 2.487 0.399
M L4 9.94 0.378 4.794 0.390 2.613 0.396 9.52 0.382 4.703 0.394 2.551 0.404
M L∞ 9.76 0.384 4.612 0.393 2.551 0.392 9.42 0.389 4.511 0.398 2.464 0.407
MU4 9.89 0.383 4.944 0.385 2.592 0.391 9.65 0.387 4.786 0.402 2.561 0.401
MU∞ 9.62 0.385 4.614 0.393 2.524 0.392 9.47 0.402 4.573 0.408 2.470 0.413

better for all pairwise comparisons between placement rules and rotation criteria. The
best solutions are produced when pieces are allowed to rotate freely. Since there is no
significant difference between the computational time in both cases, we can also conclude
that the time consumption of the Jostle heuristic is negligible and time is mainly spent
on the execution of the construction algorithm. The results are aggregating over many
problem instances. A detailed look at the instances reveals that often the number of
bins are reduced and in other cases the piece are packed more tightly across the same
number of bins. We also observed that Jostle struggles to diversify from the first local
optima, so the improvements are local with respect to the initial solution. With this
in mind, we conclude that Jostle can perform as an optimisation procedure with more
intelligence than a random search.

The final observation from these results concerns the placement rule. For random per-
mutations BL performs the best once on measuring F , ML performs the best twice and
MU performs the best three times. For Jostle the result is more consistent where ML
performs the best for each measure of K and MU performs the best for each measure
of F . This result is intuitive since K values minimising length to provide a greater
residual in the last bin, while F is based on utilisation. It is interesting to note the BL is
consistently adopted in the literature, but in our investigation, it is consistently inferior.

4.7.2 Test 02: IJS vs JS

Though Jostle yields better solutions than packing random permutations of pieces, one
major limitation is the tendency to get trapped in local optima. This occurs more
frequently with small data sets and with those with less irregularity in their shapes. As
an example, Figure 4.10(a) illustrates solution F values received at each iteration for a
certain trial for Fu data instance using JS. In this case, the Jostle heuristic has taken
9-12 cycles to reach for a better F value (i.e. F = 0.488). When jostling continues,
the solution rarely reaches F = 0.488 and the best value is recorded at the 143rd cycle.
After the 176th cycle, the Jostle has reached a steady state generating the same set of
solutions repetitively.
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 89

Figure 4.10(b) demonstrates the benefit of using a kick. Here IJS1 is run on the same
data. Introducing a kick has avoided the repetitive cycle and produced better solutions
(i.e. F > 0.499) within fewer iterations.

Figure 4.10: JS and IJS1 iterations

In Table 4.4 we present the average results obtained by JS and the three IJS proce-
dures for the 23 Nesting instances. For all placement criteria and angles of rotation,
the three IJS algorithms produce better results when compared with JS. For nesting
data instances, IJS1 provides the best results for all bin sizes when averaging over all
placement/rotation variants and, with one exception for medium bins, IJS1 contains
the overall best result. Focusing just on the IJS algorithms, there are 18 variants (3
placement rules × 2 angles of rotation × 3 IJS algorithms). Out of those 18 methods,
IJS1-MU∞ produces the best results irrespective of the bin sizes.

In Table 4.5 we present the average results obtained by JS and the three IJS procedures
for the 540 JP1 and 480 JP2 instances. Again the IJS approaches are clearly better
than the values of JS. On average IJS3 performed better for both JP1 and JP2 instances.
However, in this case, IJS3-MU4 has produced the best results. For this data the optimal
90 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Table 4.4: Performance results JS, IJS1, IJS2 and IJS3 for nesting data instances
JS IJS1 IJS2 IJS3
F T (sec) F T (sec) F T (sec) F T (sec)
Small bin BL4 0.384 304.4 0.394 322.4 0.388 310.7 0.394 319.4
BL∞ 0.389 502.7 0.408 528.5 0.394 512.1 0.404 534.8
ML4 0.386 278.6 0.393 313.4 0.390 285.8 0.397 320.5
ML∞ 0.389 509.1 0.419 526.6 0.392 533.4 0.402 575.3
MU4 0.388 281.4 0.393 294.0 0.389 286.4 0.403 335.0
MU∞ 0.407 504.9 0.437 518.6 0.427 534.4 0.433 540.5
Avg. 0.391 0.408 0.397 0.406

Medium bin BL4 0.392 287.5 0.403 298.4 0.409 298.1 0.407 315.9
BL∞ 0.398 484.7 0.409 506.1 0.403 496.3 0.404 505.4
ML4 0.394 269.5 0.409 301.8 0.400 281.2 0.402 306.2
ML∞ 0.398 476.6 0.422 515.9 0.418 488.7 0.414 530.6
MU4 0.404 267.4 0.418 281.2 0.410 272.0 0.409 306.8
MU∞ 0.408 487.9 0.436 499.0 0.415 503.4 0.429 510.1
Avg. 0.399 0.416 0.409 0.410

Large bin BL4 0.398 248.9 0.403 266.5 0.402 255.6 0.410 280.7
BL∞ 0.400 421.6 0.404 450.1 0.402 430.7 0.415 485.3
ML4 0.404 239.4 0.422 259.7 0.400 249.4 0.410 272.6
ML∞ 0.408 426.3 0.425 451.3 0.409 441.6 0.415 477.6
MU4 0.402 240.4 0.425 253.9 0.404 247.1 0.410 268.9
MU∞ 0.414 426.8 0.431 453.9 0.413 436.8 0.407 480.7
Avg. 0.404 0.418 0.405 0.411

Table 4.5: Performance results of IJS1, IJS2 and IJS3 vs JS for JP1 and JP2
instances
F values JS IJS1 IJS2 IJS3
F T (sec) F T (sec) F T (sec) F T (sec)
JP1 BL4 0.589 76.5 0.652 81.2 0.648 76.3 0.700 76.2
BL∞ 0.527 148.5 0.564 161.4 0.579 167.4 0.594 153.8
ML4 0.508 72.7 0.625 79.6 0.647 80.8 0.693 75.8
ML∞ 0.522 150.3 0.634 161.9 0.564 152.0 0.590 162.4
MU4 0.566 72.1 0.691 78.1 0.678 79.6 0.732 74.9
MU∞ 0.552 159.7 0.677 163.8 0.639 166.7 0.644 161.2
Avg. 0.544 0.641 0.626 0.659

JP2 BL4 0.572 56.3 0.652 59.3 0.682 62.3 0.725 63.8
BL∞ 0.506 98.4 0.522 101.2 0.532 102.4 0.547 103.3
ML4 0.549 55.7 0.677 61.3 0.697 61.2 0.714 61.5
ML∞ 0.505 98.1 0.632 103.4 0.607 104.6 0.516 100.1
MU4 0.617 54.9 0.701 62.1 0.680 59.2 0.747 58.3
MU∞ 0.584 97.8 0.655 107.8 0.626 112.5 0.613 101.5
Avg. 0.556 0.640 0.637 0.644

solution can only be found at the zero rotation angle, hence it is intuitive that restricting
the rotation is beneficial.

4.7.3 Test 03: Comparing with benchmark results in literature

Referring to the previous tables 4.4 and 4.5 we can compare these results with the results
from the literature. Martinez-Sykora et al. (2017) report a best value of F = 0.723
with computational time 161 seconds for JP1 and F = 0.729 with computational time
123 seconds for JP2. In both cases our algorithm (IJS3-MU4) provides better results,
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 91

specifically F = 0.732 with computational time 74.9 seconds for JP1 and F = 0.747 with
computational time 58.3 seconds for JP2. Also, our approach has achieved better results
for 11 out of 18 groups of instances for JP1. In the case of JP2 instances, our method
has achieved better results for 11 out of 16 groups of instances. A detail description
of these results are provided in table 4.6 and 4.7. The only other existing results for
this data is by Lopez-Camacho et al. (2013a), who only report results for JP1. They
achieve F = 0.690 with computational time 50 seconds. We revised the termination of
our algorithm to 50 seconds and achieved the improved result of F = 0.703.

For the nesting instances, we consider the average over all the instances for each bin
size. The only benchmark results are from Martinez-Sykora et al. (2017). For small
bins, they achieve F = 0.434 with a computation time of 1721 seconds, for medium bins
they achieve F = 0.452 with a computation time of 3172 seconds and for large bins they
achieve F = 0.403 with a computation time of 762 seconds. IJS1-MU∞ performs better
for small and large bins and takes less time for all three bin types. The overall results of
Martinez-Sykora et al. (2017) demonstrate that their algorithm has lower performance
when the average number of pieces per bin is higher. However in our case, we achieve
significantly better results for large bins in most of our approaches; i.e. IJS1, IJS2 and
IJS3. Even for small and medium sized bins, we achieved better or close results with
no significant drop in average F values, with less computational time than Martinez-
Sykora et al. (2017). Table 4.8, 4.9 and 4.10 present the results we achieved for each
data instance using IJS1 approach.
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Table 4.6: Results of the IJS1, IJS2 and IJS3 algorithm applied for JP1 instances
BVL: Best values of Lopez-Camacho et al. (2013a), BVM: Best values of Martinez-Sykora et al. (2017)
BVL(F) BVM(F) IJS1 (F values) IJS2 (F values) IJS3 (F values)
BL4 BL∞ ML4 ML∞ MU4 MU∞ BL4 BL∞ ML4 ML∞ MU4 MU∞ BL4 BL∞ ML4 ML∞ MU4 MU∞
A 0.605 0.614 0.609 0.583 0.606 0.607 0.614 0.606 0.604 0.575 0.606 0.582 0.614 0.601 0.620 0.575 0.624 0.595 0.632 0.597
B 0.929 0.878 0.770 0.675 0.744 0.795 0.890 0.801 0.675 0.709 0.745 0.712 0.830 0.796 0.827 0.739 0.817 0.724 0.880 0.812
C 0.763 0.736 0.708 0.584 0.645 0.649 0.732 0.725 0.695 0.605 0.696 0.576 0.760 0.664 0.772 0.614 0.763 0.623 0.797 0.701
D 0.579 0.591 0.593 0.576 0.595 0.587 0.593 0.582 0.587 0.573 0.58 0.567 0.594 0.587 0.603 0.576 0.602 0.584 0.610 0.576
E 0.412 0.529 0.451 0.381 0.447 0.486 0.478 0.455 0.453 0.405 0.425 0.396 0.465 0.467 0.496 0.427 0.485 0.434 0.540 0.442
F 0.496 0.564 0.517 0.483 0.481 0.524 0.532 0.520 0.509 0.486 0.498 0.481 0.513 0.507 0.538 0.493 0.533 0.501 0.548 0.498
G 0.814 0.809 0.731 0.671 0.711 0.742 0.804 0.807 0.722 0.665 0.721 0.671 0.751 0.760 0.788 0.681 0.781 0.697 0.819 0.762
H 0.928 0.868 0.759 0.683 0.743 0.771 0.877 0.870 0.765 0.696 0.768 0.707 0.808 0.793 0.836 0.719 0.822 0.716 0.883 0.812
I 0.627 0.652 0.643 0.629 0.644 0.644 0.648 0.641 0.634 0.653 0.636 0.636 0.651 0.640 0.679 0.673 0.681 0.644 0.697 0.653
J 0.665 0.701 0.672 0.656 0.670 0.665 0.672 0.668 0.667 0.644 0.667 0.653 0.675 0.668 0.690 0.652 0.689 0.665 0.698 0.661
K 0.718 0.763 0.704 0.609 0.662 0.683 0.76 0.762 0.711 0.62 0.664 0.601 0.726 0.685 0.752 0.631 0.744 0.647 0.768 0.682
L 0.512 0.619 0.550 0.409 0.539 0.583 0.596 0.588 0.587 0.446 0.592 0.425 0.586 0.552 0.618 0.477 0.621 0.489 0.630 0.551
M 0.589 0.655 0.599 0.506 0.575 0.581 0.627 0.617 0.609 0.506 0.598 0.479 0.622 0.559 0.676 0.522 0.669 0.508 0.723 0.568
N 0.503 0.668 0.534 0.493 0.519 0.497 0.534 0.509 0.517 0.494 0.501 0.492 0.519 0.525 0.543 0.497 0.531 0.496 0.549 0.495
O 0.823 0.848 0.741 0.519 0.648 0.575 0.845 0.721 0.764 0.509 0.801 0.444 0.841 0.659 0.849 0.519 0.838 0.444 0.937 0.724
P 0.678 0.852 0.638 0.445 0.605 0.639 0.704 0.683 0.661 0.542 0.654 0.487 0.677 0.615 0.696 0.559 0.689 0.511 0.717 0.628
Q 1 0.965 0.807 0.639 0.723 0.689 0.805 0.868 0.783 0.676 0.792 0.653 0.842 0.727 0.865 0.699 0.858 0.699 0.956 0.729
R 0.771 0.767 0.711 0.609 0.687 0.697 0.733 0.754 0.721 0.623 0.697 0.588 0.733 0.689 0.758 0.646 0.735 0.645 0.786 0.694
Av. 0.690 0.723 0.652 0.564 0.625 0.634 0.691 0.677 0.648 0.579 0.647 0.564 0.678 0.639 0.700 0.594 0.693 0.590 0.732 0.644
92
 Table 4.7: Results of the IJS1, IJS2 and IJS3 algorithms applied for JP2 instances
BVM: Best values of Martinez-Sykora et al. (2017)
BVM(F) IJS1 IJS2 IJS3
BL4 BL∞ ML4 ML∞ Ut4 Ut∞ BL4 BL∞ ML4 ML∞ Ut4 Ut∞ BL4 BL∞ ML4 ML∞ Ut4 Ut∞
A 0.680 0.587 0.546 0.584 0.574 0.600 0.582 0.588 0.552 0.594 0.579 0.578 0.571 0.609 0.569 0.602 0.562 0.604 0.569
B 0.760 0.715 0.635 0.737 0.753 0.743 0.755 0.748 0.645 0.743 0.684 0.729 0.701 0.768 0.656 0.768 0.694 0.804 0.701
C 0.691 0.613 0.553 0.629 0.752 0.636 0.673 0.624 0.567 0.636 0.598 0.624 0.697 0.680 0.574 0.644 0.542 0.692 0.594
F 0.551 0.495 0.477 0.503 0.529 0.511 0.529 0.495 0.473 0.512 0.489 0.503 0.493 0.533 0.489 0.518 0.486 0.525 0.489
H 0.828 0.709 0.651 0.788 0.702 0.792 0.743 0.741 0.661 0.789 0.745 0.744 0.762 0.792 0.673 0.804 0.652 0.809 0.719
L 0.576 0.546 0.418 0.541 0.532 0.539 0.532 0.536 0.422 0.547 0.497 0.547 0.511 0.582 0.449 0.581 0.435 0.588 0.521
M 0.681 0.569 0.485 0.578 0.579 0.581 0.583 0.576 0.496 0.581 0.517 0.539 0.531 0.617 0.508 0.596 0.368 0.627 0.543
O 0.788 0.708 0.469 0.769 0.651 0.803 0.672 0.761 0.474 0.800 0.636 0.781 0.672 0.823 0.484 0.790 0.699 0.826 0.624
S 0.537 0.522 0.378 0.539 0.526 0.623 0.531 0.621 0.386 0.623 0.512 0.602 0.518 0.621 0.412 0.591 0.308 0.642 0.474
T 0.921 0.829 0.538 0.875 0.483 0.908 0.883 0.865 0.55 0.901 0.532 0.874 0.747 0.907 0.563 0.906 0.435 0.959 0.747
U 0.764 0.651 0.417 0.73 0.678 0.728 0.678 0.711 0.432 0.735 0.712 0.728 0.674 0.777 0.443 0.772 0.303 0.823 0.596
V 0.989 0.775 0.462 0.803 0.593 0.922 0.593 0.811 0.482 0.838 0.589 0.894 0.548 0.915 0.503 0.878 0.582 0.993 0.645
W 0.669 0.696 0.647 0.689 0.648 0.705 0.693 0.735 0.665 0.704 0.666 0.678 0.641 0.782 0.678 0.787 0.677 0.843 0.643
X 0.632 0.588 0.496 0.589 0.603 0.599 0.604 0.599 0.51 0.599 0.512 0.573 0.519 0.632 0.538 0.641 0.497 0.649 0.581
Y 0.726 0.657 0.584 0.672 0.714 0.706 0.714 0.683 0.596 0.721 0.648 0.688 0.698 0.729 0.599 0.727 0.577 0.735 0.633
Z 0.893 0.779 0.590 0.811 0.801 0.826 0.715 0.814 0.603 0.826 0.803 0.802 0.729 0.829 0.615 0.818 0.446 0.839 0.728
Av. 0.729 0.652 0.522 0.677 0.632 0.701 0.655 0.682 0.532 0.697 0.607 0.680 0.626 0.725 0.547 0.714 0.516 0.747 0.613
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes
93
94 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

Table 4.8: Results of the IJS1 algorithm for each nesting instances with 2md
Large bins
Values include the average of F values of 10 trials, t : average time for 500 iterations
BL4 BL∞ ML4 ML∞ MU4 MU∞
F t F t F t F t F t F t
Albano 0.383 77.8 0.395 162.2 0.402 74.8 0.387 154.2 0.404 83.2 0.388 163.3
Shape2 0.437 56.2 0.424 107.5 0.446 51.2 0.447 103.7 0.440 60.7 0.451 110.3
Trousers 0.451 294.7 0.414 683.5 0.459 314.6 0.44 718.8 0.462 358.3 0.462 754.3
Shape0 0.294 245.9 0.294 504.2 0.294 248.4 0.294 521.7 0.294 263.7 0.294 548.9
Shape1 0.304 416.5 0.306 740.4 0.304 408.3 0.308 751.4 0.304 408.3 0.318 748.4
Shirts 0.647 710.9 0.651 1221.4 0.649 626.8 0.652 1143.2 0.646 618.3 0.652 1118.4
Dighe2 0.260 14.4 0.260 35.2 0.260 17.0 0.260 37.6 0.260 16.5 0.260 34.6
Dighe1 0.329 17.9 0.329 33.6 0.329 18.2 0.329 34.6 0.329 17.6 0.329 35.1
Fu 0.487 4.9 0.502 6.3 0.502 5.0 0.505 7.0 0.503 4.4 0.505 8.2
Han 0.311 83.6 0.314 172.6 0.316 86.8 0.356 176.4 0.357 79.2 0.412 144.3
Jakobs1 0.617 52.4 0.609 92.4 0.613 50.9 0.616 98.3 0.617 48.9 0.616 92.4
Jakobs2 0.455 46.0 0.452 89.8 0.45 45.8 0.454 92.2 0.456 48.4 0.448 96.8
Mao 0.256 113.4 0.256 191.3 0.61 119.2 0.61 188.3 0.61 106.7 0.610 176.7
Poly1a 0.368 22.5 0.368 41.7 0.368 23.2 0.368 43.5 0.368 24.6 0.368 42.0
Poly2a 0.376 79.5 0.384 126.3 0.382 72.4 0.384 131.7 0.381 74.3 0.386 128.8
Poly3a 0.395 149.3 0.389 247.0 0.387 131.4 0.416 258.9 0.392 133.6 0.416 246.8
Poly4a 0.399 243.3 0.416 412.5 0.394 238.6 0.407 426.4 0.399 254.9 0.418 420.7
Poly5a 0.406 311.1 0.409 526.5 0.399 336.9 0.408 556.9 0.405 328.5 0.408 589.5
Poly2b 0.443 81.8 0.455 131.0 0.454 76.6 0.45 132.9 0.454 79.4 0.459 126.3
Poly3b 0.387 171.6 0.401 252.9 0.403 177.4 0.399 249.9 0.414 156.8 0.418 239.7
Poly4b 0.402 295.8 0.405 491.2 0.397 286.8 0.401 484.7 0.393 276.1 0.406 465.7
Poly5b 0.468 433.1 0.471 652.9 0.492 428.9 0.471 671.6 0.471 408.5 0.473 701.6
Swim 0.398 2206.5 0.393 3430.0 0.404 2134.8 0.404 3396.6 0.416 1988.1 0.408 3447.3
Avg. 0.403 266.5 0.404 450.1 0.422 259.7 0.425 451.3 0.425 253.9 0.431 453.9

Table 4.9: Results of the IJS1 algorithm for each nesting instances with 1.5md
Medium bins
Values include the average of F values of 10 trials, t : average time for 500 iterations
BL4 BL∞ ML4 ML∞ MU4 MU∞
F t F t F t F t F t F t
Albano 0.495 79.2 0.504 0.504 0.506 84.5 0.523 180.5 0.525 87.3 0.532 178.0
Shape2 0.289 78.1 0.331 0.331 0.315 56.9 0.335 122.3 0.338 60.7 0.339 123.6
Trousers 0.557 374.6 0.562 0.567 0.564 349.2 0.565 783.5 0.590 379.8 0.574 814.6
Shape0 0.268 302.5 0.268 0.268 0.268 275.7 0.268 584.3 0.268 295.3 0.268 592.8
Shape1 0.270 552.6 0.369 0.369 0.270 494.0 0.376 766.5 0.275 494.0 0.369 823.2
Shirts 0.518 740.8 0.526 0.546 0.634 727.1 0.557 1337.6 0.665 680.1 0.664 1230.2
Dighe2 0.322 15.4 0.277 0.277 0.302 20.9 0.317 45.9 0.281 18.8 0.322 41.5
Dighe1 0.350 18.1 0.325 0.325 0.352 19.4 0.354 39.5 0.314 20.1 0.365 39.0
Fu 0.410 5.6 0.421 0.421 0.410 5.3 0.432 8.2 0.427 4.9 0.431 9.2
Han 0.387 94.6 0.376 0.376 0.384 89.4 0.405 215.2 0.402 84.8 0.402 161.6
Jakobs1 0.570 59.1 0.519 0.519 0.558 59.1 0.541 115.0 0.565 56.2 0.579 109.0
Jakobs2 0.408 49.0 0.401 0.401 0.398 55.8 0.417 93.1 0.413 57.1 0.438 96.8
Mao 0.494 119.1 0.493 0.493 0.496 119.2 0.494 201.5 0.495 124.8 0.501 192.6
Poly1a 0.320 23.0 0.322 0.329 0.335 24.1 0.322 46.2 0.320 26.8 0.349 46.2
Poly2a 0.361 96.9 0.358 0.358 0.353 78.9 0.368 131.7 0.346 80.2 0.366 150.7
Poly3a 0.424 164.3 0.431 0.431 0.427 142.0 0.435 310.7 0.425 147.0 0.442 273.9
Poly4a 0.398 282.2 0.412 0.412 0.392 243.3 0.433 464.7 0.401 295.7 0.437 441.8
Poly5a 0.405 345.4 0.439 0.459 0.394 397.5 0.457 640.5 0.458 335.0 0.465 613.0
Poly2b 0.378 86.7 0.381 0.381 0.381 78.2 0.389 154.2 0.396 93.7 0.402 154.1
Poly3b 0.434 185.4 0.432 0.432 0.433 195.2 0.431 249.9 0.428 175.6 0.462 292.4
Poly4b 0.461 343.2 0.465 0.465 0.453 349.9 0.465 533.2 0.401 289.9 0.483 568.2
Poly5b 0.415 441.7 0.439 0.479 0.438 450.3 0.478 698.4 0.478 412.6 0.479 834.9
Swim 0.339 2405.1 0.351 0.351 0.342 2625.8 0.351 4143.9 0.402 2246.6 0.368 3688.6
Avg. 0.403 298.4 0.409 0.413 0.409 301.8 0.422 515.9 0.418 281.2 0.436 499.0
Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes 95

Table 4.10: Results of the IJS1 algorithm for each nesting instances with 1.1md
small bins
Values include the average of F values of 10 trials, t : average time for 500 iterations
BL4 BL∞ ML4 ML∞ MU4 MU∞
F t F t F t F t F t F t
Albano 0.597 80.8 0.590 168.4 0.591 89.6 0.590 180.5 0.594 96.0 0.605 185.1
Shape2 0.300 85.1 0.300 121.9 0.293 58.0 0.300 134.5 0.300 61.9 0.300 133.5
Trousers 0.679 408.3 0.678 840.8 0.684 363.2 0.681 807.0 0.681 387.4 0.684 839.0
Shape0 0.244 296.4 0.244 560.2 0.244 295.0 0.244 654.5 0.244 292.4 0.244 616.5
Shape1 0.246 569.2 0.327 957.4 0.282 479.2 0.317 766.5 0.282 563.1 0.320 922.0
Shirts 0.601 829.7 0.590 1448.3 0.473 763.4 0.611 1377.7 0.465 680.1 0.611 1279.4
Dighe2 0.336 17.2 0.373 40.3 0.341 23.6 0.381 44.5 0.373 18.2 0.394 46.1
Dighe1 0.421 19.3 0.428 39.4 0.424 19.4 0.428 41.9 0.422 21.3 0.427 41.7
Fu 0.358 6.1 0.351 6.9 0.352 5.3 0.358 9.0 0.352 5.1 0.448 9.1
Han 0.437 105.9 0.430 206.9 0.447 99.3 0.445 241.0 0.461 90.7 0.446 168.0
Jakobs1 0.358 59.1 0.429 114.4 0.368 57.9 0.433 111.5 0.358 56.2 0.433 118.8
Jakobs2 0.365 48.5 0.415 102.8 0.401 58.1 0.415 97.8 0.401 58.8 0.418 101.7
Mao 0.437 120.3 0.450 229.1 0.447 131.1 0.450 205.5 0.449 138.5 0.472 190.6
Poly1a 0.300 25.9 0.402 44.2 0.310 27.3 0.451 49.4 0.309 29.5 0.454 47.6
Poly2a 0.349 104.7 0.353 131.4 0.352 78.9 0.363 134.3 0.367 80.2 0.452 150.7
Poly3a 0.384 185.6 0.368 320.1 0.373 154.7 0.382 326.3 0.367 148.5 0.451 304.1
Poly4a 0.402 285.0 0.349 489.7 0.384 267.7 0.351 520.5 0.383 337.1 0.409 437.3
Poly5a 0.382 335.0 0.364 578.5 0.367 437.3 0.399 704.5 0.358 355.1 0.413 692.7
Poly2b 0.371 95.4 0.412 145.7 0.37 80.5 0.412 161.9 0.339 100.3 0.419 154.1
Poly3b 0.373 207.6 0.384 292.0 0.381 218.6 0.452 247.4 0.378 172.1 0.456 304.1
Poly4b 0.406 332.9 0.407 551.5 0.407 356.9 0.418 549.2 0.405 292.8 0.428 562.5
Poly5b 0.397 503.6 0.422 766.4 0.419 490.8 0.422 726.4 0.415 416.8 0.435 935.1
Swim 0.330 2693.7 0.325 3999.4 0.326 2652.1 0.338 4019.5 0.338 2358.9 0.337 3688.6
Avg. 0.394 322.4 0.408 528.5 0.393 313.4 0.419 526.6 0.393 294.0 0.437 518.6
96 Chapter 4 Homogeneous Bin Packing Problems with Irregular Shapes

4.8 Concluding Remarks

In this chapter, we address a new solution method to solve the two-dimensional irregular
shape bin packing problem which has received little attention in the literature despite
clear industry relevance. We modify the existing use of the Jostle procedure of Dowsland
et al. (1998) applied to strip packing and convert the heuristic for the homogeneous bin
packing problem. Further, we introduced a diversification strategy into Jostle that
significantly improved the results. The proposed heuristic supports the use of restricted
rotation and unrestricted rotation of pieces.

The findings of the study are as follows:

Jostle is a promising packing heuristic that can be applied to bin packing problems. It
is demonstrated to be a low-cost search strategy that self-governs the neighbourhood
structure that modifies the solution. The designed algorithms have been shown to im-
prove on the results in the literature with lower computation time and do not require
the use of expensive MIP (Mixed Integer Programming) software, making it attractive
to small companies.

We have introduced a new way of dealing with free rotation of pieces. We have demon-
strated the value of free rotation for nesting instances and justified the results of re-
stricted rotation for the jigsaw puzzle instances.

We have designed and tested a number of placement policies and shown that despite the
extensive use of bottom-left (BL) this does not produce the best results when compared
with minimising length or maximising utilisation.
Chapter 5

Irregular shape Bin Packing in

Heterogeneous Bins

5.1 Introduction

This chapter discusses the solution methods to solve the two-dimensional irregular shape
bin packing problem with heterogeneous bins. The problem mainly applies in industries
such as sheet metal, foam cutting, furniture, plastic, paper and leather; where a set of
polygonal shaped pieces need to be cut out of rectangular stock sheets (bins). The bin
sizes are heterogeneous and more than one bin is usually required to satisfy the total
demand for pieces. All the demanded pieces must be placed in the bins in a way that the
total area of the utilised bins is minimised. We denote this problem as two-dimensional
irregular shape multiple bin size bin packing problem (2D IMBSBPP).

The problem selection is particularly motivated by the set of real-life characteristics

common for several industries. The typical form of sheet metal, foams, and timber
sheets are readily available to purchase in their standard sizes due to the convenience
of handling and storage. Most of these sheet surfaces are homogeneous and therefore
do not require to restrictions on the orientation of the pieces when placing them in the
layout. Therefore, restrictions on piece rotation are not needed.

Instead of using one standard size, companies usually use several standard sheet sizes
which are economical, in satisfying the customer demand. The aim is not only to reduce
the waste generated in the cutting process, but also to charge a competitive price to
the customers, which usually depends on the area of material used to meet the demand.
In this study, we present an algorithm which is capable of exploiting the fact of having
different standard stock sheet sizes when assigning, rotating and placing irregular pieces
so that the trim loss can be significantly reduced.

97
98 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

As reviewed in Chapter 2, the previous studies on 2D irregular shape bin packing were
limited. Only a few studies have been published recently for 2D irregular bin packing
with homogeneous rectangular bins (e.g. Lopez-Camacho et al. (2013a), Lopez-Camacho
et al. (2013b), Lopez-Camacho et al. (2014), Martinez-Sykora et al. (2017)). Out of
those, only Martinez-Sykora et al. (2017) considered the continuous rotation of pieces
that added more complexity to the problem. To the best of our knowledge, Babu and
Babu (2001) and Baldacci et al. (2014) were the only papers that had discussed packing
irregular shapes in heterogeneous bins. Both these studies presented results for packing
pieces in irregular shaped bins. In real-life, the number of applications of this problem
is limited (i.e. Leather industry) compared to the number of applications of packing
irregular pieces into heterogeneous rectangular sheets. Both papers adapted the raster
method and approximated the geometric shapes of pieces and stock sheets which cause
low accuracy in piece placement. However, the study discussed in this chapter focuses on
irregular shapes packing in heterogeneous rectangular bins with the unrestricted rotation
of pieces. Using the geometric representation method the study improves the accuracy
of solutions.

5.2 Contribution

The main contribution of this chapter is to develop methods to solve the 2D IMBSBPP
considering the unrestricted rotation of pieces.

This chapter describes three main computational methods to solve 2D IMBSBPP and
identifies an efficient method to produce good solutions. The first two methods are based
on Jostle approach discussed in Chapter 4 and the third method is inspired by the idea
of improving the overall utilisation by selecting the most appropriate bin types using a
classic bin selection heuristic.

As explained in Chapter 4, compared to other available algorithms, one of the most

important properties of the Jostle is the exploration of the solution space in a relatively
low computational effort, especially when the piece rotation is not restricted. The Jostle
heuristic explores the search space over the permutation of the pieces and it applies in
a scenario where the sequence of bins and the placement rule are fixed. We propose a
modification to the constructive algorithm discussed in Chapter 4 for the 2D ISBSBPP,
and this is described in Section 5.4. We then propose the Jostle heuristic to explore the
search space over the permutation of the pieces, within a given sequence of bins.

The first main computational method extracts the features of Jostle as well as the power
of Genetic algorithm (GA) to solve 2D IMBSBPP. Babu and Babu (2001) considered
GA to search over the solution space of the problem. They coded a long chromosome
representing the sequence of bins, the sequence of pieces and sequence of the orientation
angles of the pieces. Each chromosome represented a solution. This representation is
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 99

inefficient when the number of pieces and bin types are high. Also, the piece orientations
of this approach were limited to a restricted set of angles. In the proposed new approach,
we involve the efficient Jostle heuristic to handle the placement and orientation decisions
of pieces within the GA implementation so that the overall GA/Jostle implementation
can solve 2D IMBSBPP with either restricted rotation of pieces or unrestricted rotation
of pieces.

As the second main computational method, we propose an iterative search procedure

for the Jostle heuristic by applying kicks for the sequence of bins at frequent intervals so
that the bin sequence can be changed during the search process. In this case, we extend
the Jostle heuristic developed for 2D ISBSBPP to generate solutions for 2D IMBSBPP.

In the third algorithm, for each permutation of the pieces, the solutions are generated
using a classic bin selection heuristic and some of the features of the constructive method
proposed in Chapter 4. In this case, the quality of a solution is based on the piece
permutation and the placement rule only, as the sequence of used bins is determined
by the classic bin heuristic. By searching over different permutations of the pieces, the
overall computational method can find a good solution.

In comparison to the other approaches in the literature, to our knowledge, this is one
of the few studies available for solving 2D IMBSBPP and the first study addressing
unrestricted rotation of pieces in 2D IMBSBPP.

5.3 Problem Description

Let N 0 be the given number of rectangular bin types (stock sheets) and let Lk and Wk be,
the length and the width of bin type k ∈ {1, . . . , N 0 }, respectively. Let P = {p1 , . . . , pn }
be the set of pieces to be cut from the bins where n be the number of pieces to be cut.

A solution s = {Bks | k = 1, . . . , N 0 } is represented by a set of bins of each type used

in the solution, where each single bin bsjk (Pjk , Ojk , Xjk , Yjk ) ∈ Bks is determined by
the subset of pieces placed in the bin (Pjk ), the set of orientations used for each piece
(Ojk ) and the X and Y coordinates of the reference point of each piece (Xjk and Yjk ),
respectively. In this case, j denotes the index of each single bin of type k. A solution
S 0 S
s is a feasible solution if all the pieces are placed, N
k=1 j| bjk ∈B s P jk = P , there is no
k
overlapping between each pair of pieces placed in the same bin and no piece exceeds the
bin dimension.

Initially, we consider an unlimited number of bins of each type k, (i.e. There are enough
bins of each type to place all the pieces). We denote Nks as the number of used bins of
type k in solution s.
100 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

The position of the pieces inside the bins is given by the coordinates of the reference
point, which we assume as the bottom-left corner point of the enclosing rectangle. Note
that the reference point may change if the piece is rotated.

5.3.1 Evaluation function of a solution

We use two measurements to evaluate the quality of a complete solution s.

1) The overall utilisation Us is defined as;

Pn
Area(pi )
Us = PNi=1
0
s
k=1 Wk Lk Nk

Pn
where Area(pi ) denotes the area of the ith piece. Since i=1 Area(pi ) is constant, then
maximizing the utilization is equivalent to minimizing the total area of bins used in s.
In this chapter our aim is to maximize Us .

2) Let σs be the standard deviation of absolute bin waste of the used bins. This mea-
surement is used during the algorithm to break ties when two solutions have the same
overall utilisation. A low σ value shares the waste among the bins with low variance and
it balances the individual bin utility. A higher σs leads to a higher variation in bin space
utilisation. This effect is prominent in situations where both dense and loose packing is
possible. Therefore, within the same solution, a set of bins gets a higher utility while
another set of bins gets a lower utility. We encourage this types of solutions (i.e. a
higher σs values when there is a tie in the Us values) during the search process of the
algorithm to move the few pieces packed inside a large bin (i.e. a lower utilised bin)
to another used bin. Therefore the σ is a significant measurement at the improvement
stage of the algorithm.

In literature for 2D MBSBPP with rectangular items, the cost of bins is basically defined
in two ways; 1) The value of the bin is proportional to its area (e.g. Ortmann et al.
(2010), Hong et al. (2014)), 2) The value of bin is not proportional to its area (e.g.
Pisinger and Sigurd (2005)). In the second case, there are possibilities of incurring a
higher cost for large bins than the cost proportional to its area or incurring a relatively
higher cost for small bins than the proportional value corresponds to its area. In this
study, we selected the first option when deciding the cost of bin usage by valuing only
the usage of sheet material area. Therefore, we use Us as the main measurement when
deciding the good solutions
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 101

5.4 Packing Procedures

In this section, we introduce the three computational methods used to solve 2D IMBS-
BPP. The packing procedure discussed for 2D ISBSBPP in Chapter 4 follows the strip
packing approach while imposing division between bins. This made bin allocation im-
plicit within the placement decisions. However, 2D IMBSBPP contains rectangular bins
with different dimensions. When a sequence of bins is fixed with a set of heterogeneous
bins, then the 2D IMBSBPP requires us to address two decision problems; assigning
pieces to bins and placing those pieces in the bin, similar to the packing procedure
discussed for 2D ISBSBPP. Therefore, for a given sequence of heterogeneous bins, our
approach stimulates jostling of pieces to solve both the decision problems.

As explained in Chapter 4, the idea is to begin at a random order of pieces and apply
a constructive heuristic. The Jostle heuristic determines the placement order of pieces
for the next iteration, by scanning the piece positions of the previous layout. In order
to adapt the Jostle for 2D IMBSBPP, we first define a very long packing strip of which
the width is equal to the maximum width of the available bin types; Wmax . The length
of the strip is decided based on how the number of bins is arranged along the strip. The
P 0 0
strip length 2n × N k=1 Lk was set to arrange 2 × n × N bins along the strip as follows.

For the first step, the bin order is determined randomly and this represents a possible
random order of given heterogeneous bin types. As an example, the strip represented
in Figure 5.1 denotes a set of bins joined together. We defined bin spaces by assigning
a set of barrier lines at different intervals of length and width, to denote each bin
arranged along the strip, following the generated random order of bins. The bottom left
0 P(v−1)
corner of bin v ∈ {1, . . . , n × N } on the strip is placed at ( i, =0 Li0 , 0) (see Figure
5.1). We directly applied the Jostle procedure to the strip with additional constraints
that pieces may not be placed across a barrier and should always be placed entirely
within the bins. As illustrated in Figure 5.1, the algorithm arranges bins to facilitate
jostling between the both ends iteratively. Dowsland et al. (1998)’s method shuffles
pieces along a homogeneous strip towards left and right. However, in our case, since
the bin arrangement is heterogeneous, we considered the mirror image of one side to the
other of the strip, so that same bin configuration can be established at both ends when
the Jostle performs within a given bin configuration.

5.4.1 Constructive Algorithm 1 (CA1)

Following this procedure, for a given order of bins, we apply the constructive algorithm
described in Section 4.4 when placing pieces in the bins. The algorithm facilitates
loading of pieces into the bins, allowing feasible placements of pieces, hole-filling and
unrestricted rotation of pieces. The main difference of this constructive method is its
consideration of heterogeneous bins arranged in the Jostle strip. In each of our designed
102 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Figure 5.1: Strip to Bins

computation methods, a complete packing solution is evaluated according to the quality

of the solutions, which is measured by Us . As a part of solution construction, both Us
and σs are calculated for each solution, to differentiate between solutions with the same
Us values. For the forthcoming uses of this constructive method, we denote it in this
chapter as Constructive Algorithm 01 (CA1) for solving 2D IMBSBPP.

5.4.2 Solution improvement by Jostle heuristic

In this section we describe how we adapted CA1 within the sequence search algorithm;
Jostle. In this case, our proposed Jostle heuristic starts with a random order of pieces
and a random order of bins to construct the first layout. Subsequently, the piece permu-
tation is decided by the Jostle heuristic when jostling continues with pieces within the
selected order of bins. The placement rule sets the criteria to compare the orientation
and placement position of a candidate piece in a feasible manner. As described in Chap-
ter 4, Jostle is analogous to a local search where the search occurs over the sequence of
the pieces and each Jostle iteration generates a neighbour sequence. The constructive
algorithm takes a certain sequence of pieces and creates a layout for a certain config-
uration of heterogeneous bins. In other words, within a given configuration of bins,
Jostle allows to generate a new layout, by defining a new sequence of pieces. This is a
neighbour solution based on the arrangement of pieces of the previous packing layout.
The optimization arises from generating neighbour solutions in which the constructive
algorithm generates solution when there is a change in the piece permutation. Despite
its performance by finding the promising orientations and placements of pieces, the ap-
proach is limited to finding a good piece arrangement within a given configuration of
bins only. This is a major concern for our problem since the search cannot explore the
different configuration of bins.
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 103

We proposed two search mechanisms to explore the bin permutation and tested their
performance. The first method involves both Jostle and Genetic Algorithm to search
through the different configuration of bins considering a population of solutions at a
time. The second is a single point search method which allocates a separate kick-move
mechanism to change the bins configuration while Jostle changes the permutation of the
pieces.

5.4.3 Hybrid approach of GA and Jostle

The first computational method implements a Genetic Algorithm (GA), to work with
multiple solutions in parallel and to combine them together to generate a new solution.
In the C&P literature, GA has been implemented for SBSBPP with rectangular pieces
by Gonçalves and Resende (2013), for MBSBPP with rectangular pieces by Babu and
Babu (1999) and for packing irregular pieces in irregular stock sheets by Babu and Babu
(2001). In our implementation, a candidate solution of 2D IMBSBPP is represented by a
sequence of input bins and a permutation of the pieces. For each chromosome, the initial
input is a random permutation of bins and a random permutation of pieces. This search
algorithm applies the constructive algorithm CA1, for each decided permutation to gen-
erate solutions. The algorithm stores information related to each candidate chromosome
and their corresponding solutions; i.e. permutation of bins, pieces, piece orientations,
and associated fitness of a solution.

An example coding of a chromosome when packing 16 pieces into the bins having five
different types (sizes) are presented in Figure 5.2. The length of the bin order is equal
0
to n × N . In Figure 5.2 the numbers 2,4,4,1,...,3,4,1,2 represent the ordered bin types
(IDs) of the input bins. The second part of the code represents the sequence of pieces. A
chromosome corresponds to a certain bin packing solution that can be generated using
the constructive algorithm. The quality of each chromosome is evaluated using the
fitness function f (B ∗ , P ∗ ). In this case, the value of the function is determined using Us
and σs measurements which we discuss in detail at the solution evaluation stage of this
method.

Figure 5.2: Representation of solutions

 104 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

5.4.3.1 Outline of HGAJ

Algorithm 5 describes the general working procedure of HGAIJ. Let popSize be the
number of chromosomes in the population progresses from one generation to another.
As explained above, initially popSize number of solutions are generated using CA1 by
taking a set of random order of bins and random permutation of the pieces. In our
implementation, these set of chromosomes represents the population at first generation
(GenN = 1). Using these chromosomes, the algorithm finds their fitness values by
generating the corresponding solutions. As the next step, the crossover and mutation
operators are applied to this population of chromosomes and generates a new set of
offspring solutions. The generated offspring solutions are further processed using the
Jostle. Once this is done, all the generated offspring are added to the population so that
the total group includes the new set of offspring as well as the chromosomoes stayed at
the start of the generation. As the next step, the selection mechanism decides which
chromosomes become parents in the next generation. The selection mechanism selects
popSize number of chromosomes from the existing group of chromosomes, and those
who do not get selected are considered as discarded. In this way popSize number of
chromosomes are transfered from one generation to another. Also, the chromosomes
transferred to the next generation can contain some of the chromosomes at the start of
the generation and some of the new offspring chromosomes made by the GA operators
and improved by the Jostle heuristic. The purpose of the crossover and mutation oper-
ators is to create new solutions by changing the bin order and piece order by using the
parent chromosomes available in the current population. Likewise, the GA cycle works
with a population of solutions, and members in the population are updated or changed
at the selection stage. Next, we describe each of these steps in detail.

Algorithm 5: General structure of GA Jostle method

1 Initialise a population of chromosomes;
2 Evaluate each chromosome in the population;
3 Main loop: Generations;
4 GenN = 1;
5 while GenN < M axGen do
6 Create new chromosomes by mating chromosomes using crossover and mutation;
7 Improve offspring solutions by Jostle;
8 Evaluate and Insert new members into the population;
9 Select set of chromosomes to the next generation from new offspring and old
parents;
10 GenN = GenN + 1;
11 end
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 105

5.4.3.2 Crossover

Once parent chromosomes are selected from the chromosomes by the selection mecha-
nism, the crossover parameter Pc determines whether those selected parents will undergo
crossover or not. The parents who do not undergo crossover will be cloned (i.e copied)
as offsprings. In this case, Pc denotes the probability that a parent undergoes crossover.
The selection mechanism selects parents for mating in the current generation from the
chromosomes in the previous generation. We describe this procedure later; under the
sub topic called Selection for the next population.

The crossover operator is applied for two parent chromosomes and creates two new off-
spring chromosomes. The underlying idea of crossover is to produce offspring chromo-
somes, which can inherit common genes from their parents. Following this, our objective
was to include the key genes or sections of genes shared by many of the fittest chromo-
somes inside the offspring chromosomes. We use two forms of crossover operators in this
study; uniform crossover for the bin part of the chromosome and ordered crossover for
the piece permutation, as follows.

First, we use the uniform crossover operator, which facilitates a search through most
of the possibilities of re-combining different genes in parents (Falkenauer, 1999). An
example, Figure 5.3 illustrates the uniform crossover operation to generate a child chro-
mosome. The operator first constructs a binary mask, with random bits of 0 and 1. The
first child then inherits the genes (in the bin order) of the first parent if there is a ”1” in
the mask; otherwise, the first child inherits the genes of the second parent. The second
child is formed following the opposite way so that the genes are copied from the second
parent if there is a ”1” in the mask and from the first parent if there is a ”0” in the
mask.

The second crossover operator is used to change the piece order (in this case the bin
string related for piece permutation only). The procedure of it applies as follows. First,
from the starting point of the bit-string up to a randomly selected point is copied from
the first parent to the first offspring as illustrated in Figure 5.3. The second parent’s
piece permutation is then scanned and finds the bits which are not yet included in the
offspring to copy them to the offspring. The same procedure is used to generate the
second offspring, in which case copying genes are started from the second parent and
then from the first parent.

5.4.3.3 Mutation

After crossover, the mutation parameter Pm determines whether those selected offspring
will undergo mutation or not. The purpose of implementing mutation in this algorithm
is to maintain diversity within the population and inhibit premature convergence. In
106 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Figure 5.3: Crossover operation

our approach, we expect a major change in the bin configuration of the chromosomes
selected for mutation. In this case, we adopt inversion mutation technique to change the
bin permutation of offspring chromosomes subject to the probability parameter (Pmu ),
as described in Eiben and Smith (2013) and Mutingi and Mbohwa (2016).

If a chromosome undergoes mutation, then the inversion mutation is implemented as

follows. As described in Eiben and Smith (2013), the inversion mutation selects two
positions of the gene string randomly and reverses the order of genes in that selected
substring. In our case, we only target the bin bit array of the chromosome. As the two
positions of the gene arrays, we select one position randomly from the genes corresponds
to the used bins and one position from any other random place in the gene array corre-
sponds to the bins of the chromosome. The operator then reverses this selected string
of genes and updates the gene array of the chromosome.

By doing this, we noticed a major change in the corresponding solution as bin spaces
of the layout is dramatically changed. Accordingly, the piece placements have also been
changed. This explores solutions in the search space at a greater distance and makes
the search mechanism productive by inhibiting premature convergence.

5.4.3.4 Implementation of Jostle within the GA framework

Once a child chromosome is made using a certain permutation of bins and pieces, the
Jostle procedure is used to improve the piece layout within the sequence of bins of that
chromosome. Jostle improves the child chromosome by changing the orientation and
placement position of pieces. Since Jostle is efficient in managing orientations of pieces,
we do not propose any change for the piece orientations through genetic operators.
On the other hand, handling orientation angles by genetic operators and generating
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 107

solutions accordingly will require higher computational time as it requires us to consider

a higher number of orientation options. Instead, in Chapter 4 we already described a
new angle tuning mechanism which finds promising orientation angles efficiently once
it is embedded with Jostle. Therefore, we implemented Jostle to improve offspring
chromosomes generated by the GA operators. The number of Jostle cycles (maxJS)
used to improve the offspring which was found by a parameter evaluation study.

5.4.3.5 Selection for the next population

The next population is selected from the P ool that currently contains both the parent
and children solutions. The proposed selection strategy is set to ensure both quality and
diversity of the next population. The chromosomes in the next population is selected
from the chromosomes in P ool, using elitist and tournament selection methods.

The selection process uses three measurements; U , σ and dist to select the chromosomes
that move forward to the next generation. The chromosomes in P ool are sorted in
descending order, first by their U value and then by the σ value. The measurement d
is set to maintain the diversity among the solutions having an equal number of used
bins. If two solutions have the same bin permutation with the same number of used
bins, then our objective is to make sure one of them is not in the next population. This
provides an opportunity for another solution which is having the same Us value, in the
next generation. The following example presents such incident.

Example:

In this example, the bins in bold font denote the used bins in the bin sequence. The
first bin size is 2.5dmax in solution A; where the length and width of the bins are equal
to 2.5dmax . The measurement dmax denotes the maximum dimension of the enclosing
rectangle of each piece in their original orientation.

Solution A: Bin arrangement 2.5dmax 1.5dmax 1.0dmax 2.0dmax 1.0dmax ....2.5dmax

Solution B: Bin arrangement 2.5dmax 1.5dmax 1.0dmax 2.0dmax 1.0dmax ....2.5dmax

Solution C: Bin arrangement 2.5dmax 1.5dmax 1.25dmax 1.75dmax 1.0dmax ....2.5dmax

Both solutions A and B have the same utilisation, the same number of used bins and
probably different σ values due to the arrangement of pieces. In this example, assume
sigma of A is higher than sigma of B. The solution C also has the same Utilisation, the
same number of used bins and probably different σ, lower than σ of A and σ of B. Then,
out of those three, in the case where only two need to be selected, we set the procedure
to promote selecting A and C by defining a distance (dist) value for each chromosome
as follows.
108 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

The best solution of the pool gets a distance value ’0’, and compared to its bin configu-
ration, The other chromosomes get dist value. The distance of a chromosome x; dist(x)
is calculated as:

dist(x) = |x1 − y1 | + |x2 − y2 | + ...... + |xN x − yN x |

where xi denote the area of the ith bin located in the bin permutation of the chromosome,
yi denote the area of the ith bin located in the bin permutation of the chromosome who
got distance value ’0’. In a situation where the chromosomes x and y have a different
number of used bins (different N values), then the distance is calculated up to the N
value of chromosome x (denoted as Nx ). Following this mechanism, both Solutions A
and B gets same dist value and C gets a different dist value compared to A and B.

When selection sorts the best chromosomes, the calculated d values are considered in
order to break the ties of solutions with equal U values as explained in the above example.
In this step of the algorithm, σ values have not been considered since the priority is
assigned first to get a better Us and then to pick a higher dist value at ties in Us . In
which case, however there is an exception in picking the higher dist value only at the
selection which involves the best chromosome in pool whose dist = 0.

Elitist selection scheme

In order to keep a balance between the diversity and convergence speed, the elitist
selection scheme copies a fraction (γ%) of the best chromosomes in P ool to the new
population. γ% is a parameter of this algorithm. When arranging best chromosomes,
quality of chromosomes is evaluated by the priority order of U , dist and σ values. Once
a γ% × popSize of chromosomes are selected for the next generation, the remaining
100 − γ% × popSize are selected using the tournament selection.

Tournament selection scheme

The tournament selection makes a tournament among few of chromosomes chosen at

random from the available chromosomes and selects the best out of that tournament so
that it will be present in the next generation. The tournament size Ts is determined
usually at the parameter tuning stage.

In our study, we use both elitist and tournament selection schemes to promote the selec-
tion of fit chromosomes, ensuring a reasonable diversity in bin configurations so that we
can reduce the early convergence. Through the tournament selection we expect to pre-
serves diversity in the selected chromosomes by increasing a chance to all chromosomes
to be selected. In this case we test for a good Ts value to work with. In our studies, we
realised that higher tournament size provides less chance for weak chromosomes to be
selected and this leads possibly losing diversity.
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 109

5.4.3.6 Addressing some issues of GA Jostle approach

One major issue of this algorithm is the high computational time it takes to achieve
good solutions. The algorithm takes a longer time to transfer from one generation to
the next, hence reduces the frequency of bin-moves. At each generation, a significant
amount of computational time is devoted for Jostle to improve at each solution and this
effort is multiplied by the number of offsprings generated at each generation.

In order to reduce this computational effort, we first tried to reduce the depth of Jostle
search by reducing the M axJS. This results in a drop in the solution quality determined
by Jostle.

Next, we tried to apply Iterated Jostle procedure discussed in Chapter 4, only for the
newly generated best offspring. This minimises the computational effort of Jostle on
other offspring. The second change has improved the performance of the algorithm
by saving time with a lower compromise of its exploring ability. Accordingly, the GA
framework is implemented on a population of chromosomes to explore different bin con-
figurations, hence improving the piece arrangement of the best offspring. In summary,
while crossover and mutation operators search over the bin arrangements and the piece
permutations, the Iterated Jostle focuses on exploring permutation of pieces and orien-
tation angles. We denote this approach (the second version of HGA: HGA v2) as the
Hybrid approach of Genetic Algorithm and Iterated Jostle (HGAIJ).

5.4.4 Iterated Jostle with Random Assignment of Bins (IJRAB)

This is the second main computational method proposed in this chapter to solve 2D-
IMBSBPP. As the name implies, IJRAB uses the Jostle approach discussed in Section
5.4.2. The constructive method CA1 described in Section 5.4 is involved in generating
solutions for a given order of bins and for a given permutation of the pieces. IJRAB
is a single point search algorithm working over the sequence of bins and pieces. The
previous Hybrid approach of GA and Jostle considers multiple solutions in parallel. In
contrast, the IJRAB method involves jostling of pieces within a given order of bins and
changes the bin order at frequent intervals while Jostle continues.

The IJRAB starts with a solution generated by the constructive algorithm CA1 using
an arbitrary order of bins and an arbitrary order of the pieces. Then the Jostle heuristic
determines a neighbourhood structure by shaking pieces from left to right and right to
left along the bin order, so that each Jostle iteration generates a neighbour sequence of
the pieces. As discussed in Chapter 4, Jostle finds a local optimum after a pre-defined
number of Jostle cycles (Kp ) with no improvement. Within the same bin order, the
Iterated Jostle (IJS) is implemented to find a good solution, by applying kicks to the
piece order as described in Chapter 4. In this algorithm, we denote these kicks as piece
110 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

kicks, which are applied when the search is at a local optima (after Kp Jostle cycles).
Likewise, for a selected bin order IJS is performed iteratively until no improvement
is found for a pre-defined number of piece kicks (Kb ). As the next step of IJRAB, the
algorithm performs a perturbation to the current best solution found, instead of starting
IJS with a completely new order of bins. In this case, the next bin order is decided by
the perturbation applied to the bin order corresponds to the current best solution. We
denote this step as the application of a bin kick. Once the next bin order is decided,
IJS jostles pieces again starting from the piece permutation corresponds to the current
best solution. This procedure continues with IJS and the bin kicks. The current best
solution can be updated at any point of the algorithm by an improved solution found.

The purpose of implementing bin kicks is twofold. First, it allows to escape from one
neighbourhood landscape to another neighbourhood landscape and starts searching in
a new area of the search space. According to our experiments, this enhances the search
ability of the mechanism and allows us to explore different local optimum solutions in the
search space. Second, by changing the bin order, it removes the limitation of that Jostle
heuristic being considered only for a selected order of bins. In this implementation, a
bin kick is performed when the search process is reached to a predefined number of piece
kicks with no improvement (Kb ) in solutions within the selected bin order. In the next
set of paragraphs we describe the implementation details of these two kick types.

• Piece kick: We perform the insert move as described by Bennell and Oliveira
(2009) for a single piece (i.e. one-piece insert move) in the piece permutation of
the leading solution as follows: First, we select two random positions along the
permutation of the pieces and then remove the piece corresponds to the highest
position selected. The algorithm is set to insert this piece in front of the piece
corresponds to the other position which is selected randomly. Accordingly, the new
piece order is derived after ordering the remaining pieces, following the existing
permutation.

• Bin kick: In order to change the bin order, a random bin is selected out of the
used bins of the current best solution found and it is replaced with a random bin
selected out of the other bin types. Once this is applied the same corresponding
change is applied for the mirror bin position of the bin order as well so that the
same bin configuration can be established at both ends when performing Jostle.
Application of this bin change is critical for this algorithm to explore the solutions
related to different bin configurations.

Algorithm 6 describes execution steps of IJRAB. We use P (L,t) to denote piece order
at tth iteration, which packs pieces from left to right. P (R,t) denotes the piece order
which packs pieces from right to left at tth iteration. The bin order at tth iteration is
denoted as τ (t) . For a given solution s(t) , f (s(t) ) denotes the evaluation value by taking
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 111

into account the values of Us and σs (see Section 5.3.1). The current best solution of all
the executed iterations is denoted by s∗∗ . Also, the current best solution of a particular
bin order is denoted by s∗ . Tmax defines the maximum number of total Jostle cycles
that the algorithm which is the termination criterion of the algorithm. Similar to the
maximum number of total Jostle cycles, the termination criterion can be set run for a
certain amount of time.

5.4.5 Sequential Packing with a Bin Centric Heuristic (SPBCH)

In this approach, we build a feasible solution for a given permutation of the pieces. Our
approach is inspired by the idea of improving the utilisation of bins of classical best
fit heuristic. Our focus is to find the best bin types configuration that packs pieces
for higher utilisation. Following a given arrangement of pieces, the proposed bin-centric
heuristic determines an order of bins for the layout. We describe this procedure in section
5.4.5.1 and discuss how solutions can be improved through the local search mechanism
implements in Section 5.4.5.2.

5.4.5.1 Constructive Algorithm 2 (CA2)

The constructive algorithm initiates all pieces as an unpacked piece list (Pu ) and this
is the first step of the algorithm. Following an initial piece order, pieces are inserted
temporarily to one bin starting from a bin type k. When this progresses, if a certain
piece in the order of Pu cannot be placed within the selected bin, the next piece in
the order is tried for insertion. The process continues until all the pieces in Pu are
tried to fill the bin in a feasible manner. Then the utilisation of the bin is recorded.
The algorithm follows the same procedure for one bin from each bin type, and for each
different bin, record the utilisation it can have when pieces are inserted starting from
the initial unpacked piece list. Then the algorithm selects the highest utilised bin of
each bin type and considers that bin as the first used bin of the solution. This first
bin carries a certain number of pieces and those pieces are considered as packed pieces.
Once this is done, the other remaining pieces are updated as unpacked pieces.

With the updated unpacked list, the algorithm then tries to find the next bin to allocate
those some pieces in the unpacked list following the same procedure, by testing pieces
with the bin types given. Similarly, the process continues until pieces in the unpack
list become empty. Algorithm 7 describes how this procedure works. This algorithm
constructs a bin packing solution for a given permutation of pieces and determine the
bin order to be set in the solution through a dynamic bin selection mechanism runs
inside of it.
 112 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Algorithm 6: IJRAB
n 0
o
Input : Input Bin types k = 1, . . . , N , Placement Rule M U , Piece kick parameter (Kp ),
Bin kick parameter (Kb )
∗∗
Output: s∗∗ = {Bks | k = 1, . . . , N 0 }
1 Set t = 1, D∗ = L, Initialize the count of piece kicks and bin kicks qp = 0, qb = 0 ;
2 Set P (L,t) as a random order of pieces, Set τ (t) as a random order of bins;
3 Set f (s∗D∗ )) = − inf, Set f (s∗∗ D∗∗ )) = − inf;
4 while t ≤ Tmax do
5 Generate solution layout s(L,t) from P (L,t) and τ (t) ;
6 Evaluate f (s(L,t) );
7 Derive P (R,t) from the solution;
8 Generate solution layout s(R,t) from P (R,t) and τ (t) ;
9 Evaluate f (s(R,t) );
10 if f (s(L,t) ) > f (s(R,t) ) then
11 if f (s∗D∗ ) < f (s(L,t) ) then
12 Set s∗D∗ = s(L,t) ;
13 D∗ = L;
14 Reset qp = 0;
15 else
16 qp = qp + 1;
17 end
18 else
19 if f (s∗D∗ ) < f (s(R,t) ) then
20 Set s∗D∗ = s(R,t) ;
21 D∗ = L;
22 Reset qp = 0;
23 else
24 qp = qp + 1;
25 end
26 end
27 if f (s∗∗ ∗
D ∗∗ ) < f (sD ∗ ) then
∗∗ ∗
28 Set sD∗∗ = sD∗ ;
29 D∗∗ = D∗ ;
30 Reset qp = 0;
31 end
32 if qb < Kb then
33 if qp > Kp then
34 Apply piece kick to the piece permutation of s∗D∗ and obtain a new piece
permutation;
35 Geneate solution layout from the new piece permutation;
36 Derive next corresponding P (L,t) from the solution for jostling;
37 qb = qb + 1;
38 Reset qp = 0;
39 else
40 Derive P (R,t) from the solution;
41 end
42 else
43 Apply bin kick to the bin sequence of s∗∗ D ∗∗ and update τ
(t)
;
∗∗
44 Obtain the piece permutation of sD∗∗ ;
45 Derive next corresponding P (L,t) from the solution for jostling;
46 Reset qb = 0, qp = 0;
47 end
48 t = t + 1;
49 end
50 return s∗∗D ∗∗ ;
 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 113

When packing pieces to a bin, we adapt the features of the constructive method discussed
in Chapter 4. Since we pack pieces to only one bin at a time, the constructive method
described in Chapter 4 is implemented in a way that the pieces are packed only for one
bin space rather considering them as the sequence of bins at once.

Algorithm 7: Constructive algorithm for SPBCH

0
Input : List of Bin Types k = 1, ..., N , List of Pieces P , Placement Rule M U
Output: a feasible packing of the pieces P ∗ into the bins B ∗ and U ∗
1 Pu = P ;
2 Initialize Used bin index j = 1;
3 while (Pu is not empty) do
4 for Each bin type k do
5 for Each PIECE in Pu do
6 check for a feasible placement of PIECE in bk,j ;
7 if PIECE placement is feasible then
8 pack temporary the PIECE inside bk,j ;
9 end
10 end
11 Get the bin utilization uk ;
12 Record the PIECE s layout
13 end
14 Select the bk,i corresponds for maximum uk ;
15 Select bk,j and assign as b∗k,j with PIECE s layout;
16 Update Pu list by removing the packed PIECE s in bk,j from the list;
17 j = j + 1;
18 end
19 return packing layout P ∗ in B ∗ , U ∗ ;

5.4.5.2 Solution improvement

The solutions generated in section 5.4.5.1 are improved by implementing a local search
mechanism. We choose an initial solution S using a random permutation of the pieces
and compute the evaluation value of S (denoted as f (S)) as described in Section 5.3.1.
Initially, we assign S ∗ = S and U ∗ = U . During the search process, we generate a
0 0
neighbour solution S and compute f (S ) at each iteration. If the evaluation value of
0 0
new neighbour is better, then S and U replaces the current best S ∗ and U ∗ . Otherwise,
S ∗ is retained as the current best solution. The proposed local search improves the
current solution until reaching a local optimum when there is no further improvement
for a pre-defined number of iterations (cmax ). In order to generate a neighbour solution,
we propose following piece move mechanism.

Since the search occurs over the sequence of the pieces, each iteration generates a neigh-
bour sequence by a swap move (Bennell and Oliveira, 2009) of two random pieces of
the current best solution’s input piece order. The constructive algorithm takes this new
sequence of pieces and creates the next layout. At each iteration, only one neighbour
is generated. The complete solution is evaluated according to the evaluation criteria
discussed in section 5.3.1. At the end of the search, the best-recorded solution is kept.
114 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

In order to improve the search process, we also propose insert piece move (Bennell and
Oliveira, 2009) to disrupt the sequence to a greater extent when there is no improvement
in solution by the swap moves for a pre-defined number of iterations (cmax ). This kick is
implemented to move the local optimum to a new area of the solution space. We follow
one piece insert move to change the input piece order of the current best solution.

5.5 Results and Analysis

5.5.1 Data instances and implementation

For benchmarking, we could not find any instance of the specific problem discussed
in this chapter. To our knowledge, computational results of this specific problem are
not available in the literature. Therefore, we used the nesting instances published on
ESICUP (EURO Special Interest Group on Cutting and Packing) website for our ex-
periments. In order to evaluate the performance of the proposed algorithms, it is tested
with following instances as well as the newly created instances in order to match with
the irregular bin packing problem. The algorithms were coded in Visual C++ 2012 as a
sequential program. All computations were carried out in a personal computer Intel 2.60
GHz (Sandybridge) processor and 4GB RAM. Algorithms contain no parallel processing
and can be run using a single processor.

When selecting instances, we considered the production characteristics of sheet cutting

industries. In this case, we excluded the instances from garment industry because they
are solved as open dimension problems where the dimensions of the stock sheet; usually
the length, are not fixed. Since this problem deals with fixed dimension of stock sheets,
the garment instances are not considered in computational experiments. Another reason
for this exclusion is that garment instances are mostly used with restricted rotation while
our problem mainly considers unrestricted rotation of pieces.

Accordingly, 14 irregular shape instances (in Table 5.1) representing both convex and
non-convex polygons were used for experiments with 9 different Bin sizes: 0.5 × dmax ,
0.75×dmax 1×dmax , 1.25×dmax , 1.5×dmax , 1.75×dmax , 2×dmax , 2.25×dmax , 2.5×dmax ,
where dmax be the maximum value of length or width of each piece in their initial ori-
entation. Table 5.2 describes four different instances of bin configurations; SB, MB, LB
and Mix. In SB, we consider packing pieces with a small set of bin sizes. In comparison,
the MB and LB represent bin packing with a large set of bin sizes. In each configu-
ration, five different bin types are considered with a small bin size difference at each
configuration. Finally, the Mix option considers nine bin types representing remarkably
large bin sizes as well as remarkably small bin sizes in the input bin configuration.

We conducted our tests allowing restricted rotation (RR) and unrestricted orientation
(UR) of pieces. Since the problem is new, there are no benchmark results comparing the
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 115

solutions. However, we evaluated the proposed computational method through following

a set of tests. We set up all our experiments for 10 trials since the algorithms contain
random components and reported average output of 10 trials.

Operational details and parameters of algorithms:

Table 5.1 and 5.2 details the instances and bin configurations used for the experiments.

Table 5.1: Instances

Instance No.of Pieces Instance No.of Pieces
Shapes2 28 2×Jakobs2 50
3×Dighe1 48 Poly3a 45
3×Dighe2 30 Poly3b 45
Poly4a 60 Poly4b 60
3×Fu 36 Poly5a 75
2×Han 46 Poly5b 75
2×Jakobs1 50 Shapes 43

Table 5.2: Bin Configurations

Configuration No. of Bin types Input bin sizes Bin type ID
SB 5 0.5dmax , 0.75dmax , 1.0dmax , 1.25dmax ,1.5dmax 1,2,3,4,5
MB 5 1.0dmax , 1.25dmax , 1.5dmax , 1.75dmax ,2.0dmax 3,4,5,6,7
LB 5 1.5dmax , 1.75dmax , 2.0dmax , 2.25dmax ,2.5dmax 5,6,7,8,9
Mix 9 0.5dmax , 0.75dmax , 1.0dmax , 1.25dmax ,1.5dmax 1,2,3,4,5
1.75dmax , 2.0dmax , 2.25dmax , 2.5dmax 6,7,8,9

5.5.2 Performance of two constructive procedures (CA1 vs. CA2)

We compare the performance of CA1 and CA2 algorithms presented in Section 5.4.1
and 5.4.5.1 in Table 5.3. For each instance, an average U value was calculated for 10
trials. The values represent average results computed over the 14 instances. In each
trial, we ran each algorithm for 800 seconds under following criteria and consider the
best-recorded solution.

Rand-CA1 : Running CA1 with different random orders of bins and random orders
of pieces and pack pieces
Rand-CA2 : Running CA2 with different random orders of pieces and pack pieces

Table 5.3: Performance of construction algorithms

Rand-CA1 Rand-CA2
Bins Config.
RR UR RR UR
Avg.U of 10 trials 0.678 0.682 0.682 0.688
Mix
Avg. Time (seconds) for 100 solutions 34.66 51.23 176.52 245.54
Avg.U of 10 trials 0.665 0.675 0.668 0.677
LB
Avg. Time (seconds) for 100 solutions 37.67 56.44 141.36 188.40
Avg.U of 10 trials 0.663 0.671 0.667 0.674
MB
Avg. Time (seconds) for 100 solutions 32.78 47.15 98.41 162.49
Avg.U of 10 trials 0.650 0.657 0.656 0.665
SB
Avg. Time (seconds) for 100 solutions 29.52 41.53 74.68 120.04
116 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Considering the performance of both constructive procedures against the randomly gen-
erate permutations, Rand-CA2 performs slightly better in overall utilisation (U ) within
800 seconds for all the data sets. Also, the better solutions are produced when the piece
rotation is not restricted for both CA1 and CA2.

In comparison, CA1 and CA2 demonstrate a significant difference in computational effort

per solution. Rand-CA2 takes a longer execution time than Rand-CA1 to generate a
single solution. The bin selection process of CA2 is a time-consuming task as it checks
for the highest utilised bin type by placing pieces temporary, for each given bin type.
Since CA1 does not execute such bin section for all the given bin types and instead places
the pieces in a pre-assigned order of bins. This makes CA1 to test a higher number of
solutions than CA2 in a common runtime.

Out of the different configurations of bins, the Mix option recorded the highest utilisation
while the LB option recorded the second highest. Therefore, out of the options using a
limited number of bin types (LB, MB, SB), the LB being the best performer even though
it consumes higher computational time than others. Usually, with a set of larger bin
types, the algorithms allocate a higher number of pieces in one bin than choosing a set
of small bin types. This leads to an increase in the computational time for generating
some no-fit polygons according to the constructive method discussed in Chapter 4 as it
needs to consider the relatively large size of merged pieces for some insertion of pieces.

In the preceding sections, we investigate the performance of CA1 and CA2 procedures
when they are applicable with search methods.

5.5.2.1 Best performing GA Jostle approach:

We compared the performance of three different approaches of using GA to solve 2D

IMBSBPP. For each instance, an average U value was calculated for 10 trials over the
14 instances. In each trial, we ran the GA for 800 seconds.

GA: Running the GA approach discussed in without improving the offsprings

generated at each generation through Jostle
HGA v1 : Running the GA approach with Jostle improvement of
all the offsprings generated at each generation
HGA v2 : Running the GA approach with iterated Jostle improvement of
the best offspring generated at each generation

The parameter settings for our proposed GA approaches were determined through the
results of pilot tests conducted. We compared different values for P opSize, PCr , PM u ,
Kp , M axJS and elitist selection ratio (see in Table 5.4) and selected the combination
of parameter values that provided the best quality solutions in 800 seconds.
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 117

Table 5.4: Different parameter settings to tune GA, HGA v1 and HGA v2

P opSize : {8, 12, 16, 24, 32}

PCr : {0.7, 0.8, 0.9, 1.0}
PM u : {0.0125, 0.025, 0.05, 0.075, 0.1}
Kp : {2, 4, 6, 8, 10}
M axJS: {4, 8, 12, 16, 20, 24}
Pelitist ratio : {1/4, 1/8, 1/12, 1/16}
Ts : {2, 3, 4, 5}

Our pilot tests recommend following a set of values to run each GA approach discussed
in this study.

GA: P opSize = 24, M axJS = no Joslte improvement, PM u = 0.05

All parent solutions are crossover to generate offspring solutions
No. of solutions by elitist selection : 1/4 of popSize, Ts = 2
HGA v1 : P opSize = 12, M axJS = 8, PM u = 0.025
All parent solutions are crossover to generate offspring solutions
No. of solutions by elitist selection : 1/8 of popSize, Ts = 2
HGA v2 : P opSize = 16, Kp = 4 , M axJS = 12, PM u = 0.025
All parent solutions are crossover to generate offspring solutions
No. of solutions by elitist selection : 1/8 of popSize, Ts = 2

Table 5.5: Performance of GA and HGA approaches for 800 seconds

Restricted Rot. Avg. U Unrestricted Rot. Avg. U

Bins Config.
GA HGA v1 HGA v2 GA HGA v1 HGA v2
Mix. 0.689 0.689 0.693 0.695 0.694 0.699
LB 0.679 0.677 0.686 0.686 0.685 0.690
MB 0.678 0.675 0.681 0.684 0.679 0.687
SB 0.660 0.653 0.659 0.666 0.658 0.666

Out of the three GA-based approaches, the HGA v2 performs the best according to
the packing utilisation over the 14 instances. Table 5.5 demonstrates this for all four
bin configurations in both scenarios where piece rotation is restricted and unrestricted.
The GA approach without Jostle is also recorded the second best results by running
the algorithm for a higher number of generations within the given 800 seconds. Not
surprisingly, the computational cost of HGA v1 incurs the highest computation cost
and thereby operated with a less number of generations within the given 800 seconds.
This cause a lower utilisation than other two approaches. In comparison, HGA v2 solves
the problem within a reasonable computation time and achieves a better utilisation.

5.5.2.2 Performances of SPBCH approaches:

Two different search methods described in Section 5.4.5.2 were evaluated for the 14
instances. For each instance, we executed each algorithm for 10 trials where each trial
was run for 800 seconds. Using pilot studies, we found cmax = 15 as a good parameter
118 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

value to run SPBCH after conducting parameter tuning with different cmax = 5, 10, 15, 20
values.

LS CA2 : Local search with swap moves of pieces

ILS CA2 : Iterated local search with the swap and insert moves of pieces

Table 5.6: Performance of LS CA2 and ILS CA2 approaches for 800 seconds

Restricted Rot. Avg. U Unrestricted Rot. Avg. U

Bins Config.
LS CA2 ILS CA2 LS CA2 ILS CA2
Mix. 0.685 0.689 0.695 0.701
LB 0.673 0.681 0.675 0.694
MB 0.677 0.679 0.684 0.691
SB 0.660 0.663 0.663 0.671

Table 5.6 illustrates, there are 0.58%, 1.19%, 0.30% and 0.45% improvement in overall
utilization (on average) for Mix, LB, MB and SB bin configurations respectively by
ILS CA2 when the piece rotation is restricted. Also, it records 0.86%, 2.81%, 1.02% and
1.21% improvement in overall utilisation for Mix, LB, MB and SB bin configurations
respectively by ILS CA2 when the piece rotation is unrestricted. This indicates that
approach ILS CA2 approach generates better results and also noticed that there is no
significant difference in computation time between both approaches. Also, the Mix
option records the best utilisation in comparison to other bin configuration options (LB,
MB, SB).

5.5.2.3 Best performing IJRAB approach:

Three different versions of IJRAB were evaluated for the 14 instances in Table 5.1. For
each instance, we ran each approach for 10 trials due to the random components of the
algorithm. In each trial, we executed the algorithm for 800 seconds.

In pilot studies, we tested different values of Kp = {2, 4, 6, 8, 10}, Kb = {2, 3, 4, 5, 10}

to tune the algorithm to provide a reasonably better performance. The tuning results
demonstrated Kp = 4, Kb = 3 as good settings to run IJRAB and these settings were
used to run the algorithms.

The three versions of IJRAB tested in this study:

In comparison, the proposed IJRAB algorithm is faster than the other two computational
methods. This creates an opportunity to plug some other post-processing strategies
to improve its performance. We tested the applicability of two such post-processing
strategies with the proposed IJRAB algorithm. In Algorithm 6, each of these strategies
are applying before the bin-kick step executes at Line 42.

• Strategy 1 (S1): Try to repack pieces in the least utilised bin into the smallest
possible bin:
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 119

This is a well-known strategy suggested for rectangular bin packing problem (Friesen
and Langston, 1986; Ortmann et al., 2010). The strategy attempts to change the
bin configuration of the solution by repacking pieces inside the least utilised bin,
to the smallest possible bin type. If this move improves the solution and is better
than the leading solution, then the new arrangement of bins and pieces are ac-
cepted as the leading solution. Then s∗∗ ∗ ∗
D∗∗ , U ∗, and σ ∗ are updated according to
the new solution (see Algorithm 6).

• Strategy 2 (S2): Try to repack pieces in each of the used bins into the smallest
possible bin:
Strategy 1 is focused only on changing the least utilised bin in the order. In this
case, the same procedure is extended to repack the pieces in all the used bins
separately into the smallest possible bin. The strategy attempts to change the
bin configuration of this solution by repacking pieces inside each of the used bin
separately, to the smallest possible bin type. Each solution is compared against the
leading solution and the leading solution is updated if a better solution is found.

When any of these strategies are able to improve the leading solution, then the immediate
next step of making a bin kick will not be executed since the bin sequence is changed
by the post-processing strategy.

Considering these two strategies, we compare the results of three versions of IJRAB (see
Tables 5.7 and 5.8).

• IJRAB-A1 : Implementation of IJRAB Algorithm as it is.

• IJRAB-A2 : Implementation of IJRAB Algorithm by applying S1 for the local

optimum found in that particular bin configuration.

• IJRAB-A3 : Implement Algorithm 1 applying S2 for the local optimum found in

that particular bin configuration.

Out of three IJRAB versions, both IJRAB-A2 and IJRAB-A3 reach better packing
utilisation than IJRAB-A1 when comparing the average U values over the 14 instances
(The best average values are denoted in bold font). The IJRAB-A2 and IJRAB-A3 al-
gorithms achieve almost similar packing utilization for MB, LB, and Mix options, even
though the version IJRAB-A3 is computationally expensive than IJRAB-A2. Also,
when running IJRAB-A3, out of all the used bins, only the lowest utilised bin was fre-
quently replaced by the possible smallest bin. The improvement gap between IJRAB-A2
and IJRAB-A3 for MB, LB, and Mix reflect this by showing a minor difference. How-
ever, there is an exception for smaller bin size configuration (SB) since it demonstrates a
significant improvement. In summary, through this experiment, we identify IJRAB-A2
and IJRAB-A3 as the most promising versions of IJRAB.
120 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Table 5.7: Performance of the IJRAB algorithms for different bin configurations
(with Restricted Rotation of pieces)

Bin Config. IJRAB-A1 IJRAB-A2 IJRAB-A3

Mix Avg. U 0.692 0.700 0.703
Improvement % 1.16% 1.59%
Time (seconds for 100 Jostle cycles) 66.5 69.2 78.2
LB Avg. U 0.667 0.687 0.689
Improvement % 3.00% 3.30%
Time (seconds for 100 Jostle cycles) 71.6 74.6 76.8
MB Avg. U 0.671 0.684 0.686
Improvement % 1.94% 2.24%
Time (seconds for 100 Jostle cycles) 62.7 65.8 68.0
SB Avg. U 0.656 0.662 0.677
Improvement % 0.91% 3.20%
Time (seconds for 100 Jostle cycles) 50.5 53.1 58.6
Table 5.8: Performance of the IJRAB algorithms for different bin configurations
(with Unrestricted Rotation of pieces)

Bin Config. IJRAB-A1 IJRAB-A2 IJRAB-A3

Mix Avg. U 0.699 0.706 0.708
Improvement % 1.00% 1.29%
Time (seconds for 100 Jostle cycles) 99.7 104.6 107.5
LB Avg. U 0.680 0.695 0.696
Improvement % 2.21% 2.35%
Time (seconds for 100 Jostle cycles) 111.5 115.0 121.1
MB Avg. U 0.684 0.692 0.693
Improvement % 1.17% 1.32%
Time (seconds for 100 Jostle cycles) 98.2 101.3 103.0
SB Avg. U 0.668 0.675 0.685
Improvement % 1.05% 2.54%
Time (seconds for 100 Jostle cycles) 83.6 88.0 92.8

5.5.3 Comparison among three computational methods

Table 5.9 summarises the average U values of each computational method separately. In
order to compare the results, the best algorithm of each method was run for 800 seconds
in one trial. The values represent average results obtained over the 14 instances for 10
trials.
Table 5.9: Performance comparison of IJRAB, HGA and SPBCH methods
Restricted Rot. Avg. U Unrestricted Rot. Avg. U
Bins Config. Random Random IJRAB HGA SPBCH Random Random IJRAB HGA SPBCH
CA1 CA2 A3 v2 ILS CA2 CA1 CA2 A3 v2 ILS CA2
Mix. 0.678 0.682 0.703 0.693 0.689 0.682 0.688 0.708 0.699 0.701
Impr. % - - 3.69% 2.21% 1.03% - - 3.81% 2.49% 1.89%
LB 0.665 0.668 0.689 0.686 0.681 0.675 0.677 0.696 0.690 0.694
Impr. % - - 3.61% 3.16% 1.95% - - 3.11% 2.22% 2.51%
MB 0.663 0.667 0.686 0.681 0.679 0.671 0.674 0.693 0.687 0.691
Impr. % - - 3.47% 2.71% 1.80% - - 3.28% 2.38% 2.52%
SB 0.650 0.656 0.677 0.659 0.663 0.657 0.665 0.685 0.666 0.671
Impr. % - - 4.15% 1.38% 1.07% 4.26% 1.37% 0.90%
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 121

In general, there is a significant improvement with all three search methods. The table
also reveals best performance results by IJRAB ( A3) than HGA v2 and ILS CA2, in
all rotation variants and input bin set configurations.

Out of different input bin configurations, a mixed use of bins with a higher number
of bin types (Mix) generates significantly better solutions. A greater variety of input
bin configuration creates more opportunities in finding highly utilised bins by replacing
poorly utilised bins using smaller bin types.

The three input configurations; SB, MB and LB, with an equal number of input bin
types, demonstrate a pattern. The large bin set (LB) achieves higher utilisation whereas
small bin set (SB) gets lower utilisation. We observe that the irregular pieces placed in
large bin types tend to move, rotate and compact more when there is the higher number
of pieces inside the bin. We also noticed that the accommodation of smaller pieces in
large bin types leads to a higher utilisation of those bins.

Examining the results, IJRAB A3 and HGA v2 perform better than the SPBCH method.
This is most likely due to the time advantage of CA1 and the effective moves established
during the search process. The results also demonstrate that the piece moves from
the Jostle structure have significantly contributed to solution improvement in IJRAB
and HGAIJ. Within a reasonable computational effort (≤ 800 seconds), the IJRAB A3
demonstrates significantly better results than HGA v2. Since HGA v2 works with a pop-
ulation of solutions to ensure its bin moves, the overall method is more time-consuming
than IJRAB A3. However, in the long run, both approaches demonstrate less signifi-
cant differences in their results. In comparison, there is only a slight improvement in
the solution quality by IJRAB A3 vs. IJRAB A2.

Since our goal is to produce a computational method to generate a good solution within
a short period of time, the study involves the most efficient method; the IJRAB A3 for
the next set of experiments. For easy reference, here onward, it is simply referred as
IJRAB

5.5.4 Performance comparison: 2D IMBSBPP vs. 2D ISBSBPP

We presented Iterated Jostle algorithm in Chapter 4 to solve 2D ISBSBPP with free

rotation of pieces, which has better results in a shorter time than the other approaches
Lopez-Camacho et al. (2013a) and Martinez-Sykora et al. (2017). We use this algorithm
to solve ISBSBPP problem for each bin type separately and compared the results. Table
5.10 and 5.11 present the average U values of 10 trials of algorithm execution. We ran
an algorithm for 800 seconds at each trial.

Table 5.10 and 5.11 demonstrate the advantage of solving 2D IMBSBPP as the results
of 2D IMBSBPP outperform the results achievable from 2D ISBSBPP.
122 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

We also noticed that, with single bin sizes, most frequently there is one bin in the solution
recording poor utilization than others, especially when large bins (e.g. 2.25dmax , 2.5dmax )
are involved. This results reduction in overall utilisation as well. However using different
bin sizes maintaining a certain variability in the sizes cause improvement in utilization
by replacing such poorly utilized bins in the solution with a smaller bin type. Also, this
finding opens another path to investigate. When there is such poorly utilized bin space,
or a sheet space in sheet cutting, taking such residual spaces as a separate stock sheet
in the next orders would allow to improve the overall utilization. If a company decides
to implement such a post production policy to reuse usable residuals, still, the proposed
algorithm for 2D IMBSBPP can be applicable. We dicuss more details of this matter in
Chapter 6.

5.5.5 Best solutions

Table 5.12 and Table 5.13 present the best U values we obtained across 10 trials of
experiments by executing the three main algorithms; IJRAB A3, HGA v2 and ILS CA2,
with 9 input bin types (Mix) and 5 input bins types (LB) respectively. The tables present
the number of bins used in the solution, percentage of each bin type used and the
computation method which derived those solutions. In addition, the tables demonstrate
the average utilisation of bins for each bin type used. As an example, in Table 5.12,
the instance Fu in RR has used six bins altogether where one bin from bin types 2,3,6,9
and two bins from bin type 8 are used. The use of each of those bin types are presented
as percentages. The values within brackets below the percentages indicate the average
utilisation of each bin type used in the solution.

By examining the table, we noticed that the best results are achieved by IJRAB A3 ap-
proach in most cases except for one case in Table 5.12 and one case in Table 5.13. In some
cases, both IJRAB A3 and HGA v2 generated the same solution. Also, it demonstrate
that IJRAB A3 reaches good solutions quicker than the other two approaches.
 Table 5.10: Performance comparison: 2D IMBSBPP vs.2D ISBSBPP (with restrictions of piece rotation)

Single bin sizes Multiple bin sizes (IJRAB A3)

Instance
(1.0dmax ) (1.25dmax ) (1.5dmax ) (1.75dmax ) (2.0dmax ) (2.25dmax ) (2.5dmax ) SB MB LB MIX
Shapes2 0.589 0.518 0.576 0.605 0.648 0.640 0.518 0.662 0.671 0.690 0.698
3×Dighe1 0.689 0.630 0.612 0.562 0.574 0.680 0.551 0.689 0.705 0.701 0.713
3×Dighe2 0.680 0.653 0.680 0.666 0.510 0.605 0.490 0.727 0.712 0.737 0.730
3×Fu 0.592 0.589 0.670 0.677 0.691 0.655 0.663 0.772 0.762 0.745 0.788
2×Han 0.621 0.681 0.662 0.608 0.621 0.490 0.397 0.703 0.728 0.730 0.729
2×Jakobs1 0.681 0.713 0.778 0.667 0.613 0.605 0.653 0.799 0.800 0.805 0.811
2×Jakobs2 0.660 0.676 0.586 0.574 0.566 0.521 0.563 0.693 0.699 0.690 0.709
Poly3a 0.607 0.582 0.647 0.594 0.607 0.479 0.582 0.616 0.645 0.660 0.660
Poly3b 0.602 0.578 0.642 0.589 0.602 0.475 0.578 0.625 0.654 0.666 0.674
Poly4a 0.571 0.565 0.616 0.634 0.607 0.639 0.518 0.671 0.656 0.619 0.695
Poly4b 0.610 0.586 0.581 0.598 0.572 0.603 0.488 0.690 0.672 0.639 0.690
Poly5a 0.578 0.597 0.674 0.660 0.607 0.599 0.647 0.602 0.647 0.665 0.661
Poly5b 0.604 0.632 0.690 0.591 0.679 0.536 0.579 0.628 0.665 0.678 0.674
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Shapes 0.543 0.521 0.603 0.532 0.509 0.536 0.434 0.609 0.593 0.614 0.608
Avg. 0.616 0.609 0.644 0.611 0.600 0.576 0.547 0.678 0.686 0.689 0.703
123
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Table 5.11: Performance comparison: 2D IMBSBPP vs.2D ISBSBPP (No restriction of piece rotation)
Single bin sizes Multiple bin sizes (IJRAB A3)
Instance
(1.0dmax ) (1.25dmax ) (1.5dmax ) (1.75dmax ) (2.0dmax ) (2.25dmax ) (2.5dmax ) SB MB LB MIX
Shapes2 0.617 0.518 0.576 0.605 0.648 0.640 0.691 0.656 0.640 0.692 0.699
3×Dighe1 0.689 0.630 0.612 0.562 0.574 0.680 0.551 0.701 0.705 0.703 0.726
3×Dighe2 0.680 0.653 0.680 0.666 0.510 0.605 0.490 0.714 0.718 0.734 0.734
3×Fu 0.663 0.589 0.737 0.773 0.691 0.655 0.663 0.755 0.760 0.772 0.778
2×Han 0.677 0.681 0.662 0.608 0.621 0.490 0.596 0.705 0.734 0.731 0.740
2×Jakobs1 0.681 0.784 0.778 0.800 0.766 0.807 0.653 0.788 0.801 0.805 0.811
2×Jakobs2 0.660 0.676 0.670 0.689 0.660 0.695 0.563 0.679 0.689 0.694 0.706
Poly3a 0.607 0.582 0.647 0.594 0.607 0.479 0.582 0.630 0.665 0.663 0.674
Poly3b 0.602 0.660 0.642 0.589 0.602 0.475 0.578 0.641 0.663 0.681 0.691
Poly4a 0.607 0.621 0.616 0.634 0.607 0.639 0.518 0.686 0.678 0.644 0.689
Poly4b 0.610 0.651 0.581 0.598 0.572 0.603 0.488 0.692 0.684 0.641 0.696
Poly5a 0.607 0.647 0.674 0.660 0.607 0.599 0.647 0.636 0.670 0.684 0.676
Poly5b 0.639 0.632 0.690 0.591 0.679 0.536 0.579 0.690 0.677 0.688 0.683
Shapes 0.543 0.579 0.603 0.532 0.509 0.536 0.434 0.618 0.615 0.607 0.604
Avg. 0.634 0.636 0.655 0.636 0.618 0.603 0.574 0.685 0.693 0.696 0.708
124
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 125

Table 5.12: Best solutions received in 10 trials by running algorithm for 800
seconds -(with Mix set of bins)
Number Percentage of each type of bin is used
Instance n U Method
of bins 1 2 3 4 5 6 7 8 9
20% 60% 20% IJRAB A3,
Shapes2 28 0.718 5
[0.725] [0.723] [0.705] HGA v2
33.3% 33.3% 33.3% IJRAB A3,
3×Dighe1 48 0.725 3
[0.718] [0.650] [0.753] HGA v2
66.7% 33.3% IJRAB A3,
3×Dighe2 30 0.742 3
[0.734] [0.744] HGA v2
16.7% 16.7% 16.6% 33.3% 16.7%
3×Fu 36 0.789 6 IJRAB A3
[0.782] [0.765] [0.808] [0.812] [0.747]
33.3% 33.3% 33.3% IJRAB A3,
2×Han 46 0.740 3
[0.759] [0.763] [0.727] HGA v2
20% 20% 40% 20%
2×Jakobs1 50 0.838 5 IJRAB A3
[0.780] [0.962] [0.785] [0.904]
RR
33.3% 33.3% 16.7% 16.7%
2×Jakobs2 50 0.719 6 IJRAB A3
[0.762] [0.665] [0.730] [0.748]
25% 25% 25% 25% IJRAB A3,
Poly3a 45 0.669 4
[0.650] [0.648] [0.647] [0.690] HGA v2
33.3% 33.3% 33.3% IJRAB A3,
Poly3b 45 0.688 3
[0.580] [0.684] [0.694] HGA v2
25% 25% 50% IJRAB A3,
Poly5a 75 0.672 4
[0.540] [0.717] [0.674] HGA v2
25% 25% 25% 25% IJRAB A3,
Poly5b 75 0.684 4
[0.638] [0.700] [0.658] [0.701] HGA v2
33.3% 33.3% 33.3%
Shapes 43 0.612 3 IJRAB A3
[0.486] [0.715] [0.607]
20% 60% 20%
Shapes2 28 0.718 5 IJRAB A3
[0.717] [0.724] [0.705]
33.3% 33.3% 33.3%
3×Dighe1 48 0.720 3 HGA v2
[0.694] [0.742] [0.709]
50% 50% IJRAB A3,
3×Dighe2 30 0.753 2
[0.708] [0.781] HGA v2
42.9% 35.7% 14.3% 7.1% IJRAB A3,
3×Fu 36 0.847 14
[0.892] [0.861] [0.868] [0.779] HGA v2, LS-CA2
33.3% 33.3% 33.3% IJRAB A3,
2×Han 46 0.740 3
[0.759] [0.763] [0.727] HGA v2
33.3% 33.3% 33.3%
2×Jakobs1 50 0.852 3 IJRAB A3
[0.853] [0.852] [0.850]
UR
40% 20% 40%
2×Jakobs2 50 0.722 5 IJRAB A3
[1.000] [0.730] [0.704]
50% 50% IJRAB A3,
Poly3a 45 0.710 2
[0.573] [0.798] HGA v2
50% 50% IJRAB A3,
Poly3b 45 0.704 2
[0.688] [0.715] HGA v2
50% 25% 25% IJRAB A3,
Poly5a 75 0.696 4
[0.696] [0.670] [0.725] HGA v2
33.3% 33.3% 33.3% IJRAB A3,
Poly5b 75 0.709 3
[0.661] [0.717] [0.735] HGA v2
50% 25% 25%
Shapes 43 0.612 4 IJRAB A3
[0.602] [0.632] [0.602]
126 Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins

Table 5.13: Best solutions received in 10 trials by running algorithm for 800
seconds -(with LB set of bins)
Number Percentage of each type of bin is used
Instance n U Method
of bins 5 6 7 8 9
50% 50%
Shapes2 28 0.715 4 IJRAB A3
[0.703] [0.724]
50% 50% IJRAB A3,
3×Dighe1 48 0.740 2
[0.706] [0.756] HGA v2
50% 50% IJRAB A3,
3×Dighe2 30 0.753 2
[0.708] [0.781] HGA v2
25% 50% 25%
3×Fu 36 0.814 4 IJRAB A3
[0.774] [0.833] [0.808]
33.3% 66.7%
2×Han 46 0.727 3 IJRAB A3
[0.723] [0.727]
33.3% 66.7%
2×Jakobs1 50 0.831 3 IJRAB A3
[0.853] [0.824]
RR
33.3% 66.7%
2×Jakobs2 50 0.716 3 IJRAB A3
[0.604] [0.736]
33.3% 66.7% IJRAB A3,
Poly3a 45 0.677 3
[0.591] [0.739] HGA v2
33.3% 66.7%
Poly3b 45 0.672 3 HGA v2
[0.629] [0.702]
25% 25% 50% IJRAB A3,
Poly5a 75 0.681 4
[0.665] [0.611] [0.700] HGA v2
25% 50% 25% IJRAB A3,
Poly5b 75 0.674 4
[0.674] [0.669] [0.680] HGA v2
33.3% 33.3% 33.3%
Shapes 43 0.612 3 IJRAB A3
[0.486] [0.715] [0.607]
25% 25% 50%
Shapes2 28 0.696 4 IJRAB A3
[0.680] [0.690] [0.695]
50% 50% IJRAB A3,
3×Dighe1 48 0.671 2
[0.695] [0.657] HGA v2
50% 50% IJRAB A3,
3×Dighe2 30 0.753 2
[0.708] [0.781] HGA v2
20% 40% 40%
3×Fu 36 0.794 5 IJRAB A3
[0.698] [0.742] [0.837]
100% IJRAB A3,
2×Han 46 0.736 2
[0.736] HGA v2
33.3% 33.3% 33.3%
2×Jakobs1 50 0.852 3 IJRAB A3
[0.853] [0.852] [0.850]
UR
33.3% 66.7%
2×Jakobs2 50 0.678 3 IJRAB A3
[0.626] [0.693]
50% 50% IJRAB A3,
Poly3a 45 0.710 2
[0.573] [0.798] HGA v2
50% 50% IJRAB A3,
Poly3b 45 0.704 2
[0.688] [0.715] HGA v2
33.3% 66.7%
Poly5a 75 0.691 3 IJRAB A3
[0.642] [0.711]
33.3% 33.3% 33.3% IJRAB A3,
Poly5b 75 0.709 3
[0.661] [0.717] [0.735] HGA v2
33.3% 33.3% 33.3%
Shapes 43 0.612 3 IJRAB A3
[0.486] [0.715] [0.607]
Chapter 5 Irregular shape Bin Packing in Heterogeneous Bins 127

5.6 Concluding Remarks

In this chapter, we investigate on developing a new computational method to find good

solutions for 2D IMBSBPP in less time. The objective is to maximise the overall utili-
sation of bins. This problem is practical in several sheet cutting industries which work
efficiently with different standard sheet sizes. When considering both constructive al-
gorithms proposed, based on the computational complexity and solution quality, CA1
demonstrate the most efficient choice. Out of three main improvement methods pro-
posed, the IJRAB A3 performs better with the suggested improvement strategies applied
at frequent intervals while Jostle continues. The proposed IJRAB A3 is a low-cost solu-
tion algorithm and does not require any advanced optimisation packages to implement
it in real world problems.

The other promising approach HGAIJ also reaches equal performance to IJRAB at the
long run. Main reason for its less performance in short computational times is the design
structure of the algorithm. Since it implements as a population based metaheuristic, it
consumes a significant time to move construct population of solutions. In comparison,
the approach also take significant time to move from one search area to the other in the
search space. Compared to IJRAB, the time gap between two improvements of is less at
the early stages of a HGAIJ run. This causes considering HGAIJ as a lower performing
algorithm during short time periods.

In order to highlight the use of IJRAB algorithm, we compare the use of single sheet
(i.e. bin ) size vs. a mix of sheet sizes when cutting shapes from a given set of bin types.
Based on the results, we summarise the following key findings.

1) Use of mix of bin sizes is worth in terms of material saving than using single bin size
approaches.

2) Selection of a set of large bin sizes configuration is more advantageous than the
selection of a set of smaller bin size configuration in terms of saving material when there
is a choice of selecting from a mix of bin sizes.

The algorithm can be further used to calculate the setups based on the number of bins.
Likewise, implementation of this computation method is robust not only to evaluate the
trim loss but also to implement different objective functions such as minimising material
costs, setup costs or a combination of both, and generate solutions accordingly. This will
be our next study in this thesis as we use this efficient computational method to investi-
gate those problems in Chapter 6. We suggest extending this computational method to
solve the irregular shape bin packing problem with usable leftovers where heterogeneous
rectangular bins are possible in the form of residuals, which can be reusable in future
cut orders.
Chapter 6

Use and Reuse of Materials in

Cutting Industries
Evaluation of operational policies for
effective use of resources

6.1 Introduction

In chapters 4 and 5, we discussed solution methods applicable for 2D irregular shape bin
packing problems. The solution of the 2D irregular shape bin packing problem generates
a cutting plan that contains a set of bins (i.e. stock sheets) packed with irregular pieces.
In the case of reusing residuals, it allows the use of leftovers which are reusable as input
stock sheets for the next orders. A residual of a packed bin is considered as a usable
leftover if the area of it is larger than a certain threshold area and both length and width
of the residual area are longer than a certain threshold length. Otherwise, the residual
bin space is considered to be trim-loss which becomes a waste. In a regular production,
the Usable Leftovers (ULs) are returned to the warehouse to use them in the future
cut orders. In the C&P literature, this problem is called as 2D Irregular shape Bin
Packing Problem with ULs (2D IBPPUL) so that the cutting production is run with
reusing residuals. As reviewed in Chapter 2, Wäscher et al. (2007)’s typology denotes
this problem as 2D Irregular Residual Bin Packing Problem (2D IRBPP), in which the
objective is to minimise the material waste.

Instead of throwing away the residual stock sheet, reusing them as an input sheet for
future production orders may allow reducing the overall consumption of sheet material
throughout the sheet cutting production. In view of operations management, reuse
is a well-known production policy in manufacturing businesses. The reuse policy is
implemented in different ways within a supply chain, either as product returns; also
known as external returns (e.g. Minner (2001), French and LaForge (2006)), or internal
returns which is formed a by-product or a production scrap such as material waste

129
130 Chapter 6 Use and Reuse of Materials in Cutting Industries

generated by under-utilized processes; which are being able to recover fully or partially
(e.g. Fleischmann et al. (1997)). However, the reuse practices have also been stressed
due to the extensive effort required in handling and controlling inventory (Fleischmann
et al., 1997; Minner, 2003). For instance, reuse of materials creates variability in material
return, which causes issues in utilising standard sizes of storage shelves and trucks.
This variability generates extra effort in handling and storage, hence resulting extra
costs. From a management perspective, the importance of having the best procurement,
production and inventory policies in place is emphasised due to the costs attached to
these additional factors. To our knowledge, the existing approaches of cutting and
packing problem with 2D irregular shapes have not been able to investigate different
scenarios of the problem considering varying costs related to material, setups and off-
cuts handling. Also, as reviewed in Chapter 2, most research studies that have analysed
usable leftovers highlight the significance of conducting an economic analysis to compare
the cost of handling leftovers, as future research opportunities (Cherri et al., 2014a).
As an example, the authors of those studies believed in keeping usable leftovers in the
warehouse are economical for 1DCSP when the material cost is dominant, as the savings
in the material can compensate the costs of handling, transportation and warehouse
(Trkman and Gradisar, 2007; Cherri et al., 2014a). Similarly, the trade-off between
material saving and auxiliary costs need to be analysed with respect to the other cost
scenarios dominated by setups and dominated by labour, where cutting and packing of
2D irregular shapes is progressed.

6.1.1 Contribution

This research discusses in this chapter is an exploratory study examining the reuse
practices of residuals in sheet cutting industries. There is a lack of evidence in existing
research to justify the most suitable operational strategy for a given operational scenario,
let alone for a set of scenarios. This chapter provides guidance to make more accurate
operational decisions on procurement and production stages within the scenarios where
material prices, setup costs and handling costs vary with one another.

In achieving this, we evaluate the suitability of four operational policies in nine different
operational scenarios. With regards to saving resources, we present the potential rela-
tionships between costs and policies at different operational scenarios. In this Chapter,
we investigated two models and designed set of experiments to evaluate the effect of
different operational policies. The first model was developed with the objective function
of minimising the material waste. This is the most popular objective of existing re-
search attempts when dealing with cutting problems especially when ULs are involved.
We considered different standard bin sizes individually (Single) and the mix option of
standard bin types (Mix). Section 6.4 provides a brief description of this model. Next,
we extended our investigation and discussed the second model with the objective of
Chapter 6 Use and Reuse of Materials in Cutting Industries 131

minimising the total cost. The second model allows us to compare the performances
recorded by each policy, involving the other cost components such as setups, inventory
and handling costs of ULs when deriving solutions. We denote this as Model 02:cost
saving in Section 6.5.

In order to run multiple orders, we introduce a computation method to solve the 2D

IBPPUL problem. In this case, we present a heuristic search mechanism to solve the
problem, by modifying the IJRAB algorithm designed in Chapter 5. The problem re-
quires us to consider two procurement options of using standard stock sheets purchased.
The first option is to use a single standard bin size where only a selected type of sheet
size is purchased from the sheet supplier/s. The second option is to use multiple stan-
dard bin sizes separately, so that cut layouts can have different sizes of standard stock
sheets. In solving 2D IBPPUL, the propose two models examine the overall performance
of fulfilling the cut demands arising in successive periods. These models also consider
the fact of using usable leftovers and managing them efficiently throughout the peri-
ods without accumulating them in the warehouse. In order to satisfy this requirement,
special attention is made in reviewing different inventory and production scenarios in
relation to the residual reuse.

6.2 Overview of Using and Reusing of Material in Sheet

Cutting Process

The sheet cutting process occurs in many industries such as sheet metal, paper, tools,
and foams. During this process, a known demand of pieces is cut using multiple numbers
of sheets within a certain time period. In most cases, the process generates residuals
where some of them can be reused. If an organisation has the policy to reuse these
residuals than discarding them, it will reduce the material waste. Additionally, if the
sheet material is expensive, this will give an additional saving in cost in terms of material.
These residuals then can be used as an input material in the proceeding cut orders, in
addition to the standard sheet material of those orders. This can be continued until
all the orders of the same sheet material are completed. Thus the context of reusing
residuals expands the scope of a traditional single-order cutting problem to the multiple
consecutive periods cutting problem where residuals from one period may be used in the
following period or the periods afterwards.

The general sheet cutting problem can be defined as cutting a set of two-dimensional
pieces from a set of rectangular sheets. The cut pieces either can be regular or irregular
in shape and the rectangular sheets will have multiple sheets from one standard size or
from a mix of various sheet sizes. A cut order consists of a number of pieces assigned
to be cut within a certain time period and the time period of each order is fixed. In
other words, each order will have different cut quantities to be cut within a fixed time
132 Chapter 6 Use and Reuse of Materials in Cutting Industries

period. The cut demand for the next period is unknown while the demand for the
current period is being processed and the smallest piece dimensions are assumed to be
known in advance. The usable residuals generated from the current order can be used
in the next order/s as input material. The usability of the residuals is determined by its
size (i.e. length and width) compared to the smallest available cut piece of all orders. If
a residual dimensions are larger than the smallest of all the orders, then these residuals
are returned to the inventory, in order to be used in proceeding orders. Such residuals
in general, are referred as ”Usable Leftovers” (ULs) in cutting and packing literature.
The residuals smaller than the smallest piece are counted as trim-loss.

Compared to the other studies of irregular packing problems, irregular bin packing
with usable leftovers in a new area of research. Considering this in the multi-period
scenario is more challenging than solving irregular bin packing problem. In this thesis, we
address this problem at a basic level as an initial study. Therefore, we believe assuming
future demands are unknown as a considerable assumption which can provide substantial
improvement for minimising waste as well as substantial contribution to research in this
area. In solving this new problem, we were inspired by work on one-dimensional cutting
stock problem with usable leftovers with the multi-period scenario, which are also at
very early stage (e.g. Trkman and Gradisar (2007), Cui et al. (2016)). Those papers
have also used this assumption and justify that the assumption is reasonable to use.

Research interest in multiple order cutting problems has only gained the attention of
a few scholars. The existing methods of material reuse in relation to one-dimensional
cutting stock problems are addressed in Gradišar et al. (1999a); Trkman and Gradisar
(2007); Cherri et al. (2009); Cui and Yang (2010); Cherri et al. (2013) whereas two-
dimensional problems with regular pieces are addressed by Venkateswarlu (2001), An-
drade et al. (2014) and Andrade et al. (2016). Trkman and Gradisar (2007) discuss two
models for one-dimensional (1D) cutting stock problem where the first model is focused
on minimising the trim-loss and second on minimising the cost of material and the cost of
returning usable residuals. However, this research has remained focused on operational
aspects such as minimising cost exclusively under a selected operational policy. The
discussion of this chapter expands the scope of evaluating a set of policies and details of
these are presented in section 6.2.1.

Our study begins by identifying operational policies and different scenarios where the
sheet cutting process is executed. Policies represent the options which the organisations
follow when running the sheet cutting process, whereas scenarios represent the practical
situations that organisations face in terms of material, machinery and labour.
Chapter 6 Use and Reuse of Materials in Cutting Industries 133

6.2.1 Identifying operational policies:

When planning the cutting process, a firm has the option of either purchasing stock
sheets from one standard sheet size or as a mix of several standard sheet sizes. The
lessons from 1D cases mainly favour mixed sizes with the aim of improving material
utilisation. However, our study investigates how the purchasing decision affects the
total cost, therefore both the criteria are considered. They are 1) considering a single
standard sheet size as the purchasing option 2) considering the mix of standard sheet
sizes as the purchasing option. The next factor considered in determining the policy is
the decision the use or non-use of residuals. Combining both factors, we address below
four policies during our study:

• Policy 1: Mix bin sizes procurement policy with residual reuse (UL-Mix)

• Policy 2: Mix bin sizes procurement policy with no residual reuse (nonUL-Mix)

• Policy 3: A single bin size procurement policy with no residual reuse (nonUL-
Single)

• Policy 4: A single bin size procurement policy with residual reuse (UL-Single)

Based on the operational policy used in the production, the packing problem that will
be solved in this study is either a 2D irregular single bin size bin packing problem
(2D ISBSBPP) or 2D irregular multiple bin size bin packing problem (2D IMBSBPP)
(Wäscher et al., 2007). As an example, Policy 01 involves addressing the multiple bin
size problem while Policy 03 addresses the single bin size problem. However, even the
Policy 4 uses single bin size, the cutting and packing problem in consideration is a
multiple bin size problem because of the use of residuals, which will be in different sizes.
A detailed review of these problems can be found in Wäscher et al. (2007). In order
to solve these problems, we employ the heuristic search tools developed in this study,
which are discussed in Chapter 4 and 5.

6.3 Model for Processing Successive Cut Orders

The problem considers a sequence of equal time periods where each period is assigned
with a cut quantity. The model assumes that the cut demand for the next period is
unknown, while the demand for the current period is being processed. Therefore, the
optimisation of each order is done separately. The usable leftovers generated from the
current order are used for the future orders. The cut shapes are heterogeneous and the
smallest piece dimensions (wmin , lmin ) are assumed to be known in advance. For the
demand quantity of each period, a cutting and packing problem is solved to find a good
solution which minimises the trim loss or the total cost. As the input material, either
134 Chapter 6 Use and Reuse of Materials in Cutting Industries

standard stock sheets or existing usable leftovers or both are used to fulfill the demand
at each period. In the output of the current order, if the residual material is larger
than the smallest piece (evaluated as mentioned in Section 6.3.1), then it is considered
as candidate Usable Leftovers which can be returned to the warehouse for future use.
However, confirming them as ULs returned to the warehouse is subject to the criterion
of managing the ULs in the warehouse. Details of this criterion are described in Section
6.4 and 6.5. The residuals do not return to the warehouse as ULs are considered as trim
loss or waste.

6.3.1 Finding candidate usable leftovers

Let lmin be the minimum piece length out of all the pieces, and wmin be the minimum
piece width out of all those pieces. lmin and wmin are measured when pieces are aligned
at their initial rotation of a given instance. Let l(R) and w(R) be the length and the width
of the residual sheet (bin), after making a horizontal and vertical cut. Notations Lf and
Wf denote the length and the width of a standard bin of type f . Similarly, Lg and Wg
denote the length and the width of an OldUL bin of type g.

Figure 6.1 illustrates four possible ways of generating usable leftovers from residual space
of a packed bin.

We define following three categories of bin types for 2D IBPPUL, adopting the Cui et al.
(2016)’s terminology for leftovers defined in solving the 1D Cutting Stock problem with
Usable Leftovers (1DCSPUL).

• Standard bin types: The stock sheets in standard sizes. During the planned time
horizon of completing all order quantities, it is assumed that standard bins are
available in unlimited quantity, from each type, in the inventory.

• Old UL types: The UL stock sheet types returned to the warehouse by the cut
orders processed in earlier periods.

• New UL types: The UL return to the warehouse by the cut order processes in the
current period.

Within each category, the bins with same dimensions are considered as bins belong to
the same type. At the end of each cut order, the New UL bins are returned to the
inventory and become Old UL for the following orders.

The typical constraints of the problem applied to both models discussed in Section 6.4
and Section 6.5 can be outlined as follows. In each period, the quantity to be cut is
equivalent to the demand quantity of that period. Also, all the layouts must satisfy the
basic overlap and containment constraints of general cutting and packing problem.
Chapter 6 Use and Reuse of Materials in Cutting Industries 135

Figure 6.1: Defining usable leftovers

In order to define each model, we use following notations.

6.3.2 Notation

T : Number of time periods considered where t denotes a certain time period so that
t = 1, ..., T.

Therefore the symbols corresponding to the order processed at time period t are denoted
with subscript t:

Inputs when start processing an order quantity n at period t.

P : Pieces where P = {p1 , p2 , ...pn }

Bf,t : Set of Standard bins of bin type f

0
St : Number of Standard bin types

Nf,t : Number of Standard bins of type f in the warehouse at time period t

Bg,t : Set of OldUL bins of type g

136 Chapter 6 Use and Reuse of Materials in Cutting Industries

0
Qt : Number of OldUL bin types in the warehouse at the start of time period t

Ng,t : Number of OldUL bins of type g in the warehouse at time period t

(Ow)
Nt : Total number of Old-UL bins in the warehouse at the beginning of time period
t

Outputs after processing an order quantity n at period t.

Bh,t : Set of New UL bins of type h

0
Rt : Number of NewUL bin types return to the warehouse at period t.

Nh,t : Number of NewUL bins of type h return to the warehouse at time period t
(O)
Qt : Number of Old-UL bin types used in time period t
(S)
Nt : Total number of standard bins used in the order processed at time period t.
(O)
Nt : Total number of Old-UL bins used at period t
(R)
Nt : Total number of New-ULs bins return to the warehouse at time period t

Notations used to define the objective functions to solve the problem at

period t.
(P cs)
Areat : Total area of demanding pieces in the order processed at time period t
(S)
Areat : Total area of standard bins used in the order processed at time period t

(S) PSt0 PNf,t

Areat = f =1 j=1 (cf,j,t .Wf .Lf )

(O)
Areat : Total area of Old-UL bins used in the order processed at time period t

(O) POt0 PNg,t

Areat = g=1 j=1 (dg,j,t .Wg .Lg )

(R)
Areat : Total area of New-ULs bins return to the warehouse at the order processed
at time period t
(R) PRt0 PNh,t
Areat = h=1 j=1 (Wh .Lh )

T rimlosst : Trim loss caused in the order processed at time period t

Jmax : The maximum number of Old-UL types

where; bf,j,t and bg,j,t denote the j th bin of standard bin type f and OldUL bin of type
g respectively at time period t, and

1, if bin b
f,j,t is occupied
cf,j,t =
0, otherwise
Chapter 6 Use and Reuse of Materials in Cutting Industries 137


1, if bin bg,j,t is occupied
dg,j,t =
0, otherwise

6.3.3 Problem to be solved at each order:

Since the demand for the future periods is unknown, we do not solve an optimisation
problem over the planned time horizon. Instead, the scope of the optimisation problem
is limited only for the cut demands to be fulfilled within a period. Therefore, we solve a
cutting and packing problem individually for each period, to find a good solution within
each period for a given quantity of polygonal shapes (i.e. a heterogeneous set of pieces).
However, in both models, the ULs connects each periods by considering the up-to-date
ULs stock at each period. At any period, it is possible to use Old UL bin type in the
warehouse.
0
Let S be a given number of rectangular standard bin types and Bk be a set of bins of
0
standard bin type k where k = 1, ..., S . j denotes the index of a bin of type k so that a
bin of type k : bk,j ∈ Bk . Width and length of a bin of type k are denoted by Wk and Lk .
0
Similarly, let O be a number of rectangular shape leftover bin types made by previous
0
orders. We denote Bl be a set of bins of type l where l = 1, ..., O . j denotes the index
of a bin of type l so that a bin of type l : bl,j ∈ Bl . Width and length of a bin of type
l are denoted Wl and Ll . The newly generated number of rectangular shape leftover
0 0
bin types is denoted with R and Bm is a set of bins of type m where m = 1, ..., R . j
denotes the index of a bin of type m so that a bin of type m : bm,j ∈ Bm . Width and
length of a bin of type m are denoted Wm and Lm . The problem considers a scenario
where the number of bins from each type k is extremely large, number of bins for each
type l is finite. The input pieces consist of a set of n polygons P = p1 , ..., pn equivalent
to the demand quantity of the current order.

In each order, an input minimization problem is solved with the objective of minimising
trim-loss if Model 01 is used or minimising total cost if Model 02 is used.

A feasible solution can be denoted as a set of occupied standard bins, a set of occupied
OldULs, a set of newly generated ULs and the packed pieces. All the pieces in the order
are packed exactly to the quantity demanded in the order. Each packed piece pi , has an
orientation angle oi ∈ [0, 3600 ) and a position in its allocated bin relative to the piece’s
origin of (xi , yi ). We considered the reference point of each piece as the bottom-left
corner point of the enclosing rectangle for a given angle of rotation and defined the
origin as the bottom left corner of the first bin.
138 Chapter 6 Use and Reuse of Materials in Cutting Industries

6.3.4 Modifications proposed for IJRAB to solve 2D IRBPP

At each period, our approach requires solving an irregular shape bin packing problem
with a heterogeneous set of bins in most cases and a set of homogeneous bins otherwise.
The input bins can be either standard bins or UL bins which represent the standard
sheets or usable leftovers.

When the problem is irregular shape packing with heterogeneous bins, we check the
possibility of using IJRAB algorithm discussed in Chapter 5. The original IJRAB solves
2D IMBSBPP when the number of bins from each bin type is not limited. However,
with ULs, the number of input bins from each Old-UL is finite. Therefore, we proposed
the following modifications to the IJRAB algorithm.

Modifying IJRAB to work with a finite number of bins from each bin type:

• The original method initially arranges an arbitrary order of bins to generate a

packing layout from left to right along the strip. However, the modified algorithm
decides the bin order considering the number of bins available in each type. As an
example, if a certain bin type, for instance a bin type 7, has only one bin (an UL
bin type), then only one bin is allocated from type 7 in to the order. The number
of input bins from each Standard bin type is assumed to be a large number since
we assume no stock-outs of standard bins for each type during the planned time
horizon.

• The improvement mechanism suggested in the original work involves two types
of kicks during the search process. Out of those two, the bin kick is proposed to
alter the bin order, in which case, a used bin is replaced with any other randomly
selected bin type. In the original algorithm, any bin type other than the selected
used bin type can be used as the replacement since the number of bins for each
type is not limited. However, in the modified version, the used bin is replaced by
another bin type if the new bin type has enough number of bins to use. In fact,
this is applied to the UL bin types as their availability is finite at each time period.
Therefore, the overall computational method maintains a record of the number of
bins available for each type, at each time period.

• The original study suggests a set of post-processing strategies which replace the
under-utilised bins of the leading solutions by another bin type focusing an im-
provement in overall utilisation. Referring to the results in Chapter 5, both
IJRAB A2 and IJRAB A3 versions performed better with no significant difference
in overall utilisations than the other approaches. We decided to involve IJRAB A2
to solve the problems in this study since IJRAB A2 is more computationally effi-
cient than the other.
Chapter 6 Use and Reuse of Materials in Cutting Industries 139

6.4 Model 01: Minimizing the Trim Loss

The trim-loss is the objective measurement to evaluate the quality of a solution and it
is evaluated by the equation 6.1.

(S) (O) (P cs) (R)

T rimlosst = Areat + Areat − Areat − Areat (6.1)

The objective is to reduce the trim-loss of the individual orders. The ties in the minimum
trim-loss are broken by the solution with the maximum sum of areas of the New ULs
and the minimum number of different New UL types.

When considering usable leftovers, handling a large number of heterogeneous UL bins

in the warehouse causes additional effort for a company. This imposes managing some
practical constraints when handling leftovers. Heterogeneous ULs cause extra effort in
managing the inventory level of those leftovers since they can be accumulated in the
warehouse when time progresses.
0
In this model, we use the constraint: Qt ≤ Jmax ∀t. Once New ULs are generated, out
of which, the ULs return to the warehouse is determined according to the upper bound
Jmax . At the end of a time period, if the number of New UL types is ≤ (Jmax − Number
of OldUL types at the end of the period), then all those generated New UL types return
to the warehouse. Otherwise, the number of New ULs is determined based on the sizes
of ULs where largest ULs cuts are returned to the warehouse. In this case, we select the
largest UL types out of the newly created bin types and the Old UL bin types without
violating the constraint Jmax . Accordingly we set the model to progress with the largest
leftovers in the warehouse.

6.4.1 Computational experiments

The model presented in Section 6.4 was tested with a series of experiments to analyse
different situations arise in sheet cutting production with irregular-shaped items. All
experiments were run by an Intel 2.60 GHz processor and 4GB RAM. In the following
sub-sections, we introduce parameter values and sets of benchmark instances used to
evaluate the results generated by different production policies.

In terms of computational experiments, each instance represents a certain demand of

pieces in one period and there are 30 periods in one trial considering different demand
values. Also, for each instance with a demand, in order to find a solution, we ran
300 jostle cycles to minimise the trim loss. Since algorithm contains randomness, the
packing algorithm was executed for 10 repetitions to get an average value for each period.
Once a trial is finished after 30 periods, we summed up the result to get the total trim
140 Chapter 6 Use and Reuse of Materials in Cutting Industries

loss. Since the demand values of each period are uncertain we conducted 30 trials of
30 periods executions and finalised the average values. This effort was tested for two
operational policies; using leftovers and non-using leftovers, for different levels of Jmax ,
hence a significant number of experiments were conducted.

Instances: We used the nesting instances published on ESICUP (EURO Special Interest
Group on Cutting and Packing) website for our experiments as follows. Fu 60 pieces:
(5 x 12 pcs), Shapes2: 56 pieces (2x28 pcs), Jakobs1: 50 pieces (2x25 pcs), Poly: 60
(poly1a 15x4) pieces and Shapes: 86 pieces (43x2).

5 Bin sizes: 1.25×m, 1.5×m, 1.75×m, 2×m, 2.25×m, where m be the maximum value
of length or width of pieces of the instance when each piece in their initial orientation.

6.4.1.1 Simulation framework

In order to find the effect on each policy, we designed an empirical simulation framework
which connects inputs and outputs of each period within the planned time horizon. The
concept of using this framework was adopted by the approaches proposed by Cherri
et al. (2013) and Cui et al. (2016), to process successive orders when implementing the
residual reuse policy. Cui et al. (2016) solved each order separately since the demand
information of the next orders is not known when the current order is being processed.
In our simulation framework, we follow the same methodology to solve two-dimensional
irregular bin packing problems. Accordingly, at each period, a specific type of irregular
shape bin packing problem is solved based on the operational policy as follows.

nonUL-Single :- 2D ISBSBPP

nonUL-Mix :- 2D IMBSBPP

UL-Single :- 2D ISBSBPP at the initial period and then 2D IMBSBPP when ULs are
involved

UL-Mix :- 2D IMBSBPP

A certain experiment based on one particular policy was tested in 30 trials. In a given
trial, the test was based on a processing cut demands in successive periods to consider
the effects in short and long time horizons respectively. Random demand quantities were
processed at each time period.

We analysed the effect of using both standard and UL bin types vs. non-use of leftovers
when the production is solely focusing on minimising the material used to fulfil the
demand quantities. Each of these scenarios was evaluated based on two procurement
policies; i) using a single type of standard bins (Single S ) or ii) using a mix of standard
bin types (Mix S ). The ” S” denotes the Standard bins.
Chapter 6 Use and Reuse of Materials in Cutting Industries 141

One practical issue of using leftover bin types is the inconvenience of handling different
types of bins. Due to this reason, we evaluate how different upper bound levels of a
number of UL bin types operate with UL-Mix, UL-Single policies.

6.4.1.2 Multi-period run of orders: Period wise investigation

In this test, we consider the performance of nonUL-Mix, UL-Mix, nonUL-Single and

UL-Single policies at each time period. The upper bound Jmax limits the number of
residual types returned to the warehouse.

In the analysis, we were interested in the evaluation of the following criteria.

• The variability of total utilisation of material at the end of each period. We use the
t t
P (P cs) P (S)
measurement; cumulative usage of standard material: Ut = Areat0 / Areat0 .
t0 =1 t0 =1
This represents the utilisation of standard material occupied for the sequence of
periods. In other words, this explains the amount of material that has been utilised
against the amount that the company has paid for at the end of each time period.

• The effects of procurement decisions in the context where the sole objective is to
save material as much as possible.

The experiment settings:

We designed a set of experiments to execute the effect of each policy using the instances
mentioned in Section 6.4.1. We selected 16 consecutive periods as the number of time
periods. In each period, a random demand quantity within the range of [10-50] was taken
as the number of pieces to be cut, so that the shapes to be cut are chosen randomly
according to the required demand quantity from the instances.

Since we tested the effect of Jmax values, for the same data instances and the order
quantities, we run the tests for different levels of Jmax = {0 (with no use of ULs), 4, 16,
32}.

At each period, a cut order problem was solved using the modified IJRAB algorithm
which terminated after 300 Jostle cycles.

The summarized results are demonstrated in Figure 6.2 and details are given in Tables
A.1, A.2, A.3, A.4, A.5, and A.6. In each table, the first data column shows the periods
from 1 to 16, second to fifth columns show the Ut values at the end of each t period
when Jmax = 32, 16, 4 and 0 (i.e. there is no allowance to use ULs).
142 Chapter 6 Use and Reuse of Materials in Cutting Industries

Figure 6.2: Variation of Ut with periods

6.4.1.3 Assessing the results

As illustrated in Figure 6.2, in period wise, the overall utilization till period t (Ut )
has neither significant improvement nor drop for nonUL policies (see the Yellow lines
correspond for Jmax 0 in each graph of Figure 6.2). The Mix S option maintains the
highest values for utilisation of material than the other Single S (2.25m to 1.25m) options
when no ULs are allowed (Jmax 0); throughout the periods.

However, when UL policies are implemented, the Ut increases when time progresses,
especially for the cases starting with Single S with large bin options (see the Blue, Red,
Grey lines correspond for Jmax 32, Jmax 16, Jmax 4 in 2.25m and 2.0m graphs of the
Figure 6.2). Higher the area of the starting single bin type, steeper the improvement
of Ut gets. When starting with Mix standard bin sizes, comparatively only a little
improvement has been achieved with UL policy as the periods progresses. As a special
note, interestingly, after a certain period (in this case after the 7th period); the option
which starts with large bin sizes reaches higher Ut values than the Ut values generated
by Mix bin sizes; when the UL policy is adopted. Also, starting with small bin sizes does
not demonstrate promising improvement in Ut as starting with large bin size options or
mix bin size options.

The graphs in Figure 6.2 demonstrate reaching to a steady state after a certain period.
When analysing this at different Jmax values, it demonstrates that the higher the value
of Jmax , the period they reach to the steady states becomes longer. Therefore, when
Chapter 6 Use and Reuse of Materials in Cutting Industries 143

deriving conclusions, it is necessary to check whether the true effects of ULs are taken
into account properly after it reaches to the steady state. As an example, a short time
horizon would not favour UL-Single or UL-Mix policies since the true advantage of using
UL for saving material can only be demonstrated after reaching the steady state and
this requires a considerable number of periods. When there is a substantial time horizon
to reach the steady state, comparatively then the UL-Single policy option starting with
the largest single standard bin type demonstrated more advantage in utilising input
material by using residuals more effectively than other policies.

6.4.1.4 Multi-period run of orders: Investigation for whole time horizon

Next, we describe the overall results (i.e. The overall performance of running multi-
period cutting during the planned time horizon) for the Model 01 by processing de-
mands for consecutive periods. We conducted the tests for two planned time horizons
separately; first for 10 time-periods and then for 30 time-periods.

Area(S) ), the total

P
Table 6.1 and 6.2 demonstrate total usage of purchasing bins (
space of remaining ULs at the end of the time horizon (RemArea(U L) ) and total trim-
P
loss from each period ( T rimloss). In this study, the scope of our analysis was based on
a certain time horizon and we did not consider the continuity of operations and material
for the next periods in the next time horizon. Therefore the UL material remains at the
end of the time horizon, was assumed as a waste of material in our analysis. This implies
T rimloss + RemArea(U L) in the considered time horizon.
P
that the total waste is

In the tables, the results correspond to the nonUL-Mix and nonUL-Single policies are
denoted when jmax = 0. For UL-Mix and UL-Single policies, the tables show the
P (R)
total area of ULs return to the warehouse (in Areat ) and used the area of those
P (O)
returned ULs ( Areat ) for different jmax = 2, 4, 8, 16, 32 values. The column headers
N (S) , N (O) and N (R) denote the total number of bins of standard bins used, Old UL
bins used and New UL bins returned to the warehouse respectively. The overall material
utilisation is calculated for each instance as;

U tilization = ( Area(S) − ( trimloss + RemArea(U L) ))/ Area(S) .

P P P

The experiment settings:

We designed a set of experiments to execute the effect of each policy using the instances
mentioned in Section 6.4.1 and followed the simulation framework discussed in Section
6.4.1.1. As the number of time periods, we ran the tests for two sets of consecutive
periods; 10 and 30.

In a certain trial of the multi-period run, a random demand quantity within the range
of [10-50] was taken as the number of pieces to be cut at each period, so that the shapes
144 Chapter 6 Use and Reuse of Materials in Cutting Industries

to be cut are chosen randomly according to the required demand quantity from the
instances. The outcomes provided by the different Jmax values are tested for the same
data instances and the order quantities. In this set of tests we considered Jmax = {0
(with no use of ULs), 2, 4, 8, 16, 32}. Similar to the previous tests, at each period, an
irregular bin problem was solved using the modified IJRAB algorithm which terminates
after 300 Jostle cycles.

The summarized results are shown in Tables 6.1 and 6.2.

Table 6.1: Material Saving Scheme: Use of leftovers with different Jmax values
for 10 periods
X
Area(S) Area(O) Area(R) Rem.Area(U L) N (S) N (O) N (R)
P P P
Area(R) / Area(S) Area(O) / Area(R)
P P P P
Purchase T rimloss U tilization%
Jmax
Decision Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg %. Avg %.
Mix S 21943.1 0.0 0.0 7122.1 0.0 59.5 0.0 0.0 66.86% 0.00% 0.00%
sgl S2.25m 23853.6 0.0 0.0 9032.6 0.0 39.7 0.0 0.0 61.28% 0.00% 0.00%
sgl S2.00m 23214.6 0.0 0.0 8393.6 0.0 49.5 0.0 0.0 62.66% 0.00% 0.00%
0
sgl S1.75m 22974.8 0.0 0.0 8153.9 0.0 64.1 0.0 0.0 63.32% 0.00% 0.00%
sgl S1.50m 23360.8 0.0 0.0 8539.8 0.0 91.6 0.0 0.0 61.84% 0.00% 0.00%
sgl S1.25m 25602.7 0.0 0.0 10781.8 0.0 146.7 0.0 0.0 58.25% 0.00% 0.00%
Mix S 22012.4 243.9 555.5 6879.9 311.5 61.4 2.3 6.8 66.65% 2.83% 42.00%
sgl S2.25m 21974.3 946.9 1324.7 6775.6 377.8 37.6 4.4 7.5 66.58% 6.81% 68.04%
sgl S2.00m 22159.3 944.7 1407.6 6875.4 462.9 47.2 4.8 7.5 66.28% 6.82% 60.65%
2
sgl S1.75m 21726.3 898.5 1209.1 6594.8 310.5 61.5 5.4 8.2 66.75% 5.99% 69.69%
sgl S1.50m 22733.1 487.5 1067.5 7332.2 579.9 89.7 3.6 11.1 64.09% 4.97% 45.79%
sgl S1.25m 25291.0 426.8 758.5 10138.3 331.7 145.4 4.2 11.2 59.37% 3.30% 49.65%
Mix S 21897.9 408.8 757.8 6728.0 348.9 60.9 4.1 9.6 66.98% 3.81% 51.95%
sgl S2.25m 21836.3 1376.5 1719.2 6672.6 342.7 37.3 7.3 11.6 67.03% 8.57% 74.21%
sgl S2.00m 22036.0 1348.6 1874.1 6689.5 525.5 47.0 6.9 10.5 66.53% 9.52% 66.44%
4
sgl S1.75m 21603.6 1236.9 1563.2 6456.3 326.3 61.5 8.1 12.6 67.03% 7.87% 72.41%
sgl S1.50m 22669.3 651.2 1380.2 7119.3 729.0 89.6 5.6 17.4 64.34% 6.52% 49.74%
sgl S1.25m 25611.1 527.6 1259.2 10058.5 731.6 146.6 5.4 19.7 59.04% 4.94% 44.74%
Mix S 21921.0 474.4 918.0 6656.4 443.6 61.6 4.8 12.5 67.00% 4.95% 51.05%
sgl S2.25m 21812.2 1522.3 1924.8 6588.6 402.6 37.3 8.2 13.9 67.36% 9.76% 76.47%
sgl S2.00m 22012.9 1450.1 2039.8 6602.2 589.8 46.7 8.6 13.2 66.64% 9.87% 66.67%
8
sgl S1.75m 21607.2 1266.9 1637.2 6415.9 370.3 61.8 8.8 14.6 67.03% 8.57% 70.96%
sgl S1.50m 22700.1 779.1 1669.4 6988.8 890.3 89.9 6.6 22.8 64.32% 8.05% 51.28%
sgl S1.25m 25831.8 511.4 1582.8 9939.4 1071.4 147.2 5.9 30.8 58.88% 6.65% 44.18%
Mix S 21847.8 588.9 1023.3 6592.4 434.4 61.0 5.7 14.3 67.11% 5.54% 54.99%
sgl S2.25m 21794.2 1613.7 1963.4 6623.5 349.7 37.1 8.9 15.1 67.33% 9.99% 78.77%
sgl S2.00m 21907.4 1489.5 2072.4 6503.5 582.9 46.8 9.1 13.8 67.09% 9.94% 64.03%
16
sgl S1.75m 21661.5 1314.4 1760.5 6394.5 446.0 62.0 9.5 17.3 66.95% 9.71% 69.27%
sgl S1.50m 22758.9 849.1 1877.6 6909.4 1028.5 90.2 7.4 25.3 64.15% 8.52% 52.21%
sgl S1.25m 26019.8 602.2 2227.1 9573.8 1625.0 147.9 6.9 50.4 58.70% 9.40% 44.30%
Mix S 21862.1 603.6 1112.0 6532.6 508.5 61.3 5.9 15.5 67.13% 6.18% 54.06%
sgl S2.25m 21773.7 1632.1 1993.1 6591.7 361.0 37.2 8.3 14.5 67.52% 10.01% 79.05%
sgl S2.00m 21972.1 1479.5 2073.0 6557.6 593.5 46.6 8.9 13.3 67.06% 10.22% 68.81%
32
sgl S1.75m 21713.6 1228.7 1717.9 6403.4 489.2 62.1 8.7 17.1 66.63% 9.29% 65.93%
sgl S1.50m 22816.7 855.8 2056.5 6795.0 1200.7 90.7 7.2 30.6 63.96% 9.43% 51.40%
sgl S1.25m 26162.7 768.0 2912.9 9196.9 2144.9 148.5 8.3 66.0 58.65% 11.33% 46.37%

6.4.1.5 Assessing the results for the planned time horizon

Implementation of non-UL policies

For both time horizons as illustrated in Figure 6.3, the nonUL-Mix recorded the highest
utilisation of material than any of nonUL-Single when ULs are not allowed (Jmax = 0).
In this case, the implementation of nonUL-Single has done in five ways correspond to
the five different standard bin sizes considered. However, in terms of the number of bins
used, the nonUL-Mix has used a considerably higher number of bins than the nonUL-
Single policy with the two largest bin-sizes (see the number of Standard bins used when
Jmax = 0 in Tables 6.1 and 6.2).

Since a higher number of occupied bins leads to a higher number of setups with an order,
the advantage gained from material saving in Mix bin policy may have been traded-off
Chapter 6 Use and Reuse of Materials in Cutting Industries 145

Table 6.2: Material Saving Scheme: Use of leftovers with different Jmax values
for 30 periods
X
Area(S) Area(O) Area(R) Rem.Area(U L) N (S) N (O) N (R)
P P P
Area(R) / Area(S) Area(O) / Area(R)
P P P P
Purchase T rimloss U tilization%
Jmax
Decision Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg %. Avg %.
Mix S 65374.3 0.0 0.0 21283.4 0.0 177.1 0.0 0.0 66.62% 0.00% 0.00%
sgl S2.25m 71188.8 0.0 0.0 27097.9 0.0 118.2 0.0 0.0 61.23% 0.00% 0.00%
sgl S2.00m 69011.8 0.0 0.0 24920.9 0.0 147.0 0.0 0.0 62.68% 0.00% 0.00%
0
sgl S1.75m 68099.5 0.0 0.0 24008.6 0.0 190.3 0.0 0.0 63.44% 0.00% 0.00%
sgl S1.50m 69430.7 0.0 0.0 25339.8 0.0 272.2 0.0 0.0 61.94% 0.00% 0.00%
sgl S1.25m 76169.6 0.0 0.0 32078.7 0.0 436.1 0.0 0.0 58.23% 0.00% 0.00%
Mix S 65344.4 826.1 1566.9 20512.6 740.8 181.6 7.5 18.7 66.66% 2.67% 47.23%
sgl S2.25m 65208.9 3178.2 4008.2 20288.0 830.0 112.0 13.6 22.2 66.53% 6.92% 70.72%
sgl S2.00m 65990.2 3232.1 4636.7 20494.6 1404.6 140.7 14.9 22.8 65.92% 7.57% 58.69%
2
sgl S1.75m 64675.2 2554.2 3505.9 19632.6 951.6 183.5 15.9 24.6 66.61% 5.91% 67.52%
sgl S1.50m 67731.5 1319.9 3234.2 21726.3 1914.3 267.5 10.1 33.9 63.89% 4.99% 41.09%
sgl S1.25m 75103.0 1104.7 2266.7 29850.1 1162.0 432.1 10.9 34.1 59.40% 3.24% 47.22%
Mix S 65141.9 1302.0 1994.3 20358.6 692.3 182.0 13.0 25.5 66.94% 3.52% 59.10%
sgl S2.25m 64629.9 4492.2 5160.2 19870.9 668.0 111.2 21.5 31.8 67.32% 8.62% 79.85%
sgl S2.00m 64846.1 4670.7 5453.3 19972.6 782.6 139.1 24.1 32.2 67.30% 8.95% 77.04%
4
sgl S1.75m 64172.4 3611.7 4374.9 19318.3 763.2 183.2 24.3 35.4 66.95% 7.43% 73.88%
sgl S1.50m 67586.6 1864.1 4133.3 21226.4 2269.2 267.0 16.1 50.6 64.12% 6.60% 46.23%
sgl S1.25m 75917.5 1654.1 3560.6 29920.1 1906.5 434.9 17.2 58.7 59.11% 4.97% 48.93%
Mix S 64800.5 1780.0 2487.6 20002.0 707.6 182.1 17.3 33.3 67.27% 4.43% 66.32%
sgl S2.25m 64282.6 4859.9 5435.6 19616.1 575.7 110.8 26.9 39.3 67.79% 9.23% 81.88%
sgl S2.00m 64334.4 5268.0 5851.7 19659.9 583.6 138.5 28.0 37.0 67.77% 9.48% 78.53%
8
sgl S1.75m 64000.6 3951.8 4652.0 19209.5 700.3 183.2 28.1 42.3 67.15% 8.31% 75.61%
sgl S1.50m 67338.4 2317.1 4634.3 20930.3 2317.2 267.2 21.4 67.0 64.33% 7.71% 53.97%
sgl S1.25m 76397.5 1744.4 4183.0 29867.9 2438.6 436.2 19.1 86.5 59.07% 6.35% 50.37%
Mix S 64659.0 1897.2 2718.6 19746.7 821.4 183.3 19.6 40.0 67.51% 5.09% 68.51%
sgl S2.25m 64119.1 4914.4 5416.9 19525.7 502.5 109.9 29.0 41.8 68.22% 9.28% 86.15%
sgl S2.00m 64344.2 5242.8 5974.7 19521.4 731.9 137.7 30.4 39.6 67.96% 9.74% 82.85%
16
sgl S1.75m 63929.2 4026.0 4704.8 19159.5 678.8 183.6 28.7 45.8 67.30% 8.47% 77.31%
sgl S1.50m 67487.2 2454.1 5056.1 20794.2 2602.0 267.6 23.1 70.5 64.47% 7.91% 57.87%
sgl S1.25m 76952.9 1900.6 5974.6 28788.0 4074.0 437.7 22.0 143.0 59.11% 8.86% 51.35%
Mix S 64581.6 2056.2 2865.5 19681.4 809.3 182.0 21.2 43.3 67.52% 5.46% 69.71%
sgl S2.25m 64082.2 4898.7 5354.2 19535.8 455.5 109.0 31.2 43.3 68.27% 9.21% 88.23%
sgl S2.00m 64225.5 4576.5 5429.5 19281.7 852.9 137.3 30.1 38.3 67.98% 9.24% 78.17%
32
sgl S1.75m 63892.4 3937.7 4681.6 19057.5 743.9 183.6 29.3 48.7 67.40% 8.61% 77.26%
sgl S1.50m 67510.0 2593.1 5655.9 20356.2 3062.8 267.6 23.4 82.3 64.47% 8.64% 58.50%
sgl S1.25m 77438.8 2374.4 7982.6 27739.6 5608.2 439.1 27.2 185.7 59.16% 10.69% 54.23%

by the additional setups required. Hence analysing the nonUL-Mix policy and nonUL-
Single policy with large bin sizes would be beneficial in terms of material and setup cost.
On the other hand, nonUL-Single with small bin sizes neither provide better material
utilisation nor use less number of bins, hence smaller bin sizes demonstrate no advantage
in terms of saving material or reducing the number of setups.

Implementation of UL policies

Compared to the nonUL policies, the implementation of UL policies has resulted in an

improvement in material utilisation for both time horizons as illustrated in Figure 6.3.
The results reveal a steeper improvement by each of UL-Single policies than the UL-Mix
policy. Increasing Jmax has improved the overall material utilisation strongly at lower
values of it. This is clearly demonstrated by the rapid improvement from Jmax = 0 to
8. However, increasing Jmax at higher values do not demonstrate significant advantages
(see Jmax = 8 to 32) since there is a sufficient level of heterogeneous bin sizes to reach
a good solution with higher utilisation.

According to Figure 6.3, it is also noticeable that the substantial improvement of material
utilisation in the UL-Single policy with the largest bin size surpasses the utilisation
recorded by Mix policy at the higher values of Jmax . The Mix policies record the best
utilisation only when Jmax (< 4). Therefore, it is reasonable to say that using a UL-
Single policy with large bin sizes leads to the least usage of material and the minimum
trim loss.
146 Chapter 6 Use and Reuse of Materials in Cutting Industries

Figure 6.3: Model 01: Use of leftovers with different Jmax values

With respect to the number of bins occupied, the Tables 6.1 and 6.2 reveal that usable
P (S) P (O)
leftovers result in an increase in the total number of bins used ( Nt + Nt ) than
in non U L. In addition to the increase of setups, this causes an extra effort related
to managing the leftovers. Comparatively, for every Jmax , both UL Single policy with
large bin sizes (sgl S2.25m and sgl S2.00m) and UL Mix policies provide the minimum
number of ULs return to the warehouse. Similar to nonUL, the UL Single policy with
the largest bin size occupied the minimum number of bins.

Increasing Jmax allows the reuse of different types of bins at each period. Therefore,
we further investigate how the increase in leftovers affect material saving. The ratio
P (R) P (S)
Areat / Areat is used in measuring the degree of ULs return to the warehouse
P (O) P (R)
and the ratio Areat / Areat is used in measuring the degree of use of those
returned ULs. The results of these measurements are tabulated in the last two columns
of Table 6.1 and 6.2.

Not surprisingly, the higher the Jmax , the higher the ratios for each policy in the results.
P (O) P (R)
The UL-Single policy with the largest bin size recorded the highest Areat / Areat
ratio at higher Jmax values. Also, expanding the time horizon encourages waste reduc-
tion (i.e. Improved utilisation for 30 periods time horizon than 10 periods). The graphs
(O) P (R)
Areat / Areat , N (R) , N (O) of the Mix,
P
in Figure 6.4 illustrate how the ratio
largest (2.25m) and smallest (1.25m) standard bin options vary at each Jmax value. The
set of graphs in column one represent the results related to 10 periods whereas the other
column represents the same for 30 periods. The results demonstrate a higher ratio of
P (O) P (R)
Areat / Areat for 30 periods and this provide evidence for increasing oppor-
tunities of reusing residuals at longer time horizons. In the case of UL-Single largest
(2.25m), it has demonstrated the highest number in utilising OldUL bins in the ware-
house. Also, the results demonstrate the UL-Single smallest (1.25m) option records the
Chapter 6 Use and Reuse of Materials in Cutting Industries 147

Figure 6.4: Model 01: Use of leftovers with different Jmax values

highest number of ULs return to the warehouse than the UL-Mix and any other UL-
Single bin size options. However, the results demonstrate a lower performance by the
UL-Single smallest (1.25m) option in terms of utilising those ULs.

6.4.2 Findings and discussion

The overall results of Model 01 summaries, that both nonUL-Mix and UL-Single with
large standard bin size demonstrate significant improvements in terms of material utili-
sation than other options. While the Mix policy provides the best material saving when
there is no use of ULs, the single standard bin policy with large bin sizes provides the
148 Chapter 6 Use and Reuse of Materials in Cutting Industries

best utilisation with ULs. Therefore, when the objective is to save material, the mix
sizes policy is most favourable if the production policy is set to discourage using ULs.
Otherwise, if the operational policy is based on using ULs and there is a substantial
allowance for managing enough number of UL types, then proceeding with the largest
standard bin size with ULs is most favourable in order to achieve high utilisation of
material. In this case, it is important to set an adequate upper bound on the number of
UL bin types and proceeds the multi-period run for the substantial amount of periods
so that using ULs can be reached to a steady state level which effectively utilises the
returned ULs in the subsequent periods.

While analysing the number of bins used in each case, when the production is not
supported by the use of ULs, the single large standard bin size (nonUL-Single Large)
records the minimum number of used bins than any other option even in the context of
saving material as much as possible. Similarly, when the production is supported by the
use of ULs, the single large standard bin size (UL-Single Large) records the minimum
number of used bins than other options.

Analysis related to the number of used bins is important in terms of cost of setups which
has a significant effect on the economic production as the cost material is not always
being the dominant factor in the total cost of sheet cutting production. Therefore, we
propose our next model to further analyse the cost trade-off of material and setups arise
when implementing those most influential policies; nonUL-Mix, UL-Mix, nonUL-Single
with standard large bins and UL-Single with large bins. We also consider the labour
commitment to managing ULs since it adds a significant contribution to the total cost
in an expensive labour scenario.

6.5 Model 02: Minimizing the Cost

The findings in Section 6.4, motivate us to extend the analysis of using operational
policies considering minimising the total cost if manufacturers focus is on the economic
benefit than just saving material exclusively. In this study, we further investigate how
nonUL-Mix, UL-Mix, nonUL-Single and UL-Single policies perform to minimise the
total cost of sheet cutting.

In our investigation, we tried to identify several practical scenarios of the sheet cutting
process. The cost trade-offs have a significant impact on the economic value on deciding
which procurement or production policy to implement. Yet, this has been given less
attention in the literature and we conduct an empirical analysis of the suitability of
each operational policy using Model 2. As per practical situations, we identify that there
are three factors which have the biggest impact on sheet cutting; material, machinery
and labour. These factors contribute to four key cost factors; unit cost of material (c),
setup cost per sheet (s), holding cost per off-cut per period (h) and transportation and
Chapter 6 Use and Reuse of Materials in Cutting Industries 149

handling cost per off-cut (r). The first cost component is related to material, the second
is related to machinery and the last two are related to labour. According to the cost
structure which determines the total cost of sheet cutting, a certain situation of sheet
cutting process can be dominated by either material (i.e. material dominant), setups
(i.e. setup dominant), labours (i.e. labour dominant) or balanced. Based on these
situations, we outline nine operational scenarios considering the combinations of each
cost factor which are represented in Table 6.3 and Figure 6.5. Please note that in Figure
6.5 the scenarios are numbered in an order which supports to identify a pattern in the
performance characteristics of each operational policy.

Table 6.3: Characteristics of scenarios

Scenario Cost structure

Scenario 1 c >>> s At high level of h and r
Scenario 2 c <<< s At high level of h and r
Scenario 3 A balanced level of c & s At high level of h and r
Scenario 4 c >>> s At medium level of h and r
Scenario 5 c <<< s At medium level of h and r
Scenario 6 A balanced level of c & s At medium level of h and r
Scenario 7 c >>> s At low level of h and r
Scenario 8 c <<< s At low level of h and r
Scenario 9 A balanced level of c & s At low level of h and r

Each of these scenarios is tested for the four policies in order to determine the best
policy for each scenario.

Cost function:

The total cost function includes four cost parameters: i) Material cost of occupied bins,
ii) Setup cost proportional to the number of bins occupied with the solution, iii) Holding
cost of Old-UL bins and iv) Warehouse return cost of New-ULs.

(S) (O) (R) (S) (O) (Ow) (R)

Costt = c(Areat + Areat − Areat ) + s(Nt + Nt ) + hNt + rNt (6.2)

c : material cost per unit area

s : set up cost per bin

h : holding cost per Old-UL bin

r : inventory return cost per New-UL bin

Instead of setting an upper bound as for Model 01, the generation of UL types is con-
trolled through the return and holding costs assigned to the NewULs in the objective
function. As an example, a significant contribution of return cost can influence to re-
duce New UL bins and therefore discourage ULs to store in the warehouse. At the
150 Chapter 6 Use and Reuse of Materials in Cutting Industries

Figure 6.5: Cost Scenarios

P (S)
comparison stage, we take c × Areatfor all t values to find the total material cost
P (S) (O) P (Ow)
incurred for the planned time horizon. In addition, s(Nt + Nt ), hNt and
P (R)
rNt denote total setups cost, total holding cost and total warehouse return cost
respectively. Finally, the total cost within the planned time horizon is calculated by
P (S)
Costt + Cost of Old ULs remaining at the end of time horizon.

6.5.1 Computational experiments

The experiment settings:

Similar to Model 01, we designed a set of experiments to execute the effect of each policy
using the instances mentioned in Section 6.4.1. We selected two time horizons; 10 and 30
consecutive periods where at each period, a random demand quantity within the range of
[10-50] was taken as the number of pieces to be cut. Since we investigated 4 operational
policies in 9 different scenarios considering multi-period runs, the computation effort on
this is significant.
Chapter 6 Use and Reuse of Materials in Cutting Industries 151

Next, we set cost parameter values to analyse the variability of each paradigm not only
for the usual practical cases but also for the extreme cases, by aiming at analysing
how performances are varied from one extreme scenario to other, for each policy. As
the cost parameters, we use different values for h = {1, 10, 100, 1000, 10000} and r =
{2, 20, 200, 2000, 20000} where r = 2h and α = {0.05, 0.1, 0.5, 1, 1.5, 5, 10} where c = αs.
The selected parameter values supported to conduct an empirical analysis by changing
the relative contribution of cost components in the total cost function which was set to
be minimised at each period. The parameter α is set to change the relative contribution
of material and setup costs where higher alpha values make the cost function to be
dominated by the material cost and lower alpha values make the cost function to be
dominated by the setup cost. Similarly, higher r and h values make the cost function
to be dominated by the cost of returning and holding ULs, which act as a restriction to
return and use of ULs.

At each period, an irregular bin packing problem was solved using the modified IJRAB
algorithm which terminates after 300 Jostle cycles.

Tables 6.4, 6.5, 6.6, and 6.7 summarize the results for four policies. The same results
for each scenario are graphically represented in Figure 6.6 and 6.7 for 10 periods and 30
periods.

Table 6.4: Model 02: NonUL-Mix vs. UL-Mix results for with different α and
h values (simulation results for 10 periods)
nonUL-Mix UL-Mix
Area(S) T rimloss Cost(T otal) Area(S) Cost(T otal)
P P P P
h α Area(R) / Area(S) Area(O) / Area(R)
P P P P
U tilization% T rimloss U tilization%
Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg.
0.05 22220.2 7399.3 221041.1 0.661 22141.6 7059.9 226967.5 0.662 1.76% 33.10%
0.1 22149.8 7328.8 122129.8 0.663 22137.5 7078.1 122322.7 0.662 1.65% 34.66%
0.5 22118.3 7297.3 41962.5 0.662 22135.5 7079.8 42674.0 0.662 1.68% 36.98%
1 1 22127.2 7306.2 32033.0 0.662 22074.9 6992.2 32373.5 0.663 1.83% 35.10%
1.5 22094.0 7273.0 28774.2 0.664 22037.9 6932.8 29038.1 0.664 1.97% 34.55%
5 22000.6 7179.6 24112.2 0.667 21909.7 6699.9 24185.1 0.669 3.37% 47.26%
10 21952.3 7131.3 23026.6 0.669 21913.3 6660.4 23131.7 0.668 3.76% 47.58%
0.05 22092.2 7271.2 217788.0 0.664 22066.7 7071.6 223082.1 0.664 1.36% 41.93%
0.1 22097.7 7276.8 121186.9 0.664 22135.4 7093.2 123318.1 0.663 1.50% 33.22%
0.5 22097.0 7276.0 41899.6 0.663 22064.5 7048.2 42448.7 0.664 1.37% 35.56%
10 1 22122.8 7301.9 32050.8 0.664 22154.8 7127.7 32663.6 0.663 1.43% 34.91%
1.5 22088.9 7268.0 28860.8 0.664 22023.2 7034.6 29073.9 0.666 1.50% 49.13%
5 21976.1 7155.1 24081.3 0.668 21849.1 6803.8 24277.6 0.671 2.12% 51.47%
10 21941.5 7120.5 23031.6 0.669 21790.7 6741.6 23201.7 0.674 2.32% 54.87%
0.05 22108.3 7287.3 219271.6 0.663 22278.7 7402.8 224856.1 0.657 0.35% 29.11%
0.1 22113.3 7292.3 120600.1 0.664 22170.6 7270.6 122409.0 0.660 0.49% 27.73%
0.5 22142.1 7321.1 41729.2 0.662 22188.7 7331.8 42315.5 0.659 0.31% 48.45%
100 1 22057.5 7236.5 31994.8 0.665 22134.6 7277.5 32244.2 0.660 0.29% 44.51%
1.5 22059.4 7238.4 28732.1 0.664 22146.1 7292.4 29154.6 0.660 0.30% 50.11%
5 22025.2 7204.2 24115.7 0.667 22019.3 7187.0 24267.4 0.664 0.17% 69.18%
10 21924.9 7103.9 22991.0 0.669 21954.5 7115.5 23144.4 0.666 0.14% 42.54%
0.05 22146.4 7325.4 218608.5 0.662 22258.1 7429.1 220208.9 0.657 0.05% 34.34%
0.1 22163.1 7342.2 119783.6 0.662 22312.2 7484.6 122557.5 0.657 0.04% 21.47%
0.5 22166.1 7345.1 42009.8 0.661 22271.3 7442.9 43420.5 0.657 0.04% NA
1000 1 22083.2 7262.2 31891.8 0.664 22225.5 7401.5 33306.6 0.659 0.03% NA
1.5 22033.2 7212.3 28702.2 0.665 22160.6 7337.6 29865.8 0.660 0.03% NA
5 22012.2 7191.2 24112.5 0.668 22080.7 7258.4 24477.5 0.664 0.01% NA
10 21974.3 7153.3 23053.5 0.668 22081.3 7258.9 23535.4 0.664 0.01% NA
0.05 22126.5 7305.5 220029.4 0.662 22249.6 7417.7 229760.8 0.659 0.06% 12.01%
0.1 22151.1 7330.1 120906.5 0.663 22293.7 7470.7 129015.1 0.656 0.03% NA
0.5 22131.9 7311.0 41901.7 0.662 22282.2 7455.9 51153.5 0.658 0.06% 59.39%
10000 1 22070.1 7249.1 32107.5 0.664 22251.9 7424.5 41778.5 0.658 0.03% NA
1.5 22049.4 7228.4 28773.2 0.665 22197.7 7370.2 34873.3 0.660 0.03% NA
5 21972.0 7151.0 24093.0 0.668 22110.9 7289.5 27415.2 0.663 0.01% NA
10 21954.8 7133.8 23021.7 0.669 22062.7 7240.3 28593.1 0.664 0.02% NA

Best performing policy at each scenario

Comparison between Scenario 1 (Expensive material with higher labour cost of managing
ULs) and Scenario 7 (Expensive material with lower labour cost of managing ULs):
152 Chapter 6 Use and Reuse of Materials in Cutting Industries

Table 6.5: Model 02: NonUL-Single L vs. UL-Single L results for with different
α and h values (simulation results for 10 periods)
nonUL-Single Large UL-Single Large
Area(S) T rimloss Cost(T otal) Area(S) Cost(T otal)
P P P P
h α Area(R) / Area(S) Area(O) / Area(R)
P P P P
U tilization% T rimloss U tilization%
Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg.
0.05 23791.1 8970.1 170649.4 0.614 22429.1 6944.6 171799.9 0.654 6.22% 52.42%
0.1 23848.0 9027.0 97452.7 0.614 22381.8 6926.6 96730.9 0.654 6.15% 53.90%
0.5 23846.8 9025.8 38567.1 0.613 22333.7 6896.2 37198.3 0.655 5.87% 52.95%
1 1 23811.3 8990.3 31160.5 0.614 22374.9 6911.7 29852.3 0.655 6.22% 53.86%
1.5 23785.6 8964.6 28679.7 0.614 22293.9 6845.3 27311.5 0.658 6.37% 55.79%
5 23828.9 9007.9 25299.8 0.613 22049.5 6660.1 23672.3 0.665 8.60% 70.03%
10 23934.9 9113.9 24673.6 0.612 21938.0 6567.9 22760.1 0.669 9.54% 73.75%
0.05 23792.3 8971.4 170658.6 0.615 22344.5 6891.2 170871.5 0.656 5.94% 52.33%
0.1 23708.7 8887.7 96883.7 0.616 22393.3 6899.9 97144.5 0.655 6.27% 52.08%
0.5 23879.9 9059.0 38620.7 0.612 22336.6 6928.0 37499.9 0.657 5.65% 53.46%
10 1 23841.5 9020.5 31200.0 0.613 22400.4 6948.7 30256.8 0.655 6.10% 53.86%
1.5 23868.6 9047.6 28779.8 0.610 22240.7 6861.2 27589.4 0.659 6.16% 59.21%
5 23896.5 9075.5 25371.6 0.612 22100.0 6770.7 24026.2 0.664 7.47% 69.19%
10 23838.8 9017.8 24574.5 0.614 21958.8 6659.6 23142.9 0.668 8.74% 75.08%
0.05 23827.1 9006.1 170907.9 0.613 22305.3 7183.4 171564.1 0.653 5.17% 73.89%
0.1 24040.2 9219.2 98238.3 0.610 22400.3 7240.8 97532.7 0.651 4.94% 69.43%
0.5 23787.3 8966.3 38470.8 0.614 22355.1 7216.2 38579.9 0.652 4.97% 71.36%
100 1 23792.3 8971.4 31135.6 0.615 22380.6 7237.7 31042.6 0.652 5.02% 71.36%
1.5 23971.6 9150.6 28904.0 0.611 22326.0 7190.1 28657.4 0.654 5.16% 72.64%
5 23818.7 8997.7 25288.9 0.614 22215.3 7174.6 24777.9 0.656 5.42% 81.75%
10 23847.7 9026.7 24583.7 0.614 22271.8 7209.8 24120.5 0.654 5.35% 79.76%
0.05 23858.2 9037.2 171131.1 0.613 23675.3 8845.3 173038.2 0.620 1.05% 96.38%
0.1 23859.0 9038.0 97497.7 0.613 23670.9 8841.7 98667.6 0.620 0.92% 96.20%
0.5 23769.1 8948.1 38441.4 0.615 23632.6 8803.0 39237.1 0.619 0.75% 95.12%
1000 1 23872.9 9051.9 31241.0 0.613 23641.3 8802.3 32139.4 0.621 0.92% 91.78%
1.5 23830.6 9009.6 28734.0 0.613 23543.9 8706.5 29636.9 0.622 1.01% 93.04%
5 23829.5 9008.5 25300.4 0.614 23632.6 8800.9 25917.4 0.619 0.71% 93.63%
10 23860.2 9039.2 24596.6 0.613 23490.3 8661.7 25304.1 0.623 0.94% 96.50%
0.05 23802.5 8981.5 170731.2 0.614 23537.1 8702.9 184809.2 0.622 1.06% 94.67%
0.1 23840.6 9019.6 97422.7 0.614 23637.4 8816.4 110684.9 0.620 1.00% 100.00%
0.5 23921.4 9100.4 38687.7 0.612 23538.0 8708.9 50401.9 0.622 1.09% 96.84%
10000 1 23908.0 9087.1 31287.0 0.612 23594.8 8765.5 40277.9 0.621 0.74% 95.19%
1.5 23871.3 9050.3 28783.1 0.613 23537.6 8708.0 40309.9 0.622 0.95% 96.14%
5 23823.6 9002.6 25294.2 0.613 23663.8 8842.8 35407.5 0.620 0.84% 100.00%
10 23893.2 9072.3 24630.7 0.612 23558.1 8728.2 35482.1 0.622 0.90% 95.80%

Table 6.6: Model 02: NonUL-Mix vs. UL-Mix results for with different α and
h values (simulation results for 30 periods)
nonUL-Mix UL-Mix
Area(S) T rimloss Cost(T otal) Area(S) Cost(T otal)
P P P P
h α Area(R) / Area(S) Area(O) / Area(R)
P P P P
U tilization% T rimloss U tilization%
Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg.
0.05 65911.5 21820.6 653298.2 0.662 65782.8 21050.6 684220.2 0.663 1.79% 45.59%
0.1 65740.5 21649.6 358053.8 0.664 65679.8 21064.2 373612.5 0.665 1.62% 50.76%
0.5 65766.3 21675.4 124576.1 0.663 65728.5 21054.7 128274.2 0.664 1.81% 51.01%
1 1 65702.1 21611.2 95437.4 0.664 65624.5 20928.8 96871.7 0.664 1.82% 49.26%
1.5 65654.2 21563.3 85716.5 0.665 65299.9 20650.3 86682.2 0.667 1.88% 54.53%
5 65435.0 21344.1 71740.9 0.668 64895.2 20032.2 71958.7 0.672 2.98% 60.04%
10 65375.7 21284.8 68575.9 0.669 64727.7 19844.1 68571.5 0.673 3.68% 66.74%
0.05 65853.5 21762.6 654636.0 0.662 65652.6 21047.1 689980.7 0.664 1.59% 50.79%
0.1 65937.4 21846.5 357939.5 0.662 65590.7 21069.0 374576.8 0.665 1.41% 53.57%
0.5 65832.5 21741.6 124116.0 0.663 65543.2 21012.2 128554.1 0.665 1.50% 55.18%
10 1 65740.7 21649.8 95241.3 0.663 65513.9 20997.1 98234.8 0.666 1.42% 54.25%
1.5 65610.6 21519.6 85494.5 0.664 65306.8 20800.0 87915.9 0.668 1.55% 59.02%
5 65351.5 21260.6 71613.1 0.668 64847.9 20230.4 73318.5 0.673 2.25% 63.98%
10 65321.8 21230.9 68521.6 0.668 64783.7 20217.7 70044.1 0.674 2.36% 68.91%
0.05 65958.2 21867.2 648241.5 0.662 66037.3 21840.5 667451.9 0.660 0.43% 62.63%
0.1 65869.2 21778.3 358044.7 0.662 65973.1 21801.8 366609.2 0.660 0.35% 65.35%
0.5 65773.0 21682.0 124026.6 0.663 66040.7 21849.0 127864.6 0.659 0.41% 62.73%
100 1 65781.4 21690.5 95399.9 0.663 65874.0 21719.3 97597.5 0.660 0.31% 68.37%
1.5 65687.1 21596.2 85664.5 0.665 65706.0 21556.2 87060.2 0.662 0.29% 68.65%
5 65349.7 21258.8 71617.2 0.668 65425.9 21301.1 72391.4 0.665 0.22% 76.17%
10 65281.2 21190.3 68484.1 0.669 65279.6 21163.3 69177.6 0.666 0.25% 84.67%
0.05 65953.8 21862.9 654711.7 0.662 66182.9 22076.1 663699.5 0.658 0.08% 69.73%
0.1 65838.3 21747.4 359385.8 0.663 66286.7 22171.9 370571.1 0.657 0.07% 49.07%
0.5 65884.5 21793.6 124604.6 0.662 66294.6 22187.2 131664.6 0.657 0.05% 44.92%
1000 1 65716.2 21625.3 95324.8 0.664 66134.8 22034.6 101748.1 0.658 0.04% 67.24%
1.5 65580.2 21489.2 85476.2 0.665 65930.4 21833.5 90779.7 0.661 0.03% NA
5 65335.1 21244.2 71551.0 0.668 65670.2 21575.7 75948.8 0.663 0.02% NA
10 65247.7 21156.7 68450.6 0.669 65665.0 21571.1 71076.4 0.664 0.02% NA
0.05 65916.3 21825.4 652014.7 0.662 66227.1 22120.8 727967.8 0.658 0.04% 43.52%
0.1 65930.6 21839.7 357314.7 0.662 66344.8 22246.6 408833.8 0.657 0.03% NA
0.5 65765.9 21675.0 124235.9 0.663 66290.2 22185.0 185529.1 0.657 0.03% NA
10000 1 65833.3 21742.4 95278.2 0.663 66209.7 22112.8 171594.1 0.658 0.04% 77.91%
1.5 65623.7 21532.8 85540.5 0.665 66053.3 21956.9 124502.1 0.660 0.02% NA
5 65414.6 21323.7 71727.2 0.668 65785.9 21693.0 102210.1 0.662 0.02% NA
10 65368.4 21277.4 68583.1 0.668 65693.5 21600.8 101311.5 0.664 0.01% NA

In Scenario 1, the nonUL-Mix has demonstrated the lowest cost as illustrated in the top
right graphs of Figures 6.6 and 6.7. In this scenario selecting a nonUL option is cost
beneficial than UL options since there is a high cost associated with managing leftovers
Chapter 6 Use and Reuse of Materials in Cutting Industries 153

Table 6.7: Model 02: NonUL-Single L vs. UL-Single L results for with different
α and h values (simulation results for 30 periods)
nonUL-Single Large UL-Single Large
Area(S) T rimloss Cost(T otal) Area(S) Cost(T otal)
P P P P
h α Area(R) / Area(S) Area(O) / Area(R)
P P P P
U tilization% T rimloss U tilization%
Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg.
0.05 71288.4 27197.5 511340.5 0.612 66020.3 20699.1 510074.1 0.660 5.27% 64.66%
0.1 71032.8 26941.9 290269.9 0.613 66044.3 20724.9 288580.1 0.660 5.24% 64.50%
0.5 71225.6 27134.7 115192.0 0.612 65998.1 20716.5 110741.2 0.660 5.32% 66.08%
1 1 71136.6 27045.7 93092.4 0.612 65961.6 20738.4 88530.2 0.660 5.36% 67.99%
1.5 71149.8 27058.9 85789.7 0.612 65654.5 20514.4 80848.6 0.664 5.59% 71.42%
5 70988.6 26897.7 75370.6 0.613 64839.0 19897.9 69792.2 0.675 7.87% 83.33%
10 71208.3 27117.4 73406.1 0.612 64725.4 19743.3 67352.4 0.676 8.79% 84.34%
0.05 71218.3 27127.4 510837.5 0.612 66020.2 20739.9 511910.0 0.660 5.08% 64.57%
0.1 71052.7 26961.8 290351.3 0.613 65984.5 20718.4 289590.1 0.661 5.04% 64.63%
0.5 71208.5 27117.5 115164.3 0.612 65959.2 20836.3 112577.7 0.661 5.10% 69.34%
10 1 71182.0 27091.1 93151.7 0.612 65844.4 20683.7 90327.4 0.662 5.28% 69.21%
1.5 71055.4 26964.5 85675.9 0.613 65674.9 20555.0 82746.6 0.664 5.54% 71.73%
5 71191.5 27100.6 75586.0 0.612 64969.0 20126.3 71537.4 0.671 6.90% 83.22%
10 71280.4 27189.5 73480.4 0.612 64669.4 19900.0 68698.9 0.675 7.77% 86.49%
0.05 71174.5 27083.6 510523.2 0.612 66228.4 21601.5 513315.8 0.654 4.50% 82.00%
0.1 71154.0 27063.1 290765.0 0.612 66337.2 21634.2 294034.8 0.653 4.50% 79.50%
0.5 71119.1 27028.2 115019.7 0.612 66303.6 21678.2 116485.9 0.653 4.65% 82.68%
100 1 71055.9 26965.0 92986.7 0.613 66079.7 21531.7 93514.2 0.655 4.48% 84.56%
1.5 71191.2 27100.3 85839.6 0.612 66000.7 21547.9 85462.3 0.655 4.60% 88.08%
5 71054.1 26963.2 75440.1 0.613 65883.0 21473.3 74313.4 0.656 5.17% 90.64%
10 71087.4 26996.4 73281.4 0.613 65801.5 21433.0 71684.7 0.656 5.10% 91.74%
0.05 71091.9 27001.0 509931.1 0.613 70420.4 26326.3 513893.1 0.620 0.87% 99.49%
0.1 71168.8 27077.9 290825.7 0.612 70399.7 26299.8 293564.0 0.620 0.92% 98.62%
0.5 70870.7 26779.8 114618.0 0.614 70233.6 26131.5 117822.5 0.621 0.96% 98.34%
1000 1 71187.9 27097.0 93159.5 0.612 70400.6 26300.7 95962.2 0.620 0.95% 98.66%
1.5 71112.2 27021.3 85744.3 0.613 70281.7 26165.0 88496.2 0.621 0.94% 96.11%
5 71200.8 27109.8 75595.9 0.612 70342.8 26243.2 77960.8 0.620 0.86% 98.57%
10 71110.6 27019.7 73305.4 0.613 70238.5 26127.5 75849.5 0.621 0.90% 96.83%
0.05 71171.0 27080.1 510498.1 0.612 70441.1 26341.3 541109.1 0.619 0.84% 98.50%
0.1 71303.1 27212.2 291374.3 0.611 70386.2 26285.4 324609.6 0.620 0.88% 98.40%
0.5 71273.9 27182.9 115270.1 0.612 70429.7 26329.5 147876.7 0.620 0.91% 98.54%
10000 1 71094.2 27003.3 93036.9 0.613 70424.8 26322.6 122525.0 0.619 0.86% 98.12%
1.5 71164.2 27073.2 85807.0 0.612 70456.8 26357.2 120329.6 0.620 0.91% 98.65%
5 71313.0 27222.1 75715.0 0.612 70316.8 26216.8 108960.3 0.620 0.94% 98.63%
10 71229.6 27138.7 73428.1 0.612 70472.6 26364.1 102128.9 0.619 0.83% 97.00%

(r). Therefore, at each period, the algorithm selects the solutions with less number of
NewULs, in order to minimise the total cost. In this scenario, the choice of selecting
single bin size vs. mix bin size policy leads us to choose the mix bin size policy due to
its ability to save material through a variety of bin sizes and thereby reducing the total
cost. As the scenario is significantly dominated by material cost compared to the setup
cost, the impact of setup cost by using a higher number of bins with a mix of bin sizes
is negligible. In summary, within the context of implementing nonUL, the results reveal
that the mix of bin sizes is able to save expensive material better than any of the single
bin size options when setup cost is not making significant effect against the cost saving
achieved by the material.

We noticed a similar outcome in Model 1, where the mixed bin sizes provide the highest
saving than any single bin size if the production policy is based on no use of ULs. In
fact, the context of not using ULs in Model 1 (i.e. Jmax = 0) emulates the features of
Scenario 1 in Model 2, by discouraging the occurrence of ULs. Additionally, in Scenario
1 in Model 2, since materials are expensive, it also encourages saving material as much
as possible as in Model 1.

The results related to the Scenario 7 illustrated by the bottom right graphs in Figures
6.6 and 6.7, shows the lowest total cost when the UL-Large policy is imposed. Scenario
7 motivates minimising material usage due to the dominance of the material cost in the
cost function. The scenario also encourages more returns of ULs to the warehouse due
to the low labour cost of managing ULs, while encouraging saving material as much
154 Chapter 6 Use and Reuse of Materials in Cutting Industries

Figure 6.6: Model 02: Total cost graphs by running simulation for 10 periods
for each policy

as possible. According to Tables 6.4, 6.5, 6.6, and 6.7, it is evident that the UL-Large
policy results an increased number of leftovers when h and r are lower and α is higher.
In order to reduce the total cost in Scenario 7, the results demonstrate more returns of
ULs and using them, to save more material since the material cost dominates the total
cost. We have observed similar outcome when assessing results of Model 1 in which the
highest usage of ULs was recorded in UL single bin size policy with larger bin size than
any other policy.

Comparison between Scenario 2 (Expensive setups with higher labour cost of managing
ULs) and Scenario 8 (Expensive setups with lower labour cost of managing ULs):

In Scenario 2 (see the graphs on the top left in Figures 6.6 and 6.7), the cost environment
Chapter 6 Use and Reuse of Materials in Cutting Industries 155

Figure 6.7: Model 02: Total cost graphs by running simulation for 30 periods
for each policy

is dominated by setup cost and labour cost, therefore results demonstrate solutions with
less number of bins by encouraging minimising the number of used bins and the returning
of ULs to the warehouse. Therefore, the policies with single standard bins with smaller
sizes and mix sizes of bins are not favoured in this case since those options result in a
higher number of bins. Instead, using large bin sizes which use a lower number of setups
are preferred in this scenario. This situation is also supported by the experiment results
of Model 1 in the case of nonUL which records the lowest number of used bins with
largest bin sizes.

However, the scenario simultaneously motivates using ULs due to the low cost of labour
in managing ULs. The simulation results for Scenario 8 show almost equal minimum
costs for nonUL-Single Large and UL-Single Large policies than the other options (see
156 Chapter 6 Use and Reuse of Materials in Cutting Industries

bottom left graphs of Figures 6.6 and 6.7). As in Scenario 2, Scenario 8 motivates
solutions with a lower number of bins (i.e. setups). Set-up cost dominates the cost
function and therefore supports the use of large bin sizes which reduces the number of
used bins. Even though there is an allowance for ULs, the higher dominating power
of setup cost in this scenario forces to use less number of bins from standard bins, UL
bins or both. The UL-Single Large shows improvement in saving material than nonUL-
Single Large though it doesn’t demonstrate a significant advantage in the view of total
cost since both policies tend to have no significant difference in amount of setups used.
Also, in Scenario 8, achieving low-cost solutions with smaller standard bin sizes with or
without using ULs is not likely due to the limitation in their sizes (i.e. smaller standard
bins and smaller ULs making a higher number of used bins in the solutions which are
not advantageous in terms of setup costs).

Scenario 4 (Expensive material with medium level of labour cost for managing ULs)

Scenario 4 is an intermediate situation between Scenario 1 and Scenario 7. Scenario 4

motivates minimising material usage as material cost dominates the total cost function,
however, the impact of the labour cost of managing ULs can be varied from high to
low. As a transitional scenario, it is worth discussing how the appropriate policies are
determined by the variability in labour costs in the total cost function.

The three graphs on the right side of the Figures 6.6 and 6.7 demonstrate how the
variability in labour costs affects the total cost when the material is expensive. When
moving from Scenario 7 to 1 (from low to high of r and h), the best applicable policy
is changed from UL-Large to nonUL-Mix. During this transition, the total cost of the
UL-Large policy with respect to the other three policies is increased. On the other
hand, the total cost of the nonUL-Mix policy with respect to the other three policies
is decreased. This implies that the appropriate policy for Scenario 4 depends on the
relative bias towards Scenario 7 or Scenario 1. However, the applicable policies in this
scenario where the expensive material is being used can only be UL-Large or nonUL-Mix.

In summary, it is evident that both UL-Large policy and nonUL-Mix policies are more
sensitive to variations in labour costs of managing ULs when expensive sheet materials
are used. It also proved that when unit costs of h and r are moving from low to high,
the best applicable policies are shifting from UL Large policy to nonUL-Mix policy.

Results of Scenario 5 (Expensive setups with medium level of labour cost for managing
ULs)

Scenario 5 is the intermediate situation between Scenario 2 and Scenario 8. Scenario

5 motivates minimising number of used bins as setup cost dominates the total cost
function, while the impact of the labour cost of managing ULs varies from high to low.
As a transitional scenario, we discuss the appropriate policy alignment for the variability
in labour cost in the total cost function.
Chapter 6 Use and Reuse of Materials in Cutting Industries 157

The three graphs in the left side of the Figures 6.6 and 6.7 demonstrate how the variabil-
ity in labour costs (h and r) affects to the total cost when the setup cost is high. When
moving from Scenario 8 to 2 (from low to high of r and h), the best applicable policy
is changed from UL-Large to nonUL-Large. During this transition, the total cost of the
UL-Large policy with respect to the other three policies is increased. In contrast, the
total cost of nonUL-Large policy with respect to the other three policies is decreased.
This implies that the appropriate policy for Scenario 5 depends on the relative bias to-
wards Scenario 8 or Scenario 2. However, the applicable policies in this scenario where
expensive setups are being used can only be UL-Large or nonUL-Large.

In summary, this demonstrates that both UL-Large policy and nonUL-Large policies
are more sensitive to variations in labour costs of managing ULs when expensive setups
occur. It is also evident that when unit costs of h and r are moving from low to high,
the best applicable policies are shifting from UL Large policy to nonUL-large policy.

Results of Scenario 3 (Balanced cost level of material and setup at high labour cost
factors for managing ULs)

Scenario 3 is an intermediate situation between Scenario 2 and Scenario 1. Scenario

3 discourages using ULs due to the high cost of labour. On the other hand, the total
cost is mostly determined by the variability of the relative ratio between setup cost and
material cost.

The test results illustrated in top three graphs of the Figures 6.6 and 6.7 demonstrate
how the variability in the ratio between unit setup and unit material cost affects the total
cost in Scenarios 2, 3 and 1. When moving from Scenario 2 to 1 (from relatively lower
material cost to higher material cost and from relatively higher setup cost to lower setup
cost), the best applicable policy is changed from nonUL-Large to nonUL-Mix. During
this transition, the total cost of nonUL-Large policy relative to nonUL-Mix policy is
increased. In contrast, the total cost of nonUL-Mix policy relative to nonUL-Large
policy is decreased. This implies that the appropriate policy for Scenario 3 depends
on the relative bias towards Scenario 2 or Scenario 1. However, the applicable policies
swing between nonUL-Large or nonUL-Mix.

Results of Scenario 9 (Balanced cost level of material and setups at low labour cost
factors for managing ULs)

Scenario 9 is an intermediate situation between Scenario 8 and Scenario 7. Scenario

9 motivates using ULs to save material due to the low cost of labour, however, the
total cost is determined by the variability of the relative ratio between setups cost and
material cost.

The test results illustrated in bottom three Figures 6.6 and 6.7 demonstrate how the
variations in c and s affect the total cost in Scenarios 8, 9 and 7. When moving from
Scenario 8 to 7 (from relatively lower material cost to higher material cost ad from
158 Chapter 6 Use and Reuse of Materials in Cutting Industries

relatively higher setup cost from lower setup cost), surprisingly, the best applicable
policy remains unchanged in UL-Large, due to the dominance of leftover generation
motivated by the low cost of h and r. Therefore, it is interesting to analyse the change
that would happen when leftover generation is limited by increasing h and r.

Results of Scenario 6 (Balanced cost level of material and setups at medium level of h
and r)

Scenario 6 demonstrates a balanced state of all other eight scenarios. Along the hor-
izontal level of Scenario Matrix illustrated in Figure 6.5, material cost and setup cost
dominates the decision of selecting the appropriate policy (see the variation of three
cost graphs from left to right; in the middle of the Figures 6.6 and 6.7). In other words,
the results correspond to this scenario demonstrate minimum cost by using UL-Large
or nonUL large policies when setup cost dominates the cost function. On the other
hand, when the material cost dominates the cost function, the nonUL-Mix and UL-Mix
provides the minimum cost.

Along the vertical level of the Matrix, variability in labour costs of managing ULs affects
the total cost. In this case, Scenario 6 acts as a transitional stage between Scenario 3 and
Scenario 9. When unit costs of h and r are moving from low to high (see the variation
of three cost graphs from bottom to top; in the middle of the Figures 6.6 and 6.7, the
best applicable policies shift from UL-Large policy to nonUL-Large policy when setup
cost is dominated or if not, to nonUL-Mix policy when material cost is dominated.

6.5.2 Findings and discussion

Summarising the results of the direct cost model, we develop a supportive tool which
companies can use as a guide in deciding the appropriate policy for their cost scenario.
The graphical illustration of the framework is shown in Figure 6.8.

If a company follows one of the four extreme scenarios; 1,2,7 and 8, this framework
provides direct guidance in choosing the appropriate policy. For example, the framework
guides a company to use UL-Large policy if their practical situation is similar to Scenario
7. One of the key findings of the data analysis of Model 2 is that when the sheet cutting
process acts on a transitional scenario, then there is an intersection of policies that can
be used in that scenario. For example, in Scenario 5, there is a shift in policy which
depends on the variability of the labour cost. Similarly, in Scenario 3, the policy shift
can be identified based on the variability of the ratio between material cost and setup
cost. This is clearly represented in the supportive tool by showing the intersections of
policies in transitional scenarios.

By reviewing existing literature, we found two main practices that companies believe
when dealing with residuals. First one is that they believe that material waste can
Chapter 6 Use and Reuse of Materials in Cutting Industries 159

Figure 6.8: Supportive tool for policy selection

be minimised by reusing residuals. Cui et al. (2016) claim that multiple bar types in
one-dimensional cutting are useful for improving material utilisation as they expand
the solution space. Furthermore, they claim that different sizes of residuals can improve
utilisation when there is a small number of standard bar types to purchase. The findings
of this study support this argument, even for the 2D irregular sheet cutting process, by
providing economic outputs in Scenarios 7,8 and 9. However, the other scenarios failed
to provide economical solutions mainly due to the domination of other cost factors.

Manufacturers also prefer reducing the number of setups, due to the assumption of high
setup cost in manufacturing compared to material cost. According to our analysis, we
found that this is highly applicable in Scenarios 2,5 and 8, but is not applicable or
Scenarios 1,4 and 7.

This sums up that practical situations of using residuals depend on different cost factors.
Therefore, it is important to evaluate those different factors in deciding the appropriate
operational policy, to achieve cost-effective sheet cutting process with residuals.
160 Chapter 6 Use and Reuse of Materials in Cutting Industries

6.6 Concluding Remarks

For practical applications, it is believed that the use of a combination of sheet sizes can
reduce the overall consumption of material. In order to achieve this, a manufacturer
tends to follow one of four operational policies: 1) use a mix of standard sheet sizes with
no use of residuals, 2) use a mix of standard sheet sizes with the use of residuals, 3) use
a single size standard stock sheet with no use of residuals or 4) use single size standard
stock sheets with the use of residuals. In this chapter, we investigate the outcome of those
four operational policies. The study explicitly considers managing the maximum number
of UL-types if a production policy is based on using residuals; by considering the handling
and storage limitation of ULs arise in a production environment. In comparison, the
policies based on ULs often lead to the significant reduction in the trim-loss and improve
the overall utilisation of material. In this study, we demonstrate that the largest standard
bin sizes perform well in saving material than other mentioned policies, by allowing an
adequate upper bound on number of UL bin types for an adequate number of periods,
so that multi-period run can be reached to a steady state level which effectively returns
a substantial amount of ULs to the warehouse and effectively utilizes those ULs in
subsequent periods. Otherwise, the study recommends the NonUL Mix policy to operate
in order to achieve the highest advantage in saving material.

Next, we extend the multi-period analysis to consider cost of material as well as setup
cost and cost of managing ULs related to the sheet cutting process. In this case, we
produce results to provide useful insights for manufacturers to tune their production
and procurement policies using the cost results generated. Our investigation is based
on a series of multi-period runs, which apply the above-mentioned four policies on nine
different cost scenarios which are illustrated in Table 6.3 and Figure 6.5. The newly
proposed cost model fills the gap in the literature by simulating multi-period cut order
processing with different operational policies, particularly related to the 2D irregular
sheet cutting problems. The study addresses the issues in cost trade-off and provides
economical recommendations applicable in different scenarios.

By analysing the results, we summarise the outcomes of each policy at different cost
scenarios and present them as a supportive tool which is presented in Figure 6.8. This
shows that using Standard Single Large bins with ULs (UL-Single Large) is economical
in low labour cost scenarios of managing ULs. If the cost of managing ULs is expensive,
our findings recommend to use NonUL-Mix policy if material cost is dominant than
setup cost; or use Non-UL Single policy with large bin sizes if setup cost is dominant
than the material cost in the total cost function.

The findings in Figure 6.8 support operations managers to identify the practical scenario
of a given sheet cutting process by assessing the characteristics such as the price of sheet
material, the cost of machinery and the cost of labour, and then to select the most
suitable policy. Furthermore, the tool provides an in-depth analysis of the transitional
Chapter 6 Use and Reuse of Materials in Cutting Industries 161

scenarios which explains the variability of cost factors affect to the total cost. We
believe that the overall methodology described in this chapter can be applied for other
cutting and packing problems to analyse the same behavioural aspects and decide the
appropriate policies for different practical situations.
Chapter 7

Conclusion and Future Work

7.1 Conclusion

The main objective of this study is to develop efficient computational methods to solve
three types of irregular shape bin packing problem. Based on the findings of this study,
the concluding remarks are summarised as follows.

First, in Chapter 4 we consider the two-dimensional irregular shape single bin size bin
packing problem, without restricting the final orientation angle of pieces at the point
of placement. The newly developed constructive method embedded within the Jostle
heuristic, finds competitively local optimal solutions within reasonable computation time
compared to the other computational methods in the literature. As a search method
which is applicable for irregular bin packing problems, we demonstrate that Jostle has
the ability to self-governing the neighbourhood structure that modifies the solution. The
proposed angle tuning mechanism can find promising placement-orientation angles on a
continuous scale. We also propose an iterated jostle procedure to explore the solution
space more broadly rather than being stuck at one local optimum. We show that the
Iterated Jostling procedure enhances the searchability and improves its efficiency in
finding good solutions.

In comparison, the tests conducted based on different pre-ordering, such as sorting pieces
by decreasing area, did not reach the solution quality obtained by the local search meth-
ods. In addition, the generated layouts were significantly better than a random search
process. After implementing and testing several different placement techniques we re-
alised that the Minimum length and Maximum utilisation placement rules are better
than the traditional Bottom Left rule for irregular bin packing with unrestricted rota-
tion. We also test the results generated for unrestricted rotation of pieces vs. restricted
rotation of pieces. Not surprisingly, packing pieces with no rotation restriction outper-
forms the results generated through restricted rotation of pieces apart from the jigsaw

163
164 Chapter 7 Conclusion and Future Work

instances. Overall, the proposed algorithms have generated better results within less
computation time than the results published in the literature.

In Chapter 5, we consider the two-dimensional irregular shape multiple bin size bin
packing problem with the unrestricted rotation of pieces. The problem is new to the
literature and is a more general version of the 2D irregular bin packing problem. In
this study, we introduce two constructive algorithms to generate feasible solutions for
the problem. The first algorithm; denoted as CA1, is an extended method of the con-
structive method discussed in Chapter 4 which is supported by the Jostle heuristic in
finding the input permutation of pieces. The second constructive algorithm; denoted as
CA2, generates a feasible solution for a given order of pieces and uses a traditional bin
selection heuristic in allocating pieces to the available bin types. Three main solution
methodologies; Hybrid method of Genetic Algorithm/Jostle (HGA), Iterated Jostle with
Random Assignment of Bins (IJRAB) and Sequential Packing with Bin Centric Heuris-
tic (SPBCH), are designed to search for a solution. Each method consists of a unique
way of changing the permutation of bins during the solution searching stage so that
overall utilisation of bins is maximised. Out of the two constructive algorithms, CA1
supported by the Jostle heuristic demonstrates better results in terms of solution quality
and computational complexity. Out of three main computational methods, HGA and
IJRAB generate the better results and both used the approach of jostling pieces within
a certain configuration of bins arranged in a certain sequence. The two methods fol-
low different techniques to change the bin configuration, hence change the permutation
pieces as well.

We also test different search improvement strategies and identify one version of IJRAB
(IJRAB A3) outperforms all the other search techniques in terms of finding quality
solutions in quicker time. The IJRAB implements kicks to change the bin order and piece
permutation when the search process becomes stuck at a local optimum. By running
IJRAB with the proposed post-processing strategy; repacking of pieces in each of the
occupied bins into the possible smallest bin, makes the search process more efficient. We
also test different bin size configurations to test our algorithms and identify that a mix of
bin sizes including very small bin sizes and very large bin sizes with moderate sized bins,
leads to the best results in utilising the overall usage of bin area. The major findings
of this study are applicable when a company requires making procurement decisions in
a situation when there is a set of standard sheets available in multiple sheet sizes for
the sheet cutting process. The study also emphasises its use when there is an option to
use off-cuts to cut irregular shapes in the cutting process. In both situations, the cut
order is processed considering a heterogeneous set of input bin types. According to the
findings in Chapter 5, the use of a mix of bin sizes is worthwhile in saving material rather
than using a single bin size. Moreover, given a range of input bin types, selecting large
bin sizes is advantageous than smaller bin sizes in terms of saving material. Clearly,
having large bins in the pool of input bin types is also advantageous for irregular shape
Chapter 7 Conclusion and Future Work 165

bin packing, not only to improve the utilisation but also to reduce the number of bins
occupied in the final solution. In Chapter 6, the newly proposed IJRAB algorithm is
used to minimise the bin waste and to minimise the total cost of setups, sheet material
usage and off-cuts handling and storage costs, when residuals are considered for reusing.

In Chapter 6, our investigation is extended toward the implementation of the above algo-
rithms in practical cases. Usually, the first problem addressed here is how a manufacturer
can manage combinations of sheet sizes to reduce overall consumption of materials. In
general, this can be achieved by adopting one of the following four operational policies;
1) use a mix of standard sheet sizes (nonUL-Mix), 2) use a single standard sheet size
(nonUL-Single), 3) use a single standard sheet size with residual reuse (UL-Single) and
4) use a mix of standard sizes of sheets with residual reuse (UL-Mix). We experiment
with sequences of cut orders occurring in consecutive time periods as data instances,
to evaluate the outcomes of the above four operational policies. The general problem
addressed here is the multi-period irregular bin packing problem. Each policy specifies
the specific problem type; whether it is a multi-period irregular shape single bin size
bin packing problem, multi-period irregular shape multiple bin size bin packing prob-
lem or multi-period irregular shape multiple bin size bin packing problem with usable
leftovers. We design a computational method to solve these problems considering the
multi-period sheet cutting optimisation with unknown future demands. The developed
algorithmic tools are based on the best performing Iterated Jostle heuristic and IJRAB
search algorithm described in chapters 4 and 5. When the objective is to minimise the
trim loss, in comparison, the policies based on reusing residuals (i.e. off-cuts) lead to
a significant reduction in trim-loss. The study of reusing residuals explicitly considers
the constraint on the number of Usable Leftover (UL) types since it requires significant
effort to handle and store the usable off-cuts which are heterogeneous in sizes. If the UL
policy is imposed with a reasonable upper bound on the number of UL bin types, using
only the largest standard bin sizes, performs best in terms of saving material. This is
true when the considered time span has higher number of time periods (i.e.> 9), since
it requires a certain amount of time periods to be progressed to have a substantial mix
of heterogeneous set of ULs to achieve better utilization than from a single standard
bin type. Otherwise, for the situations where shorter time spans are considered, using
a mix of standard bin types is advantageous. Even if the UL policy is imposed, in a
multi-period run, it requires a substantial number of periods, a substantial number of
ULs and their size heterogeneity to reach the steady state level which effectively utilises
returned ULs to the warehouse. Otherwise, the Non-UL Mix policy achieves the best
advantage in terms of saving sheet material.

Next, we extend this multi-period analysis by considering other cost factors involved,
to analyse the effect of the same four different operational policies on nine possible
operational cost scenarios. The nine scenarios are based on the relative values of cost
166 Chapter 7 Conclusion and Future Work

variables, so that four scenarios represent the situations where the value of one cost vari-
able is significantly higher than values of the other cost variables. Also, five intermediate
scenarios were defined considering the cost sensitivity and trade-off arising in different
economic situations. This covers material cost, setups cost and handling and storage
cost of off-cuts, which affect the total cost of the process. Figure 6.8 summarises the
findings of this study which works as a guide to select the best performing policy suitable
for each scenario. Selecting the most appropriate policy in an intermediate scenario is
critical since there is an intersection of policies. A shift of policies from one to another
was noted when one cost paradigm is varying from a lower value to a higher value.
Therefore, findings recommend selecting the most appropriate policy in such cases by
analysing the cost variability.

In summary, the thesis has achieved its three main objectives in solving the target prob-
lems; 2D ISBSBPP, 2D IMBSBPP and 2D IBPP with ULs. The algorithms developed
in this thesis have demonstrated their applicability in those problems by finding reason-
ably good solutions in quicker time. Using the algorithms developed, we have presented
how solving multi-period irregular shape bin packing problem with usable leftovers can
be applied to identify the most suitable operational policy for a given sheet cutting pro-
cess. While contributing to the C&P field with a new set of solution methods, we fulfill
the main purpose of this study, which is to develop computational methods for a set of
practical problems arise in sheet cutting industries.

As key contributions, all the computational methods in this study facilitate packing ir-
regular pieces by identifying the promising rotation of pieces when they are being placed
in the bins rather than restricting the piece rotation to a predefined finite set of orien-
tation angles. Compared to the other approaches, the approach proposed in the study
does this more efficiently on a continuous scale of angles range from 0 to 360 degrees.
Second, the application of Jostle heuristic in bin packing problems (2D ISBSBPP and
2D IMBSBPP) is new and includes additional features such as unrestricted rotation of
pieces which was not available in the previous implementation of Jostle algorithms. We
introduce this method as a fast heuristic for solving irregular shape bin packing problems.
Compared to some computational techniques applied in the same context, the proposed
computational methods does not require expensive software such as CPLEX to imple-
ment these algorithms. Third, the study has presented a computational approach for
another variant of IBPP which has not been considered in the literature. While solving
this problem, we achieve one of our aims on evaluating the reusing practices in material
waste arise in sheet cutting processes. Also, we discussed the cost effectiveness of a set
of well-known operational policies within nine different operational cost scenarios.
Chapter 7 Conclusion and Future Work 167

7.2 Future Work

There are several improvement areas related to the scope of this study that can be
investigated as future studies.

The computational methods introduced in this study are based on the search frameworks
designed to find good solutions. The proposed Jostle, Iterated Jostle and the other meta-
heuristic approaches are subject to problem-specific parameter values. Even though we
present better results compared to the results in the existing literature, still there is room
for fine-tuning the parameters or altering the moves to enhance the search procedures.

One significant part of all these algorithms is a generation of no-fit polygons. Throughout
the experiments, we realised, approximately 70% of the computational time is consumed
for no-fit polygon generation. Therefore, it is always better to improve procedures of
generating no-fit-polygons in future research.

Due to the use of geometric representation, the proposed methods always find an accu-
rate representation and placement of the pieces. However, one limitation of this study
is that every piece needs to be a polygonal shape. In future research, we recommend ex-
tending this to the pieces with curved edges that can enhance the scope of applicability
to cases in sheet cutting industries where pieces have curved shapes.

Another important component of the developed algorithms is the generation of the ini-
tial solution. In this study, we mainly relied on random sequence of pieces as the initial
permutation of pieces. However, further investigation of different pre-processing mech-
anisms such as clustering pieces, or an investigation of new rules to order pieces may
be able to find efficient methods to improve the solution approach. As one of the ma-
tured research areas, the algorithms developed for rectangular shapes multiple bin size
bin packing problems provide greater intuition for the irregular bin packing problem.
We believe investigation on those computational techniques or their solution improve-
ment strategies would help to enhance the solution methods developed for irregular bin
packing.

Another main investigation of this study is on managing the reuse of residuals. In our
study, we considered differentiating waste from usable leftovers if leftover size is larger
than the smallest piece type. However, it would be better if further research is carried
out to define an economical size. Considering leftovers which are smaller than this
economical size, would become a cost burden to the company.

Sheet cutting process was the main common application area of the problems we con-
sidered in this study. Based on cut orders processed in different periods, we develop our
analysis so that the outcome of this analysis can be incorporated into decision support
tool to determine the most appropriate operational policy. In this study, our analysis
was limited to major policies such as reusing residuals, purchasing multiple sheet sizes;
168 Chapter 7 Conclusion and Future Work

which are common to most of the sheet cutting industries. However, this effort can be
further extended by being biased towards the organization-specific constraints. For ex-
ample, manufacturers who follow philosophies such as lean manufacturing would impose
different constraints to the storage and work-in-progress stocks of sheet material and
residuals so that the overall multi-period model can be changed.

As another future area of research, it would be interesting to consider other variants

of multi-period irregular bin packing problems since it has potential applications in
manufacturing industries. We considered particularly the variant where demand cut
quantities of a future period are unknown. This allowed us to solve irregular bin packing
problem individually for each period since the future qualities to be cut is unknown.
However, there are situations where manufacturers know the future cut quantities in
advance or at least predicts order demand for future time periods. In such cases, the
solutions have to be searched considering the optimisation throughout the time span
rather than optimising within individual periods.

7.3 Summary

This thesis has studied the irregular shape bin packing problem for homogeneous and
heterogeneous bin types. Several algorithms have been designed and tested. These have
produced competitive results and as a result, we have been able to draw insight about
the problem.
Appendix A

Period wise Results for Model 01

Table A.1: Multi period cutting with residual use (starting with Single standard
2.25m bins)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.6015 0.6080 0.6010 0.5965
2 0.6145 0.6170 0.6002 0.5958
3 0.6399 0.6453 0.6082 0.5934
4 0.6374 0.6461 0.6288 0.5946
5 0.6545 0.6530 0.6317 0.5929
6 0.6558 0.6535 0.6332 0.5913
7 0.6602 0.6580 0.6409 0.5921
8 0.6655 0.6664 0.6456 0.5927
9 0.6682 0.6710 0.6496 0.5930
10 0.6685 0.6716 0.6474 0.5922
11 0.6718 0.6732 0.6475 0.5934
12 0.6721 0.6738 0.6451 0.5930
13 0.6745 0.6756 0.6500 0.5936
14 0.6723 0.6745 0.6482 0.5929
15 0.6720 0.6753 0.6482 0.5926
16 0.6739 0.6759 0.6479 0.5917

169
170 Appendix A Period wise Results for Model 01

Table A.2: Multi period cutting with residual use (starting with Single standard
2.00m bins)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.5972 0.5925 0.5944 0.5907
2 0.6414 0.6428 0.6406 0.5925
3 0.6442 0.6560 0.6517 0.5942
4 0.6543 0.6592 0.6590 0.5947
5 0.6590 0.6658 0.6577 0.5970
6 0.6676 0.6648 0.6610 0.5970
7 0.6666 0.6700 0.6640 0.5978
8 0.6702 0.6719 0.6635 0.5983
9 0.6753 0.6723 0.6665 0.5993
10 0.6759 0.6730 0.6656 0.5982
11 0.6738 0.6731 0.6652 0.5968
12 0.6756 0.6739 0.6656 0.5970
13 0.6781 0.6781 0.6671 0.5966
14 0.6786 0.6767 0.6677 0.5969
15 0.6800 0.6771 0.6669 0.5971
16 0.6816 0.6786 0.6670 0.5970

Table A.3: Multi period cutting with residual use (starting with Single standard
1.75m bins)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.6109 0.6145 0.6092 0.6115
2 0.6463 0.6389 0.6381 0.6145
3 0.6531 0.6492 0.6527 0.6108
4 0.6520 0.6548 0.6553 0.6102
5 0.6612 0.6622 0.6551 0.6109
6 0.6605 0.6655 0.6571 0.6110
7 0.6622 0.6654 0.6622 0.6133
8 0.6680 0.6684 0.6638 0.6164
9 0.6698 0.6689 0.6636 0.6169
10 0.6713 0.6683 0.6611 0.6152
11 0.6726 0.6696 0.6633 0.6139
12 0.6725 0.6706 0.6636 0.6136
13 0.6733 0.6721 0.6646 0.6128
14 0.6735 0.6730 0.6640 0.6128
15 0.6725 0.6741 0.6620 0.6131
16 0.6734 0.6739 0.6624 0.6122
Appendix A Period wise Results for Model 01 171

Table A.4: Multi period cutting with residual use (starting with Single standard
1.50m bins)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.6231 0.6248 0.6281 0.6257
2 0.6408 0.6527 0.6552 0.6214
3 0.6473 0.6553 0.6580 0.6225
4 0.6584 0.6594 0.6611 0.6214
5 0.6609 0.6586 0.6646 0.6216
6 0.6616 0.6608 0.6643 0.6209
7 0.6624 0.6651 0.6657 0.6220
8 0.6638 0.6666 0.6668 0.6227
9 0.6646 0.6668 0.6660 0.6235
10 0.6643 0.6662 0.6670 0.6219
11 0.6636 0.6674 0.6676 0.6213
12 0.6653 0.6676 0.6674 0.6207
13 0.6669 0.6676 0.6677 0.6206
14 0.6684 0.6689 0.6679 0.6207
15 0.6687 0.6698 0.6681 0.6209
16 0.6687 0.6703 0.6682 0.6211

Table A.5: Multi period cutting with residual use (starting with Single standard
1.25m bins)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.5917 0.5891 0.5936 0.5941
2 0.6081 0.5906 0.5920 0.5977
3 0.6111 0.6023 0.5922 0.5970
4 0.6123 0.6025 0.5921 0.5980
5 0.6158 0.6058 0.5934 0.5979
6 0.6178 0.6044 0.5970 0.5984
7 0.6197 0.6038 0.5954 0.5992
8 0.6216 0.6047 0.5981 0.5996
9 0.6217 0.6030 0.5972 0.5994
10 0.6225 0.6038 0.5976 0.5989
11 0.6231 0.6040 0.5989 0.5987
12 0.6241 0.6046 0.5991 0.5993
13 0.6243 0.6052 0.5998 0.5994
14 0.6255 0.6049 0.6008 0.5990
15 0.6255 0.6056 0.6010 0.5989
16 0.6257 0.6059 0.6010 0.5986
172 Appendix A Period wise Results for Model 01

Table A.6: Multi period cutting with residual use (starting with Mix of standard
sizes)

Ut
Period
Jmax = 32 Jmax = 16 Jmax = 4 Jmax =0
1 0.6539 0.6516 0.6507 0.6557
2 0.6608 0.6617 0.6596 0.6517
3 0.6626 0.6671 0.6656 0.6584
4 0.6639 0.6625 0.6690 0.6567
5 0.6642 0.6650 0.6691 0.6572
6 0.6669 0.6681 0.6682 0.6563
7 0.6669 0.6670 0.6676 0.6543
8 0.6684 0.6697 0.6682 0.6540
9 0.6707 0.6707 0.6696 0.6543
10 0.6704 0.6712 0.6692 0.6541
11 0.6686 0.6700 0.6656 0.6503
12 0.6703 0.6710 0.6661 0.6503
13 0.6703 0.6713 0.6667 0.6503
14 0.6714 0.6721 0.6668 0.6500
15 0.6722 0.6722 0.6667 0.6499
16 0.6721 0.6721 0.6667 0.6499
References

Abuabara, A. and Morabito, R. (2009). Cutting optimization of structural tubes to

build agricultural light aircrafts. Annals of Operations Research, 169(1):149–165.

Adamowicz, M. and Albano, A. (1972). A two-stage solution of the cutting stock prob-
lem. Information Processing, 71:1086–1091.

Adamowicz, M. and Albano, A. (1976). Nesting two-dimensional shapes in rectangular

modules. Computer Aided Design, 8:27–33.

Agarwal, P., Flato, E., and Halperin, D. (2002). Polygon decomposition for efficient
construction of minkowski sums. Computational Geometry Theory and Applications,
21:39–61.

Albano, A. and Sapuppo, G. (1980). Optimal allocation of two-dimensional irregular

shapes using heuristic search methods. IEEE Transactions on Systems, Man and
Cybernetics, 10(5):242–248.

Alvarez-Valdes, R., Martinez, A., and Tamarit, J. M. (2013). A branch & bound al-
gorithm for cutting and packing irregularly shaped pieces. International Journal of
Production Economics, 145(2):463–477.

Andrade, R., Birgin, E. G., and Morabito, R. (2016). Two-stage two-dimensional guil-
lotine cutting stock problems with usable leftover. International Transactions in Op-
erational Research, 23(1-2):121–145.

Andrade, R., Birgin, E. G., Morabito, R., and Ronconi, D. P. (2014). Mixed integer
programming models for two-dimensional non-guillotine cutting problems with usable
leftovers. Journal of the Operational Research Society, 65(11):1649–1663.

Arbib, C. and Marinelli, F. (2005). Integrating process optimization and inventory plan-
ning in cutting-stock with skiving option: An optimization model and its application.
European Journal of Operational Research, 163(3):617–630.

Art, R. C. (1966). An approach to the two-dimensional irregular cutting stock problem.

Technical Report Report 36-Y08, IBM Cambridge Scientific Centre.

173
174 REFERENCES

Babu, A. R. and Babu, N. R. (1999). Effective nesting of rectangular parts in multiple

rectangular sheets using genetic and heuristic algorithms. International Journal of
Production Research, 37(7):1625–1643.

Babu, A. R. and Babu, N. R. (2001). A generic approach for nesting of 2-D parts in 2-D
sheets using genetic and heuristic algorithms. Computer-Aided Design, 33(12):879–
891.

Balaprakash, P., Birattari, M., and Stützle, T. (2007). Improvement Strategies for the
F-Race Algorithm: Sampling Design and Iterative Refinement, volume 4771 of Lecture
Notes in Computer Science, pages 108–122. Springer.

Baldacci, R., Boschetti, M. A., Ganovelli, M., and Maniezzo, V. (2014). Algorithms for
nesting with defects. Discrete Applied Mathematics, 163, Part 1:17–33.

Baum, E. (1986). Iterated descent: A better algorithm for local search in combinatorial
optimization problems. Technical report, Caltech, Pasadena, CA.

Baum, E. B. (1987). Towards practical ‘neural’ computation for combinatorial optimiza-

tion problems. In AIP Conference Proceedings 151 on Neural Networks for Computing,
pages 53–58, Woodbury, NY, USA. American Institute of Physics Inc.

Bennell, J. A. (1998). Incorporating problem specific knowledge into a local search

framework for the irregular shape packing problem. PhD thesis, University of Wales,
Swansea.

Bennell, J. A. and Dowsland, K. A. (1999). A tabu thresholding implementation for

the irregular stock cutting problem. International Journal of Production Research,
37(18):4259–4275.

Bennell, J. A. and Dowsland, K. A. (2001). Hybridising tabu search with optimisation

techniques for irregular stock cutting. Management Science, 47(8):1160–1172.

Bennell, J. A., Dowsland, K. A., and Dowsland, W. B. (2001). The irregular cutting-
stock problem : a new procedure for deriving the no-fit polygon. Computers & Oper-
ations Research, 28(3):271–287.

Bennell, J. A. and Oliveira, J. F. (2008). The geometry of nesting problems: A tutorial.

European Journal of Operational Research, 184(2):397–415. 00128.

Bennell, J. A. and Oliveira, J. F. (2009). A tutorial in irregular shape packing problems.

The Journal of the Operational Research Society, 60:s93–s105.

Bennell, J. A., Scheithauer, G., Stoyan, Y., and Romanova, T. (2010). Tools of math-
ematical modelling of arbitrary object packing problems. Annals of Operations Re-
search, 179(1):343–368.
REFERENCES 175

Bennell, J. A. and Song, X. (2008). A comprehensive and robust procedure for ob-
taining the nofit polygon using Minkowski sums. Computers & Operations Research,
35(1):267–281.

Bennell, J. A. and Song, X. (2010). A beam search implementation for the irregular
shape packing problem. Journal of Heuristics, 16(2):167–188. 00037.

Bertrand, J. W. M. and Fransoo, J. C. (2002). Operations management research method-

ologies using quantitative modeling. International Journal of Operations & Production
Management, 22(2):241–264.

Birattari, M. (2009). Tuning Metaheuristics: A Machine Learning Perspective. Springer

Science & Business Media.

Birattari, M., Stützle, T., Paquete, L., and Varrentrapp, K. (2002). A Racing Algo-
rithm for Configuring Metaheuristics. In Proceedings of the Genetic and Evolutionary
Computation Conference, GECCO 02, pages 11–18, San Francisco, CA, USA. Morgan
Kaufmann Publishers Inc.

Birattari, M., Yuan, Z., Balaprakash, P., and Stützle, T. (2010). F-Race and Iterated
F-Race: An Overview. In Experimental Methods for the Analysis of Optimization
Algorithms, pages 311–336. Springer, Berlin, Heidelberg.

Blażewicz, J., Hawryluk, P., and Walkowiak, R. (1993). Using a tabu search approach for
solving the two-dimensional irregular cutting problem. Annals of Operations Research,
41(4):313–325. 00109.

Bohme, D. and Graham, A. (1979). Practical experiences with semiautomatic and

automatic part nesting methods. Computer Applications in Automation of Shipyard
Operations and Ship Design III, pages 213–220.

Bounsaythip, C. and Maouche, S. (1997). Irregular shape nesting and placing with
evolutionary approach. In Computational Cybernetics and Simulation 1997 IEEE
International Conference on Systems, Man, and Cybernetics, volume 4, pages 3425–
3430.

Brooks, R. L., Smith, C. A. B., Stone, A. H., and Tutte, W. T. (1940). The dissection
of rectangles into squares. Duke mathematical Journal, 7:312–340.

Brown, A. R. (1971). Optimum packing and depletion: The computer in space - and
resource-usage problems. American Elsevier, London, New York.

Burke, E., Hellier, R. S. R., Kendall, G., and Whitwell, G. (2006). A new bottom-left-
fill heuristic algorithm for the two-dimensional irregular packing problem. Operations
Research, 54(3):587–601.
176 REFERENCES

Burke, E. K., Gendreau, M., Hyde, M., Kendall, G., Ochoa, G., Özcan, E., and Qu, R.
(2013). Hyper-heuristics: a survey of the state of the art. Journal of the Operational
Research Society, 64(12):1695–1724.

Burke, E. K., Hellier, R. S. R., Kendall, G., and Whitwell, G. (2007). Complete and
robust no-fit polygon generation for the irregular stock cutting problem. European
Journal of Operational Research, 179(1):27–49.

Burke, E. K., Hellier, R. S. R., Kendall, G., and Whitwell, G. (2010). Irregular packing
using the line and arc no-fit polygon. Operations Research, 58(4-Part-1):948–970.

Carravilla, M. A., Ribeiro, C., and Oliveira, J. F. (2003). Solving nesting problems with
non-convex polygons by constraint logic programming. International Transactions in
Operational Research, 10(6):651–663.

Charalambous, C. and Fleszar, K. (2011). A constructive bin-oriented heuristic for the

two-dimensional bin packing problem with guillotine cuts. Computers and Operational
Research, 38:1443–1451.

Chernov, N., Stoyan, Y., Romanova, T., Pankratov, A., Chernov, N., Stoyan, Y., Ro-
manova, T., and Pankratov, A. (2012). Phi-Functions for 2D Objects Formed by
Line Segments and Circular Arcs, Phi-Functions for 2D Objects Formed by Line Seg-
ments and Circular Arcs. Advances in Operations Research, Advances in Operations
Research, 2012, 2012:e346358.

Cherri, A. C., Arenales, M. N., and Yanasse, H. H. (2009). The one-dimensional cut-
ting stock problem with usable leftover—a heuristic approach. European Journal of
Operational Research, 196(3):897–908.

Cherri, A. C., Arenales, M. N., and Yanasse, H. H. (2013). The usable leftover one-
dimensional cutting stock problem—a priority-in-use heuristic. International Trans-
actions in Operational Research, 20(2):189–199.

Cherri, A. C., Arenales, M. N., Yanasse, H. H., Poldi, K. C., and Gonçalves Vianna,
A. C. (2014a). The one-dimensional cutting stock problem with usable leftovers – A
survey. European Journal of Operational Research, 236(2):395–402.

Cherri, A. C., Arenales, M. N., Yanasse, H. H., Poldi, K. C., and Gonçalves Vianna,
A. C. (2014b). The one-dimensional cutting stock problem with usable leftovers – A
survey. European Journal of Operational Research, 236(2):395–402.

Cherri, L. H., Mundim, L. R., Andretta, M., Toledo, F. M. B., Oliveira, J. F., and Car-
ravilla, M. A. (2016). Robust mixed-integer linear programming models for the irregu-
lar strip packing problem. European Journal of Operational Research, 253(3):570–583.

Cui, Y., Song, X., Chen, Y., and Cui, Y.-P. (2016). New model and heuristic solution
approach for one-dimensional cutting stock problem with usable leftovers:model and
REFERENCES 177

heuristic for 1D cutting stock problem with leftovers. Journal of the Operational
Research Society.

Cui, Y., Yang, L., Zhao, Z., Tang, T., and Yin, M. (2013). Sequential grouping heuristic
for the two-dimensional cutting stock problem with pattern reduction. International
Journal of Production Economics, 144(2):432–439.

Cui, Y. and Yang, Y. (2010). A heuristic for the one-dimensional cutting stock problem
with usable leftover. European Journal of Operational Research, 204(2):245–250.

Cuninghame-Green, R. (1989). Geometry, shoemaking and the milk tray problem |new
scientist.

Dagli, C. H. and Hajakbari, A. (1990). Simulated annealing approach for solving stock
cutting problem. In 1990 IEEE International Conference on Systems, Man, and
Cybernetics Conference Proceedings, pages 221–223.

Dagli, C. H. and Tatoglu, M. Y. (1987). An approach to two-dimensional cutting stock

problems. International Journal of Production Research, 25(2):175–190.

Dighe, R. and Jakiela, M. J. (1995). Solving Pattern Nesting Problems with Genetic
Algorithms Employing Task Decomposition and Contact Detection. Evolutionary
Computation, 3(3):239–266.

Dowsland, K. A. and Dowsland, W. B. (1995). Solution approaches to irregular nesting

problems. European Journal of Operational Research, 84(3):506–521.

Dowsland, K. A., Dowsland, W. B., and Bennell, J. A. (1998). Jostling for position:
local improvement for irregular cutting patterns. Journal of the Operational Research
Society, 49(6):647–658. 00051.

Dowsland, K. A., Vaid, S., and Dowsland, W. B. (2002). An algorithm for polygon
placement using a bottom-left strategy. European Journal of Operational Research,
141(2):371–381.

Dréo, J. (2006). Metaheuristics for Hard Optimization: Methods and Case Studies.
Springer Science & Business Media.

Dyckhoff, H. (1990). A typology of cutting and packing problems. European Journal of

Operational Research, 44(2):145–159.

Egeblad, J., Nielsen, B. K., and Odgaard, A. (2007). Fast neighborhood search for two-
and three-dimensional nesting problems. European Journal of Operational Research,
183(3):1249–1266. 00097.

Eiben, A. E. and Smith, J. E. (2013). Introduction to Evolutionary Computing. Springer

Science & Business Media.
178 REFERENCES

Elkeran, A. (2013). A new approach for sheet nesting problem using guided cuckoo search
and pairwise clustering. European Journal of Operational Research, 231(3):757–769.
00019.

Erjavec, J., Gradišar, M., and Trkman, P. (2012). Assessment of stock size to mini-
mize cutting stock production costs. International Journal of Production Economics,
135(1):170–176.

Falkenauer, E. (1999). The worth of the uniform [uniform crossover]. In Proceedings

of the 1999 Congress on Evolutionary Computation-CEC99 (Cat. No. 99TH8406),
volume 1, page 782 Vol. 1.

Feo, T. A. and Resende, M. G. C. (1989). A Probabilistic Heuristic for a Computationally

Difficult Set Covering Problem. Operational Research Letters, 8(2):67–81.

Fischetti, M. and Luzzi, I. (2008). Mixed-integer programming models for nesting prob-
lems. Journal of Heuristics, 15(3):201–226.

Fleischmann, M., Bloemhof-Ruwaard, J. M., Dekker, R., van der Laan, E., van Nunen,
J. A. E. E., and Van Wassenhove, L. N. (1997). Quantitative models for reverse
logistics: A review. European Journal of Operational Research, 103(1):1–17.

Foerster, H. and Wäscher, G. (2000). Pattern reduction in one-dimensional cutting stock

problems. International Journal of Production Research, 38(7):1657–1676.

Freeman, H. and Shapira, R. (1975). Determining the Minimum-area Encasing Rectangle

for an Arbitrary Closed Curve. Commun. ACM, 18(7):409–413.

French, M. L. and LaForge, R. L. (2006). Closed-loop supply chains in process industries:

An empirical study of producer re-use issues. Journal of Operations Management,
24(3):271–286.

Friesen, D. K. and Langston, M. A. (1986). Variable-sized bin packing. Computers &

Operations Research, 15(1):222–230.

Fujita, K., Akagi, S., and Hirokawa, N. (1993). Hybrid approach for optimal nesting
using a genetic algorithm and a local minimization algorithm. In in Proc. of the 19th
Annual ASME Design Automation Conference, part 1, pages 477–484.

Gendreau, M. and Potvin, J., editors (2010). Handbook of Metaheuristics. Springer,

New York, 2nd edition.

Ghosh, P. (1991a). An algebra of polygons through the notion of negative shapes.

CVGIP: Image Understanding, 54(1):119–144.

Ghosh, P. K. (1991b). An algebra of polygons through the notion of negative shapes.

CVGIP: Image Understanding, 54(1):119–144.
REFERENCES 179

Glover, F. (1986). Future paths for integer programming and links to artificial intelli-
gence. Computers & Operations Research, 13(5):533–549.

Gomes, A. M. (2013). Irregular Packing Problems: Industrial Applications and New

Directions Using Computational Geometry. In 11th IFAC Workshop on Intelligent
Manufacturing Systems, pages 378–383.

Gomes, A. M. and Oliveira, J. F. (2001). Automatic marker making. In 4th Metaheuris-

tics International Conference, pages 47–52.

Gomes, A. M. and Oliveira, J. F. (2002). A 2-exchange heuristic for nesting problems.

European Journal of Operational Research, 141(2):359–370.

Gomes, A. M. and Oliveira, J. F. (2006). Solving irregular strip packing problems

by hybridising simulated annealing and linear programming. European Journal of
Operational Research, 171(3):811–829.

Gonçalves, J. F. and Resende, M. G. C. (2013). A biased random key genetic algorithm

for 2D and 3D bin packing problems. International Journal of Production Economics,
145:500–510.

Gonzalez, T. F. (2007). Handbook of Approximation Algorithms and Metaheuristics.

CRC Press.

Gradišar, M., Jesenko, J., and Resinovič, G. (1997). Optimization of roll cutting in
clothing industry. Computers & Operations Research, 24(10):945–953.

Gradišar, M., Kljajić, M., Resinovič, G., and Jesenko, J. (1999a). A sequential heuristic
procedure for one-dimensional cutting. European Journal of Operational Research,
114(3):557–568.

Gradišar, M., Resinovič, G., and Kljajić, M. (1999b). A hybrid approach for optimization
of one-dimensional cutting. European Journal of Operational Research, 119(3):719–
728.

Gradisar, M. and Trkman, P. (2005). A combined approach to the solution to the

general one-dimensional cutting stock problem. Computers & Operations Research,
32(7):1793–1807.

Grinde, R. B. and Cavalier, T. M. (1995). A new algorithm for the minimal-area convex
enclosure problem. European Journal of Operational Research, 84(3):522–538.

Guide, J. V. D. R. (2000). Production planning and control for remanufacturing: indus-

try practice and research needs. Journal of Operations Management, 18(4):467–483.

Guo, B., Peng, Q., Cheng, X., and Dai, N. (2015). Free-form contour packing based on
material grid approximation and lowest-gravity-center methods. Expert Systems with
Applications, 42(4):1864–1871.
180 REFERENCES

Han, W., Bennell, J. A., Zhao, X., and Song, X. (2013). Construction heuristics for two-
dimensional irregular shape bin packing with guillotine constraints. European Journal
of Operational Research, 230(3):495–504.

Heckmann, R. and Lengauer, T. (1995). A simulated annealing approach to the nest-

ing problem in the textile manufacturing industry. Annals of Operations Research,
57(1):103–133.

Henn, S. and Wäscher, G. (2013). Extensions of cutting problems: setups. Pesquisa

Operacional, 33(2):133–162.

Hifi, M. and M’Hallah, R. (2003). A hybrid algorithm for the two-dimensional layout
problem: the cases of regular and irregular shapes. International Transactions in
Operational Research, 10(3):195–216.

Holland, J. H. (1992). Adaptation in Natural and Artificial Systems: An Introductory

Analysis with Applications to Biology, Control and Artificial Intelligence. MIT Press,
Cambridge, Mass, new ed edition edition.

Hong, S., Zhang, D., Lau, H. C., Zeng, X., and Si, Y.-W. (2014). A hybrid heuristic algo-
rithm for the 2d variable-sized bin packing problem. European Journal of Operational
Research, 238(1):95–103.

Hopper, E. (2000). Algorithms and other Meta-Heuristic Methods. PhD thesis, Univer-
sity of Wales, UK.

Hopper, E. and Turton, B. C. H. (2001). A Review of the Application of Meta-Heuristic

Algorithms to 2D Strip Packing Problems. Artificial Intelligence Review, 16(4):257–
300.

Hu, Y., Hashimoto, H., Imahori, S., and Yagiura, M. (2015). Efficient implementations
of construction heuristics for the rectilinear block packing problem. Computers &
Operations Research, 53:206–222.

Hutter, F., Hoos, H. H., Leyton-Brown, K., and Stützle, T. (2009). Paramils: An
Automatic Algorithm Configuration Framework. Journal of Artificial Intelligence
Research, 36:267–306.

Huyao, L., Yuanjun, H., and Bennell, J. A. (2007). The Irregular Nesting Problem: A
New Approach for Nofit Polygon Calculation. The Journal of the Operational Research
Society, 58(9):1235–1245.

Imamichi, T., Yagiura, M., and Nagamochi, H. (2009). An iterated local search algorithm
based on nonlinear programming for the irregular strip packing problem. Discrete
Optimization, 6(4):345–361.

Jakobs, S. (1996). On genetic algorithms for the packing of polygons. European Journal
of Operational Research, 88(1):165–181.
REFERENCES 181

Junior, B. A., Pinheiro, P. R., and Saraiva, R. D. (2013). A Hybrid Methodology

for Nesting Irregular Shapes: Case Study on a Textile Industry. IFAC Proceedings
Volumes, 46(24):15–20.

Kallrath, J. (2008). Cutting circles and polygons from area-minimizing rectangles. Jour-
nal of Global Optimization, 43(2-3):299–328.

Kantorovich, L. V. (1960). Mathematical methods of organizing and planning produc-

tion. Journal of Management Science, 6:366–422.

Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983). Optimization by simulated

annealing. SCIENCE, 220(4598):671–680.

Koch, S., König, S., and Wäscher, G. (2009). Integer linear programming for a cutting
problem in the wood-processing industry: a case study. International Transactions in
Operational Research, 16(6):715–726.

Konopasek, M. (1981). Mathematical treatments of some apparel marking and cutting

problems. U.S. Department of Commerce Report, 99-26-90857-10.

Leao, A. A. S., Toledo, F. M. B., Oliveira, J. F., and Carravilla, M. A. (2016). A semi-
continuous MIP model for the irregular strip packing problem. International Journal
of Production Research, 54(3):712–721.

Leung, S. C. H., Lin, Y., and Zhang, D. (2012). Extended local search algorithm based
on nonlinear programming for two-dimensional irregular strip packing problem. Com-
puters & Operations Research, 39(3):678–686.

Li, Z. and Milenkovic, V. (1995). Compaction and separation algorithms for non-convex
polygons and their application. European Journal of Operations Research, 84:539–561.

Liu, D. and Teng, H. (1999). An improved bl-algorithm for genetic algorithm of the or-
thogonal packing of rectangles. European Journal of Operational Research, 112(2):413–
420.

Liu, H. and He, Y. (2006). Algorithm for 2D irregular-shaped nesting problem based on
the NFp algorithm and lowest-gravity-center principle. Journal of Zhejiang University
SCIENCE A, 7(4):570–576.

Lopez-Camacho, E., Ochoa, G., Terashima-Marı́n, H., and Burke, Edmund, K. (2013a).
An effective heuristic for the two-dimensional irregular bin packing problem. Annals
of Operations Research, 206(1):241–264.

Lopez-Camacho, E., Terashima-Marı́n, H., Ochoa, G., and Conant-Pablos, S. E. (2013b).

Understanding the structure of bin packing problems through principal component
analysis. International Journal of Production Economics, 145(1):488–499.
182 REFERENCES

Lopez-Camacho, E., Terashima-Marin, H., Ross, P., and Ochoa, G. (2014). A unified
hyper-heuristic framework for solving bin packing problems. Expert Systems with
Applications, 41(15):6876–6889.

López-Ibáñez, M., Dubois-Lacoste, J., Cáceres, L. P., Birattari, M., and Stützle, T.
(2016). The irace package: Iterated racing for automatic algorithm configuration.
Operations Research Perspectives, 3:43–58.

Lourenço, H. R., Martin, O. C., and Stützle, T. (2010). Iterated Local Search: Frame-
work and Applications. In Gendreau, M. and Potvin, J., editors, Handbook of Meta-
heuristics, number 146 in International Series in Operations Research & Management
Science, pages 363–397. Springer US. DOI: 10.1007/978-1-4419-1665-5 12.

Luke, S. (2013). Essentials of Metaheuristics. lulu.com.

Mahadevan, A. (1984). Optimization in Computer-aided Pattern Packing (Marking,

Envelopes). PhD thesis, North Carolina State University.

Martin, R. R. and Stephenson, P. C. (1988). Putting objects into boxes. Computer-Aided

Design, 20(9):506–514.

Martinez-Sykora, A. (2013). Nesting Problems: Exact and Heuristic Algorithms. PhD

thesis, University of Valencia, Spain.

Martinez-Sykora, A., Alvarez-Valdes, R., Bennell, J., and Tamarit, J. M. (2015). Con-
structive procedures to solve 2-dimensional bin packing problems with irregular pieces
and guillotine cuts. Omega, 52:15–32.

Martinez-Sykora, A., Alvarez-Valdes, R., Bennell, J. A., Ruiz, R., and Tamarit, J. M.
(2017). Matheuristics for the irregular bin packing problem with free rotations. Eu-
ropean Journal of Operational Research, 258(2):440–455.

Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., and Teller, E.
(1953). Equation of state calculations by fast computing machines. Journal of Chem-
ical Physics, 21(6):1087–1092.

Milenkovic, V., Daniels, K., and Li, Z. (1991). Automatic marker making. In Third
Canadian Conference on Computational Geometry, pages 243–246.

Minner, S. (2001). Strategic safety stocks in reverse logistics supply chains. International
Journal of Production Economics, 71:417–428.

Minner, S. (2003). Multiple-supplier inventory models in supply chain management: A

review. International Journal of Production Economics, 81-82:265–279.

Mladenović, N. and Hansen, P. (1997). Variable neighborhood search. Computers &

Operations Research, 24(11):1097–1100.
REFERENCES 183

Morabito, R. and Belluzzo, L. (2007). Optimising the cutting of wood fibre plates in the
hardboard industry. European Journal of Operational Research, 183(3):1405–1420.

Mutingi, M. and Mbohwa, C. (2016). Grouping Genetic Algorithms: Advances and

Applications. Springer.

Oliveira, J. F. and Ferreira, J. A. S. (1993). Algorithms for Nesting Problems, pages

255–273. Lecture Notes in Economics and Mathematical Systems. Springer Berlin
Heidelberg.

Oliveira, J. F., Gomes, A. M., and Ferreira, J. S. (2000). TOPOS-A new constructive
algorithm for nesting problems. OR-Spektrum, 22(2):263–284.

Ortmann, F. G., Ntene, N., and van Vuuren, J. H. (2010). New and improved level
heuristics for the rectangular strip packing and variable-sized bin packing problems.
European Journal of Operational Research, 203(2):306–315.

Peng-Yeng, Y. (2012). Modeling, Analysis, and Applications in Metaheuristic Comput-

ing: Advancements and Trends. IGI Global.

Pisinger, D. and Sigurd, M. (2005). The two-dimensional bin packing problem with
variable bin sizes and costs. Discrete Optimization, 2(2):154–167.

Qu, W. and Sanders, J. L. (1987). A nesting algorithm for irregular parts and factors
affecting trim losses. International Journal of Production Research, 25(3):381–397.

Resende, M. G. C. and Ribeiro, C. C. (2010). Greedy Randomized Adaptive Search Pro-

cedures. In Handbook of Metaheuristics, International Series in Operations Research
& Management Science, pages 219–249. Springer, Boston, MA.

Resende, M. G. C. and Ribeiro, C. C. (2016). GRASP: The basic heuristic. In Opti-

mization by GRASP, pages 95–112. Springer New York.

Ribeiro, C., Carravilla, M. A., and Oliveira, J. (1999). Applying constraint logic pro-
gramming to the resolution of nesting problems. Pesquisa Operacional, 19(2):239–247.

Rocha, P. (2014). Geometrical models and algorithms for irregular shapes placement
problems. PhD thesis, INESC,Faculty of Engineering, University of Porto, Portugal.

Rocha, P., Rodrigues, R., Toledo, F. M. B., and Gomes, A. M. (2013). Circle Covering
using Medial Axis. IFAC Proceedings Volumes, 46(7):402–407.

Rocha, P., Rui, R., Gomes, A. M., Andretta, M., and Toledo, F. M. B. (2014). Circle
covering representation for nesting problems with continuous rotations. In Proceedings
of the 19th World Congress. The International Federation of Automatic Control, pages
5235–5240, Cape Town, South Africa.

Roodman, G. M. (1986). Near-optimal solutions to one-dimensional cutting stock prob-

lems. Computers & Operations Research, 13(6):713–719.
184 REFERENCES

Scheithauer, G. (1991). A note on handling residual lengths. Optimization, 22(3):461–

466.

Segenreich, S. A. and Braga, L. M. P. F. (1986). Optimal nesting of general plane figures:

A monte carlo heuristical approach. Computer Graphics, 10(3):229–237.

Sinuany-Stern, Z. and Weiner, I. (1994). The one dimensional cutting stock problem
using two objectives. The Journal of the Operational Research Society, 45(2):231–236.

Song, X. and Bennell, J. A. (2014). Column generation and sequential heuristic proce-
dure for solving an irregular shape cutting stock problem. Journal of the Operational
Research Society, 65(7):1037–1052.

Stoyan, Y., Pankratov, A., and Romanova, T. (2015a). Cutting and packing problems for
irregular objects with continuous rotations: mathematical modelling and non-linear
optimization. Journal of the Operational Research Society.

Stoyan, Y., Pankratov, A., and Romanova, T. (2015b). Quasi-phi-functions and optimal
packing of ellipses. Journal of Global Optimization, pages 1–25.

Stoyan, Y., Scheithauer, G., Gil, N., and Romanova, T. (2001). Phi-functions for primary
2D-objects. Studia Informatica Universalis, 2:1–32.

Stoyan, Y., Scheithauer, G., Gil, N., and Romanova, T. (2004). Phi-functions for com-
plex 2D-objects. Quarterly Journal of the Belgian, French and Italian Operations
Research Societies, 2(1):69–84.

Sweeney, P. E. and Paternoster, E. R. (1992). Cutting and Packing Problems: A Catego-

rized, Application-Orientated Research Bibliography. The Journal of the Operational
Research Society, 43(7):691–706.

Takahara, S., Kusumoto, Y., and Miyamoto, S. (2003). Solution for textile nesting
problems using adaptive meta-heuristics and grouping. Soft Computing, 7(3):154–
159.

Talbi, E. (2009). Metaheuristics: From Design to Implementation. John Wiley & Sons,
Inc.

Terashima-Marı́n, H., Ross, P., Farı́as-Zárate, C. J., López-Camacho, E., and

Valenzuela-Rendón, M. (2010). Generalized hyper-heuristics for solving 2D regular
and irregular packing problems. Annals of Operations Research, 179(1):369–392.

Toledo, F. M. B., Carravilla, M. A., Ribeiro, C., Oliveira, J. F., and Gomes, A. M.
(2013). The dotted-board model: A new MIP model for nesting irregular shapes.
International Journal of Production Economics, 145(2):478–487.

Trkman, P. and Gradisar, M. (2007). One-dimensional cutting stock optimization in

consecutive time periods. European Journal of Operational Research, 179(2):291–301.
REFERENCES 185

Umetani, S., Yagiura, M., Imahori, S., Imamichi, T., Nonobe, K., and Ibaraki, T. (2009).
Solving the irregular strip packing problem via guided local search for overlap mini-
mization. International Transactions in Operational Research, 16(6):661–683.

Venkateswarlu, P. (2001). The trim-loss problem in a wooden container manufacturing

company. Journal of Manufacturing Systems, 20(3):166–176.

Wang, F. K. and Liu, F. T. (2014). A new decision model for reducing trim loss and
inventory in the paper industry. Journal of Applied Mathematics, 2014(Article ID
987054).

Wäscher, G. (2012). Cutting and packing typology: part1, part2, part3. EURO Summer
Institute on Cutting and Packing (ESI-CP), Porto, Portugal.

Wäscher, G., Haußner, H., and Schumann, H. (2007). An improved typology of cutting
and packing problems. European Journal of Operational Research, 183(3):1109–1130.

Watson, P. D. and Tobias, A. (1999). An efficient algorithm for the regular w1 packing of
polygons in the infinite plane. Journal of the Operational Research Society, 50:1054–
1062.

Xu, Y. (2016). An Efficient Heuristic Approach for Irregular Cutting Stock Problem in
Ship Building Industry. Mathematical Problems in Engineering, 2016:e8703782.

Xu, Y., Yang, G., and Pan, C.-c. (2013). An Approach Based on Evaluation Particle
Swarm Optimization Algorithm for 2D Irregular Cutting Stock Problem. In Tan, Y.,
Shi, Y., and Mo, H., editors, Advances in Swarm Intelligence, number 7928 in Lecture
Notes in Computer Science, pages 168–175. Springer Berlin Heidelberg.