THE HONG KONG POLYTECHNIC UNIVERSITY
DEPARTMENT OF INDUSTRIAL AND SYSTEMS
ENGINEERING

A SUBOPTIMUM- AND PROPORTION-BASED
HEURISTIC GENERATION METHOD FOR
COMBINATORIAL OPTIMIZATION
PROBLEMS

Fan XUE

A Thesis Submitted in Partial Fulfilment of
the Requirements for the Degree of
Doctor of Philosophy

September 2012

CERTIFICATE OF ORIGINALITY

I hereby declare that this thesis is my own work and that, to the best of
my knowledge and belief, it reproduces no material previously published
or written, nor material that has been accepted for the award of any other
degree or diploma, except where due acknowledgement has been made in
the text.

XUE Fan

ii

Abstract

Automated heuristic selection and heuristic generation have increasingly attracted attention in solving combinatorial optimization problems
emerging from both theory and practice. This thesis presents a heuristic
generation algorithm, called Suboptimum- and Proportion-based On-thefly Training (SPOT), which can enhance existing heuristics with the aid of
instance-specific information.
By making use of the proposed “sample-learn-generate” framework,
SPOT samples small-scale subproblems, initially. Then, it collects the instancespecific information from the suboptima of the subproblems by the means of
machine learning. Lastly, it generates new heuristics by modifying existing
heuristics and data structures.
In the development of SPOT, two standards were incorporated to regulate
the problem input and the machine learning data. The software implementation was done in Java, with two external development libraries, the HyFlex
and the Weka. In terms of testing, two well-known NP-Complete combinatorial optimization problem domains were employed: the Traveling Salesman
Problem (TSP) and the permutation Flow-Shop scheduling Problem (FSP).
Each generated heuristic was tested with the TSP and the FSP domains.
iii

iv
To verify the result of using SPOT, one of the winners of the international
hyper-heuristic competition CHeSC 2011, named PHunter, was tested with
the generated heuristics by SPOT. In the TSP, adapting SPOT gave little
betterment, but in FSP, the improvements were significant. It increased
the overall score of the PHunter from 20.5 to 43 (out of 50). Indeed, it also
outperformed the best records in CHeSC 2011: 32 by AdaptHH, 29.5 by ML
and 26 by VNS-TW.

Publications

Journal Papers
Published
1. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2011). A learningbased variable assignment weighting scheme for heuristic and exact
searching in Euclidean traveling salesman problems. NETNOMICS:
Economic Research and Electronic Networking, 12, 183õ207.
doi:10.1007/s11066-011-9064-7
2. Chan, C. Y., Xue, F., Ip, W. H., & Cheung, C. F. (2012) A Hyper-heuristic
inspired by Pearl Hunting. In Y. Hamadi & M. Schoenauer (Eds.),
Learning and intelligent optimization (pp. 349–353). Lecture Notes in
Computer Science. Springer-Verlag. doi: 10.1007/978-3-642-34413-8 26

Ready for submission
1. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. A hybrid variable depth
search approach to personnel scheduling problems.

v

vi

Conference Presentations
1. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2010) Towards a learningbased heuristic searching reform scheme. 24th European Conference
on Operational Research (EURO2010), July 11–14, Lisbon, Portugal.
2. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2010) A learningbased variables assignment weighting scheme for heuristic and exact
searching. 1st International Conference on Computational Logistics
(ICCL2010), September 20–22, Shanghai, China.
3. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2011) Pearl Hunter: a
cross-domain hyper-heuristic that compiles iterated local search algorithms. The OR Society’s 53rd Annual Conference(OR53), September
6–8, Nottingham, United Kingdom.
4. Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2012) A Hyper-heuristic
inspired by Pearl Hunting. 6th Learning and Intelligent OptimizatioN
Conference (LION6), January 16–19, Paris, France.

Acknowledgements

My most sincere thanks go to my PhD supervisors and mentors, Dr.
Chan, Dr. Ip and Dr. Cheung, for supporting me in these past five years.
Dr. Chan is a very sincere man and devotes himself to both theoretical
and practical engineering. He gave me the freedom to pursue my interests
without restriction. I am grateful to Dr. Ip, too. He introduced me to this
research and inquired, discussed and enlightened me on the research very
often. Dr. Cheung is supportive on both research and daily life and also
provided many in-depth and insightful discussions about the research. I
hope that I could be as lively and liberal as Dr. Chan and be as enthusiastic
and energetic as Dr. Ip and Dr. Cheung someday. I also thank the members
of my PhD committee for their helpful questions and suggestions in general.
I would also like to thank Prof. Fan and Prof. Zhu at Civil Aviation
University of China. They inspired me in some related areas before the
research, and have been supporting me on my daily life up-to-date. I also
thank my colleagues Jackson Tang, Wang Lei, Qian Chen and Chen Qing for
many interesting and insightful discussions. I will forever be thankful to my
wife and my family for long-time and unconditional support, encouragement
and love.

vii

viii
I also thank the developers, contributors and maintainers of the algorithm
development libraries HyFlex and Weka, which provided stable and effective interfaces and functions on hyper-heuristic manipulation and stochastic
machine learning. I also appreciate the developers, contributors and maintainers of the open source projects Notepad++ and PSPP, which offered rich
and free functionality in the implementation and the data analysis of this
research.
The work described in this research project was substantially supported
by a grant from the Department of Industrial and Systems Engineering, The
Hong Kong Polytechnic University (Grant No. RP1Z).

Table of Contents

Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

i

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

v

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv
List of Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxiii
1

Introduction to the SPOT Algorithm . . . . . . . . . . . . . . . . .

1

1.1

Optimizations from Ancient to Modern Times . . . . . . . . .

2

1.2

Combinatorial Optimization . . . . . . . . . . . . . . . . . . . .

6

1.3

Research Motives and Background . . . . . . . . . . . . . . . . 15

1.4

Objectives and Achievements of Thesis . . . . . . . . . . . . . 24
1.4.1

Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.4.2

Achievements and contributions . . . . . . . . . . . . . 25

ix

x
1.5
2

Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1

2.2

2.3

3

Outline of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 26

Meta-Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.1.1

Well-known algorithms . . . . . . . . . . . . . . . . . . 31

2.1.2

Sampling in meta-heuristics . . . . . . . . . . . . . . . . 33

Hyper-Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.1

Heuristic selection approaches . . . . . . . . . . . . . . 43

2.2.2

Heuristic generation approaches . . . . . . . . . . . . . 47

Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . 49
2.3.1

Well-known classification techniques . . . . . . . . . . 50

2.3.2

Sampling and attribute selection for supervised learning 54

2.3.3

Meta-learning . . . . . . . . . . . . . . . . . . . . . . . . 56

The SPOT Hyper-Heuristic . . . . . . . . . . . . . . . . . . . . . . . 59
3.1

An Overview of the SPOT Approach . . . . . . . . . . . . . . . 60

3.2

Formal Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.3

3.2.1

Combinatorial optimization problem . . . . . . . . . . 66

3.2.2

Hyper-heuristics . . . . . . . . . . . . . . . . . . . . . . 68

3.2.3

The U/EA and the U/EA2 standards . . . . . . . . . . 70

3.2.4

The SPOT hyper-heuristic . . . . . . . . . . . . . . . . . 72

3.2.5

Heuristic selection and heuristic generation . . . . . . . 74

Methodology of Developing a SPOT Hyper-Heuristic . . . . . 78

xi

3.4
4

The design phase . . . . . . . . . . . . . . . . . . . . . . 78

3.3.2

The run phase . . . . . . . . . . . . . . . . . . . . . . . . 85

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Java Implementation of the SPOT Hyper-Heuristic . . . . . . . . . 89
4.1

Supporting Libraries . . . . . . . . . . . . . . . . . . . . . . . . 90
4.1.1

HyFlex . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

4.1.2

Weka . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

4.2

The Class Design . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4.3

Implementation in Java . . . . . . . . . . . . . . . . . . . . . . . 94

4.4
5

3.3.1

4.3.1

The class SPOT ML . . . . . . . . . . . . . . . . . . . . . 94

4.3.2

The class SPOT . . . . . . . . . . . . . . . . . . . . . . . 97

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

Application I: The Traveling Salesman Problem Domain . . . . . 101
5.1

An Introduction to the TSP Domain . . . . . . . . . . . . . . . 102

5.2

Implementation of the TSP Domain in HyFlex . . . . . . . . . 103

5.3

Development of the SPOT for TSPs . . . . . . . . . . . . . . . . 105

5.4

5.3.1

P1: Transformations and sampling . . . . . . . . . . . . 106

5.3.2

P2: Parameter determination . . . . . . . . . . . . . . . 110

5.3.3

P3: The generation of new LLHs . . . . . . . . . . . . . 117

Experiments and Observations . . . . . . . . . . . . . . . . . . 117
5.4.1

On the individual LLHs . . . . . . . . . . . . . . . . . . 119

xii
5.4.2
5.5
6

Comparisons with other hyper-heuristics . . . . . . . . 123

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

Application II: The Permutation Flow-Shop Scheduling Problem
Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1

An Introduction to the FSP Domain . . . . . . . . . . . . . . . 130

6.2

Implementation of the FSP Domain in HyFlex . . . . . . . . . 131

6.3

The Development of a SPOT for the FSP . . . . . . . . . . . . . 133

6.4

6.5
7

6.3.1

P1: Transformations and sampling . . . . . . . . . . . . 133

6.3.2

P2: Parameter determination . . . . . . . . . . . . . . . 138

6.3.3

P3: The generation of new LLHs . . . . . . . . . . . . . 144

Experiments and Observations . . . . . . . . . . . . . . . . . . 146
6.4.1

On the individual LLHs . . . . . . . . . . . . . . . . . . 147

6.4.2

Comparisons with other hyper-heuristics . . . . . . . . 152

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . 159
7.1

7.2

Discussion on the SPOT Methodology and Findings . . . . . . 160
7.1.1

Principal findings . . . . . . . . . . . . . . . . . . . . . . 160

7.1.2

Interpretation and implications of findings . . . . . . . 163

7.1.3

Interpretation in the context of the literature . . . . . . 166

7.1.4

Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 167

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

xiii
Appendix A

Suboptima of Training Instances . . . . . . . . . . . . . 175

Appendix B

Examples of Data and Results of Learning . . . . . . . 179

Appendix C

Original Data of Figures . . . . . . . . . . . . . . . . . . 195

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

xiv

List of Tables

1.1

Examples of the characteristic D on the map . . . . . . . . . . 16

2.1

Examples of records with known labels . . . . . . . . . . . . . 50

2.2

References for further reading on the classification techniques
not selected in this thesis . . . . . . . . . . . . . . . . . . . . . . 52

3.1

Main phases in the design phase in developing a SPOT hyperheuristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1

A list of generated attributes of the i-th assignment from one
raw attribute . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

4.2

The default values of main parameters of the classification
and attribute selection methods in Weka . . . . . . . . . . . . . 97

5.1

Ten benchmark instances used in the implementation of TSP
in HyFlex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

5.2

The LLHs in the implementation of TSP in HyFlex . . . . . . . 105

xv

xvi
5.3

Three benchmark TSP instances used in the development of
the SPOT algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 108

5.4

Average test results on different aspect ratios in the rectangular
selection sampling method, where n0 = 400 and the aspect
ratio of the minimum boundary rectangle is assumed as 1:1
(100 runs, best performance in bold) . . . . . . . . . . . . . . . 115

5.5

Average results and significants of tests on different subproblems sampled, where n0 = 400 (100 runs, significant p in bold) 116

5.6

Test results on the tour quality of the generated LLHs and
their base algorithms (100 runs, n = 400, Depth of search = 1,
improvements on average performance in bold) . . . . . . . . 120

5.7

Average time of generating new LLHs (100 runs, n = 400,
Depth of search = 1) . . . . . . . . . . . . . . . . . . . . . . . . 122

5.8

Average run time of each LLHs (100 runs, n = 400, Depth of
search = 1, instrument error < 0.16 ms) . . . . . . . . . . . . . 123

5.9

The top three hyper-heuristics in the TSP domain in CHeSC
2011 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

5.10 Median tour length of the PHunter, the SPOT-PHunter and
three other hyper-heuristics in the TSP domain in HyFlex (31
runs, best results in bold for each instance) . . . . . . . . . . . 125
6.1

Twelve benchmark instances used in the implementation of
FSP in HyFlex . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

6.2

The LLHs in the implementation of FSP in HyFlex . . . . . . . 132

xvii
6.3

Four benchmark FSP instances used in the development of the
SPOT algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

6.4

The 5m + 3 raw attributes designed for a job J i against another
job J j in the FSP domain (i , j) . . . . . . . . . . . . . . . . . . . 137

6.5

Average results and significants of r in tests on different subproblems sampled, where n0 = 30 (100 runs, significant values
of p in bold) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

6.6

Average results and significants of r in tests on different combination of parallel subproblems sampled, where n0 = 30 (100
runs, significant values of p in bold) . . . . . . . . . . . . . . . 143

6.7

Average results and significants of time of generation (timeoverall )
in tests on different combination of parallel subproblems sampled, where n0 = 30 (100 runs, significant values of p in bold) . 144

6.8

The modified LLHs for FSPs and modifications by the SPOT
heuristic generation . . . . . . . . . . . . . . . . . . . . . . . . . 146

6.9

Test results on the makespan of the generated LLHs and their
base algorithms (10 runs, 100 iterations in each run, Depth
of search = 1, Intensity of mutation = 1, improvements on
average performance in bold, significant values of p in bold) . 149

6.10 Average overall time (timeoverall ) of generating new LLHs (10
runs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.11 Average run time of each LLHs (10 runs, 100 iterations in each
run, instrument error < 0.16 ms) . . . . . . . . . . . . . . . . . 151

xviii
6.12 The top three hyper-heuristics in the FSP domain in CHeSC 2011152
6.13 Median makespan of the PHunter, the SPOT-PHunter and
three other hyper-heuristics in the FSP domain in HyFlex (31
runs, best results in bold for each instance) . . . . . . . . . . . 153
6.14 A flawed transformation of FSPs that resulting in a very low
approximability . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
A.1 Suboptima of the training instances in the FSP . . . . . . . . . 178
B.1 The group and raw attributes of assignments in the 20-nearestneighbor-first candidate set of two subproblems of the TSP instance rat783 (15840 rows × (1 column of group + 2 columns
of raw attributes), n0 = 400, sampled by a random selection
and a rectangular selection, respectively) . . . . . . . . . . . . 180
B.2 The full training table consisting of normalized attributes of
assignments and the label of suboptima of two subproblems
of the TSP instance rat783 (15840 rows × (86 columns of
attributes + 1 column of labels), attributes normalized from
Table B.1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
B.3 Results of the attribute selection for the full training table of
Table B.2 (15840 rows × (8 columns of attributes + 1 column
of labels)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

xix
B.4 Eight selected columns of test data and the summation of
predicted labels of the assignments in the 20-nearest-neighborfirst candidate set of the TSP instance rat783 (15660 rows ×
(8 columns of selected attributes + 1 column of summations of
labels)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
B.5 The group and raw attributes of assignments of the “following”
relationship in two subproblems of the FSP instance TA082
(3480 rows × (1 column of group + 103 columns of raw attributes), n0 = 30, sampled by two random selections, respectively) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
B.6 The full training table from two subproblems of the FSP instance TA082 (3480 rows × (206 columns of attributes + 1
column of labels), attributes normalized from Table B.5) . . . . 188
B.7 The test data and the summation of predicted labels of the assignments of the “following” relationship in the FSP instance
TA082 (19800 rows × (206 columns of normalized attributes +
1 column of summations of labels)) . . . . . . . . . . . . . . . . 192
B.8 The weights (or summations of Σ label) according to Table B.7
(which generated the figure (i) in Figure 6.2) . . . . . . . . . . 193
B.9 An example of adding the weights of a predicted permutation
and an NEH permutation (where the three columns of weights
present the figures (i), (ii) and (iii) in Figure 6.2, respectively) . 194
C.1 Original data of Figure 1.4 (D. S. Johnson & McGeoch, 2002,
p. 376) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

xx
C.2 Original data of Figure 5.1 . . . . . . . . . . . . . . . . . . . . . 196
C.3 Original data of Table 5.10 . . . . . . . . . . . . . . . . . . . . . 197
C.4 Original data of Figure 5.3 . . . . . . . . . . . . . . . . . . . . . 198
C.5 Original data of Figure 6.3 . . . . . . . . . . . . . . . . . . . . . 199
C.6 Original data of Figure 6.4 (Significant values of p in bold) . . 200
C.7 Original data of Table 6.13 (Significant values of p in bold) . . 201
C.8 Original data of Figure 6.5 . . . . . . . . . . . . . . . . . . . . . 202

List of Figures

1.1

Sun’s placement for T’ien Chi’s horses racing . . . . . . . . . .

5

1.2

The brachistochrone problem . . . . . . . . . . . . . . . . . . .

6

1.3

Euler diagram for classes of P, NP, NP-Complete and NP-hard

8

1.4

Examples of non-dominated and dominated heuristics for
solving the 10000-city Random Uniform Euclidean Instances
in the DIMACS TSP Challenge (D. S. Johnson & McGeoch,
2002, p. 376) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.5

Britain fractal coastline . . . . . . . . . . . . . . . . . . . . . . . 17

1.6

TSP Examples from different data sets . . . . . . . . . . . . . . 19

1.7

Euler diagram of instance-specific adjustments of algorithms,
hyper-heuristics and meta-heuristics . . . . . . . . . . . . . . . 20

2.1

A tree taxonomy of some well-known classification techniques
(Hilario, Kalousis, Nguyen, & Woznica, 2009) . . . . . . . . . . 51

2.2

Pseudo codes of the C4.5 classifier . . . . . . . . . . . . . . . . 53

xxi

xxii
3.1

A comparison between the conceptual diagram of the SPOT
methodology and that of the conventional hyper-heuristic
framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.2

The mapping from feasible solutions to objective function
values in a combinatorial optimization problem . . . . . . . . 67

3.3

Schematic diagram of a hyper-heuristic . . . . . . . . . . . . . 69

3.4

The mapping of U/EA transformation from the search space
of a combinatorial optimization problem . . . . . . . . . . . . . 71

3.5

Schematic diagram of SPOT hyper-heuristics . . . . . . . . . . 73

4.1

The class diagram of the SPOT hyper-heuristic . . . . . . . . . 93

4.2

The pseudo codes of function SPOT.run() . . . . . . . . . . . . 98

5.1

Average test results for determining n’ in the development of
the SPOT algorithm, on the three TSP test instances (100 runs) 113

5.2

A comparison of two candidate sets of the TSP instance rat783118

5.3

Scores of the PHunter, the SPOT-PHunter and three other
hyper-heuristics in the TSP domain in HyFlex . . . . . . . . . 126

6.1

An example of an FSP schedule with a permutation ( J 1 , J 2 , J 3 ) 130

6.2

Examples of a predicted permutation, an NEH permutation
and a weighted permutation for the FSP instance TA082 . . . 139

xxiii
6.3

Average test results for determining n’ in the development of
the SPOT algorithm, on the four FSP training instances (100
runs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

6.4

Comparisons between the new LLHs and their base algorithms for solving the 100 × 20 FSP instances TA81, TA82,
TA83, TA84 and TA85 . . . . . . . . . . . . . . . . . . . . . 151

6.5

Scores of the PHunter, the SPOT-PHunter and three other
hyper-heuristics in the FSP domain in HyFlex . . . . . . . . . . 154

A.1 Suboptima of the training instance rat783 in the TSP . . . . . 175
A.2 Suboptima of the training instance pcb1173 in the TSP . . . . 176
A.3 Suboptima of the training instance d1291 in the TSP . . . . . 177
B.1 Instance-specific result of the J48 classifier for the TSP instance
rat783, training data shown in Table B.3 . . . . . . . . . . . . 183
B.2 Instance-specific result of the JRip classifier for the TSP instance rat783, training data shown in Table B.3 . . . . . . . . 183
B.3 Instance-specific result of the NaiveBayes classifier for the TSP
instance rat783, training data shown in Table B.3 . . . . . . . 184
B.4 A set of promising edges selected from the 20-nearest-neighborfirst candidate set of the training instance rat783 in the TSP,
according to the Σ label column in Table B.4 (where black
edges in the second figure were the assignments with Σ label

≥ 2 and the gray edges were those with Σ label = 1) . . . . . . 186

xxiv
B.5 Instance-specific result of the J48 classifier for the FSP instance
TA082, training data shown in Table B.6 . . . . . . . . . . . . . 189
B.6 Instance-specific result of the JRip classifier for the FSP instance TA082, training data shown in Table B.6 . . . . . . . . . 190
B.7 Instance-specific result of the NaiveBayes classifier for the FSP
instance TA082, training data shown in Table B.6 . . . . . . . . 191

List of Symbols

Notation

Description

~
A

as the set of all the possible attribute vectors
of solutions to a combinatorial optimization
problem. 68, 69, 71–74

a
as an attribute in supervised learning. 52
as the cardinal number of a set of a limit number of paramters. 77

~a
as a vector of the attributes in supervised
learning. 52, 55
as a vector of attributes of a solution to a
combinatorial optimization problem. 68, 69,
71, 72, 74
ag

as an attribute generation for a hyperheuristic. 68, 69, 72, 73

ℵ0

as the cardinality of the natural numbers. 75–
77, 87, 88

xxv

xxvi
Notation

Description

C

as the set of all countable numbers. 76, 77

C
as the number of cities in a TSP. 7, 37, 106
as the class in supervised learning. 52
c

as a city in a TSP. 7, 36, 37, 106, 108, 109, 111

cov

as the percentage of the edge candidate sets
covering the edges in a known suboptimum.
111–117, 121, 127, 196

D
as a constant of “characteristic of a frontier”.
15, 16
as a discrete domain of a variable in a combinatorial optimization problem. 66, 72
d¯

as the average depth of optimal assignments
in a set of predicted priority lists of assignments. 112–117, 121, 196

d

as the distance between two cities in a TSP. 7,
36, 106, 109, 111

F

as the F value of one-way ANOVA. 115, 116,
120, 143, 144, 148, 149, 197, 200, 201

xxvii
Notation

Description

f

as the objective function in a combinatorial
optimization problem. 66, 67

fl

as the relationship of following between two
jobs in an FSP. 134, 135, 137, 138, 140

G

as a variable of the measurement scale of fractal coastline. 15, 16

Γ

as the set of possible values of confidence
levels of predicting a assignment. 74, 77

γ

as the predicted value (confidence) of appearing in the optimum for an assignment. 74,
111

H

as the set of low-level heuristics for a hyperheuristic. 68, 69, 73–77

h

as a low-level heuristic for a hyper-heuristic.
68, 69, 73, 74

hc

as a hard constraint in a combinatorial optimization problem. 67

J

k

as a job in an FSP. xvii, xxii, 130, 134, 137, 140

xxviii
Notation

Description
as the number of maximum swaps for a tour
for a TSP. 33
as the size of k-quardrant candidate set. 103

L

as the estimate of length of seacoast. 15

l

as the learning in a hyper-heuristic. 68, 69,
72, 73

M

as a positive constant prefactor of fractal
coastline. 15, 16

m

as the number of machines in an FSP. xvii, 7,
132, 135–138, 150, 155

n
as the number of cities in a TSP. xvi, 7, 10–12,
22, 24, 35, 36, 38, 40, 57, 71, 106–110, 120, 122,
123
as the number of jobs in an FSP. 7, 130, 132,
134–137, 145, 150, 155
as the number of variables in a combinatorial
optimization problem. xvi–xix, xxii, xxiii, 10,
48, 63–67, 72–75, 77, 79, 81–83, 85, 87, 88, 94,
95, 104, 108, 110, 112–116, 119, 138, 140–144,
147, 156, 180, 187, 196, 199

xxix
Notation

Description
as the size of a program in genetic programming. 49
as the number of attributes. 52, 55

na

as the number of attribute vectors for a hyperheuristic. 68, 72, 74, 77, 87, 88

nc

as the number of hard constrains in a combinatorial optimization problem. 67

nh

as the number of low-level heuristics in a
hyper-heuristic. 68

O

as the big O notation, bounded above by another function asymptotically. 10, 24, 67

Θ

as the big Θ notation, bounded by another
function asymptotically. 48, 49, 77

P

as all of the NP optimization problems. 9

P

as the function of probability. 34, 36, 37, 52,
85

p
as the number of possible assignments in a
group of competitor. 94–96, 108, 109, 136
as the significance of one-way ANOVA. xvi,
xvii, xx, 115, 116, 120, 121, 125, 126, 143, 144,
148, 149, 153, 197, 200, 201

xxx
Notation

Description

π

as a performance measure for hyperheuristics. 68, 69, 73

R

as the set of all real numbers. 67–69, 71, 73

ra

as the number of raw attributes defined for
each assignment. 94–96, 109, 137

r
as the appriximation factor. 9, 10
as the function of resemblance between two
solutions. xvii, 25, 79, 80, 83, 107, 108, 111,
136, 140–144, 148, 155, 156, 161, 162, 199

S

as the solution space (set of feasible solutions)
to a combinatorial optimization problem. 66–
69, 71–73, 79

s

as a feasible solution to a combinatorial optimization problem. 67, 69, 71, 73, 79

sam

as the sampling procedure in the SPOT hyperheuristic. 72, 73

Σ

as a set of U/EA2 solutions mapping from
the fesiable solutions to a combinatorial optimization problem. 71–73

σ

as a U/EA2 value of a fesiable solution to a
combinatorial optimization problem. 71, 73

xxxi
Notation

Description

T

as a bijection of transformation from a fesible
solution to a U/EA2 solution. 70–73

t

as the processing time of a job on a machine
in an FSP. 7, 135, 137

τ

as the quantity of pheromone on an edge in
ant colony approaches. 36, 38

V

as the set of all possible solutions in a combinatorial optimization problem. 66, 67

ν

as an assignment (valuation) for a variable in
a combinatorial optimization problem. 66

v

as the set of all assignments for a variable in
a combinatorial optimization problem. 66

X

as the set of variables in a combinatorial optimization problem. 66, 70

x
as a variable in a combinatorial optimization
problem. 66, 70
as a value of raw attibute. 95, 96
as the x coordinate of a city in a TSP. 109

y

as the y coordinate of a city in a TSP. 109

xxxii
Notation

Description

Z

as the set of all integers. 66, 76

List of Acronyms

Notation

Description

APX

Approximable. 9

BBO

Black Box Optimization. 13, 43

CHeSC

Cross-domain Heuristic Search Challenge. iv,
xvi, xviii, 21, 45, 90, 99, 103, 104, 118, 119,
123–125, 131, 132, 146, 147, 152, 153, 167–169,
172, 198, 202

DIMACS

Center for Discrete Mathematics and Theoretical Computer Science. xxi, 14, 17, 33, 40

DSAFO

Dynamic Scheduling Agents with Federation
Organization. 39

EDAs

Estimation of Distribution Algorithms. 32–36,
39, 54, 61, 63, 166

xxxiii

xxxiv
Notation

Description

EHBSA

Edge Histogram Based Sampling Algorithm.
35

FSP

permutation Flow-Shop scheduling Problem.
iii, iv, xvi–xix, xxii–xxiv, 7, 9, 25, 26, 88, 129–
139, 141, 145, 146, 151–155, 160–163, 167, 168,
171, 172, 178, 187–192

GCC

GNU Compiler Collection. 98

GRASP

Greedy Random Adaptive Search Procedure.
32, 34, 39, 40, 57

HK

Held-Karp. 12, 14, 33, 102

HyFlex

Hyper-heuristics Flexible framework. iii, viii,
xv, xvi, xviii, xxii, xxiii, 89–92, 97–99, 101, 103–
105, 109, 112, 114, 119, 122, 123, 125–127, 129,
131, 132, 135, 145, 147, 150, 152–154, 167, 169,
170, 172, 196

LKH

Lin-Kernighan-Helsgaun. 33

xxxv
Notation

Description

LLH

low-level heuristic. xv–xvii, xxiii, 41, 42, 44–
46, 48, 49, 57, 61, 62, 65, 68, 69, 73–79, 83–86,
90–93, 98, 99, 104, 105, 108, 110–112, 117–125,
127, 128, 131–133, 138, 144–157, 165, 169, 173,
200

NFL

No-Free-Lunch. 13, 15, 23, 43, 164, 173

NP

Non-deterministic Polynomial time. 7–9

NP-Complete Non-deterministic
Complete.

Polynomial

time-

iii, 7–10, 12, 21, 33, 35, 38,

41, 67, 68, 87, 88, 99, 102, 130, 159, 160, 164,
171, 172
NP-hard

Non-deterministic Polynomial time-hard. 8,
9, 12, 23, 64, 102, 130, 164

P

Polynomial-time. 7–9

PHunter

Pearl Hunter. iv, xvi, xviii, xxii, xxiii, 21, 42,
45, 97, 99, 111, 118, 119, 123–126, 138, 139, 146,
147, 152–154, 157, 167, 172, 197, 198, 201, 202

SAT

Boolean Satisfiability Problem. 44, 45, 49

xxxvi
Notation

Description

SPOT

Suboptimum- and Proportion-based On-thefly Training. iii, iv, xvi–xviii, xxii, xxiii, 15,
24–26, 29, 59–62, 65, 66, 70–74, 77–81, 83, 85–
87, 89, 92–94, 96–99, 101, 105–111, 113, 114,
117–119, 121–129, 133, 136–142, 144, 146–148,
151–157, 159–167, 169–173, 186, 193, 194, 196–
199, 201, 202

TSP

Traveling Salesman Problem. iii, iv, xv, xvi,
xviii, xix, xxi–xxiii, 7, 9–12, 14, 17, 22–26, 33,
35, 36, 38, 40, 47, 49, 57, 60, 71, 80, 83, 84, 86,
87, 101–110, 113, 117, 118, 121–126, 128, 130,
131, 140, 145, 160–163, 165–169, 172, 175–177,
180, 181, 183–186

U/EA2

Unconstrained and with Equinumerous Assignments and Equinumerous Attributes. 66,
71, 73, 78, 79, 81, 106, 107, 110, 133, 136, 137,
161

U/EA

Unconstrained and with Equinumerous Assignments. xxii, 61, 63, 66, 70–73, 78–80, 83,
87, 92, 106, 107, 133, 135, 155, 160–162, 164,
169

xxxvii
Notation

Description

VLSI

Very-Large-Scale Integration. 102

Weka

Waikato Environment for Knowledge Analysis. iii, viii, xv, 51–54, 82, 89–93, 95–97

xxxviii

Chapter 1
Introduction to the SPOT Algorithm

Many [geographical curves] are
statistically “self-similar,” meaning that
each portion can be considered a
reduced-scale image of the whole.
How Long is the Coast of Britain?
Benoı̂t B. Mandelbrot

Many — if not all — human activities involve searching for ways to
achieve particular goals. Individuals and organizations are called rational
if they maximize their benefits or minimize inconvenience in equivalent
measures. As a result, optimal decisions are highly demanded in many
practical as well as theoretical problems.

1

2

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

1.1

Optimizations from Ancient to Modern Times

Since antiquity, many problems on decision making emerged from human social life, and were in the form maximizing benefits. Those successful
solvers became well-known representatives of human intelligence and rationality. The Roman poet Virgil (29-19B.C./1904, lines 1.365–1.368) recorded
a legendary story in the foundation of Carthage by Queen Elissa (Dido of
Carthage) around 800 B.C.:
Dēvēnēre Locōs, ubi nunc ingentia cernis
moenia surgentemque novae Karthāginis arcem,
mercātı̄ue solum, factı̄ dē nōmine Byrsam,
taurı̄nō quantum possent circumdare tergō.

[English (Virgil, 29-19B.C./1893, p. 14):
So to the place they came, where now thou spyest
The lofty walls and rising citadel
of new-built Carthage, and of land they bought —
Called Byrsa from their bargaining — so much
As with a bull’s hide they might compass round.]

This isoperimetric problem of enclosing the maximum area with a given perimeter is now also known as Dido’s problem (e.g., Merrill, 1919). The solution
for this case, with the Mediterranean coast as a given edge, is a semicircle
(Hackley, 1847, Appendix I, Theorem IV). Someone said that the cunning
Queen Elissa cut the bull hide into very narrow strips and circumscribed a

1.1. OPTIMIZATIONS FROM ANCIENT TO MODERN TIMES

3

maximized size of land in a semicircle1 . Some others, such as R. B. Smith
(1913, p. 14), argued that the Greek legend was based on the superficial
resemblance of the Phoenician word “Bozra” (a fortress) to the Greek “Byrsa”
(an ox-hide). In addition, archaeological evidence shows that the city wall of
Carthage2 was not of circular arc shape. Although the authenticity of the oxhide legend is debatable, it is still clear that the idea of Dido’s maximization
appeared for more than two millennia.
Another historical story from the Kingdom Ch’i in the Warring States
Period, about 350 B.C., of ancient China was recorded by Sima (91B.C./2010,
pp. 01.761–01.762):
RŽ›

*úf9Å-

"9fgÙ2vØ$ƒì§2kþ!

¥!eŒ"ß´9f4XR

µ7

R&,ƒ§›

Z7"9ŠŒ§9f

9*úfÅ

ƒee›*þe§

3-

þe›*¥e§

9nŒª§ XR˜ØÒ 2Ò§w

§—U-

Ò"8X
µ78±

¥e›*ee"8Q
Z7"

[English (Ssu-ma, 91B.C./1994, pp. 39–40): [Ch’i’s general] T’ien
Chi XR raced horses and gambled heavily with the Noble Scions
of Ch’i several times. [His guest] Sun Tzu [Sun Pin 9

] noticed

that the horses’ speed was not much different and that the horses
fell into high, middle and low grades. After this, Sun Tzu told
T’ien Chi, “Just bet heavily, My Lord, and I can make you the
winner.”
1 See Elissar, Dido, the Queen of Carthage and her city, http://www.phoenicia.org/

elissardidobio.html.
2 See History Channel’s TV program Engineering an Empire: The Carthage.

4

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
T’ien Chi confidently agreed and bet a thousand chin3 with King
[Wei %, r. *378–343 B.C.] and the Noble Scions [of Ch’i] on a race.
Just before the wager Sun Tzu said, “Now match their high-grade
horses with your low-grade horses, take your high-grade horses
to match their middle-grade horses and take your middle-grade
horses to match their low-grade horses.”
After they raced the three grades [of horses], T’ien lost once but
won twice and eventually gained the king’s thousand chin.]

The military strategist Sun Pin (or Bin), author of Sun Bin’s Art of War, suggested to General T’ien Ch’i a discrete placement strategy for horse-racing,
as shown in Figure 1.1. Sun took advantage of the best combination4 of
matching and made T’ien the winner. After that, King Wei questioned Sun
on the arts of war and made him his counselor.
Minimization or maximization problems solved with modern mathematics can be traced back to the 17th century. One well known problem was
Bernoulli (1696)’s brachistochrone (Woodhouse, 1810/1964, p. 3), i.e. fastest
descent curve without friction, as shown in Figure 1.2:
Datis in plano verticali duobus punctis A et B, assignare mobili
M viam AMB, per quam gravitate sua descendens, et moveri
incipiens a puncto A, brevissimo tempore perveniat ad alterum
punctum B.
[English (D. E. Smith, 1929, p. 645): If two points A and B are
3 Note: Weight unit, 1 chin = 256.26 g.
4 It should be noted that this matching plan is not an equilibrium.

1.1. OPTIMIZATIONS FROM ANCIENT TO MODERN TIMES

Middle

Low

$

King’s

n

N

$

n

$

n

T’ien’s
Faster

High

T’ien’s

N

"

King’s

n

N

"

n

$

n

N

N

(i) common matching:
T’ien lost (0/3 scores)

Slower

Grade

5

N

(ii) Sun’s placement:
T’ien won (2/3 scores)

Figure 1.1: Sun’s placement for T’ien Chi’s horses racing
given in a vertical plane, to assign to a mobile particle M the path
AMB along which, descending under its own weight, it passes
from the point A to the point B in the briefest time.]
The name brachistochrone was derived from Greek βράχıστoς (brachistos, the
shortest) and χρóνoς (chronos, time or delay). Six mathematicians, Newton,
Leibniz, Von Tschirnhaus, L’Hôpital, his elder brother Jacob Bernoulli and
Bernoulli himself (Gerdts, 2012, p. 1; Stillwell, 2010, p. 259), (independently)
found a solution — the cycloid5 . Stillwell (2010, p. 259) regarded Jacob
Bernoulli’s solution as the most profound and the first major step in the
development of the calculus of variations.
During the last decades, the widespread use of computers enabled a lot
of optimization methods to be used that otherwise would be too tedious and
5 Galileo proved a circular arc is better than strait line for brachistochrone in 1638, but

failed to recognize a cycloid (Babb & Currie, 2008).

6

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
A

M

B
Figure 1.2: The brachistochrone problem
time-consuming for human planners. With a steadily developing computer
technology nowadays, the increasing machine power becomes an indispensable tool (Kleunen, 1976) and makes the optimization process more and more
efficient. In the contemporary world, almost all branches of engineering
have been increasingly dependent on computerized techniques (Kassim &
Cadbury, 1996).

1.2

Combinatorial Optimization: Problems and Algorithms

This thesis focuses on a class of problems encoded in discrete values
called combinatorial optimization problems. In a combinatorial optimization
problem, the aim is to look for an object from a finite (or possibly countably
infinite) set (Papadimitriou & Steiglitz, 1982, p. 2). A typical object can be an
integer number, a subset, a permutation or a graph structure (C. Blum & Roli,
2003). The best object is called the optimal solution or optimum, and the objects
very close to the optimal solution are called suboptimal solutions or suboptima.

1.2. COMBINATORIAL OPTIMIZATION

7

Examples of combinatorial optimization problems are the Traveling Salesman
Problem (TSP) and the permutation Flow-Shop scheduling Problem (FSP). The
definitions of the TSP and the FSP are quoted as follows, respectively.

In the traveling salesman problem, a set C = {c1 , c2 , . . . , cn } of
cities is given, and for each pair (ci , c j ) of distinct cities, a distance
d(ci , c j ). The target is to find a Hamiltonian cycle (also known
as a Hamiltonian circuit) of all the cities that minimizes the tour
length (D. S. Johnson & McGeoch, 1997).

Given n jobs to be processed on m machines in the same order,
the process time of job i on machine j being ti,j (i = 1, 2, . . . , n;
j = 1, 2, . . . , m), find the sequence of jobs such that the total
elapsed time (makespan) is minimized (Nawaz, Enscore, & Ham,
1983).

The TSP and the FSP seem easy to describe and validate, but they have
been proved to be in the NP-Complete (Non-deterministic Polynomial timeComplete) class — a very difficult class of problems (Papadimitriou, 1977;
Garey, D. S. Johnson, & Sethi, 1976). In computational complexity theory, NP6
(Non-deterministic Polynomial time) is the set of decision problems solvable
in polynomial time on a non-deterministic Turing machine. P (Polynomialtime) is the complexity class containing decision problems which can be
solved by a deterministic Turing machine using a polynomial amount of
computation time. Equivalently, solutions to NP problems can be verified
6 See http://qwiki.stanford.edu/index.php/Complexity Zoo:N.

8

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
NP-hard

NP-hard

NP-Complete

P = NP = NP-Complete

P

(ii) P , NP

Complexity

NP

(i) P = NP

Figure 1.3: Euler diagram for classes of P, NP, NP-Complete and NP-hard
by a deterministic Turing machine in polynomial time. A problem is NPComplete, if and only if every other NP problem is reducible to this problem
(Garey & D. S. Johnson, 1979, p. 13). Another common term, NP-hard (Nondeterministic Polynomial time-hard), represents the class of problems that
are at least as hard as problems in NP-Complete, as shown in Figure 1.3
(Wikipedia, 2012b). If any one NP-Complete problem can be generally solved
in polynomial-time on a deterministic Turing machine, so can every NP
problem.
The relationship “P versus NP?” is still unsolved, and the Clay Mathematics Institute is offering7 a US$ one million prize for the first correct proof or
disproof. However, most researchers believe P , NP. For example, in Aaronson (2005)’s poll, 61 out of 100 theorists thought P , NP, 9 thought P = NP,
7 As one of the seven Clay Mathematics Institute’s Millennium Prize Problems, “P versus

NP?” problem was announced on May 24, 2000 at Collège de France and is still open in 2012.
See http://www.claymath.org/millennium/.

1.2. COMBINATORIAL OPTIMIZATION

9

22 offered no opinion and 8 offered other opinions. In this thesis, P , NP
is assumed. As a result, no polynomial-time (A. L. Blum & Rivest, 1992) or
sub-exponential time (Woeginger, 2003) algorithm can exactly solve any NPComplete problem, in the worst case. In other words, for an NP-Complete
problem, the solution process will typically face either
• an explosive running time for an optimal solution, or
• suboptimal solutions for a reasonable running time.
In fact, to guarantee finding a suboptimum for an NP-Complete problem is
also very difficult. Ausiello et al. (1999, p. 93) defined APX (Approximable) as
the class of all NP optimization problems P such that, for some factor r ≥ 1,
there exists a polynomial-time r-approximation algorithm for P . Ausiello et
al. (1999, p. 94) proved that a TSP holding triangular inequality does not belong
to APX unless P = NP, and Papadimitriou and Vempala (2006) proved such
a TSP is NP-hard to approximate, if the ratio is less than 220/219. Williamson
et al. (1997) proved that an FSP is NP-hard to approximate with a factor less
than 5/4.
Exact algorithms for solving combinatorial optimization are those which
guarantee finding an optimal solution in a finite amount of time. Examples
of exact algorithms include:
branch-and-bound which systematically enumerates all feasible solutions
in a tree and drops subtrees on-the-fly by lower and upper bounds
(Lawler & Wood, 1966),
Lagrangian relaxation based methods such as Held and Karp (1970)’s 1-

10

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
tree for TSP,

integer (linear) programming based methods 8 such as the cutting-plane
algorithm (Nemhauser & Wolsey, 1988, pp. 367–378) and the branchand-price (Barnhart, Johnson, Nemhauser, Savelsbergh, & Vance, 1998),
based on branch-and-bound and column generation,
exponential neighborhood search with an exponential setup time followed
by a polynomial (e.g., in O(n3 )) time local search (Ahuja, Ergun, Orlin,
& Punnen, 2002; Gutin & Glover, 2005),
recursive exhaustive search which enumerates all possible combinations
and is not the best option usually.
The exact approaches mainly encounter the first difficulty. However, when
the number of variables is limited, most exact algorithms — sometimes
even exhaustive search algorithms — can be very effective and efficient for
problem solving, see Applegate, Bixby, Chvátal, and Cook (2001)’s branchand-cut for small-scale TSPs (n < 103 ).
Heuristic (also known as approximate9 and incomplete) optimization algorithms only concentrate on a (small) subset of the solution space, usually
by experience-based techniques, and cannot guarantee the optimality of the
results. However, many heuristic optimization algorithms can find “good
enough” solutions in a “small enough” computation time. Examples of
heuristic optimization algorithms are:
8 Modeling or transforming a combinatorial optimization problem into integer program-

ming does not help on the complexity. The integer programming itself is NP-Complete in
the strong sense (Karp, 1972), even each integer belonging to the set {0,1} (Sahni, 1974).
9 It should be noted that an approximate algorithm is called r-approximation if and only
if it guarantees its results are no more than r times the optimum.

1.2. COMBINATORIAL OPTIMIZATION

11

first-come, first-served (or first-in-first-out): always selects the first available
component or assignment for the eventual result,
greedy construction: a step-by-step incremental construction, with a greedy
rule in each step, such as the multiple-fragment merge for TSPs (Bentley,
1992),
alpha-beta pruning: an approximate version of branch-and-bound with inexact branch reduction, see (Knuth & R. W. Moore, 1975),
meta-heuristics: high-level strategies for guiding a search (Glover, 1986)
according to feedback from the objective function, previous decisions
and prior performance (Stützle, 1998). More details can be found in
Section 2.1,
random: a return of a randomly generated solution, not the best option
usually.
The computations of heuristic optimization algorithms are usually in polynomialtime. However, there might possibly exist some exponential heuristic optimization algorithms, such as an impractical local search for the best 0.5nedge-swap solution in a TSP.
Integrations of exact and heuristic optimization algorithms (Puchinger &
Raidl, 2005) can be either exact or heuristic. An example of exact integrations
was given by French, Robinson, and Wilson (2001)’s branch-and-bound for
maximum satisfiability. They employed a steady-state meta-heuristic (evolutionary algorithm) to repeatedly suggest the fittest solutions from hanging
nodes in order to guide branching. The new integration still guaranteed to

12

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

find an optimal solution and was reported to be better than a single branchand-bound or an evolutionary algorithm. A heuristic example by Cotta
and Troya (2003) also integrated the branch-and-bound and evolutionary
algorithm, though in a different way. They embedded branch-and-bound
as a low-level operator in the evolutionary algorithm. Their approach was
heuristic, because its main framework was a heuristic and did not guarantee
an optimal result.
In large-scale NP-Complete optimization, exact and heuristic optimization algorithms are two indispensable forces for solving problems. The
optimum of a problem can be found by the squeeze theorem, if the greatest
lower bound by exact algorithms and the least upper bound by heuristic optimization algorithms become equal. For example, Applegate, Bixby, Chvátal,
Cook, et al. (2009) improved the lower bound to a known feasible solution
found by heuristics and thus certificated an optimum for a large-scale TSP
instance (n = 85, 900) in 286.2 CPU-days for a 2.4 GHz CPU.
One algorithm is said to be dominated by another if it generally returns
a worse (or the same) solution while costs more computation time, and
vice versa for non-dominated. Only a few non-dominated exact algorithms
exist for a problem, at the most, because every exact algorithm guarantees
the same optimum. On the other hand, there are a lot of non-dominated
heuristic algorithms for each NP-hard combinatorial optimization problem.
For example, Figure 1.4 shows some non-dominated heuristic optimization
algorithms for the 10000-city random Euclidean TSP, where the Held-Karp
(HK) bound is a lower bound. In practice, the economic importance of many
combinatorial optimization problems stimulates increasing demands for

1.2. COMBINATORIAL OPTIMIZATION

13

more non-dominated algorithms. This is also a cause leading to the thriving
of heuristics studies.
Although many heuristics aim at being cross-domain, besides being nondominated, there is a deadly pitfall from the No-Free-Lunch (NFL) theorems
(Wolpert & Macready, 1997). There are two NFL theorems relating to the set
of all possible problems and the set of all possible algorithms, respectively:
Theorem 1 (No-Free-Lunch Theorem 1). Over all possible problems, no search
algorithm is better than any other algorithm, on average.
Theorem 2 (No-Free-Lunch Theorem 2). Over all possible search algorithms, no
search problem is harder than any other problem, on average.
The NFL theorems have been proved in different ways, such as in Wolpert
and Macready (1997) and Culberson (1998). Though the theorems can be
criticized as too ideal, such as the condition of “all possible problems”. The
NFL theorems mean that no algorithm can perform statistically better than
a random search of possible solutions on average, if no problem-specific
information is considered (also known as the Black Box Optimization, or
BBO).
In fact, handling the problem-specific characteristics is important for a
general-purpose optimization algorithm to stay away from the pitfall of
NFLs (Smith-Miles, 2008a). There are at least two options for an optimization
algorithm, theoretically, to keep a safe distance from the pitfall of the NFLs:
• To become a well-designed problem domain-specific (ad-hoc) algorithm
(Burke, Kendall, et al., 2003, pp. 459–460), and

14

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

Non-dominated algorithms
Dominated algorithms

Spacefill

35

Pareto frontier
Strip
Karp

30

Excess over the HK bound (%)

NI
NN

25

CHCI
20
Greedy
15
FI
Savings
AppChristo

CCA

Christo

10

GENI
2-Opt

5

3-Opt
LK
0
0.01

0.1

1

MLLK CLK
10

100

Helsgaun
1000

Tabu

10000

Normalized computation time (s)

Figure 1.4: Examples of non-dominated and dominated heuristics for solving
the 10000-city Random Uniform Euclidean Instances in the DIMACS TSP
Challenge (D. S. Johnson & McGeoch, 2002, p. 376)

1.3. RESEARCH MOTIVES AND BACKGROUND

15

• To discover and to make use of domain-specific, problem data setspecific and/or instance-specific characteristics dynamically, from case
to case,
where most of the successful cross-domain heuristics as well as exact methods
belong to the latter class (e.g., Poli & Graff, 2009).
In this study, the SPOT methodology is proposed for cross-domain combinatorial optimization problems. The SPOT methodology obtains instancespecific information by machine learning to generate new heuristics or to
modify existing heuristics towards a non-dominated direction. Meanwhile,
the instance-specific information keeps the SPOT method safe from the pitfall
of the NFL theorems.

1.3

Research Motives and Background

Some extremely difficult real-world problems can be successfully approximated to the best results very efficiently by investigating instance-specific
information in a portion of the whole problem. One example is the selfsimilarity in fractals (Hutchinson, 1981). A self-similar object is exactly or
approximately similar to a part of itself. Mandelbrot (1967) presented such a
well-known approach to measure the length of the coastline of Great Britain,
see Figure 1.5 (Wikipedia, 2012a). Mandelbrot cited Richardson (1961)’s estimate of length
L( G ) = MG1− D ,

(1.1)

16

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

Table 1.1: Examples of the characteristic D on the map
D

Geographic boundary

1.00
1.25

An extreme value, a frontier that looks straight on the map
The west coast of Great Britain, which looks like one of the
most irregular seacoasts in the world
The land frontier of Germany in about 1899 A.D.
The land frontier between Spain and Portugal
The Australian coast
The coast of South Africa, a coast selected as looking one of
the smoothest in the atlas.

1.15
1.14
1.13
1.02

where M is a positive constant prefactor, G is a variable of the measurement
scale and D ≥ 1 is an instance-specific constant of the fractal dimension or a
“characteristic of a frontier, may be expected to have some positive correlation
with one’s immediate visual perception of the irregularity of the frontier”.
Hutchinson (1981) listed the characteristics of some geographic boundaries,
as shown in Table 1.1. Geographic curves in the real world are so involved in
their detail that their lengths are often infinite (or indefinable). Now, however,
the whole coastline can be estimated by measuring the “characteristic” D
through investigation of a small portion of the whole seacoast.
Another example is the sampling. In statistics and survey methodology,
sampling consists of selecting some part of a population to observe so that
one may estimate something about the whole population (Thompson, 2002,
p. 1). For example, the mean and the standard deviation of a population of
real numbers following an unknown normal distribution can be unbiased
estimatedly through the mean and the standard deviation of a random sample, respectively. A sample is representative only when the sampling process
is both accurate (with a mean of errors no more than a given threshold) and
reproducible (with a variance of errors no more than another given threshold)

1.3. RESEARCH MOTIVES AND BACKGROUND

G = 200km
L( G ) ≈ 2, 400km

G = 100km
L( G ) ≈ 2, 800km

17

G = 50km
L( G ) ≈ 3, 400km

Figure 1.5: Britain fractal coastline
(Petersen, Minkkinen, & Esbensen, 2005). The success of the estimations lies
in the representativeness, which relies on the characteristics of the population.
Some well-known applications of sampling include census and production
testing (e.g., Milor & Sangiovanni-Vincentelli, 1994).
What attracts interest in this thesis are the instance-specific characteristics
(usually) hidden in combinatorial optimization problems and the methods
of finding them from a proportion of a large-scale problem. Take two benchmark instances of TSP as an illustration: P1 (the E100.0 from the DIMACS
TSP Challange10 ) with randomly generated cities and P2 (the xqf131 from
10 The eighth DIMACS (Center for Discrete Mathematics and Theoretical Computer Sci-

ence) Implementation Challenge (2001), see http://dimacs.rutgers.edu/Challenges/ and
http://www2.research.att.com/∼dsj/chtsp/.

18

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

the VLSI data set11 ) from the print circuit industry with visually regular
cities, as shown in Figure 1.6. Since the visual difference in Figure 1.6 is clear,
one can hope that the instance-specific characteristics can probably unveil
more information than common knowledge, such as the triangular inequality. Another question is how to find the characteristics from a proportion, if
possible.
In the literature, various approaches have employed instance-specific information as “characteristics” to adjust both exact and heuristic optimization
algorithms for better performances for decades. There are three typical ways
of adjustment, including:
algorithm selection: returns a preferred subset of ready-to-use heuristics,
parameter optimization: (also known as parameter tuning and algorithm configuration) returns numerical parameters for heuristics,
algorithm generation: returns interesting12 (most likely unknown previously, potentially useful) heuristics.
According to the nature of the object algorithm, the three typical ways can
be further divided, as shown in Figure 1.7. Figure 1.7 also shows the relationships between the adjustments and two popular concepts of heuristics,
meta-heuristics and hyper-heuristics13 .
Typical algorithm selection for exact algorithms are: Lobjois and Lemaı̂tre
(1998)’s branch-and-bound approach which determines branching from a
11 See http://www.tsp.gatech.edu/vlsi/.
12 Nevertheless, some interesting results could be dominated by existing research, such as

Kotsiantis, Zaharakis, and Pintelas (2006).
13 See Section 2.2 and Section 3.2.5.

1.3. RESEARCH MOTIVES AND BACKGROUND

(i) A 2D TSP P1 with 100 randomly generated cities

(ii) A 2D TSP P2 with 131 cities from print circuit industry
Figure 1.6: TSP Examples from different data sets

19

20

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
Meta-heuristics
Hyper-heuristics
Generation
of exact methods?

Complexity of control

Instance-specific
parameter optimization

Heuristic
generation
Instance-specific
heuristic generation

for exact methods

for heuristics

for exact methods

for heuristics

Instance-specific selection

Heuristic
selection

Figure 1.7: Euler diagram of instance-specific adjustments of algorithms,
hyper-heuristics and meta-heuristics
number of (heuristic) options by performance prediction for each particular
problem instance, and Leyton-Brown, Nudelman, Andrew, McFadden, and
Shoham (2003)’s algorithm portfolios. Examples of parameter optimization
for exact algorithms are Kadioglu, Malitsky, Sellmann, and Tierney (2010)’s
instance-specific algorithm configuration for CPLEX, the best known mixed integer programming solver and Hutter, Hoos, Leyton-Brown, and Stützle (2009)’s
local search-based method named ParamILS. As far as is known, no approach
has been directly developed for exact algorithm generation. However, there
may be some possible indirect ways, such as Ibaraki and Muroga (1970)’s
linear programming for discovery of classifiers, and later works such as Brown,
Pittard, and Park (1996)’s optimum split on tree nodes. Such an indirect
approach can be called an “exact generation” if:

1.3. RESEARCH MOTIVES AND BACKGROUND

21

• there exists a one-to-one transformation between instances in an optimization problem domain to a set of classification problems; and
• the results of linear programming can guarantee no error (optimum)
for any instance of classification.
Furthermore, it must cost at least exponential time (Woeginger, 2003) in
general for any exact generation, if it exists, in solving an NP-Complete
optimization problem. In fact, if there is a non-zero error between the training
labels (from the optimum) and the predicted values of the labels (usually
represented as a solution), the generation can be considered as heuristic
generation in practice.
Heuristic selection, including instance-specific algorithm selection and
parameter optimization for heuristics, has received increasing attention in
recent years. For example, there were twenty heuristic selection algorithms
submitted to the hyper-heuristic competition CHeSC (Cross-domain Heuristic Search Challenge) 201114 from all over the world. Each competitor had a
particular strategy of making use of instance-specific and domain-specific information. Some cross-domain heuristic selection methods proved their nondominance in certain domains, such as Chan, Xue, Ip, and Cheung (2012)’s
Pearl Hunter (PHunter) hyper-heuristic which found new best-known solutions (upper bounds) for six benchmark instances in the Employee Scheduling Benchmark Data Sets15 . The space of results of a heuristic selection is far
smaller than that of heuristic generation, as discussed in Section 3.2.5.
14 The first Cross-domain Heuristic Search Challenge, see http://www.asap.cs.nott.ac.uk/

external/chesc2011/.
15 See http://www.cs.nott.ac.uk/∼tec/NRP/.

22

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
In the literature, many heuristic generation methods are based on stochas-

tic learning techniques. Conventional stochastic learning methods belong
to supervised learning, which builds a concise model of the distribution of
class labels in terms of decision attributes (or predictor features) (Kotsiantis
et al., 2006). Typical examples include genetic programming and decision
tree (or rules) learning. Genetic programming constructs a heuristic in an
online evolutionary way (Koza, 1992, pp. 79–120). The latter method typically builds a heuristic in a hierarchical structure of rules that recursively
classify the class labels by construction and pruning (Murthy, 1998). Genetic
programming has been the most common (Burke, Hyde, Kendall, Ochoa,
Ozcan, & Qu, 2010) heuristic generation methodology.
However, many heuristic generation approaches could not generate nondominated new heuristics in comparison to existing ones. For example, as
“the best over all genetic programming work done on TSP”, Keller and Poli
(2007)’s genetic programming hyper-heuristic was limited to very small-scale
(n ≤ 102 ) TSP instances. In comparison, a meta-heuristic (e.g., Helsgaun,
2000) could almost guarantee to find an optimal solution for n ≤ 103 in a few
seconds. There are two main problems that make the feasibilities of many
heuristic generation methods very low:
Computation time: The time cost is, in fact, always crucial for non-dominance
and practicability of any algorithm. (See Figure 1.4)
Generalization: The high cross-domain capability was “the biggest challenge” (Burke, Hyde, Kendall, Ochoa, Ozcan, & Qu, 2010).
The main reason for the first difficulty is probably related to the explosive

1.3. RESEARCH MOTIVES AND BACKGROUND

23

search space of heuristics to generate, especially in large-scale NP-hard
combinatorial optimization problems, see Section 3.2.5. Other reasons for
the first difficulty depend on the learning methods, such as the (probably
possible) non-convergence of the canonical evolution (Rudolph, 1994) for a
genetic programming and the very expensive time for preparing a number of
training examples for a decision tree or rule learning. Reasons for the latter
difficulty also depend on the learning, such as the encoding in some problem
domains (Burke, Hyde, Kendall, Ochoa, Ozcan, & Qu, 2010) for a genetic
programming and the definitions of attributes and mutually exclusive classes
(Bose & Mahapatra, 2001) for a decision tree or rule learning. It should be
noted that the NFL theorems would be included in the second difficulty,
when all the domains were claimed by an algorithm.
Xue, Chan, Ip, and Cheung (2011) presented some ideas for handling
the first difficulty. They proposed a novel sampling method to resolve the
first problem. The supervised class association rules learning was employed
to discover the probability of assignments of being “promising” for each
decision making in choosing two edges starting from a known city. It was
found that the rules learned from a sampled subproblem were generally
consistent with those from the master problem in the Euclidean TSP. The
approach consists of four main steps:
• sampling a proportion (subset of variables with corresponding constraints) with a bounded size as a subproblem,
• solving the subproblem,
• finding new stochastic heuristic rules for the portion, and

24

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM
• applying new heuristic rules back to the given problem.

The sampling stood for selecting a subset of variables from all the given
variables, as well as the corresponding constraints. The bounded size on
the proportion dramatically reduced the time complexity (to O(n log n)).
The experimental results of the effectiveness of the approximate method in
Euclidean TSPs (3.16 × 103 . n ≤ 106 ) seemed very satisfactory. The time
cost of the method was less than a typical iteration of the resulting heuristic
when n ≥ 3.16 × 104 . However, the second problem was not mentioned
in Xue, Chan, et al. (2011). The decision attributes were experience-based,
specific for the Euclidean TSP, and fixed. The limited attributes and finite
values for each attribute made the method a heuristic selection, though
stochastic learning was involved. See Section 3.2.5 for the definitions.

1.4

Objectives and Achievements of Thesis

This study aims at presenting the SPOT method, which is expected to
be a practicable, instance-specific and cross-domain heuristic generation
methodology. Therefore, the difficulties on time cost and generalization
addressed in Section 1.3 would inevitably be encountered. The objectives
are given in Section 1.4.1 in order to try to tackle the two difficulties. Section
1.4.2 presents major achievements of the thesis.

1.4.1

Objectives

The primary objective of this research is

1.4. OBJECTIVES AND ACHIEVEMENTS OF THESIS

25

• To develop an efficient and instance-specific methodology that carries
out on-the-fly supervised learning.
There are four particular supporting objectives to fulfill the primary objective,
including:
• To standardize the input combinatorial optimization problem in the
methodology,
• To regulate a systematic way of compiling decision attributes for various measures in different domains,
• To define an indicator to guide the development procedure and to
predict the effectiveness approximately, and
• To explore effective ways of making use of learned instance-specific
information to generate new heuristics.

1.4.2

Achievements and contributions

All the objectives listed in Section 1.4.1 were achieved in this thesis.
Chapter 3 presents the instance-specific heuristic generation approach SPOT
(Suboptimum- and Proportion-based On-the-fly Training) as an extension of
Xue, Chan, et al. (2011). The standard of the input problem and an indicator
r of “resemblance” are given in Section 3.2.3 and examined in the two domains of the TSP and the FSP. A set of equations for generating new decision
attributes based on domain-specific measures are designed to normalize the
data for learning, as shown in Section 4.3. Applications and experiments on
the TSP and the FSP domains are represented in Chapter 5 and Chapter 6.

26

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

Some plans for modifying existing heuristics, including mutations and local
search, are also illustrated in Chapter 5 and Chapter 6.
Besides the SPOT methodology, another important theoretical contribution is the formulation of hyper-heuristics in Section 3.2.4 and the formal
definitions of heuristic selection and heuristic generation in Section 3.2.5. In
this thesis, the boundary between the two subclasses of hyper-heuristics is
defined on the basis of the countability of the maximum number of possibly
generated heuristics. The formulations, especially the formal distinction
between heuristic selection and heuristic generation, are proposed for the
first time, as far as is known.

1.5

Outline of Thesis

Besides this chapter, there are six more chapters in this thesis. As a continuation of the research background, state-of-the-art of hyper-heuristics,
related meta-heuristics and other approaches are reviewed in Chapter 2.
Among all the approaches, heuristic generation methods are the focus. Chapter 3 proposes the concept of the SPOT method and the supporting standards
and indicators for it. Formal definitions of the combinatorial optimization
problem, hyper-heuristics, the two subclasses of hyper-heuristics and the
SPOT algorithm are also introduced in Chapter 3. An implementation is introduced on the basis of two well-known algorithm development frameworks
in Chapter 4. Two test domains, the TSP and the FSP, are involved in Chapter
5 and Chapter 6, respectively. The whole methodology of developing the
SPOT hyper-heuristic to the two typical combinatorial optimization prob-

1.5. OUTLINE OF THESIS

27

lems can be found in Chapter 5 and Chapter 6. Discussion and concluding
remarks are given in Chapter 7.

28

CHAPTER 1. INTRODUCTION TO THE SPOT ALGORITHM

Chapter 2
Literature Review

Zij die wensen dat de toekomst anders is dan
het verleden, moeten het verleden bestuderen.
If you want the present to be different
from the past, study the past.
Ethica (Ethics)
Baruch Spinoza

Important approaches related to the SPOT algorithm, such as heuristic
generation, sampling and machine learning, are recalled in this chapter. The
families of typical meta-heuristics are reviewed in Section 2.1. Some metaheuristic methods involving sampling techniques are introduced in detail.
Section 2.2 surveys hyper-heuristic approaches, including heuristic selection
and heuristic generation methods. The supervised machine learning, or
classification, and related techniques in machine learning are reviewed in
Section 2.3.

29

30

CHAPTER 2. LITERATURE REVIEW

2.1

Meta-Heuristics

The term meta-heuristic, introduced by Glover (1986), derives from the
composition of Greek εὺρίσκω (heuriskō, to find) and a suffix µετά (meta,
behind or in an upper level). Although this term has been widely adopted,
there is still not an extensively accepted definition. Some definitions from
previous researchers are quoted as follows. It can be synthesized that a metaheuristic usually combines the basic heuristics (such as first-come, first-served)
and dynamically controls searching by feedback in a problem solving.
A meta-heuristic is a heuristic method within which another
heuristic local search procedure can be used at each step. (de
Werra & Hertz, 1989)

A meta-heuristic is a high-level and problem-independent definition of a search strategy that can then be specialized to the specific
optimization problem at hand. (Birattari, 2009, p. 28)

Meta-heuristics are typically high-level strategies which guide
an underlying, more problem specific heuristics, to increase their
performance. . . . Many of the methods can be interpreted as introducing a bias such that high quality solutions are produced
quickly. This bias can be of various forms and can be cast as descent bias (based on the objective function), memory bias (based
on previously made decisions) or experience bias (based on prior
performance). (Stützle, 1998)

2.1. META-HEURISTICS
2.1.1

31

Well-known algorithms

As a subclass of heuristics, meta-heuristics cannot guarantee finding the
optimal solutions. However, various meta-heuristic approaches have been
widely adopted and can be found in thousands16 of published works. Some
early meta-heuristics were developed decades before the proposal of the
term itself, such as the local search (Croes, 1958; Lin, 1965). Comprehensive
surveys of meta-heuristics can be found in C. Blum and Roli (2003) and Gendreau and Potvin (2010). Some well-known meta-heuristics (in chronological
order) are listed as follows.
Local search (or hill-climbing) starts from a feasible solution and repeatedly
replaces it with a better neighbor, until no better neighboring solutions
can be found (Croes, 1958; Lin, 1965). Successors of local search include
iterated local search (Lourenço, Martin, & Stützle, 2003), variable-depth
search (van der Bruggen, Lenstra, & Schuur, 1993) and variable neighborhood search (Mladenović & Hansen, 1997).
Simulated annealing solves problems in a way similar to annealing used by
metallurgists (Kirkpatrick, Gelatt, & Vecchi, 1983; Černý, 1985).
Genetic algorithm is a population-based meta-heuristic with repeated mutation, crossover, and selection (Holland, 1992). The genetic algorithm and
closely related genetic programming (Koza, 1992, pp. 79–120), artificial
immune system (Hunt & Cooke, 1996) and Memetic algorithm (Moscato &
16 Google ScholarTM (http://scholar.google.com/) returned 1,200 results with “metaheuris-

tics” in title, 1,130 for “metaheuristic”, 400 for “meta-heuristics” and 587 for “meta-heuristic”,
on April 23rd, 2012.

32

CHAPTER 2. LITERATURE REVIEW
Cotta, 2003) are also called evolutionary algorithms (Eiben & J. E. Smith,
2003, pp. 15–35).

Tabu search employs a short-term memory to prevent recent moves, an intensification mechanism for focused explorations, and a diversification
mechanism for search balance (Glover, 1986; Glover & Laguna, 1997).
Greedy randomized adaptive search procedures (GRASP) begin with multiple heuristic (greedy random) initializations (Feo & Resende, 1989,
1995).
Ant colony optimization holds a positive feedback mechanism based on the
density of “pheromone” and a “random proportional transition rule”
(Colorni, Dorigo, & Maniezzo, 1992; Bonabeau, Dorigo, & Theraulaz,
1999).
Particle swarm optimization evolves from the collective behavior of an animal school (Kennedy & Eberhart, 1995; Kennedy, Eberhart, & Shi,
2001).
Estimation of distribution algorithms (EDAs) estimate the relations from
the preceding population and manipulating population evolution according to the relations involved (Mühlenbein & Paaß, 1996; Mühlenbein, Bendisch, & Voigt, 1996).
Cross entropy method is a general Monte Carlo approach initially developed
by Rubinstein (1997) for estimating the probability of rare events in
complex stochastic networks, and adopted for solving combinatorial
optimization problems later (Rubinstein, 1999).

2.1. META-HEURISTICS

33

As two well-known meta-heuristics which are closely related to the algorithm in this thesis, EDAs and ant colony optimization usually intends
to model the variables as independent values or values with very low correlations. Some neighborhoods of local search consist of a certain amount
of neighboring solutions with a limited number of variables to be flipped
(for binary ones) or changed. The neighborhoods in this form implicitly
refer to low correlations between variables. In fact, local search methods on
such neighborhoods can be highly competitive in many problem domains.
For example, the Lin-Kernighan-Helsgaun (LKH) (Helsgaun, 2000, 2009)
which is an iterated k-Opt (k ≤ 8 usually) local search kept most of the
records of open benchmark instances17 for TSPs. At the same time, the
greedy heuristics, such as Bentley (1992)’s multiple fragment merging greedy,
could find a solution within about 15% over HK bound very quickly for a
TSP, see DIMACS TSP Challenge. The idea of greedy construction heuristics,
i.e., solving a problem by assigning each variable with a limited number
of values greedily and independently, is also based on the hypothesis of
the independency of variables. The effectiveness of such neighborhoods
(and greedy heuristics) seems to undermine, at least in certain NP-Complete
problems, the hypothesis of high correlations among variables.

2.1.2

Sampling in meta-heuristics

There are a lot of explicit or implicit sampling methods in many metaheuristic approaches, such as the evolutionary algorithms, ant colony op17 See http://www.tsp.gatech.edu/vlsi/summary.html for VLSI TSP Data Set. See http://

www.akira.ruc.dk/∼keld/research/LKH/DIMACS results.html for random and clustered
instances in DIMACS TSP Challenge.

34

CHAPTER 2. LITERATURE REVIEW

timization, particle swarm optimization, GRASP, EDAs and cross entropy
method. A typical example is the EDAs, also known as probabilistic modelbuilding genetic algorithms. EDAs were originally designed as an alternative for evolutionary algorithms (Mühlenbein & Paaß, 1996; Mühlenbein,
Bendisch, & Voigt, 1996). However, EDAs inherit neither crossover nor mutation operators. Instead, the essence of the method is in estimating the
relations from the preceding sample and manipulating sample evolution.
Take the OneMax problem (maximizing the Manhattan norm of a binary
vector ~x) as an example:
n

max k~x k1 = ∑ | xi |

(2.1)

i =1

s.t. xi ∈ {0, 1}

1≤i≤n

(2.2)

The first stage consists in generating an initial sample S0 of a number (nonzero) of individuals by the probability distribution:
n

P0 (~x ) = ∏ P0 ( xi ),

(2.3)

i =1

where P0 ( xi ) is the initial probability of xi = 1. The second stage consists
in selecting “good” (with higher objective values) individuals to form a
good

new sample S0

⊆ S0 . In the third stage, the probability distribution is

re-estimated:
n

P1 (~x ) = ∏ P( xi |S0

good

).

(2.4)

i =1

The three steps can loop until a stop criterion is met. Products in Equations
2.3 and 2.4, an implicit local search-like diversification, require that each

2.1. META-HEURISTICS

35

variable is independent of the others. For those cases in which variables are
dependent on others, Mühlenbein and Mahnig (1999) proposed the factorized
distribution algorithm and Pelikan, Goldberg, and Cantú-Paz (1999) developed
the Bayesian optimization algorithm driven by Bayesian networks. Both of these
two algorithms and their successors inherited the probability-assignments
sampling essence of EDAs.
However, the impact of EDAs in solving NP-Complete combinatorial
optimization problems seemed limited. For example, Ceberio, Irurozki,
Mendiburu, and Lozano (2012) showed that the Edge Histogram Based Sampling Algorithm (EHBSA), one of the best EDAs, could find satisfactory
solutions (0.37% above optima on average) for very small (n < 150) benchmark instances of TSP in 100n iterations. Tsutsui, Pelikan, and Goldberg
(2003) incorporated a 2-Opt local search to EHBSA and showed a very high
chance to return an optimal solution in tens of thousands of iterations for a
TSP with n = 439 cities, under certain configurations of parameters. Nevertheless, the state-of-the-art EDAs seemed not able to compete with other
non-dominated heuristics. The main difficulty when applying the EDAs is
the estimation of the probability distribution. EDAs also have at least four
parameters to be determined: population size, selection size, offspring size
and selection method.
In the ant colony optimization, Colorni et al. (1992)’s roaming ant was
another algorithmic concept close to the individual in EDAs. Ant colony
algorithms employ similar probability-assignments sampling, and some
different probability updating rules as their main mechanisms. A basic ant
colony algorithm called ant system was proposed by Colorni et al. (1992) for

36

CHAPTER 2. LITERATURE REVIEW

solving TSPs. In each iteration, a number of ants traverse the graph and
build complete paths of n edges. For an ant at a city ci , the probability of
choosing a path (ci , c j ), where i , j, is according to the “random proportional
transition rule” (Bonabeau et al., 1999, pp. 42–43):

Pij =





τ ij


α

· d ( ci , c j )


− β

(τ ) · (d(ci , cl ))

 ∑l not visited il
0
α

if c j was not visited (2.5a)

−β

otherwise,

(2.5b)

where α and − β are two parameters controlling the importance of the trail
intensity (τ ij , quantity of pheromone) and the distance18 , respectively. It can
be found that the intensity τ bears a resemblance to the probability P in
EDAs. The difference is that the nearness is also considered in the sampling.
Hence the sampling is closer to the nearness-based neighborhoods of a local
search. A selected ant leaves a certain amount of pheromone ∆τ ij on its entire
course after each iteration, the quantity of which depends on the quality of
the solution found by the ant:

∆τ ij =




Q
Tour length found by the ant

if the ant chose (ci , c j ) (2.6a)



0

otherwise,

(2.6b)

where Q a fixed parameter. A “forgetting-bad-solution” mechanism, evaporation with a factor ρ, was also introduced to update the τ ij to:

(1 − ρ)τ ij +

∑ ∆τij .

(2.7)

all ants

There were also a number of derivations of ant colony approaches. The
18

d ( ci , c j )


 −1

is also known as the visibility.

2.1. META-HEURISTICS

37

ant system in Dorigo, Maniezzo, and Colorni (1991) employed an elitist
ants mechanism, in which the best ant (that which traversed the shortest
path) deposits a large quantity of pheromone, with a view to increase the
probability of the other ants of exploring the most promising solution. In
Gambardella and Dorigo (1995)’s Ant-Q, the rule of local update was the
Q-learning, a reinforced learning. However, no improvement compared with
the ant system could be demonstrated. Besides, even in the opinion of the
authors, this algorithm is not more than a pre-version of the ant colony system.
Dorigo and Gambardella (1997a, 1997b) presented the ant colony system,
and later with a 3-Opt local search, to improve the performance for problems
of higher dimensions, by introducing a few mechanisms:
• A parameter q0 ∈ (0, 1) for a balance between diversification and
intensification for each ant at a city ci :

cnext =




 arg max Pij by Equation 2.5a

if q ≤ q0

(2.8a)

c j ∈C,i,j



the city chosen in the usual way

otherwise (2.8b)

where q ∈ [0, 1] is a uniformly distributed random variable,
• A global update procedure, in which only the best ant would do a
pheromone update in the global update,
• A list of candidates of the closest cities, which was partly inherited
from local search.
In fact, the first mechanism within the ant colony system brought more intensification (focusing on the most “promising” candidate) into the sampling.

38

CHAPTER 2. LITERATURE REVIEW

Stützle and Hoos (1997) proposed a MAX-MIN ant system based on the ant
system and presented some notable differences:
• The values of the trails are limited by τ min and τ max ;
• Only the best ant updates a trail of pheromone;
• The trails are initialized with the maximum value τ max ;
• The updating of the trails is made in a proportional manner, the
strongest trails being less reinforced than the weakest;
• A re-initialization of the trails can be carried out.
The final results are obtained by updating the best-so-far solution with an
increasingly strong frequency, during the running of the algorithm. The first
mechanism within the MAX-MIN ant system, compared to that of the ant
system, made more diversifications in the sampling.
However, the experiments of ant colony optimization approaches only
showed a limited impact in solving NP-Complete combinatorial optimization problems. Early approaches without local search could be easily dominated (Dorigo, Maniezzo, & Colorni, 1991), even in small-scale benchmark
instances. For example, the results of ant colony methods for TSPs were usually around 0.3-2% (Stützle & Hoos, 1997) and were limited within n < 103
due to (relatively) expensive tour construction and a great number of iterations. Some later ant colony approaches employing local search methods
could improve and solve TSP instances beyond 103 cities, such as in Dorigo
and Gambardella (1997a) and L. Li, Ju, and Zhang (2008).

2.1. META-HEURISTICS

39

Another typical example was GRASP. The multi-start of GRASP, usually greedy randomized solutions, can be regarded as a biased Monte Carlo
sampling. One difference between the multi-start of GRASP and the probabilistic selection in EDAs and ant colony approaches is that there is no
learning or evolving for the sampling. Feo and Resende (1995) presented
an example of GRASP with random selection on greedy subsets of assignments. A restricted candidate list was defined for each variable to exclude
the unpromising candidates and a component-by-component construction
heuristic chose solution components randomly from the restricted candidate
lists to form a solution.
Another example was a probabilistic greedy in a multi-agent implementation of GRASP called Dynamic Scheduling Agents with Federation Organization (DSAFO) for an airport ground service scheduling problem at Beijing
Capital International Airport (Fan & Xue, 2006; Xue & Fan, 2007). In DSAFO,
resources such as baggage trucks were divided into groups. A representative
agent was designed to manage a group of resources, to query tasks from
the blackboard and to try to assign the published task with its own resource
according to the earliest finish time heuristic. The tasks were listed one by
one on the “blackboard” according to an earliest due date heuristic. The only
randomness was a priority of communication for each representative agent.
Some researchers also embedded a local search in the GRASP to improve
the results further (Feo & Resende, 1995; Marinakis, Migdalas, & Pardalos,
2005b).
GRASP is flexible and easy to implement. The results of GRASP can be
generally competitive with basic heuristics, such as the DSAFO outperformed

40

CHAPTER 2. LITERATURE REVIEW

the earliest due date heuristic, the earliest ready time heuristic and a MAXMIN ant system in the bi-objective optimization of airport ground service
scheduling. However, the non-dominance was not always guaranteed. For
example, Marinakis et al. (2005b) showed a comparison on solution quality
between an iterated local search, involving GRASP and a 3-Opt local search,
and other fifty four algorithms submitted in the DIMACS TSP Challenge
in some benchmark TSP instances (103 ≤ n ≤ 2.4 × 103 ). Marinakis et al.
(2005b)’s method generally returned an average 1.181% excess over optima
and was ranked as the thirteenth. However, their result was still not able
to be competitive with other iterated 3-Opt local search methods such as
D. S. Johnson and McGeoch (1997)’s (with an average 0.413% excess). In
addition, the time cost was not reported in Marinakis et al. (2005b). Marinakis,
Migdalas, and Pardalos, 2005a presented another GRASP approach and
showed a twelfth rank.
Differing from the methods mentioned above, many other meta-heuristics
that employ sampling treat one solution as an individual and the whole
search space, or a promising part of it, as the population for a given problem.
Different methods were developed to estimate the promising (hopefully
suboptimal) solutions by investigating the promising subset in the sample.
A number of new algorithms have been developed, and a lot of interesting
findings have been achieved to guide and to control the heuristic search
procedure. The resulting meta-heuristics returned, in general, more satisfactory results than conventional simple heuristics. However, for a certain
problem domain, the ad-hoc methods such as a local search algorithm with
a sophisticated and well-designed neighborhood usually were more com-

2.2. HYPER-HEURISTICS

41

petitive with the sampling-based methods, especially in large-scale problem
instances. A possible reason was the distributions of suboptima (or local
minima/maxima) in an NP-Complete problem were extremely irregular and
difficult to estimate. In addition, the decision procedure of the optimal sampling plan is, ironically, another problem to optimize, for example, Raschke,
Krishen, Kachroo, and Maheshwari (2012) discussed a combinatorial optimization model for identification and optimization of sampling in certain
conditions.

2.2

Hyper-Heuristics

A hyper-heuristic is a search method or learning mechanism for selecting or
generating heuristics to solve computational search problems (Burke, Hyde,
Kendall, Ochoa, Ozcan, & Qu, 2010). The term itself was introduced by
Denzinger, Fuchs, and Fuchs (1997) and the definition became well-accepted
by Burke, Kendall, et al. (2003). The prefix was from ὺπέ% (hyper, above or
with many dimensions). In fact, the relation between hyper-heuristics and
meta-heuristics, i.e. “hyper-heuristics versus meta-heuristics”, is not very
clear due to the different definitions of meta-heuristics. For example, hyperheuristics * meta-heuristics could be held in de Werra and Hertz (1989)’s
definition, as quoted in Section 2.1. However, hyper-heuristics ⊆ metaheuristics, which could be held in other definitions such as Stützle (1998)’s,
is assumed in this thesis, as shown in Figure 1.7. What distinguishes hyperheuristics from the meta-heuristics is the search space. A hyper-heuristic
generally solves a problem by calling given or generated heuristics, which
are called low-level heuristics (LLHs), and does not directly manipulate the

42

CHAPTER 2. LITERATURE REVIEW

assignments of variables within a solution. Therefore, the search space of
a hyper-heuristic is a space of combinations of LLHs. On the other hand,
a meta-heuristic can directly manipulate the meta-data in a solution. The
search space of a meta-heuristic can be a space of solutions.
Because of the “off-the-peg” (Burke, Hyde, Kendall, Ochoa, Ozcan, & Qu,
2010) nature, hyper-heuristics are generally easier to be adapted to different
problem domains than meta-heuristics. As a side effect accompanying the
high adaptability, the resulting selected (and some generated) heuristics, in
many cases, were dominated by those meta-heuristics carefully fine-tuned
for a problem domain. However, some hyper-heuristics (mostly heuristic
generation) could be non-dominated and outperformed existing heuristics
such as Chan et al. (2012)’s PHunter hyper-heuristic, as mentioned in Section
1.3. It was believed that one of the main contributors to the new upper
bounds found by the PHunter was a newly introduced LLH which had not
been implemented before or carefully fine-tuned in any other approaches.
If it was proved true, a heuristic generation methodology which improved
existing LLHs could notably increase the competitiveness of a general hyperheuristic.
Burke, Hyde, Kendall, Ochoa, Ozcan, and Qu (2010) surveyed comprehensive hyper-heuristic approaches and classified all the hyper-heuristics
into two categories: heuristic selection and heuristic generation. In this thesis,
the two categories are formally redefined according to the countability of
the maximum space of LLHs, as shown in Section 3.2.5. As a result, both
extensions of heuristic selection and heuristic generation are extended. Because Burke, Hyde, Kendall, Ochoa, Ozcan, and Qu (2010)’s review was so

2.2. HYPER-HEURISTICS

43

general and comprehensive, it is not necessary to repeat every well-known
hyper-heuristic approach in this thesis. The approaches that are critically
important, newly developed or missed in Burke, Hyde, Kendall, Ochoa,
Ozcan, and Qu (2010) are introduced in the following.

2.2.1

Heuristic selection approaches

The algorithm selection problem was first formalized by Rice (1976). The
question in an algorithm selection problem (Xu, Hutter, Hoos, & LeytonBrown, 2008) is: “which algorithm(s) should run now in order to minimize
some performance objective, such as the expected runtime?” Although the
resulting selected algorithm seems more effective than an arbitrary algorithm on average, the algorithm selection problem, itself, is still governed by
Theorem 1 of NFL if it is regarded as a BBO, which means that no specific information is considered. In fact, determination of the best algorithm to solve
one algorithm selection problem is, ironically, another algorithm selection
problem (Smith-Miles, 2008a).
Early hyper-heuristic approaches, similar to the early meta-heuristics,
were proposed far earlier than the term. For example, Crowston, Glover,
Thompson, and Trawick (1963) sampled a set of decision rules and updated
the probabilities of the rules to learn (or converge) to improve the search
process. Crowston et al. (1963) showed the quality of the solution found
by learning was more satisfactory than those by single isolated rules or by
random sampling among the rules. That was also one of the early heuristic
selection applications.
Burke, Hyde, Kendall, Ochoa, Ozcan, and Qu (2010)’s heuristic selection

44

CHAPTER 2. LITERATURE REVIEW

concept referred to the methodologies for choosing or selecting existing
heuristics. In this thesis, heuristic selection is defined as the hyper-heuristics
of which the possible results (space) of LLHs are countable for any problem,
including those with countably infinite variables. Therefore, some research
areas used to be “close” to but outside the heuristic selection in Burke, Hyde,
Kendall, Ochoa, Ozcan, and Qu (2010) are regarded as heuristic selection
methods in this thesis, such as the parameter optimization approaches for
heuristics and the (heuristic) algorithm portfolios. See Section 3.2.5 for more
details.
Tuning parameter set automatically for heuristics could be considered as
early practice in hyper-heuristics. An early example was the use of Rechenberg’s 1/5 success rule for changing the mutation rate by evolution strategies in an evolutionary algorithm presented in 1973 (Burke, Hyde, Kendall,
Ochoa, Ozcan, & Qu, 2010). Some simple heuristics as well as many metaheuristics have a number of parameters to be determined in each problem
domain. Population-based meta-heuristics received much attention, such
as Grefenstette (1986), Eiben, Hinterding, and Michalewicz (1999) and Lobo
and Goldberg (2004)’s different methods of controlling parameters for evolutionary algorithms, respectively. For local search heuristics, an example was
Boyan (1998) and later Boyan and A. Moore (2000)’s parameter optimization
method “STAGE” in seven problem domains, including bin-packing and
the Boolean Satisfiability Problem (SAT). STAGE employed a fixed number
of experience-based attributes of the current solution to estimate the performance of each given local search heuristics by some linear, quadratic,
cubic and other regression methods. The best dynamically estimated local

2.2. HYPER-HEURISTICS

45

search won the turn to run. STAGE was proven effective and efficient and
could be compared to up-to-date heuristics, such as Kim and Wy (2010)’s
results in 1D bin packing. Many recent advances of hyper-heuristics, such as
Mısır, Verbeeck, Causmaecker, and Vanden Berghe (2012)’s approach, which
won the CHeSC 2011 competition, had involved parameter optimization for
manipulating LLHs.
A portfolio of algorithms is a collection of different algorithms and/or different copies of the same algorithm running on different processors (Gomes
& Selman, 1997; Huberman, Lukose, & Hogg, 1997), or equivalently on fractions of CPU cycles on a single CPU. The first solution (for satisfiability-style
problems) or the best solution (for minimization or maximization problems)
is chosen as the result of the portfolio. Xu, Hutter, et al. (2008)’s portfolio of solvers, SATzilla, which won three gold medals, one silver medal
and one bronze medal in the 2007 SAT competition19 , is a typical example.
SATzilla predicts performances (such as computation time) of solvers by tens
of features (or attributes) and determines the best subset of solvers to run.
Chan et al. (2012)’s PHunter was another example. Given a specific problem,
PHunter tests all given LLHs in crossover, mutation and ruin-and-recreate
diversification classes, and constructs three portfolios with the best three
LLHs in each class, respectively. The iterations of the three LLHs are 3, 2
and 1, respectively, for each portfolio. The final combination of portfolios
was determined by pre-trained decision tree prediction according to tens of
attributes. In addition, parallel execution of heuristic and exact algorithms
are also closely related to heuristic algorithm portfolios, such as Denzinger
19 See http://www.satcompetition.org/2007/.

46

CHAPTER 2. LITERATURE REVIEW

and Offermann (1999)’s multi-agent based approach for achieving cooperation between different search-systems and Talukdar, Baerentzen, Gove, and
De Souza (1998)’s asynchronous team of cooperative agents.
Some hyper-heuristics returned previously unknown heuristics, and
hence might look like heuristic generation methods. However, many of
them are actually heuristic selection methods, when the spaces of LLHs
are countable or even finite for an input problem with countably infinite
variables. Even though the result may seem to be new, powerful and difficult to discover for experienced human managers, such a hyper-heuristic is
equivalent to a selection process on ready-to-use LLHs by a one-to-one transformation. An early example was Bernstein (1987)’s method “SH”. SH builds
a population of new heuristics from a carefully chosen vocabulary, collects
statistics on their effectiveness, identifies those that work very well, or those
most likely to improve a bad schedule radically. It is not hard to find that SH
was a populate-and-select style heuristic selection method, while the space of
possible algorithms was limited by a constant number. Vázquez-Rodrı́guez
and Petrovic (2010)’s method searches simultaneously for the best sequence
of 13 dispatching rules and the number of operations to be handled by each
dispatching rule. Vázquez-Rodrı́guez and Petrovic (2010)’s approach was
equivalent to setting up a priority parameter belonging to {0, 1, . . . , 13}20 to
each dispatching rule and a positive integer to a pseudo parameter “length”.
Other instance-specific selection examples that employed machine learning techniques include Koonce and Tsai (2000)’s decision-rule-like heuristic
for dispatching priority in production scheduling, Kwak and Yih (2004)’s
20 Where 0 stands for not selected.

2.2. HYPER-HEURISTICS

47

discovery of the decision tree approach (competitive decision selector) in
production control, X. Li and Olafsson (2005) and later Olafsson and X. Li
(2010)’s decision tree for dispatching jobs, Burke, Petrovic, and Qu (2006)’s
cased-based reasoning heuristic selection method in timetabling problems
and Xue, Chan, et al. (2011)’s class association rule learning in Euclidean
TSPs. There also were data set-specific (or family-specific) examples, such
as Lee, Piramuthu, and Tsai (1997)’s dispatching rule discovery embedded
in a genetic algorithm and Sha and C.-H. Liu (2005)’s discovery of a new
due date assignment heuristic (as a decision tree) for the dynamic job shop
problem, respectively.
In many heuristic selection methods, learning was regarded as the essence
of hyper-heuristics. That was especially obvious in the heuristic discoveries
mentioned above, where machine learning techniques were generally involved (Koonce & Tsai, 2000; Kwak & Yih, 2004; Sha & C.-H. Liu, 2005; X. Li
& Olafsson, 2005; Olafsson & X. Li, 2010). These methods revealed a promising class of applications of heuristic selection (or generation) — tremendous
machine learning algorithms in certain conditions, as shown in Section 2.3.

2.2.2

Heuristic generation approaches

Burke, Hyde, Kendall, Ochoa, Ozcan, and Qu (2010) described heuristic
generation as “generating new heuristics from the components of existing
ones”. However, some heuristic discovery approaches are precisely equivalent to other heuristic selection processes, such as Bernstein (1987)’s populateand-select method shown in Section 2.2.1. Such an algorithm could probably
possibly be misclassified by Burke, Hyde, Kendall, Ochoa, Ozcan, and Qu

48

CHAPTER 2. LITERATURE REVIEW

(2010)’s classification. This section presents a number of typical heuristic
generation approaches, which are so sophisticated and may actually return
one or many LLHs out of an uncountable number of possible LLHs.
A typical example of heuristic generation is Joslin and Clements (1999)’s
“squeaky wheel” optimization. The squeaky wheel implements an order of priority for building components such as the variables waiting for assignments.
There is a certain construction process, such as a greedy heuristic, that builds
a solution according to a given priority ordering. Trouble spots, “bottleneck”
components or the key components to change for a better solution, are identified through analyzing the constructed solution. Then the priority ordering
is updated by giving the trouble spots higher priorities and the next iteration
starts unless the termination condition is met. The LLH space of “squeaky
wheel” optimization is bounded in Θ(n!) generally, where n is the number
of the components (variables usually). However, only a portion of the space
can be reached in practice if the number of the variables is great enough.
Aickelin, Burke, and Li (2009) presented an updated version, evolutionary
“squeaky wheel” optimization, in which selection and mutation operators
were involved between analysis of the solution and prioritization of priority
ordering.
Another example is genetic programming (Koza, 1992, pp. 79–120). Genetic programming is an evolutionary algorithm-based methodology. The
goal of genetic programming is evolving computer programs in a natural
selection way for certain problems. Particularly, the chromosomes in general
evolutionary algorithms are programs. For example, there are two equations
“3 × x + 4” and “1 − 2 ÷ y” selected from a population in a certain iteration.

2.3. SUPERVISED LEARNING

49

The results of a mutation operation may be “3 × x + 5” and “1 − 1 ÷ y”,
respectively. The results of a crossover operator may be “3 × x + 2 ÷ y” and
“1 − 4”. The size of the space of possible programs is bounded in Θ(2n ), if
there are two free “characters” in the vocabulary and the equation is a 1D
string, where n is the maximum size of the program. The size of the space
will increase dramatically, if there are more characters and more complex
forms of equations. Therefore, in practice, some constraints like the “maximum size of program” are usually introduced to restrict the space of LLHs.
It is not necessary for every genetic programming to be a heuristic generation, because there are the programs that can be seen as solutions directly.
Examples of heuristic generation of genetic programming approaches can be
found in Ho and Tay (2005), Jakobović, Jelenković, and Budin (2007) and Tay
and Ho (2008) in production scheduling, Burke, Hyde, and Kendall (2006)
and Burke, Hyde, Kendall, and Woodward (2007) in bin packing, Fukunaga
(2008) in SAT, Keller and Poli (2008) in TSP, and Pillay and Banzhaf, 2007
and Bader El Den and Poli (2009) in timetabling.

2.3

Supervised Learning

Supervised learning, supervised machine learning, or classification, is the
search for algorithms that reason from externally supplied records (or instances) to produce general hypotheses, which then make predictions about
future instances (Kotsiantis, 2007). In other words, supervised learning aims
at building a concise model of the distribution of class labels in terms of predictor features (or decision attributes) (Kotsiantis et al., 2006). An attribute
can be continuous, binary or categorical, but a label usually accepts discrete

50

CHAPTER 2. LITERATURE REVIEW

Table 2.1: Examples of records with known labels
Case T’ien’s horse
(attribute 1)

King’s horse
(attribute 2)

Date
(attribute 3)

Was T’ien’s faster?
(label)

1
2
3
4
5
6
...

high-grade
high-grade
high-grade
middle-grade
middle-grade
middle-grade
...

day 1
day 2
day 3
day 1
day 2
day 3
...

no
yes
no
yes
yes
yes
...

high-grade
high-grade
high-grade
high-grade
high-grade
high-grade
...

or categorical values only. Table 2.1 shows some examples of instances with
known labels from T’ien’s possible records of horse-racing. Supervised learning is considered as a classification problem due to the discrete labels and
the result of a supervised learning is a classifier. The results are usually
classification rules, regression functions or known examples. A possible
classification rule learned from the data in Table 2.1 is “if T’ien’s horse is
high-grade and the king’s is middle-grade then T’ien’s horse is faster.”

2.3.1

Well-known classification techniques

A taxonomy of some well-known classification techniques can be organized as a tree, as shown in Figure 2.1. Classification methods can be
divided into more and more specialized and well formalized methods such
as those on the right side of Figure 2.1. The discriminant function methods,
for instance, subsume recursive partitioning, which further includes three
algorithms: LTREE, CART and C4.5.
Three of the most efficient methods among those listed in Figure 2.1,
Naı̈ve Bayes normal, C4.5 and Ripper, are employed in this thesis. Therefore,

Classification
methods

Discriminant function methods

-a

is

Discriminative
methods
is-a

-a

is

Generative
methods

a

is-a
is-a
is-a
is-a

is-a

is-

is-a

is-a

is-a

is-a

is-a
is-a
is-a

Set covering

is-a

k nearest neighbors

Neural networks

Linear SVM

Nonlinear SVM

C4.5

CART

LTREE

CN2

Ripper

Weka IBk

MLP backpropagation

is-a

is-a

is-a

Naı̈ve Bayes discretized

Naı̈ve Bayes kernel

Naı̈ve Bayes normal

Complement Naı̈ve Bayes

Naı̈ve Bayes multinomial

is-a
is-a

is-a
is-a
is-a
is-a
isa

Support vector machines

Fisher’s linear
discriminant

Recursive partitioning

Logistic Regression

Normal quadratic discriminant analysis

Normal linear discriminant analysis

Naı̈ve Bayes

2.3. SUPERVISED LEARNING
51

Figure 2.1: A tree taxonomy of some well-known classification techniques
(Hilario, Kalousis, Nguyen, & Woznica, 2009)

52

CHAPTER 2. LITERATURE REVIEW

Table 2.2: References for further reading on the classification techniques not
selected in this thesis
Classification method

Reference

Naı̈ve Bayes multinomial
Complement Naı̈ve Bayes
Naı̈ve Bayes kernel
Naı̈ve Bayes discretized
Normal LDA
Normal QDA
Logistic Regression
Fisher’s LD
LTREE
CART
Support vector machines
Neural networks
Weka IBk
CN2

Rennie, Shih, Teevan, and Karger (2003)
McCallum and Nigam (1998)
John and Langley (1995)
John and Langley (1995)
Lachenbruch and Goldstein (1979)
Lachenbruch and Goldstein (1979)
Le Cessie and Van Houwelinge (1992)
Fisher (1936)
Gama (1999)
Breiman, Friedman, Olshen, and Stone (1983)
Suykens and Vandewalle (1999)
G. P. Zhang (2000)
Aha, Kibler, and Albert (1991)
Clark and Niblett (1989)

they are introduced in detail below. There are two criteria to select classification techniques to use. One criterion is the efficiency (or quick running) of
each algorithm. Another is the overall diversity, i.e., to choose from different
categories. Suggested literature for further reading on other classification
techniques is listed in Table 2.2.
Naı̈ve Bayes normal is a probabilistic classifier based on Bayes’ theorem

P( A| B) =

P( B| A) P( A)
.
P( B)

(2.9)

In addition, all attributes (variables) are considered independent (naive) of
each other. Although the assumption of independency seems poor and naive,
Naı̈ve Bayes usually competes well with other sophisticated methods (Rish,
2001). Given attributes ~a = { a1 , a2 , . . . , an } and a class C, the likelihood of C
is
n

P(~a|C ) = P( a1 , a2 , . . . , an |C ) = ∏ P( ai |C ),
i =1

(2.10)

2.3. SUPERVISED LEARNING

53

1. Check for base cases
2. For each attribute a
2.1. Find the feature that best divides the training data such as information
gain from splitting on a
3. Let abest be the attribute with the highest normalized information gain
4. Create a decision node node that splits on abest
5. Recurse on the sub-lists obtained by splitting on abest and add those nodes
as children of node

Figure 2.2: Pseudo codes of the C4.5 classifier
because of the assumption of independency. “Normal” means each attribute
is estimated like a normal distribution, not in a kernel estimator way (John &
Langley, 1995) and there is no supervised discretization to convert numeric
attributes to categorical values. Examples of successful applications of the
Naı̈ve Bayes normal include text classification (McCallum & Nigam, 1998)
and bio-informatics (Yousef et al., 2006). An open source Java implementation
of Naı̈ve Bayes normal provided in Weka (See Section 4.1) is employed, as
described in Section 4.3.
C4.5 generates a decision tree based on the concept of information entropy
(Quinlan, 1993, pp. 17–24). Kotsiantis (2007) summarized the pseudo codes of
C4.5 concisely, as shown in Figure 2.2. The decision tree is created in a greedy
way by iteratively maximizing the normalized information gain on each
node branching. Quinlan later developed improved commercial successors21
C5.0 (for Unix/Linux) and See5 (for Windows). This thesis employs J48,
which is an open source Java implementation of C4.5 provided in Weka.
Cohen (1995) proposed the Ripper method, or repeated incremental pruning
to produce error reduction. There are two main phases in Ripper. The first
phase is a building phase, where two actions, a growing action and a pruning
21 See http://www.rulequest.com/see5-info.html.

54

CHAPTER 2. LITERATURE REVIEW

action, are repeatedly performed. In the growing, conditions (antecedents)
of rules are greedily (with highest information gain) increased until 100%
confidence is met. In the pruning, conditions are pruned in a certain metric.
The second phase is an optimization phase, where the size of the set of
built rules is reduced to fit the training records. JRip, an open source Java
implementation of Ripper provided in Weka is employed in this thesis.

2.3.2

Sampling and attribute selection for supervised learning

The amount of training records may be very large in some applications of
supervised learning. For example, the Network Intrusion Data used in the
1999 KDD Cup22 was raw TCP dump data collected for nine weeks. There
were 41 attributes and one label defined for 4,940,000 training records and
3,110,290 test records. The sizes of the training data and the test data were
743 MB and 430 MB, respectively. Sampling of records is usually introduced
in such large-scale data sets. The goal of sampling is to minimize the sample
size while maintaining the quality of learning results (H. Liu & Motoda, 2001,
p. 6). Reinartz (2002) summarized the most well-known samplings as, similar
to those in the EDAs, two groups:
Random sampling which randomly chooses a subset, and
Stratified sampling which selects the records of the minority class more
frequently in the imbalanced data sets, so that the minority class will
not be ignored due to the vanishing probability.
22 See http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html.

2.3. SUPERVISED LEARNING

55

Characteristics of data are also important in sampling. Typical characteristics range from simple statistics up to information on the data quality
(Reinartz, 2002). The usual way to examine sampling is by comparing the differences in the indicators defined on a classifying procedure and the learning
results, such as computation time, memory cost, correctness and recall, on the
sampled records and on the original data set, such as in Zaki, Parthasarathy,
Li, and Ogihara (1997).
Besides the reduction of records (rows) by sampling, the attributes (columns)
can also reduce some classifier indicators, such as the computation time and
the memory used. An attribute selection (or feature selection) algorithm aims
at finding the optimal subset in a given set of training attributes. An evaluation function or algorithm, that approximately indicates a promising value
for a proposed attribute subset, is usually attached, too. Clearly, an attribute
selection may run up to 2n − 1 times and cost too much time to numerate all
non-empty subsets of given attributes, where n = k~ak is the number of given
attributes. So a termination condition is usually developed. An example
of attribute selection techniques was Hall (1999)’s correlation-based feature
selection, which tended to choose subsets with low correlation internally and
high correlation with the label.
Because the correlations between attributes are usually not zero (partially
or fully dependent), some attribute manipulating methods also generate new
attributes to try to eliminate the correlations before attribute selection, such
as (Markovitch & Rosenstein, 2002)’s FICUS algorithm. This has been called
feature construction or transformation (Kotsiantis, 2007). The generated attributes might lead to more concise and accurate classification and might help

56

CHAPTER 2. LITERATURE REVIEW

people understand some new concepts. In fact, such a transformation can
also be finished by principal component analysis or weighted principal component
analysis (Kriegel, Kröger, Schubert, & Zimek, 2008).
In terms of computation cost, the sampling approaches in supervised
learning are considerably different from those in the meta-heuristics that consider a solution as an individual record. In supervised learning, a sampling
approach generally reduces the number of the records notably and hence
the overall computation time is usually reduced remarkably. For such a
meta-heuristic, a sampling approach commonly brings extra, sometimes tens
of times more, computation cost to prepare the training records to identify the
“good” ones. It could be synthesized that the variables are assumed highly
correlated to others in such a meta-heuristic.

2.3.3

Meta-learning

If finding the optimal classifier for a set of given records could be regarded
as an optimization problem, the choice of the best classification technique
for finding the optimal classifier becomes an algorithm selection problem.
Particularly, such a choice of the best classification technique is called metalearning (Vilalta & Drissi, 2002). The prefix “meta”, which means behind or
in an upper level, is exactly the same as that of meta-heuristics. Vilalta and
Drissi (2002) proposed a framework of meta-learning consisting of a set of
“base-learners” in the lower level and a “meta-learner” in the upper level.
For example, Lindner and Studer (1999) employed case-based reasoning to
govern and select supervised learning algorithms, and Xu, Krzyżak, and
Suen (1992) developed a method consisting of Bayesian formalism for com-

2.3. SUPERVISED LEARNING

57

piling predictions from a number of classifiers. In fact, the “upper level”
controlling processes of meta-learning, similar to those in hyper-heuristics,
are not necessarily machine learning techniques.
Meta-learning is naturally closely related (Smith-Miles, 2008b) to hyperheuristics (especially heuristic selection). For instance, Kanda, Carvalho,
Hruschka, and Soares (2011) proposed a meta-learning framework, which
is actually a hyper-heuristic, to select optimization algorithms for solving
TSPs. Meta-attributes were defined in the problem and the efficiency of the
optimization algorithms examined. The meta-attributes included the number
of cities, the number of edges, the lowest, the highest and the average costs of
edges, and the most frequent edge cost. Four meta-heuristics, including tabu
search, GRASP, simulated annealing and genetic algorithm, were employed
as LLHs in small-scale (n ≤ 100) TSP instances. In fact, Cruz-Reyes et
al. (2012) stated an on-going trend on the shifting of (heuristic) algorithm
selection methods from meta-learning to hyper-heuristics.

58

CHAPTER 2. LITERATURE REVIEW

Chapter 3
The SPOT Hyper-Heuristic

+¥ã ,

g˜€"

See a spot when looking at a leopard
through a bamboo tube.
5-~#r#• 6
(A New Account of the Tales of the World)
Liu, Yiqing

The quoted adage, “see a spot when looking at a leopard through a
bamboo tube,” actually has had two contradictory metaphorical meanings
through long use:
• (Original meaning) having a limited view thus missing the big picture;
• (Transferred meaning) being able to visualize the whole animal.
The SPOT algorithm, which involves sampling proportions, may possibly encounter the contradictory issue, too. This chapter mainly presents the SPOT
algorithm. Standards and indicators are also introduced to explain why and
59

60

CHAPTER 3. THE SPOT HYPER-HEURISTIC

how the presented approach can “visualize the whole” effectively instead
of having “a limited view”. Section 3.1 describes an overview of the SPOT.
Formal definitions of important concepts are given in Section 3.2. Particularly, heuristic selection and heuristic generation are formally redefined in
Section 3.2.5. The SPOT algorithm can accordingly be classified as a heuristic
generation method in general. Section 3.3 presents a typical methodology
for developing the SPOT hyper-heuristic, including the determination of
important parameters. Finally, Section 3.4 provides some discussion on the
presented algorithm and on the formal definitions.

3.1

An Overview of the SPOT Approach

The SPOT (Suboptimum- and Proportion-based On-the-fly Training) approach is a heuristic generation methodology which modifies or creates the
heuristics and/or data structures for solving combinatorial optimization
problems according to the instance-specific information that is learned from
sampled small-scale subproblems and their suboptimum23 . The main framework of the SPOT can be described as “sample-learn-generate”. The SPOT
algorithm employs a standard for unconstrained problems in sampling and
a number of supervised learning methods for learning instance-specific information. The attributive phrase “suboptimum-based” stands for an important
idea in the SPOT — to discover knowledge about suboptima instead of
optima. The attributive phrase “proportion-based” stands for another impor23 In thesis, a proportion (sample) of variables is a finite set, such as 30 cities in a 3,000-city

TSP; a sampled subproblem is a problem (finding a solution for some objective) defined on
a sample of variables, with all the soft constraints defined on the sample, such as finding
the shortest tour for the 30 cities; a suboptimum of such a subproblem is a solution not
exceeding the optimum very much, such as a tour within 105% length of the optimal tour.

3.1. AN OVERVIEW OF THE SPOT APPROACH

61

tant idea, sampling a portion of variables to form a subproblem for further
investigation and learning. Based on the two ideas, the SPOT approach
can estimate all assignments in all suboptima from the suboptima of some
subproblems. Thus the SPOT method can be implemented as an “on-the-fly”
preprocessor within the routine of a heuristic or (possibly exact) algorithm.
The variables, similar to those in Naı̈ve Bayesian methods and in EDAs, are
regarded as relatively independent in the SPOT methodology.
The conceptual infrastructures of the SPOT method and a conventional
hyper-heuristic are compared, as shown in Figure 3.1. There are two algorithmic objects, a set of LLHs and a hyper-heuristic algorithm, in ellipse vertices
in the first figure. The data objects are shown in rectangles. The conventional
hyper-heuristic algorithm dynamically controls the set of LLHs according
to the feedback of some performance attributes. In comparison, the SPOT
method introduces one more data object — a small-scale subproblem (and its
suboptima). In the latter figure, the feedback is not directly compiled from a
feasible solution to the given problem. Instead, the feedback is about the competitions among assignments of variables and collected from the suboptima
of its subproblem. The SPOT algorithm creates or modifies existing LLHs
to generate new LLHs according to the feedback. The generation of new
LLHs is highly associated with the meta-data of the sampled subproblem
and the given problem. Therefore, the SPOT algorithm is an instance-specific
algorithm.
The most important concept in the SPOT methodology is the standard
(or model) of the input combinatorial optimization problem, which is called
U/EA (Unconstrained and with Equinumerous Assignments). A precise

62

CHAPTER 3. THE SPOT HYPER-HEURISTIC

Performance
attributes
(feedback)

A hyperheuristic
algorithm

Se
le
ge ctio
n
ne
rat or
ion

LLH
algorithms
ing

lv
So

A combinatorial
optimization problem
& its solutions
(i) Framework of a conventional hyper-heuristic

g
tin ts
e
n
p
m me k)
o
C ign ac
b
s
as eed
f
(

A small-scale
subproblem &
its suboptima

The SPOT
algorithm

In

sta
n
cre cem ati spe
od on cifi
ifi
ca or c
tio
n

Solving

Sa
m

pl
in
g

LLH
algorithms
g
in

lv
So

A combinatorial
optimization problem
& its solutions
(ii) The framework of SPOT
Figure 3.1: A comparison between the conceptual diagram of the SPOT
methodology and that of the conventional hyper-heuristic framework

3.1. AN OVERVIEW OF THE SPOT APPROACH

63

definition of the U/EA standard can be found in Section 3.2.3. The U/EA
standard regulates the input problem so that there is no inter-variable hard
constraints. The constraint-free condition means that any arbitrary change
of any assignment in a feasible solution still results in a feasible solution.
Another assumption is that the correlation coefficients among the variables
are almost zero. The constraint-free and the correlation-free conditions
enable proportion-based learning and suboptimum-based learning, as briefly
described as follows.
First, some methods of solving a U/EA problem can be approximately
estimated as some classifiers. In this thesis, an assignment (or valuation)
refers to associating a variable (or a set of variables sometimes) with a certain
constant value. Two assignments are called competitors, if their target variables are actually the same while their associated values are different. The
set consisting of all the competitors relating to a variable is called the group
of competitor assignments for the variable. Given a solution, there is usually
one and only one assignment for each variable, thus there is one and only
one “winner” assignment in each group. Such a “winner” assignment is
called an optimum assignment if there is (at least) an optimum that contains
the assignment. Henceforth, the objective of finding a solution in a search
space becomes a classification of n optimum assignments out of n groups. In
fact, the idea of transforming searching in solution space to classification or
estimation of variables is not new. A typical existing example is EDAs.
Furthermore, the classifier can be approximately learned from representative subproblems (proportions) and their optima. Because of the constraintfree and the correlation-free conditions, the classification of the n optimum

64

CHAPTER 3. THE SPOT HYPER-HEURISTIC

assignments can then be approximately trained by a sampled representative
subset of n0 groups, like the ordinary sampling in supervised learning, where
n0  n. The variables, instead of solutions, are regarded as “individuals” in
the sampling, as shown in Figure 3.1. Each assignment generates one row
of a training record. A training record stands for the decision attributes of a
competitor assignment, and the label denotes “whether the assignment is an
optimum assignment or not”. The subset of groups should be representative,
which is also concerned in general sampling as shown in Section 2.3.2. The
optimum assignments in the n0 groups are directly relating to their optima.
The classifier can be approximately learned from the suboptima of representative subproblems, too. Because of the NP-hardness, it could be too
difficult to classify all the n0 optimum assignments correctly and quickly
for a NP-hard subproblem. On the other hand, the number of common assignments between a suboptimum and an optimum, as well as between two
suboptima, should be a dominatingly large fraction due to the correlationfree condition. For instance, changing an arbitrary assignment in an optimum
can probably generate a considerably large number of suboptima. Therefore,
the classification label can be transformed from being optimum assignments
to being suboptimum assignments in the n0 groups. In practice, multiple subproblems can be sampled to reduce the sampling error in certain problems.
The classifier can also be approximately learned on-the-fly, if the size
n0 and the set of decision attributes are well restricted. The time cost of
the whole learning process mainly consists of two parts: (i) the time for
preparing the training records (suboptima) and (ii) the time for training the
classifier. For most algorithms, the first part is highly associated with the

3.2. FORMAL DEFINITIONS

65

size n0 of the subproblem. The latter part is mainly associated with the rows
(number of assignments) and columns (decision attributes) for stochastic
machine learning techniques. If both can be controlled to appropriate scales,
the overall time cost will be well restricted to an on-the-fly level. A possible
useful technique for controlling the columns is the attribute selection which
speeds up learning. However, more time spent on learning usually means
a more precise and more general classifier. Hence a sufficient amount of
time is still necessary for learning. The balance between the precision of
the classifier and the time cost should be carefully considered during the
development of the SPOT hyper-heuristic. Similar decision attributes can
also be found in recent advances of hyper-heuristics, such as Kanda et al.
(2011)’s meta-features and the features of SATzilla (Xu, Hutter, et al., 2008).
Finally, the classifier can be transformed to heuristics. When the classifier
is applied back to the given problem, each assignment can be associated with
the approximate value (probability) of being a “winner” (or (sub)optimum
assignment). Types of the value may include distance, weight, job demand
and penalty of dissatisfactions. Then many, if not all, LLHs can be modified
in their searching behavior with the values. A typical example for local search
heuristics is a reordered search sequence for the (probably reconstructed)
neighborhood. Some completely new LLHs might also be generated according to the values. For example, the most favorite assignment in each group
is chosen to form a solution. It seems that modifying existing LLHs seems
more competitive than creating completely new LLHs in practice.

66

CHAPTER 3. THE SPOT HYPER-HEURISTIC

3.2

Formal Definitions

This section aims at precisely describing the concepts involved in this
thesis and improving the reproducibility of the presented approach. Section 3.2.1 redefines the combinatorial optimization problem, in which the
extension includes the objective problem domains of the SPOT methodology. Section 3.2.2 formulates hyper-heuristics based on the concepts in the
combinatorial optimization problem. The U/EA and the U/EA2 standards
are formatted in Section 3.2.3. The presented SPOT algorithm is defined in
Section 3.2.4. The definitions of heuristic selection and heuristic selection
are given in Section 3.2.5. Accordingly, the SPOT algorithm is regarded as a
hyper-heuristic and a heuristic generation method in general.

3.2.1

Combinatorial optimization problem

The SPOT method aims at solving different domains of combinatorial
optimization problems. A formal definition is extended from C. Blum and
Roli (2003) in this thesis and can be given as follows.
Definition 1 (Combinatorial optimization problem). A combinatorial optimization problem (S, f ) can be defined by :
• a countable set of variables X = { x1 , x2 , . . . , x n }, n ∈ Z+ ;
• discrete variable domains D1 , D2 , . . . , D n ;
• a set of possible solutions V = v1 × v2 × ... × vn , where vi = {( xi , ν)|ν ∈
Di } is the set of all assignments (valuations) for variable xi ;

3.2. FORMAL DEFINITIONS

67

s∈S

f (s)

f ∈ R+ ∪ { 0 }

Search space

Objective function

Objective
function values

Figure 3.2: The mapping from feasible solutions to objective function values
in a combinatorial optimization problem
• (hard) constraints hc1 , hc2 , . . . , hcnc , where hci : V 7→ {0, 1}, 1 ≤ i ≤ nc;
• a set of feasible solutions S = {s|s ∈ V, ∏inc
=1 hci ( s ) = 1};
• a real value objective function f to be minimized24 , where f : S 7→ R+ ∪

{0}.
The main mapping in a combinatorial optimization problem is the objective
function, as shown in Figure 3.2. To solve a combinatorial optimization
problem one has to find a (feasible) solution s∗ ∈ S with minimum objective
function value, that is, f (s∗ ) ≤ f (s) for any s ∈ S. s∗ is called a global optimum
(or globally optimal solution) of the problem; and the set S∗ ⊆ S consisting
of all possible s∗ is called the set of global optima.
Some combinatorial optimization problems are not difficult to solve, especially for those having an optimal substructure. For example, the problem of
finding the shortest path in a finite graph can be solved by Dijkstra (1959)’s
algorithm in O(n2 ) time. However, many other combinatorial optimization problems are in the NP-Complete class. Any NP-Complete problem,
theoretically, can be transformed to any other NP-Complete problem in
(deterministic) polynomial time. For example, Ruiz-Vanoye et al. (2011)
24 As maximizing an objective function f is the same as minimizing − f , this thesis mainly

focuses on minimization problems without loss of generality.

68

CHAPTER 3. THE SPOT HYPER-HEURISTIC

summarized transformations between hundreds of NP-Complete problems,
where many problems were combinatorial optimization problems.

3.2.2

Hyper-heuristics

Hyper-heuristics aim at selecting or generating heuristics to solve problems. The process of selecting or generating heuristics is regarded as a machine learning procedure in this thesis. The definition of the hyper-heuristics
is given as Definition 2, which is partially based on Rice (1976)’s definition
of algorithm selection. A similar mapping model can be found in Ochoa,
Vázquez-Rodrı́guez, Petrovic, and Burke (2009).

~ l ) in the
Definition 2 (Hyper-heuristics). A hyper-heuristic is a 3-tuple ( ag, A,
model of Figure 3.3, where:
• S is the set of search space of a given combinatorial optimization problem as
defined in Definition 1;

~ = {~a1 ,~a2 , . . . ,~ana } is a set of attribute vectors;
• A
~ is a mapping of attribute generation, sometimes ag can return
• ag : S 7→ A
multiple attribute vectors;
• H = {h1 , h2 , . . . , hnh } is a set of low-level heuristics (LLHs);
• l is a machine learning procedure that predicts the best LLH or generates a
new LLH according to all vectors of attributes;
• π : H × S 7→ R+ ∪ {0} is a mapping of performance measure.

3.2. FORMAL DEFINITIONS

69

s∈S

π ∈ R+ ∪ { 0 }

Search space

Performance
measure space
π (h, s)
Performance
mapping

ag(s)
Attribute generation

~ or A
~0 ⊆ A
~
~a ∈ A

l

h∈H

Attribute space

Learning

LLH space

A hyper-heuristic
Figure 3.3: Schematic diagram of a hyper-heuristic
In Definition 2, the learning l is the core of a hyper-heuristic. However,
the resulting LLHs of a hyper-heuristic do not necessarily guarantee their
solutions to be optima. In other words, the best algorithms in the LLH
space that maximize π are usually not guaranteed by a hyper-heuristic. The
hyper-heuristics that returns better LLH h in general are preferred naturally
in optimization practice. If someone considers deciding on an optimum
hyper-heuristic as a new optimization problem, the objective is to minimize

~ and l being three variables to determine.
π with the ag, A
In many heuristic selection approaches, properties of LLHs such as the
computation time were chosen as decision attributes. The values of objective
functions were considered as the labels (or part of labels) for supervised
learning. A typical result of learning may look like “if a particular condition
is met, return the first LLH; otherwise return the second LLH.” It should be
noted that hyper-heuristics can optimize continuous optimization problems.
However, the problems are assumed as combinatorial optimization in the

70

CHAPTER 3. THE SPOT HYPER-HEURISTIC

following, because of the objective domains of the SPOT method.

3.2.3

The U/EA and the U/EA2 standards

A U/EA (Unconstrained and with Equinumerous Assignments) standard
(or model) is proposed for the combinatorial optimization problem to enable
the approximation in the SPOT algorithm. There are two restrictions in the
U/EA standard. The first restriction is that there are no inter-variable hard
constraints. In other words, changing the assignment of each variable is unconstrained by others. The other restriction is the equinumerous assignments,
which means the domain of each variable has the exactly same cardinal
number.
Definition 3 (Equinumerous assignments). Two variables xi and x j in Definition
1 have equinumerous assignments, if and only if k Di k = k D j k, i , j, xi , x j ∈ X.
Definition 4 (U/EA problem). A combinatorial optimization problem in Definition 1 is called Unconstrained and with Equinumerous Assignments (U/EA), if
and only if:
1. there are no hard constraints among different variables in the given problem;
and
2. all variables have equinumerous assignments in the given problem.
Definition 5 (U/EA transformation). In the model of Figure 3.4, the injective
function T from the solution space of a combinatorial optimization problem is a
U/EA (Unconstrained and with Equinumerous Assignments) transformation, if
and only if the image of T fulfills the definition of U/EA in Definition 4.

3.2. FORMAL DEFINITIONS

71

s∈S

T (s)

σ∈Σ

Search space

An injective
function

A new
search space

Figure 3.4: The mapping of U/EA transformation from the search space of a
combinatorial optimization problem
As a mapping, a U/EA transformation can be either an equivalent (usually bijective) transformation or a relaxation (usually injective) technique. In
general, relaxation is much easier to implement in practice. For example,
one naive and effective relaxation technique is directly removing all the
constraints among the variables.
An example of a U/EA problem can be given in the TSP domain. The
solution (with n edges) to a TSP is required to be a Hamiltonian cycle (connected graph) as a hard constraint. A possible relaxation is “to find n edges
to form a graph so that each city has a degree of 2”. In the relaxed model, the
graph is still with n vertices and n edges, but the connectivity of the graph is
not necessarily held.
Another standard (or model) called U/EA2 is presented to regulate the
attributes in the SPOT algorithm further. In the U/EA2 standard, each

~ has the same number of columns (dimensions), and
attribute vector ~a ∈ A
all the attributes in one column are consistent and have the same units.
Definition 6 (Equinumerous attribute vectors). Two attribute vectors ~ai and

~a j are called equinumerous if and only if they have equal dimensions (amount of
columns) over a certain field, i.e. dimR ~ai = dimR ~a j .
Definition 7 (U/EA2 ). An attribute space in Definition 2 is called Unconstrained
and with Equinumerous Assignments and Equinumerous Attributes (U/EA2 ), if

72

CHAPTER 3. THE SPOT HYPER-HEURISTIC

and only if:
1. all the attribute vectors are generated from a U/EA combinatorial optimization problem; and
2. all the attribute vectors are equinumerous attributes.

3.2.4

The SPOT hyper-heuristic

Definition 8 (SPOT hyper-heuristic). The SPOT hyper-heuristic is a 7-tuple

~ l ) as shown in the model of Figure 3.5, where:
(sam, S0 , T, Σ, ag, A,
• sam : S 7→ S0 is a mapping of sampling which returns a small-scale (n0  n)
subproblem;
• S0 is the solution space of the subproblem;
• T : S0 7→ Σ is a mapping of U/EA transformation;
• Σ is the solution space of the transformed problem, kΣk  kSk;

~ = {~a1 ,~a2 , . . . ,~ana } is a set of attribute vectors, where each vector stands
• A
for one specified situation for one possible assignment competing with all other

~ covers all the possible situations, na ≥ n  n0 ; and
competitors, and A
0

~ ∑n Di is a mapping of assignment-based attribute generation
• ag : Σ0 7→ A
0

which generates a set of ∑n Di (as many as the size of all possible assignments
in the subproblem) attribute vectors from one solution, the attributes in each
vector are fully determined by the situation in regard to the competition of an
assignment.

3.2. FORMAL DEFINITIONS

73

s∈S
(Size=n)
Search space
sam(s)
Sampling
s0 ∈ S0
Search space of
a subproblem

(Size=n0  n)

T (s0 )
U/EA transformation
σ∈Σ

π ∈ R+ ∪ { 0 }

A U/EA
search space

Performance
measure space

ag(σ)
Assignment-based
attribute generation

π (h, s)
Performance
mapping

~0 ⊆ A
~
A

l

h∈H

U/EA2
attribute space

learning

LLH space

The SPOT hyper-heuristic
Figure 3.5: Schematic diagram of SPOT hyper-heuristics
The other notions are the same as those in Definition 2.

Theorem 3. The SPOT algorithm is a special case of hyper-heuristics.
Proof. Define a mapping ag0 (s) = ag ( T (sam(s))), then the SPOT hyper-

~ l i is a hyper-heuristic h ag0 , A,
~ l i.
heuristic hsam, S0 , T, Σ, ag, A,



74

CHAPTER 3. THE SPOT HYPER-HEURISTIC

~ stands
In the SPOT hyper-heuristic, each possible attribute vector ~a ∈ A
for a particular possible situation of an assignment competing with all other
competing assignments. The returned algorithm h can generally assign
each attribute vector ~a with a prediction of a (computable, usually finite)
real-valued value γ ∈ Γ = [0, 1], where Γ is the set of possible predictions.
An example is the “likelihood of appearing in suboptima”. Therefore, the
general cardinal number of H is k H k = kΓkna . In other words, there could
be as many as kΓkna possible LLHs in total. In this thesis, na ≥ n is assumed
to be true. See Section 3.4 for more details on this assumption.

3.2.5

Heuristic selection and heuristic generation

Distinguishing heuristic selection and heuristic generation is actually not
easy due to various transformations (and equivalences) between different
problems. For example, the parameter optimization methods mentioned in
Section 1.3 can be regarded as complicated algorithm selection procedures.
An example of parameter optimization is “to determine two 0-1 integer parameters for an algorithm”. There are four possible configurations, including
(0, 0), (0, 1), (1, 0) and (1, 1). The problem can be seen as selecting a single
best algorithm from four. An algorithm generation approach, on the other
hand, seems difficult in becoming an equivalent selection procedure. An
example of algorithm generation is “to find the best propositional formula on
n (Boolean) propositional variables for fitting a Boolean function”. There are
2n irreducible formulae standing for 2n different functions. If every function
is expected to be tried in a test-and-select fashion, the problem can be seen
as selecting the best algorithm from 2n algorithms.

3.2. FORMAL DEFINITIONS

75

A meaningful boundary between selecting an algorithm from four and
selecting one from 2n is the computability of computers. Because the size n of
a combinatorial optimization problem is countable, as defined in Definition
1, i.e., n ≤ ℵ0 . With the growth of n, the LLH space of a generation method
grows explosively and becomes incomputable so that it cannot be enumerated and tested singly on a Turing machine. “Generation” or “make-up”
behavior, instead of the test-and-select method, is necessary for the feasibility
of returning heuristics in practice, and is called “generation”. Otherwise, an
exact (full) exploration of the LLH space must cost explosive time25 .
Definition 9 (Heuristic selection). A hyper-heuristic in Definition 2 is a heuristic
selection algorithm in a domain of combinatorial optimization problems, if and only
if k H k ≤ ℵ0 can be held for all the problems in that domain, necessarily including
those with n = ℵ0 , where H is the LLH space of the hyper-heuristic.
Definition 10 (Heuristic generation). A hyper-heuristic in Definition 2 is a
heuristic generation algorithm in a domain of combinatorial optimization problems,
if and only if k H k > ℵ0 can be held for no less than one problem with n = ℵ0 in
that domain, where H is the LLH space of the hyper-heuristic.
Typical examples of heuristic selection algorithms and heuristic generation algorithms can be found in Section 2.2.1 and Section 2.2.2, respectively.
The definitions given in this chapter clearly show the intensions and outline
the extensions of heuristic selection and heuristic generation. The definitions
are compatible with previous research, such as Burke, Hyde, Kendall, Ochoa,
Özcan, and Woodward (2010). In fact, every heuristic selection algorithm
25 This could probably be a reason why there was no exact algorithm generation method

in literature, as far as concerned. See Figure 1.7.

76

CHAPTER 3. THE SPOT HYPER-HEURISTIC

can be equivalently transformed to the optimization of a parameter tuning
problem with finite independent parameters according to the definitions,
while every heuristic generation algorithm can be equivalently transformed
to the optimization of a parameter tuning problem with infinite independent
parameters. See Section 3.4 for more details.
Definition 11 (Turing (1936)’s computable number). A number is computable
if it differs by an integer from the number computed by a circle-free machine.
Lemma 1. The cardinal number of the set C of all computable numbers is ℵ0 .
Proof. There are countably many Turing machines. Thus C is countably
infinite at most (kCk ≤ ℵ0 ). The set of all integers is a subset of the set
of all computable numbers (Z ⊆ C). Thus kCk ≥ kZk = ℵ0 . Therefore,

kCk = ℵ0 .
For example, all rational numbers,



√

2, log5 8, π and e are computable

numbers. However, in fact, most irrational numbers are incomputable. It
should be noted that incomputable irrational numbers are rarely involved in
the applications and research of combinatorial optimization. Therefore, algorithms involving incomputable numbers are not discussed in the following.
Turing (1936), Davis (1965), Ko and Friedman (1982), Soare (1996) provide
more information about computability and computable numbers.
Theorem 4. An algorithm selection method for heuristics with a finite (or countably
infinite) set of ready-to-use heuristics is a heuristic selection algorithm.
Proof. The LLH set H is the same as the set of given ready-to-use heuristics.
Therefore, k H k ≤ ℵ0 .



3.3. METHODOLOGY OF DEVELOPING A SPOT HYPER-HEURISTIC 77
Theorem 5. A parameter optimization method for heuristics with a finite number
of computable real parameters is a heuristic selection algorithm.
Proof. Let a be the cardinal number of all parameters, k H k ≤ kCk a ≤

(ℵ0 ) a = ℵ0 .



Theorem 6. The SPOT hyper-heuristic is a heuristic generation algorithm.
Proof. The SPOT hyper-heuristic is a hyper-heuristic (Theorem 3). It is clear
that na ≥ n is assumed to be true26 . For the set of possible predictions for
each attribute vector, kΓk ≥ 2 can be generally held, where Γ is the set of
possible values for each attribute vector in the learning result. The cardinal
number of H is k H k = kΓkna ≥ 2na ≥ 2n = 2ℵ0 > ℵ0 , for all n = ℵ0 .



The incomputability (or undecidability) involved in heuristic generation
algorithms, in fact, has two sides. On one hand, the incomputability of the
(very large-scale) LLH space makes it possible to return new results that
have not been unexamined by previous researchers. On the other hand, the
incomputability brings difficulties (mainly on complexity) in the practice of
optimization. The incomputability can, sometimes, be constrained manually
so that a “promising” part of the LLH space can be focused. One example
of restrictions to the “squeaky wheel” optimization is “to determine the top
five ‘bottleneck’ spots” (in Θ(n5 ) class, where n is the number of variables)
instead of “determining the full priority sequence” (in Θ(n!) class) in each
iteration. Clearly, the new constrained “squeaky wheel” optimization focuses
on a “promising” part and becomes a heuristic selection according to the
definitions.
26 See Section 3.2.4 and Section 3.4.

78

CHAPTER 3. THE SPOT HYPER-HEURISTIC

3.3

Methodology of Developing a SPOT Hyper-Heuristic

In general, there are two major phases in developing the SPOT hyperheuristic: the design phase and the run phase. In the first phase, the input
problem is remodeled to meet the standards of U/EA and U/EA2 . Domainspecific parameters, including numeric and non-numeric parameters, are
determined and fine-tuned. The plans for applying the results of supervised
learning are also designed in the first phase. The latter phase is much more
straightforward. Given an input problem, the SPOT hyper-heuristic can
generate a number of new LLHs for solving it. The instance information is
learned in the latter phase.

3.3.1

The design phase

There are three phases in developing a SPOT hyper-heuristic, as shown
in Table 3.1. A seemingly unusual thing is that the development should stop
in certain cases after the first phase. That means the SPOT algorithm cannot
handle the input problems well, because of three possible reasons:
• The input problems do not meet the requirements of the U/EA standard;
• The attributes for supervised machines learning do not meet the requirements of the U/EA2 standard; and/or
• The correlations among variables are too high so that the suboptimumand proportion-based learning always fails to approximate the correct
classifiers to identify (sub)optimum assignments.

3.3. METHODOLOGY OF DEVELOPING A SPOT HYPER-HEURISTIC 79

Table 3.1: Main phases in the design phase in developing a SPOT hyperheuristic
Phase Brief content
P1
P2
P3

3.3.1.1

Successful results

To do when fails

To try to transform the
A U/EA problem &
To quit for other methods
given problem
a U/EA2 attribute space
To determine parameters
A set of parameters
To try different plans
To modify or to create LLHs New LLHs
To try different plans

P1: Transformations

In the first phase (P1), the main concern is on transformation. One transformation aims at converting an input combinatorial optimization problem
to a U/EA problem. This step can be omitted if the given problem has occasionally met the two requirements of U/EA. The other transformation aims
at determining decision attributes and generating one attribute vector for
each possible assignment. In certain cases, the two transformations cannot
be conducted, which means the SPOT algorithm would not be able to handle
the problem well.
Another requirement besides the two standards is the low-correlation
assumption. An indicator r : S × S 7→ [0, 1] of “resemblance” is introduced
to measure the percentage of same assignments of two different solutions.

r (s, s0 ) =

Number of same assignments
.
n

(3.1)

According to the assumption, the r value between any two suboptima of a
given instance should be close to 1. Furthermore, the average resemblance of
every (different) problem instance in the same domain should be very close
to 1, in the ideal cases.

80

CHAPTER 3. THE SPOT HYPER-HEURISTIC

Definition 12 (Compatibility of U/EA problems). A domain of problems which
meet the U/EA standard are called “compatible” with the SPOT hyper-heuristic,
if and only if for given a set {r̄1 , r̄2 , . . .} of average resemblances of suboptima
of different instances (one average resemblance for each instance) and the overall
average resemblance r̄ for all instances, the following conditions can generally be
held:
• (Generality) r̄i ≈ r̄ j ≈ r̄, where r̄i , r̄ j ∈ {r̄1 , r̄2 , . . .},
• (Approximability) r̄ ≈ 1.
A high r̄, such as 0.8, denotes the assignments in suboptima are hopefully
well approximable through investigating the subproblems and suboptima.
A low average resemblance (e.g., 0.15) means it is probably not reliable to
approximate in this way. A high level of generality means that the difference
between learning from a problem instance and learning from its subproblems
is small. A high level of approximability denotes the difference between
learning from optima and learning from suboptima is small. The compatibility of the U/EA standard enables27 feasible and reliable suboptimumand proportion-based learning. In many perturbative meta-heuristics, such
as the Lin-Kernighan local search for TSP mentioned in Section 2.1.1, new
suboptima are expected to be reached from a known suboptimum by modifying a few assignments. A high level of compatibility is usually closely
related to a domain with successful perturbative methods. Generally, either
a low approximability or a low level of generality, or both, denote that the
27 In fact, the compatibility could be the assumption of the proposed methodology. How-

ever, it is introduced as a “soft” indicator other than a “hard” assumption for the purpose of
a friendly cross-domain foundation.

3.3. METHODOLOGY OF DEVELOPING A SPOT HYPER-HEURISTIC 81
input problem is barely compatible with the assumptions of the SPOT hyperheuristic. The developer should try another transformation or change to
another algorithm.
The other standard to be determined in P1 is the U/EA2 model. A set of
attributes needs to be designed to represent the competition features of the
assignments, so that the sampled subproblem and its suboptimum can be
transformed into training data. A label column is determined by the training
suboptimum. The trained classifiers can predict the unlabeled assignment
in the given problem. The principle of the design of this model is to include
enough information for supervised learning. The state-of-the-art supervised
learning can precisely remove the useless and redundant decision attributes
in the results. It is usually not necessary to validate attributes design carefully
with indicators.

3.3.1.2

P2: Parameter determination

The second phase (P2) aims at determining numeric and non-numeric
parameters. There is one important numeric parameter and one important
non-numeric parameter to determine and to fine-tune the approach. Both are
sampling parameters. The numeric parameter is the size n0 of the subproblem
to be sampled. The non-numeric parameter is the method of sampling. With
the two parameters, a subproblem can be generated. Other parameters are
“secondary” parameters, for example,
• an effective problem algorithm28 which can find one or more suboptima
28 Determination of an algorithm can also be regarded as a determination of a non-numeric

parameter.

82

CHAPTER 3. THE SPOT HYPER-HEURISTIC
for the subproblem and its parameters, and
• a supervised machine learning method and related parameters to discover instance-specific knowledge.
Above parameters are called “secondary”, because they usually have been

well studied. The effective methods and suggested values of parameters are
well-known in many domains. In development, the secondary parameters
should be determined before the two important parameters of sampling.
Some secondary parameters, such as the problem solving algorithm, can
be strongly domain-dependent. The design and fine-tuning of the domaindependent secondary parameters can be referred to the domain literature.
The machine learning is relatively independent of the problem domain, if
the training data (including decision attributes) is conducted for a generalpurpose machine learning. In this thesis, three machine learning methods,
including C4.5, Naı̈ve Bayes and Ripper, are employed together to discover
knowledge simultaneously. The three methods are very efficient according
to literature and come from the three main branches of supervised learning
techniques, as shown in Figure 2.1. The final result of classification is the
average value of the three methods for each unlabeled row (assignment). It is
believed that the average result of the three is more robust. The parameters
for the three methods are borrowed from suggested values by Weka, as
shown in Section 4.3.1. The result of machine learning can predict “what
assignment ought to be in optima or suboptima”.
In P2, the parameter n0 is determined and fine-tuned first. The way
method sampling is set to “random” selection, having an average effective
n0 in general. A verification of this development can be found in Section 6.5.

3.3. METHODOLOGY OF DEVELOPING A SPOT HYPER-HEURISTIC 83
With determined standards and secondary parameters (inclining subproblem
solving and learning), one most favored assignment can be identified from
a group of competitors for each variable. The resemblance r is used to
measure the resemblance between the identified most favorite assignments
and a known suboptimum. A higher level of r denotes, hopefully, more
improvements by the final generated LLHs. Different values of n0 can be
tested for selecting one value that maximizes r. Generally, when n0 increases,
r increases and the time cost of the subproblem solving and the learning
increases (dramatically), too. Therefore, the developer should balance the
size n0 to make sure r is high enough, and the time cost is limited. In practice,
if n0 is set to a constant value, the overall time cost of the SPOT heuristic
generation will probably be in polynomial-time.
Random selection should generally be a feasible and effective sampling
method. The reason lies in the correlation-free condition among the variables in the U/EA standard. In random selection, a proportion of the given
problem is made in terms of the designed group randomly29 . Furthermore,
certain sampling methods can also be investigated to fine-tune the sampling
and the classifiers, especially when some inter-variable hard constraints have
been ignored in P1. Examples are a cropped area of cities in an Euclidean TSP,
and a cropped (nearest) segment of job permutations. Different sampling
methods can be tested. One or more methods can be selected to maximize
the indicators r in general. It should be noted that some groups of competitor
assignments would partially change after the sampling.
29 The random choices were based on the pseudo random number generator in computer

programs. Therefore, the generated numbers were not completely random in theory.

84

CHAPTER 3. THE SPOT HYPER-HEURISTIC

3.3.1.3

P3: Heuristic modification

The third phase (P3) is the “last mile” and aims at transforming the
instance-specific results of learning (classifiers) back to LLHs. There basically
are two main ways of generating LLHs: modifying and creating. Modifying
LLH means that the generated LLH is fully or partially based on existing
heuristics. If taking a heuristic as a search sequence that traverses a limited
scope of the search space, one can principally modify the heuristic in two
ways:
• to change the search scope and
• to change the search sequence.
An example for the first is improving the 2-Opt local search to the 3-Opt
local search in the TSP domain. An example of the latter is reordering the
Lin-Kernighan local search for the TSP, which originally prefers to try highpriority (short edges) swaps first. A modified version prefers to try the most
favored edge determined by learning results first. Creating LLH denotes
that the generated LLH is fully constructed by the classifiers. A naive (but
sometimes ineffective or infeasible) LLH is: to select a variable randomly,
to choose the most favorite assignment from the group regarding the hard
constraints, and to repeat until a solution is built. The indicators in P3 are
the traditional solution quality and the time cost. According to experience,
modifying methods are more efficient than creating in general, therefore, the
thesis focuses on modifying LLHs.

3.3. METHODOLOGY OF DEVELOPING A SPOT HYPER-HEURISTIC 85
In conclusion, the first and the last phases are about transformation. The
two phases are relatively difficult to handle. There are only two main parameters to be determined, one after the other, in the second phase. The expected
final performance can be measured by a numeric indicator. Therefore, the
second phase is relatively easier. Some important components, such as the
standards and subproblem solving algorithm, are domain-specific. In other
words, the domain-specific parameters must be redesigned when the SPOT
hyper-heuristic is applied to a new domain.

3.3.2

The run phase

It is more straightforward to solve a problem than to design and fine-tune
the SPOT algorithm. The procedure of solving can be briefly described as
follows. Given a problem P, There are four phases for generating new LLHs:
• To start with a problem P and its unlabeled attributes about assignments,
• To sample for a subproblem P0 with determined size n0 and to solve the
subproblem efficiently to a suboptimum with a determined algorithm,
• To generate labeled attributes referring to the suboptimum,
• To learn instance-specific knowledge (as classifiers) on what assignments are likely to appear in suboptima with three methods, C4.5,
Naı̈ve Bayes and Ripper, to predict the labels of unlabeled attributes in
phase 1., and to modify certain heuristics for new LLHs according to
the labels.

86

CHAPTER 3. THE SPOT HYPER-HEURISTIC

It should be noted that the last step consists of a process of finding instancespecific knowledge. Therefore, the generated LLHs become instance-specific
algorithms and might not be useful in solving other problem instances.

3.4

Discussion

There are some similarities between the proposed SPOT algorithm and
the methodology in Xue, Chan, et al. (2011). For example, they have similar
concepts in the sampling subproblem and suboptimality, parameters, indicators and heuristic generation processes. Both generate instance-specific
algorithms. The main difference is that Xue, Chan, et al. (2011) presented
a heuristic selection approach, though the instance-specific information was
obtained by a class association rules learning. There were fourteen primafacie decision attributes and one label defined. Each attribute has a finite
set of values. The classifier was in the form of class association rules with
confidences. The number of possible classifiers was countable. Thus, the
number of possible LLHs was countable. The method in Xue, Chan, et al.
(2011) was also specialized in regard to the Euclidean TSPs. In comparison,
the SPOT algorithm employs a semi-automatic attribute generation (See Section 4.3.1) and three different supervised learning methods for cross-domain
development. The learning module in the SPOT hyper-heuristic is expected
to be more open and robust.
The idea of dividing (sampling) and analyzing a proportion for solving
a given problem is not novel. It can be found in many heuristic and exact
algorithms, such as the first-come, first-served. It can also be found in the

3.4. DISCUSSION

87

early days of differential calculus, ancient art of war and activities even earlier.
It is a natural thinking in problem solving. In this thesis, a standard U/EA is
given to regulate the input for the sampling proportion. The standard and
another property, compatibility, guarantee high approximability and low
correlations of variables for sampling.
If a given domain holds a good approximability to the U/EA standard,
heuristic (or at least perturbative) algorithms and their decision attributes
should not be too difficult to find in literature. The initial raw attribute can
be set to a possibly maximal set. The attribute selection procedure in SPOT
can help in choosing the best raw attributes for your domain.
It should be noted that the number na of all possible attribute vectors
is determined by the definition of the attributes. Generally, there are three
cases when n = ℵ0 , listed as follows:
• If there is at least one attribute having countably infinite possible values,
such as an integer or a real number, then na must be countably infinite.
• If there are a countably infinite number of attributes, and each attribute
has a finite definition domain, then na must be countably infinite.
• If the number of attributes is finite and the definition domain of each
attribute is also finite, such as three Boolean attributes, then na must be
finite.
The inequality na ≥ n may not be held in the last case. In fact, a natural
attribute for an assignment in an NP-Complete problem usually has no
less than n or infinite possible values, such as the distance in TSP and the

88

CHAPTER 3. THE SPOT HYPER-HEURISTIC

processing time in FSP. In addition, attributes are defined to distinguish the
scenario of the problem so that some classifiers can be built. Therefore, a
design in the last case probably cannot meet the demand of resolution and
distinguishability for NP-Complete problems. Therefore, na ≥ n is assumed
to be true in this thesis.
The definitions of heuristic selection and heuristic generation are consistent with the existing taxonomy in the literature, such as Burke, Hyde,
Kendall, Ochoa, Özcan, and Woodward (2010) in most cases. In fact, the
extensions of both concepts are extended. For example, the parameter tuning
techniques which used to be out of hyper-heuristics are regarded as heuristic
selection in this thesis. It should be noted that the two concepts are only
distinguishable on problem domains that can be extremely large (n = ℵ0 ).
That means, if a problem domain contains a strictly finite (or limited number
of) variables, any hyper-heuristic on the domain cannot be considered as
either heuristic selection or heuristic generation in this study. However, this
situation is rare in typical combinatorial optimization problems. Therefore,
the issue is not addressed in detail, and further research could be carried out
in the future. A possible plan is to define the hyper-heuristic approaches on
an always finite problem domain with finite variables as heuristic selection,
according to computability.

Chapter 4
Java Implementation of the SPOT HyperHeuristic

Die Philosophie ist keine Lehre, sondern eine
Tätigkeit.
Philosophy is not a body of doctrine but
an activity.
Tractatus Logico-Philosophicus
Ludwig J. J. Wittgenstein

The design and implementation of the proposed SPOT algorithm are
presented in this chapter. Two algorithm development libraries, HyFlex and
Weka, are briefly introduced in Section 4.1. Section 4.2 represents the design
classes and public functions in the SPOT algorithm. The implementation
of two cross-domain classes and their functions, such as the training of
classifiers and the general procedure of problem solving, are detailed in
Section 4.3. Section 4.4 presents some overall discussion.

89

90

CHAPTER 4. JAVA IMPLEMENTATION

4.1

Supporting Libraries

This section introduces two algorithm development libraries: HyFlex and
Weka. HyFlex is a cross-domain hyper-heuristic development library and
Weka is a machine learning software and machine learning development
library. Both are written in Java language and both support cross-platform
development.

4.1.1

HyFlex

HyFlex30 (Hyper-heuristics Flexible framework) is a Java cross-domain
hyper-heuristic development platform (Ochoa, Hyde, et al., 2012). HyFlex
was also the competition platform in the hyper-heuristic competition CHeSC
2011 which focused on heuristic selection approaches. This thesis adopts
the same version of HyFlex as in CHeSC 2011. There are two main parts
designed in HyFlex for general heuristic selection approaches (Ochoa, Hyde,
et al., 2012):
• cross-domain part of hyper-heuristic algorithms, and
• domain-specific part of LLHs.
There is a “domain barrier” assumed between the two parts. The crossdomain part, which is capable of calling ready-to-use LLHs through normalized public functions, is considered the active part. The domain-specific part
is considered the passive part. Particularly, there are two optional parame30 See http://www.asap.cs.nott.ac.uk/external/chesc2011/hyflex description.html.

4.1. SUPPORTING LIBRARIES

91

ters, depth of search and intensity of diversification, for each LLH. An LLH
may use none, one or two of the parameters.
Two super classes are defined as interfaces for the hyper-heuristics. One
is the HyperHeuristic which aims at encapsulating the data and operations
in the cross-domain part. Another is the ProblemDomain which emphasizes
on including the domain-specific LLHs and their data structures. The data
structures are completely inaccessible for the HyperHeuristic class. A hyperheuristic approach implemented in HyFlex can interact with the domainspecific LLHs through public function calls. Public interface functions in
HyperHeuristic are mainly designed for supporting high-level strategies,
such as setting a time limit, tracing changes of objective values and getting the
best-so-far objective values. Public interface functions in the ProblemDomain
include getting and setting the parameters, loading problem instances, listing
and calling available LLHs and initializing and copying solutions in the
memory. More details about HyFlex can be found in Ochoa, Hyde, et al.
(2012).

4.1.2

Weka

Weka31 (Waikato Environment for Knowledge Analysis) is a machine
learning software and a development library in Java (Witten & Frank, 2005).
The version used in this thesis is 3.6. Weka contains many well implemented
algorithms for data analysis and machine learning, such as most of the
supervised learning methods shown in Figure 2.1. Weka also provides
visualization tools and graphical user interfaces to make the algorithms easy
31 See http://www.cs.waikato.ac.nz/ml/weka/.

92

CHAPTER 4. JAVA IMPLEMENTATION

to use. The algorithms in Weka accept training data and test data as tables
of values, where each column stands for a decision attribute (or a label)
and each row denotes a training (or test) record. This study employs three
learning methods, J48, NaiveBayes and JRip, from Weka. The three classifiers
are Java implementations of the well-known C4.5, Naı̈ve Bayes and Ripper
methods introduced in Section 2.3.1. The parameters of the three methods
are the default values given by Weka. A heuristic attribute selection method
BestFirst is also used to choose a few of the most informative attributes when
the amount attributes are too large.

4.2

The Class Design

The operations in the SPOT algorithm are designed and encapsulated
in three classes, ProblemDomain UEA2, SPOT ML and SPOT, as shown as
n the thick borders in Figure 4.1. The three classes are introduced in the
following.
The class ProblemDomain UEA2 implements the super class ProblemDomain in HyFlex for supporting the domain-specific operations in the SPOT
hyper-heuristic, such as the sampling for subproblems (genSubprob()), the
U/EA transformation, the generations of raw attributes32 and the modifications (or creations) of LLHs. The operations on transformation and attributes
are encapsulated in public functions, such as manipulating competitive
groups (getCompetitiveGroups()), raw attributes (getRawAttributes())
and labels (getIntLabels() and setIntLabels()).
32 The first step of attribute generation, see Section 4.3 for more details.

4.2. THE CLASS DESIGN

93

HyFlex.ProblemDomain
...
...
ProblemDomain UEA2
genSubprob(size, seed)
Domain
getCompetitiveGroups() : int[]
getRawAttributes() : double[][]
getIntLabels() : int[]
setIntLabels(lbls)
works on
1

TSP UEA2
...
...

FSSP UEA2
...
...
1

HyFlex.HyperHeuristic
...
run()
...

SPOT
new llhs: LLH []
uses

1

1

SPOT ML
genNormAtts(groups, rawAtts) : double[][]
trainModelAndTest(atts, lbls, attsToBeLabeled) : int[]
trainModelAndTestAfterSel(atts, lbls, rawAtts, groups) : int[]
1

uses

1

Weka.attSelection.BestFirst
...
...
1

Weka.classifiers.NaiveBayes
...
...

1

Weka.classifiers.J48
...
...

1

Weka.classifiers.JRip
...
...

Figure 4.1: The class diagram of the SPOT hyper-heuristic
The class SPOT ML includes the cross-domain operations about attribute
generation and training and applying classifiers. genNormAtts() is a function that populates a number of attributes by comparing a raw attribute
of an assignment to those of its competitors, see Section 4.3. SPOT ML
encapsulates three effective classification techniques, i.e. J48, NaiveBayes
and JRip, in the function trainModelAndTest(). The attribute selection

94

CHAPTER 4. JAVA IMPLEMENTATION

method BestFirst is encapsulated into trainModelAndTestAfterSel(),
which extends the function trainModelAndTest().
The class SPOT is the high-level controller. It inherits the standard functions from the super class HyperHeuristic. The class SPOT accepts an instance of ProblemDomain UEA2 as its input problem and solves it step by
step, see Section 4.3 for the details of the steps. The learning in the class
SPOT is powered by the functions of SPOT ML.

4.3

Implementation in Java

The three classes were implemented in Java 2 Platform Standard Edition
(SE) version 1.6.0 32 (64-bit). The compiler was the default compiler provided in Java 2 SE. The machine was an Intel Core i7 975 3.33 GHz33 . The
operating system was a Windows 7 64-bit with Service Pack 1. The class
ProblemDomain UEA2 is relatively domain-specific. The examples of its
implementation can be found in Section 5.2 and Section 6.2. The classes
SPOT ML and SPOT are cross-domain. Their implementations are detailed
as follows.

4.3.1

The class SPOT ML

There are three main public functions to implement in SPOT ML. Define
a problem with n groups, p competitor assignments in each group and
ra different raw attributes (measures) for each assignment. The function
33 The official test program showed that 1.00 CPU second in four parallel thread mode for

this CPU was equivalent to 1.00 CPU second for a Pentium 4 3.0 GHz.

4.3. IMPLEMENTATION IN JAVA

95

genNormAtts() generates a table of normalized attributes with ra × (2p + 3)
columns and n × p rows. Each row stands for the competition between one
assignment and its competitors. To illustrate the generation of the columns,
it is assumed that there is only n = 1 group of p assignments and ra = 1
raw attribute (measure). The raw attributes of the assignments are measured
and put in an ordered set { x1 , x2 , . . . , x p }, where x1 ≤ x2 ≤ . . . ≤ x p . For
each i-th assignment, Table 4.1 shows 2p + 3 normalized attributes (columns)
which describe the competition. A function bound100 is used to limit the
ranges. The first generated attribute is the real value of the raw attribute.
The remaining attributes are counts and comparisons without any units. A
column of Boolean labels can be attached to the table, where true denotes the
assignment can be found in a suboptima (or optima) and false otherwise.
Particularly, when there are two assignments in each group, some columns
can be removed due to redundancy. For example, the second row of attribute
“repeat counter” in Table 4.1 can be omitted.
The function trainModelAndTest() consists of training of the classifiers and classifying of unlabeled data. It is a part of the last step in Section
3.3.2. First, the table consisting of normalized attributes and labels is used
to train the three classifiers. Each training is carried out with the default
parameters suggested by Weka, as listed in Table 4.2. The three trained
classifiers predict the unlabeled data which is generated from the given
problem by measuring the same raw attributes and generating normalized
attributes. The average (or summation) value of the labels predicted by
the three classifiers is regarded as the final label (prediction). In practice,
non-numeric Boolean values, such as true and false, can be converted to

96

CHAPTER 4. JAVA IMPLEMENTATION

Table 4.1: A list of generated attributes of the i-th assignment from one raw
attribute
Attribute(s)

Equation

Absolute value
Repeat counter
Equal to the first?
Equal to the last?
Relative values

xi
†
∑ra
j=1 same ( x i , x j )
same( xi , global‡ minimum of x1 )
same( xi , global‡ maximum of x p )
bound100\ ( xi /x ]j ), 1 ≤ j ≤ p

Count
1
1
1
1
p

Relative increments bound100( x j /x ]j+1 ), 1 ≤ j ≤ p − 1
p−1
(
1 if x = y,
†: same( x, y) =
0 if x , y.
‡: The minimum or
maximum is available if there are multiple groups.

if x ≥ 100,
100
\: bound100( x ) = x
if − 100 < x < 100,


−100
 if x ≤ 100.

if x < 0,
100
]: When y = 0, x/y = 1
if x = 0,


−100 if x < 0.

numeric values, such as 1 and 0, in order to do arithmetic operations.
A preprocessing of attribute selection generally can select the most informative columns to reduce the computation cost of training and to enhance the generalization of the classifiers. The attribute selection is very
useful when the number of columns ra × (2p + 3) is too large (e.g., 2000).
The function trainModelAndTestAfterSel() in the class SPOT extends
trainModelAndTest() with an attribute selection preprocess. The preprocess calls the BestFirst search and the CfsSubsetEval evaluator provided by
Weka. Details about the selection method and the correlation-based evaluator
can be found in Hall (1999, Chap. 4).

4.3. IMPLEMENTATION IN JAVA

97

Table 4.2: The default values of main parameters of the classification and
attribute selection methods in Weka
Method

Main parameter

J48

Confidence
Min number of objects
Number of folds

0.25
2
3

Quinlan (1993)

JRip

Number of folds
Min number of objects
Optimizations
Seed

3
2
2
1

Cohen (1995)

NaiveBayes

Using kenel estimator

false

BestFirst

Lookup cache size
Search termination

4.3.2

Default value

1
5

Reference

John and Langley (1995)
Hall (1999)

The class SPOT

The class SPOT mainly implements one public interface run() inherited
from HyFlex.HyperHeuristic, as mentioned in Section 3.3.2. The function
run() consists of a full execution of the problem solving of the presented
algorithm. Given a set of determined parameters and a problem to solve, the
function run() can be briefly described as pseudo codes in Figure 4.2.
First, unlabeled attributes of possible assignments for classification are
generated in Lines 1 and 2 in Figure 4.2. Secondly, a subproblem is sampled
and is solved by a determined algorithm, as shown in lines 3 and 4. The
PHunter (Pearl Hunter), which is an effective heuristic selection algorithm developed by Chan et al. (2012) on HyFlex, is employed as the input algorithm.
The execution time of PHunter can be usually limited to a very short period,
for example, a few seconds. Suboptima can be found in most cases, because
the size of the subproblem is also limited. Then, the attributes and labels
about the suboptima of the subproblem can be generated, as shown as lines

98

CHAPTER 4. JAVA IMPLEMENTATION

1.
2.
3.
4.
5.
6.
7.
8.

ProblemDomain UEA2 i1 := load problem();
atts1 := SPOT ML.genNormAtts(i1 .getCompetitiveGroups(), i1.getRawAttributes());
ProblemDomain UEA2 i2 := i1 .genSubprob();
i2 .solved by algorithm(an input algorithm);
atts2 := SPOT ML.genNormAtts(i2 .getCompetitiveGroups(), i2.getRawAttributes());
lbls2 := i2 .getIntLabels();
predLbls := SPOT ML.trainModelAndTest(atts2 , lbls2 , atts1 );
i1 .setIntLabels(predLbls);

Figure 4.2: The pseudo codes of function SPOT.run()
5 and 6. Finally, the assignments of the given problem can be predicted, and
some LLHs can be modified, as shown in lines 7 and 8. In the 7th line, there
is also an alternative learning function trainModelAndTestAfterSel()
consisting of a preprocess of attribute selection, as described in Section 4.3.1.

4.4

Discussion

The variable assignment weighting scheme presented in Xue, Chan, et al.
(2011) was implemented in C language and was compiled by GCC (GNU
Compiler Collection). The processes of class association rules learning and
weighting assignments were designed as domain-specific and static calls. In
comparison, the processes of the presented SPOT algorithm, which aim at
robust and cross-domain applications, are designed as open interfaces. For
example, the generation of normalized decision attributes on the basis of raw
attributes can provide more possible correlations for the models of stochastic
machine learning. Another example is the combination of three classification
techniques from different branches of supervised learning.
There are several advantages to develop the SPOT algorithm on the
HyFlex library. Although the version involved in this thesis is mainly used

4.4. DISCUSSION

99

for heuristic selection, HyFlex presents a concise and precise infrastructure
of interactions between domain-specific LLHs and cross-domain algorithms.
Secondly, there are six domains of NP-Complete benchmark problems as well
as their LLHs already included in HyFlex. Finally, a public hyper-heuristic
competition, CHeSC 2011, already validated the effectiveness and robustness
of HyFlex. The competition also attracted tens of hyper-heuristic researchers
worldwide and publicized their computational results on HyFlex. In other
words, HyFlex provides adequately reliable LLHs and hyper-heuristics for
reference.
In this thesis, another hyper-heuristic PHunter is employed as the solver
for the subproblem. There were three main reasons. First, as a cross-domain
hyper-heuristic, PHunter has fewer structures and codes compared to a number of algorithms in which each one is a specific solver for one domain. Fewer
structures and codes also mean fewer possible sources of implementation
errors. The second reason is the effectiveness. PHunter was the fourth overall
and the first in hidden domains34 out of twenty hyper-heuristics in CHeSC
2011. The last reason is that PHunter and SPOT share the same development
library HyFlex. Therefore, it is convenient to incorporate PHunter.
The reason of using a 64-bit Java instead of a 32-bit one lies on the maximum heap size. In a 32-bit Java environment on windows, the maximum
heap size is typically 1.8 G bytes (IBM35 Java) or 4 G bytes (Oracle/Sun
Java36 ). However, the class SPOT ML sometimes needs more memory for
storing data and training the classifiers, such as a 3 to 5 G bytes of memory or
34 See http://www.asap.cs.nott.ac.uk/external/chesc2011/results.html.
35 See http://publib.boulder.ibm.com/infocenter/javasdk/tools/index.jsp?topic=/com.

ibm.java.doc.igaa/ 1vg00014884d287-11c3fb28dae-7ff6 1001.html.
36 According to the “Xmx” parameter.

100

CHAPTER 4. JAVA IMPLEMENTATION

even more in the large-scale problems in Chapter 5 and Chapter 6. Therefore,
a 64-bit Java environment was chosen as the environment.

Chapter 5
Application I: The Traveling Salesman Problem Domain

Example is the school of mankind, and
they will learn at no other.
Letters on a Regicide Peace
Edmund Burke

In this Chapter, the presented SPOT hyper-heuristic is applied to a wellknown combinatorial optimization domain, the Traveling Salesman Problem
(TSP), in order to try to replicate Xue, Chan, et al. (2011)’s modeling approximately. Section 5.1 briefly overviews the background and some well-known
algorithms and data structures for TSPs. The implementation of TSP domain in HyFlex is introduced in Section 5.2. The domain-specific design and
parameters of the SPOT hyper-heuristic are given in Section 5.3 in detail.
Section 5.4 shows the results of experiments and some discussion is given in
Section 5.5.

101

102

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

5.1

An Introduction to the TSP Domain

As defined in Section 1.2, a TSP aims at finding the shortest possible tour
in which each given city is visited exactly once. Examples of TSPs can be
found in Figure 1.6 and Figure A.1. TSP is one of the most comprehensively
studied problems in computer science, operational research and engineering
practice. A common form of TSPs is Euclidean TSP, where the distance
between cities is the Euclidean distance. The industrial applications of TSPs
include vehicle routing, electric power cable networks, and VLSI (Very-LargeScale Integration) chip fabrication (Punnen, 2004; Applegate, Bixby, Chvátal,
& Cook, 2006).
Both TSP (Garey & D. S. Johnson, 1979) and Euclidean TSP (Papadimitriou, 1977) were proven in the NP-Complete class. Papadimitriou and
Vempala (2006) showed that TSPs with triangular inequality are NP-hard to
approximate, with a ratio less than 220/219. Nevertheless, TSPs, especially
Euclidean TSPs, can be regarded as well-handled in practice, after decades of
research. Algorithms that guarantee finding optimal tours are now capable
of solving relatively large-scale instances. For example, the Concorde code
solved the 85,900 cities in a time of 136 CPU-years for a 2.4 GHz processor
(Applegate, Bixby, Chvátal, & Cook, 2006). Besides, efficient heuristics can
quickly find tours within a few percent of optimal on real-world instances
(D. S. Johnson & McGeoch, 1997). For example, Bentley (1992)’s greedy could
find a solution about 15% over the HK lower bound for a TSP with 85,900
cities in three seconds on a 500 MHz EV6 Compaq Alpha processor. More
sophisticated heuristics can get within 1% of optima in reasonable amounts

5.2. IMPLEMENTATION OF THE TSP DOMAIN IN HYFLEX

103

of time (Helsgaun, 2000; D. S. Johnson & McGeoch, 2002).
It is believed that one of the major reasons for the efficiencies of heuristics,
especially local search, are the highly effective edge candidate set structures.
The main concern of a candidate set is that it is impossible for most of very
long edges to appear in optimal tours. So a candidate set usually contains a
limited number of promising (short or holding particular properties) edges
for each city according to the geometry. Popular candidate sets include the
nearest-neighbor-first (Lin & Kernighan, 1973), the Delaunay (Reinelt, 1994),
the k-quadrant (Miller & Pekny, 1995), and the α-nearness (Helsgaun, 2000).
There also are other techniques for modifying search procedures according
to suboptima, such as the backbone (W. Zhang & Looks, 2005) and the tourmerging (Cook & Seymour, 2003). According to the definitions in this thesis,
both the backbone and the tour-merging can be considered as heuristic
generation methods in regard to the amount of possible modifications on
heuristics, though they did not include complicated and explicit machine
learning techniques.

5.2

Implementation of the TSP Domain in HyFlex

The problem domain of TSP was implemented in HyFlex and used as a
hidden test domain37 in CHeSC 2011. The implementation employed ten
widely used benchmark instances from the TSPLIB data sets38 , as shown
in Table 5.1. All the problems are 2D Euclidean instances. The six drilling
instances are from the print circuit board industry, and the two largest in37 Not public before the competition.
38 Available at: http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/.

104

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.1: Ten benchmark instances used in the implementation of TSP in
HyFlex
No.

Id

Size (n)

Referential optimum

0
1
2
3
4
5
6
7
8
9

pr299
pr439
rat575
u724
rat783
pcb1173
d1291
u2152
usa13509
d18512

299
439
575
724
783
1173
1291
2152
13509
18512

48191
107217
6773
41910
8806
56892
50801
64253
19982859
645238

In CHeSC?

Notes

X

Drilling problem
Drilling problem
Rattled grid
Drilling problem
Rattled grid
Drilling problem
Drilling problem
Drilling problem
Map of USA
Map of Germany

X

X
X
X

stances are national maps. The distance implemented in HyFlex was the
real-value Euclidean distance. However, the optima in Table 5.1 are based on
the rounded integer distance defined in TSPLIB. Because the error caused
by the rounding operation is limited, the optima can be used as referential
optima to measure the excesses of suboptima approximately. After a random
selection, five instances, No. 0, 2, 6, 7 and 8, were selected in the test beds to
benchmark the hyper-heuristics in CHeSC 2011.
Twelve LLHs were implemented, as shown in Table 5.2. The LLHs could
be categorized into four classes: mutation, ruin-and-recreate, local search
and crossover. Algorithms in the crossover class started from two tours and
created one new tour for return. Other algorithms started from one tour and
return a new (or same) return. All the implemented LLHs were carried out
on the nearest-neighbor-first candidate set. Eight nearest neighboring cities
were considered as candidates of the preceding or succeeding cities. In other
words, two out of eight shortest edges departing from one city were expected
to be in the optimum case.

5.3. DEVELOPMENT OF THE SPOT FOR TSPS

105

Table 5.2: The LLHs in the implementation of TSP in HyFlex
Type

Id

Parameters used
Intensity of
mutation

Mutation

Notes

Depth of
search

0
1
2
3
4

X
X

Random reinsertion
Swap of two cities
Randomization
Partial randomization
A number of swaps of cities

Ruin-and-recreate

5

X

Greedy partial reconstruction

Local search

6
7
8

Crossover

9
10
11
12

5.3

X
X
X

First improvement 2-Opt
Steepest descent 2-Opt
First improvement 3-Opt
Order crossover
Partially mapped crossover
Precedence preservative crossover
One point crossover

Development of the SPOT for TSPs

This section presents the application of the SPOT heuristic generation
for the domain TSP. The development can be divided into the three phases
mentioned in Section 3.3:
P1: transformations for standards and sampling,
P2: determination of parameters, and
P3: generation of new LLHs (by modification).
All the domain-specific operations developed in phases P1 and P3 were
encapsulated in the interface functions within the class TSP UEA2, as shown
in Figure 4.1. The determined parameters were stored and used by the class
SPOT.

106

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

5.3.1

P1: Transformations and sampling

The designs of the U/EA and the U/EA2 standards are described in
this section. The first standard is actually the problem model of the SPOT
algorithm so that the given combinatorial optimization problems must be
transformed to meet the standard. The latter standard regulates the generation of the labeled training data and the unlabeled test data, so that
meaningful classifiers can be trained and are compatible with the unlabeled
test data.

5.3.1.1

Design and validation of transformations

The hard constraint in a TSP is the requirement for a Hamiltonian cycle.
The Hamiltonian cycle is a graph cycle (i.e., closed loop) that visits each node
exactly once. In the perspective of a city, the Hamiltonian cycle means there
are only two edges connected to itself. Alternatively, each city has a degree
of two, in a formal description. The only inter-variable hard constraint is that
all the n edges must form a closed loop.
In the U/EA standard, there is no inter-variable hard constraint. Therefore, the inter-variable hard constraints were directly removed. The innervariable constraint of “having a degree of two” was kept. Then the transformed problem was actually a relaxed TSP, which could be stated as follows.
In a transformed TSP, a set C = {c1 , c2 , . . . , cn } of cities is given,
and for each pair (ci , c j ) of distinct cities there is a distance d(ci , c j ).
The target is to find n pairs so that (i) the summation of the n
lengths is relatively short and (ii) each city connects two and only

5.3. DEVELOPMENT OF THE SPOT FOR TSPS

107

two cities.
It might seem confusing to find “relatively short” targets. In fact, training
examples of “relatively short” edges could be obtained from a suboptimum
(of a subproblem). The instance-specific information about “what kind of
edge seems like ‘relatively short’ in a similar problem” could be learned, too.
The cardinal number of possible assignments for each city was always n − 1
originally. Henceforth, the new relaxed TSP meets the U/EA standard.
Three instances listed in Table 5.3 were selected to test the compatibility
of the relaxed TSP in the U/EA standard. The selection was mainly based
on the problem size. Small-scale problems, such as those with tens of cities,
might bring a biased geographic structure into the learning. Very large-scale
problems, such as those with thousands or millions of cities, cost too much
time due to the scale. Therefore, three problems with around one thousand
cities were selected. In each instance, two arbitrary suboptima were selected
randomly to check the resemblance between the two and to estimate the
average resemblance of each instance approximately. The approximability, i.e.
the overall average resemblance r̄, was 70.40%. Both the approximability and
generality (degree of approximation among r of instances) were acceptable.
Therefore, the U/EA model was said to be compatible to the SPOT hyperheuristic.

5.3.1.2

Design of decision attributes

The last task in P1 was to design attributes to meet the U/EA2 standard.
The design involved the definitions of the groups of competitors, the raw

108

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.3: Three benchmark TSP instances used in the development of the
SPOT algorithm
No.

4
5
6

Id

Size (n)

% excess the referential optimum

783
1173
1291

rat783
pcb1173
d1291

Resemblance r (%)

Suboptimum 1

Suboptimum 2

0.7609
0.5590
2.9279

0.8086
0.4394
3.2397

95.27
60.70
55.23

Approximability r̄ (%)

70.40

attributes and the label(s). The definitions were important. Because both the
labeled training data and the unlabeled test data would be generated according to the definitions, and the training data determined the classification and
the final LLHs.
The group of competitors for a city ci was designed as the set of all the
n − 1 edges connecting to the city ci :


(ci , c j )|1 ≤ j ≤ n, j , i .

(5.1)

Each group was marked with a unique identity (integer). In this case, the
function getCompetitiveGroups() returned an array of n × (n − 1) integers, where each value stood for the identity of an edge group, and there
were p = n − 1 assignments in each group.
Nevertheless, effective candidate sets were also considered during the
design. Given a candidate set, the group for a city ci became


(ci , c j )|c j ∈ candidate set of(ci , sizec ) ,

(5.2)

where sizec (2 ≤ sizec ≤ n − 1) was a constant integer denoting the size of

5.3. DEVELOPMENT OF THE SPOT FOR TSPS

109

the edge candidates for each city. The candidate set employed in HyFlex
was the nearest-neighbor-first set, and the sizec was 8. In the design of
the SPOT hyper-heuristic, the same candidate set was used to generate
candidates, and the sizec was set to 20 to include more edges that could
possibly be suboptimum assignments. As a result, the implemented function
TSP UEA2.getCompetitiveGroups() returned an array of 20 × n integers,
where each group consisted of p = 20 assignments.
Besides the length of the edges, angular properties were also considered during the design of the attributes, because all the instances were 2D
Euclidean TSPs. For each assignment (edge) (ci , c j ) in each group, there
were ra = 2 raw attributes designed. The first was the (Euclidean) distance, i.e. d(ci , c j ). The other raw attribute was the angle of the edge
atan2(y j − yi , x j − xi ), where atan2 is a variation of the arctangent function:

atan2(∆y, ∆x ) =




arctan (∆y/∆x )








arctan (∆y/∆x ) + π







arctan (∆y/∆x ) − π



+π/2








−π/2







0

∆x > 0
∆y ≥ 0, ∆x < 0
∆y < 0, ∆x < 0

(5.3)

∆y > 0, ∆x = 0
∆y < 0, ∆x = 0
∆y = 0, ∆x = 0

Therefore, the function TSP UEA2.getRawAttributes() returned a 2D
array (in 20 × n rows and 2 columns) of real values. The class SPOT then
generated ra × (2p + 3) = 86 columns of normalized attributes and 20 × n
rows of data through the function genNormAtts(). In fact, the 86 columns of

110

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

attributes included 10 of the 14 decision attributes in Xue, Chan, et al. (2011).
The tables generated from subproblems were compatible to the unlabeled
data in terms of classification. The designed attributes meet the standard of
U/EA2 .
The design of the label was, in fact, determined by the design of the
group. Because the group was the set of 20 nearest-neighbor-first candidates
for each city, the label appeared in the suboptimum for each edge candidate
(assignment). The label of each row was set to 1, if the edge appeared in
the optimum (or optima); to 0 if not. The labels were obtained through the
function TSP UEA2.getIntLabels() which returned an array of 20 × n
integer values consisting of 2 × n 1s and 18 × n 0s.
In summary, the transformation and the designs were relatively straightforward and partially used Xue, Chan, et al. (2011)’s model. The objective
of the SPOT hyper-heuristic was designed as “generating LLHs to focus on
the edge candidates that would likely appear in suboptima”. Examples of
generated data tables of labeled training data and unlabeled test data can be
found in Table B.1 and Table B.4 in Appendix B.

5.3.2

P2: Parameter determination

There were two important parameters to be determined, one by one, in
this phase, as described in Section 3.3.1. The first parameter was the size
of the subproblems n0 , which was numeric. The other was the method of
sampling for generating subproblems, and it was non-numeric.

5.3. DEVELOPMENT OF THE SPOT FOR TSPS
5.3.2.1

111

The size n0 of subproblems

The tests of n0 were set up as follows. The function SPOT.run() was
executed without the final step of generating the LLHs. It was because the
indicators were measurable before the generation of LLHs. The sampling
method was set to the random selection. The attribute selection was enabled
by calling the function trainModelAndTestAfterSel() for classification.
The usual number of selected attributes was around ten, see Table B.1 and
Table B.3 in Appendix B for a comparison. The computation time of the embedded subproblem solver PHunter was set to six seconds. After training the
classifiers and labeling the test data, the 20-nearest-neighbor-first candidate
sets were sorted by the weighted distance suggested by Xue, Chan, et al. (2011):


γidx 
1−
× d ( c i , c j ),
3

(5.4)

where c j was the idx-th edge in the candidate set of ci and γidx was the
summation of predicted labels for the assignment (ci , c j ). Then sizec was set
to 8 for generating a comparable data structure. In fact, weighting the edge
candidates is not a new idea, and relevant examples could be can in different
heuristic and exact algorithms, such as Nemhauser and Wolsey (1988, pp.
494–495)’s branch-and-bound tree.
The indicator r was decomposed into two indicators to measure the
suboptimality of the generated candidate set. The first indicator cov was
the average coverage of suboptimum assignments, which was the percentage of
the suboptimum assignments covered by the candidate set. It is clear that

112

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

0 ≤ cov ≤ 1 and a greater cov usually stood for a more expected39 result for
learning and a better suboptimality. The other indicator d¯ was the average
depth, which measured the average index of the optimum assignments in
the candidate set. In other words, d¯ showed how many non-suboptimum
assignments were “wrongly” placed before suboptimum ones, on average. A
lower level of d¯ denoted a better accessibility of the promising assignments
when the cov was not changed.
Six values of n0 were tested, as shown in Figure 5.1. One hundred independent tests were carried out for each value and each instance. Figure 5.1
shows the trends of the average coverage cov, the average depth d¯ and the
time cost against the value of n0 . The “8-nearest-neighbor-first” was the data
set implemented in HyFlex.
It can be seen from Figure 5.1 that the new candidate set included more
suboptimum assignments when n0 ≥ 50. A decreasing trend of average
depth d¯ meant that the suboptimum assignments were gradually moving
towards the heads of the ordered groups when n0 increased. Comparing to
the candidate set implemented in HyFlex, the new candidate set generally
covered more suboptimum assignments but placed them slightly deeper. For
instance, when n0 = 800 the new candidate set covered +0.142%40 more on
the fraction but placed the suboptimum assignments +0.017 deeper when
n0 = 800. It was acceptable, because cov, which related to the scope of search
of LLHs, was slightly more important than d¯ which related to the depth of
search.
39 A cov ≈ 1 could also possibly mean the learning was overfitting in some cases.
40 See Table C.2 for data.

5.3. DEVELOPMENT OF THE SPOT FOR TSPS

113

Mean of cov (%)

99
98.8

× SPOT

98.6

×

×

×

×
98.4

×

8-nearest-neighbor-first

98.2
98
25

50

200

100

400

800

n0

Mean of d¯

2×
1.95

×

SP

OT

×

1.9

×

×

1.85

×

8-nearest-neighbor-first
1.8
25

50

200

100

400

800

n0
10
Time cost (s)

×
8
6×

×

SPOT

×

×

×

4
2
0
25

50

200

100

400

800

n0

Figure 5.1: Average test results for determining n’ in the development of the
SPOT algorithm, on the three TSP test instances (100 runs)

114

CHAPTER 5. APPLICATION I: THE TSP DOMAIN
The size n0 was chosen as 400 after the tests. The main reason was to

balance performance and time cost. The time spent on learning seemed to
increase considerably when n0 > 400. If too much time was spent on learning,
it would be difficult to call the application an “on-the-fly” hyper-heuristic.

5.3.2.2

The methods of sampling for subproblems

Besides the general random selection, another method, rectangular selection, was also tested to try to fine-tune the SPOT algorithm. For the 2D
Euclidean instances in HyFlex, the geography of the cities can be regarded as
being in a minimum boundary rectangle. A smaller rectangle can be selected
within the boundary rectangle to choose a block (subset) of cities. Given the
size n of the input problem and the size n0 of the subproblems, the area of
the small rectangle is n0 /n of that of the minimum boundary rectangle, if the
cities are assumed to be distributed uniformly in space. In practice, a few
small rectangles can be randomly placed and the one with the closest density
of cities can be selected, in order to reduce errors from possible irregular
distribution of cities.
The aspect ratio of the small rectangle was considered as a parameter.
Letting the aspect ratio of the minimum boundary rectangle be a constant
1:1, five ratios were tested, as shown in the first column in Table 5.4. It can be
seen that both the best (greatest) cov and the best (least) d¯ were generated
in the ratio 1:1, which meant keeping the same aspect ratio as the minimum
boundary rectangle. Therefore, the aspect ratio in the rectangular selection
was set to be the same as the boundary box in the following.
Tests were carried out to examine what sampling method was able to

5.3. DEVELOPMENT OF THE SPOT FOR TSPS

115

Table 5.4: Average test results on different aspect ratios in the rectangular
selection sampling method, where n0 = 400 and the aspect ratio of the
minimum boundary rectangle is assumed as 1:1 (100 runs, best performance
in bold)
Aspect ratio

cov (%)

d¯

1:16
1:4
1:1
4:1
16:1

98.570
98.613
98.615
98.612
98.563

1.936
1.874
1.869
1.879
1.876

Timeoverall (s)

TimeL (s)

6.30
6.65
6.94
6.66
6.31

0.28
0.61
0.89
0.62
0.29

generate a sample that led to the best candidate set in terms of the two
indicators. Table 5.5 shows the average results and significance by one-way
ANOVA tests41 . The first column of Table 5.5 is the indicator, and the second
column denotes the test instance. Results of the rectangular selection, the
random selection and two parallel subproblems where each method sampled
one of them are, listed in the third column, the fourth column and the fifth
columns, respectively. Given the method of sampling as categories, the F
values and the significance p show the variances of mean values and the
statistical significance of null hypothesis, as listed in the last two columns,
respectively.
It can be observed from Table 5.5 that the average values on cov were
not always significant, though the cov resulted by the “Parallel” column
seems slightly higher. The average values on d¯ were clearly significant, as
written in bold. Therefore, it was statistically significant that the d¯ relating
to the parallel runs of both selections was the best (least). The time spent
on the parallel mode was approximately twice much as that spent on one
41 The one-way ANOVA (analysis of variance) is employed to compare means of two or

three samples in this thesis. The null hypothesis is that two or more groups are drawn from
the same population. When there are only two means to compare, the relation between
ANOVA and the t-test is given by F = t2 .

116

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.5: Average results and significants of tests on different subproblems
sampled, where n0 = 400 (100 runs, significant p in bold)
Indicator

Instance

cov

Rectangular

Random

selection

selection

Parallel

4
5
6

99.617
98.892
97.332

99.617
98.892
97.343

99.617
98.892
97.345

Average

98.613

98.617

98.618

4
5
6

2.103
1.893
1.652

2.107
1.906
1.618

2.103
1.895
1.599

Average

1.883

1.877

1.866

d¯

ANOVA
F

p

0.51
—†
3.12

0.60
—
0.05
—

6.65
92.79
19.80

0.00
0.00
0.00
0.00

†: No variance.

subproblem. In the following, two parallel subproblems were sampled by
random selection and rectangular selection, respectively. Each subproblem
and its suboptimum contributed half of the training data for supervised
learning.
With the determined transformations, design of attributes, sampling
and other secondary algorithms and parameters, a predicted label could be
generated, and a weighted distance could be calculated for each assignment
(in the candidate set) on-the-fly. However, it should be noted that the 8nearest-neighbor-first candidate set already achieved an effective pair of
indicators, cov = 98.477% and d¯ = 1.842, as shown in Figure 5.1. The average
results in the “Parallel” column in Table 5.5 only shows a slightly higher
coverage (∆cov = +0.141%) but a slightly weakened depth (∆d¯ = +0.024).
Therefore, the new candidate set, as well as newly generated heuristics,
should be capable of providing slightly higher qualified solutions, while it
might need in-depth search capability to find them.

5.4. EXPERIMENTS AND OBSERVATIONS
5.3.3

117

P3: The generation of new LLHs

This part was the last mile of the development and of the function
TSP UEA2.setIntLabels(). The new candidate set was obtained after
P1 and P2. Figure 5.2 shows an example of the new candidate set. In the
second figure in Figure 5.2, thick lines are the assignments with Σ labels ≥ 2,
the thin lines are those with Σ labels = 1 and the gray lines are those with Σ
labels = 0. Although there were priorities given in the new candidate set, the
difference between the two figures in Figure 5.2 is actually not significant, as
indicated by cov and d.¯ More details of generating the example can be found
in Appendix B.
The new candidate set generated by the SPOT algorithm was only applied
to the three local search LLHs in Table 5.2. The reason was the local search
LLHs had stronger in-depth search capabilities than those in other categories.
The algorithms number 6, 7 and 8 listed in Table 5.2 were called number 6’,
7’ and 8’ after change of the candidate set. Although the codes of all local
search LLHs remained the same, the scope and traces of their exploration of
the search space changed due to the change of the candidate set. The original
algorithm numbers 6, 7 and 8 were called “base algorithms” of the new ones.
Generally, the three new LLHs can be regarded as generated by the SPOT
hyper-heuristic.

5.4

Experiments and Observations

This section presents two groups of tests. In the first group of experiments, the individual performance of each new LLH was examined, with

118

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

(i) The 8-nearest-neighbor-first

(ii) A new candidate set by the SPOT

Figure 5.2: A comparison of two candidate sets of the TSP instance rat783
comparisons to its base algorithm. In the other group of tests, the three
new LLHs replaced their base algorithms. The heuristic selection method
PHunter solved the updated domain consisting of new LLHs. The results
were compared to the solving of PHunter with the original LLHs and the
winners in the TSP domain in CHeSC 2011.
All the tests were carried out with the same instances which had been
selected for the competition in CHeSC 2011, i.e. the numbers 0, 2, 6, 7 and

5.4. EXPERIMENTS AND OBSERVATIONS

119

8 shown in Table 5.1. The size n0 of the subproblems was set to 400 and
two subproblems were generated in parallel to provide training data, as
described in Section 5.3.2. The overall computation time was set to be the
same as the normalized time in CHeSC 2011, which was equivalent to 10
CPU minutes for a Pentium 4 3.0 GHz computer. All the time costs of the
processes in the SPOT hyper-heuristic, such as sampling subproblems, solving the subproblems, learning instance-specific knowledge and modifying
LLHs, were fully included in the overall time in order to make the results
comparable. For the embedded subproblem solver PHunter, the time limit
for solving each subproblem was 1% of the overall time, i.e. equivalent to 6
CPU seconds on a Pentium 4 3.0 GHz machine. The time of the rest processes
in the SPOT was not directly restricted.

5.4.1

On the individual LLHs

Table 5.6 shows the results on tour quality of the newly generated LLHs
in terms of percentage excess over the referential optima42 . There were
100 runs for each configuration and each problem instance. In each run, a
random initial tour was generated for all local search LLHs and all LLHs
started from the same initial tour independently. One-way ANOVA tests
were also included to check the significance of the changes. The second
column is the test instance. The third and the fourth columns show the
excesses of the shortest tour length and the average length found by the base
algorithms in HyFlex, respectively. The fifth and the sixth columns stand
for the shortest tour length and the average length found by the new LLHs,
42 It should be noted that the referential optima do not necessarily present the true optima.

120

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.6: Test results on the tour quality of the generated LLHs and their
base algorithms (100 runs, n = 400, Depth of search = 1, improvements on
average performance in bold)
Group

No. 6
versus
No. 6’

No. 7
versus
No. 7’

No. 8
versus
No. 8’

Inst

0
2
6
7
8

0
2
6
7
8

0
2
6
7
8

% excess by
the base LLH

% excess by
the new LLH

Min

Avg

Min

Avg

6.944
6.329
9.690
10.905
10.027

12.941
8.151
17.066
14.679
11.840

6.944
6.329
9.900
10.729
10.027

12.993
8.156
17.172
14.660
11.840

Avg improvement (%)
6.273
5.915
9.683
10.957
9.515

ANOVA
F

p

0.052
0.005
0.106
-0.019
0.000

0.01
0.00
0.05
0.01
0.00

0.93
0.97
0.83
0.93
—

-0.190

Avg

0.92

12.107
8.559
16.900
14.240
11.589

-0.164
-0.001
0.150
-0.166
0.000

0.07
0.00
0.05
0.05
0.00

0.79
0.99
0.76
0.44
—

Avg improvement (%)

0.322

Avg

0.75

7.045
4.292
13.984
10.387
6.985

-0.119
-0.009
-0.099
-0.120
0.000

0.11
0.01
0.04
0.31
0.00

0.74
0.91
0.84
0.58
—

Avg improvement (%)

0.743

Avg

0.77

12.271
8.560
16.750
14.407
11.589

3.548
3.113
6.657
7.346
5.844

∆ avg

7.164
4.301
14.083
10.507
6.985

5.909
5.915
10.038
10.707
9.515

2.535
3.113
6.683
7.108
5.844

respectively. The seventh column denotes the change on the average excess,
where improvements (negative numbers) are written in bold. The last two
columns present the F values and the confidence p examined by the one-way
ANOVA tests. The last row in each group shows the average improvement
in percentage and the mean of p in each group. The average improvement of
each new LLH is defined as

Avg improvement = −

new % excess − base % excess
× 100%.
base % excess

(5.5)

It can be observed from the seventh column that the generated No. 7’ and

5.4. EXPERIMENTS AND OBSERVATIONS

121

No. 8’ LLHs found slightly better (0 < average improvement < 1%) tours
on average than their base algorithms. However, the generated No. 6’ local
search returned many inferior (average improvement < 0) results on average,
compared to its base algorithm. The improvements by No. 8’ seems more
steady (lower variance in the seventh column) and slightly more in amount
(+0.743% against +0.322%) compared to those of No. 7”s. However, all the
changes, no matter whether improved or weakened, were not statistically
significant43 (p  0.05).
Nevertheless, the observations are consistent with the interpretation and
prediction according to the indicators cov and d¯ in Section 5.3.2. In fact, No.
6 LLH was the first improvement 2-Opt and No. 7 LLH was the steepest
descent 2-Opt, as shown in Table 5.2. The only difference between them was
that the first one accepts any swap of two cities if it can improve the tour
and the second one only accepts the swap with best improvement. Hence,
it can be observed that the differences between the first and the second
groups in the seventh column in Table 5.6 were considerable. No. 8 LLH
was the first improvement 3-Opt, which explored a higher level of swaps
than No. 6. The observation of the differences between the corresponding
groups in Table 5.6 verified that in-depth search algorithms were slightly
better in this particular application of the SPOT heuristic generation. In
fact, Xue, Chan, et al. (2011) also observed a similar phenomenon that an
α-nearness-like candidate set determined by machine learning made obvious
improvements on a 5-Opt local search but encountered significant failure on
43 It should be noted that the insignificant and slight improvements did not necessarily

lead to the conclusion that the SPOT algorithm was useless in the TSP domain. In fact, the
same implementation could generate much more powerful LLHs by changing the rule of
sorting edge candidates. See Section 5.5 for more details.

122

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.7: Average time of generating new LLHs (100 runs, n = 400, Depth
of search = 1)
Instance

0
2
6
7
8

Size n

299
575
1291
2152
13509

Time (s)
Min

Avg

Max

Std dev

12.70
13.95
13.10
12.95
19.47

14.10
14.92
14.35
13.49
23.03

16.75
17.41
16.43
14.66
28.69

0.66
0.59
0.48
0.32
2.25

a 2-Opt heuristic for Euclidean TSPs.
The time cost of generating the LLHs is shown in Table 5.7. The third to
the sixth column show the minimum, the average, the maximum values and
the standard deviation of the time cost of the whole SPOT hyper-heuristic
for each instance. It should be noted that a 12 equivalent CPU seconds for
solving two subproblems was included in the data. It could be observed that
the maximum time cost in the instance with 13509 cities was no more than
about 30 seconds. Therefore, the SPOT could facilitate an instance-specific
heuristic generation in an “on-the-fly” fashion.
Table 5.8 shows the run time of the LLHs in solving problems in milliseconds. The computation time of No. 6, No. 7 and No. 8 LLHs which
was provided by HyFlex is shown in the second, the fifth and the eighth
columns, respectively. The time of No. 6’, No. 7’ and No. 8’ LLHs generated
is shown in the third, the sixth and the ninth columns, respectively. The
fourth, the seventh and the last columns show the changes of average time
in the three groups. It can be observed that the changes of time in the first
four instances were almost unnoticeable, especially when referring to the
level of the instrument error. In the last row, No. 7’ and No. 8’ LLHs took

5.4. EXPERIMENTS AND OBSERVATIONS

123

Table 5.8: Average run time of each LLHs (100 runs, n = 400, Depth of search
= 1, instrument error < 0.16 ms)
Instance

0
2
6
7
8

Average run time of LLHs in the three group (ms)
No. 6

No. 6’

∆

No. 7

No. 7’

∆

No. 8

No. 8’

∆

0.6
1.3
3.0
6.2
103.6

0.6
1.3
2.8
6.2
97.2

0.0
0.0
-0.2
0.0
-6.4

0.5
1.4
2.8
6.2
100.8

0.5
1.3
2.8
6.6
104.4

0.0
-0.2
0.0
0.3
3.6

8.7
15.9
52.7
111.1
4162.9

8.4
16.2
53.5
111.1
4191.6

-0.3
0.3
0.8
0.0
28.7

about 4% and 1% more time than their base algorithms, respectively, and
No. 6’ one saved about 6% of time. In general, the computation time of the
new LLHs was not considerably different from those of their base algorithms.
With consideration of the tour qualities of the new LLHs, No. 7’ and No. 8’
could possibly replace their base algorithms in the HyFlex framework.

5.4.2

Comparisons with other hyper-heuristics

Many existing hyper-heuristics, especially heuristic selection methods,
can benefit from the newly generated LLHs by the SPOT hyper-heuristic by
employing the new LLHs into their LLH sets. For example, the generated
instance-specific LLHs, especially No. 7’ and No. 8’, could replace their base
algorithms in HyFlex. The heuristic selection methods, such as the PHunter,
can run on the basis of an updated LLH set. Such a combination of SPOT
preprocessor and the body of the PHunter is noted as the SPOT-PHunter in
the following.
The HyFlex framework and the CHeSC 2011 competition provided public
and reliable results44 for twenty hyper-heuristics. In the TSP domain in
44 See http://www.asap.cs.nott.ac.uk/external/chesc2011/raw-results.html.

124

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

Table 5.9: The top three hyper-heuristics in the TSP domain in CHeSC 2011
No

Algorithm

Score

Author

1
2
3

AdaptHH
EPH
PHunter

40.25
36.25
26.25

Mısır et al. (2012)
David Meignan
Chan et al. (2012)

CHeSC 2011, the top three hyper-heuristics are listed in Table 5.9. The
scoring system45 was the Formula one points system, in which the winner
got 10 points, the runner-up got 8 points, and the third got 6 in each instance.
There were five instances selected. Therefore, the maximum of possible score
was 50.
The SPOT-PHunter was tested as a normal algorithm player in CHeSC
2011. The parameters of the PHunter in the body were the same as determined in Chan et al. (2012). All the three new local search LLHs were used to
replace their base algorithms, even though No. 6’ might be less competitive
than its base algorithm. The tests were conducted for the five instances, and
each was tested 31 times. The time cost in each test was equivalent to 10
minutes on a Pentium 4 3.0 GHz CPU. The median values46 were used for
scoring, as shown in Table 5.10. The second to the fourth columns show the
results of the top three hyper-heuristic approaches in CHeSC 2011. The fifth
and the sixth columns present the median solutions found by PHunter and
the SPOT-PHunter. Because the tests in CHeSC 2011 were carried out on a
32-bit47 Java environment, the third and the fifth columns of PHunter data
in the table were labeled as PHunter32bit and PHunter64bit , respectively. The
last column shows the confidence of one-way ANOVA test on comparing the
45 See http://www.asap.cs.nott.ac.uk/external/chesc2011/scoring.html.
46 The median value of a set {1, 1, 3} is 1, while the average value is 5/3.
47 Confirmed by private communication in 2011.

5.4. EXPERIMENTS AND OBSERVATIONS

125

Table 5.10: Median tour length of the PHunter, the SPOT-PHunter and three
other hyper-heuristics in the TSP domain in HyFlex (31 runs, best results in
bold for each instance)
Inst

Best results† in CHeSC 2011
AdaptHH

.

0
2
6
7
8

Test results in thesis

EPH PHunter32bit PHunter64bit SPOT-PHunter ANOVA p

48194.9
48194.9
6810.5
6811.9
53099.8
52925.3
66879.8
66756.2
20822145.7 21064606.3

48194.9
6813.6
52934.4
67136.8
21246427.7

48194.9
6812.6
52835.2
66826.0
21221065.6

48194.9
6802.4
52872.7
66606.4
21419868.0

—‡
0.00
0.28
0.00
0.00

†: Results on a 32-bit Java.
‡: Exactly the same.

mean tour qualities between PHunter64bit and SPOT-PHunter. The best results for each instance are written in bold. Two entries, the PHunter64bit and
the SPOT-PHunter, were added to the set of twenty hyper-heuristics. The updated scores, according to the scoring system, of the top five hyper-heuristics
are shown in Figure 5.3.
First, it should be noted that the median tour length returned by the
PHunter on a 64bit Java environment was equal to or slightly better (less)
than that on a 32bit Java in every test instance, as shown in Table 5.10, though
the codes (and complied binaries) of the PHunter algorithm remained exactly
the same. Therefore, the results of the PHunter64bit and the SPOT-PHunter
are more comparable.
The involving of new LLHs generated by the SPOT increased the score of
the PHunter from 27 to 34, as shown in Figure 5.3. Besides No. 0 instance
in which every algorithm returned identical results, the SPOT-PHunter beat
the PHunter64bit in two instances (No. 2 and No. 7), and both improvements are significant. The PHunter64bit obtained better results in two other
instances, and one improvement is significant and another is insignificant

126

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

40

T 34
PO

30

By

S
the

32

27

27

20

18

10

PHunter64bit SPOT-PHunter

AdaptHH†

EPH†

†
PHunter32bit

(†: Results on a 32-bit Java)

Figure 5.3: Scores of the PHunter, the SPOT-PHunter and three other hyperheuristics in the TSP domain in HyFlex
(p = 0.28 > 0.05). In summary, it can be observed that the SPOT hyperheuristic insignificantly and slightly improved the PHunter algorithm. The
result is consistent with the observation and interpretation on the insignificant and slight improvements in Section 5.4.1. What made the SPOT-PHunter
have better scores, as shown in Figure 5.3, was the scoring system. For example, in the instance No. 8 in Table 5.10, the PHunter64bit did not gain
much advantage in scoring after beating the SPOT-PHunter. Its score in this
instance was low, because it was still dominated by the AdaptHH and the
EPH.

5.5. DISCUSSION

5.5

127

Discussion

Compared to the algorithm in Xue, Chan, et al. (2011), the SPOT heuristic
generation implemented in this study was only able to make insignificant and
slight improvements in both theory and practice. Although the observations
partially verified the theoretical prediction by the indicators, there were certain
reasons that led to the insignificant and slight improvements. One main
reason is related to the capability of in-depth search of the local search LLHs
for modifying. For example, one of the major LLHs for modifying was an
efficient 5-Opt in Xue, Chan, et al. (2011), and the base LLHs in this study
were 3-Opt and 2-Opt local search. Another possible reason was the size of
the candidate set. When the size was 8 in HyFlex, the indicator cov showed
that a 8-nearest-neighbor-first candidate set had an almost 99% coverage
of suboptimum assignments. That meant there was very limited room for
the SPOT hyper-heuristic to improve. In comparison, the coverage of the
5-nearest-neighbor-first candidate set was about 97% and the maximum
room to improve was then about 3%.
The HyFlex is a great platform to test new hyper-heuristics. It is open
and flexible, and encapsulates a lot of algorithms and data structures in its
implementation. Nevertheless, there is still some room to extend and/or
to improve its LLHs and data structures further. For example, the 2-Opt
in HyFlex returned tours with more than a 13% excess over the referential
optima on average, as shown in Table 5.6, when the 2-Opt local search in
Helsgaun (2000) could find tours within a 4% excess, in general, in the same
data set of benchmark instances. Some other candidate sets could also be

128

CHAPTER 5. APPLICATION I: THE TSP DOMAIN

considered. For example, according to experiments, a k-quadrant candidate
set noticeably reduced the average excess of tours by a considerable 35%.
The same implementation of the SPOT hyper-heuristic generated new LLHs,
that further reduced the average excess of the k-quadrant candidate set by
around 1%, by sorting the edges in the k-quadrant order. The SPOT heuristic
generation in this study can probably generate stronger LLHs in an updated
implementation of the TSP domain.

Chapter 6
Application II: The Permutation Flow-Shop
Scheduling Problem Domain

µεταβoλὴ δὲ πάντων γλυκύ.
Change in all things is sweet.
The Nicomachean Ethics
Aristotle

This chapter presents an application of the SPOT heuristic generation for
another well-known combinatorial optimization domain, the permutation
Flow-Shop scheduling Problem (FSP). The structure of this chapter is similar
to that of Chapter 5. A brief introduction to FSPs is given in Section 6.1, and
Section 6.2 describes the implementation of the FSP domain in HyFlex. The
adaptation and design of the SPOT heuristic generation are presented in Section 6.3. Experiments and observations are shown in Section 6.4, concluding
with general discussion in Section 6.5.

129

130

CHAPTER 6. APPLICATION II: THE FSP DOMAIN
Makespan

t1,3

Machine 3
t1,2

Machine 2
Machine 1

t1,1

t2,3

t2,2
t2,1

t3,3
t3,2

t3,1
Execution time

Figure 6.1: An example of an FSP schedule with a permutation ( J 1 , J 2 , J 3 )

6.1

An Introduction to the FSP Domain

FSP, sometimes referenced as F/permu/Cmax or n/m/p/Cmax notationally, aims at finding a permutation of n jobs that minimizes the makespan.
Figure 6.1 shows an example of an FSP schedule. FSP has been comprehensively studied since the 1950s and still attracts many researchers even
now, see S. M. Johnson (1954), Heller (1960), Ignall and Schrage (1965), Taillard (1993), Reza Hejazi and Saghafian (2005). FSP and its variants can be
applied to the traditional domains such as the surface mount technology and
printed circuit manufacturing as well as some innovative domains such as
multimedia object scheduling in the World Wide Web (Allahverdi & Al-Anzi,
2002).
As stated in Chapter 1, FSP was proven to be NP-Complete (Garey, D. S.
Johnson, & Sethi, 1976) and NP-hard to approximate with a factor less than
5/4 (Williamson et al., 1997). However, S. M. Johnson (1954)’s polynomialtime two-machine scheduling had a strong influence in the later algorithms,
such as Ignall and Schrage (1965)’s Branch-and-Bound, Campbell, Dudek,
and Smith (1970)’s CDS and Gupta (1975)’s approach. Similar to the TSP

6.2. IMPLEMENTATION OF THE FSP DOMAIN IN HYFLEX

131

domain, there are effective heuristics for FSPs, too. A well-known heuristic is
Nawaz et al. (1983)’s NEH heuristic which employs a “curtailed-enumeration”
and can obtain very satisfactory solutions — usually within a 5% excess over
the best-known solutions in large-scale FSPs. Other exact and heuristic
algorithms, such as Reeves (1995)’s genetic algorithm and Taillard (1990)’s
tabu search, are also popular in solving FSPs. Some hyper-heuristic research
has also been carried out. For example, Vázquez-Rodrı́guez and Ochoa (2011)
investigated evolving ranking functions for NEH to prioritize operations.

6.2

Implementation of the FSP Domain in HyFlex

HyFlex implemented the FSP domain and publicized a part of the FSP
instances for open tests in CHeSC 2011. There were twelve FSP instances, as
shown in Table 6.1. The instances were selected from a well-studied benchmark data set48 by Taillard (1993). All the problems were randomly generated
instances. The best-known solutions in the fourth column in Table 6.1 were
collected from Zobolas, Tarantilis, and Ioannou (2009). Three instances, Nos.
1, 3 and 8, were randomly selected into the test beds. Two more instances,
Nos. 10 and 11, which were not publicized before the competition, were also
included for evaluating the hyper-heuristics in CHeSC 2011.
Table 6.2 shows fifteen LLHs implemented in FSP domain. The LLHs
could also be categorized into the four classes: mutation, ruin-and-recreate,
local search and crossover. Similar to the crossover methods for TSPs,
crossover algorithms for FSPs started from two tours and created one new
48 Available at: http://mistic.heig-vd.ch/taillard/problemes.dir/ordonnancement.dir/

ordonnancement.html.

132

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.1: Twelve benchmark instances used in the implementation of FSP in
HyFlex
No.

Id

0
1
2
3
4
5
6
7
8
9
10
11

TA081
TA082
TA083
TA084
TA085
TA092
TA093
TA111
TA112
TA114
TA101
TA113

Size (n × m)

Best known

100 × 20
100 × 20
100 × 20
100 × 20
100 × 20
200 × 10
200 × 10
500 × 20
500 × 20
500 × 20
200 × 20
500 × 20

6202
6183
6271
6269
6314
10480
10922
26059
26520
26456
11195
26371

In CHeSC?
X
X

X
X
X

Notes
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated
Randomly generated

Table 6.2: The LLHs in the implementation of FSP in HyFlex
Type

Id

Parameters used

Notes

Intensity of Depth of
mutation
search
Mutation

Ruin-andrecreate

0
1
2
3
4
5
6

Local search 7
8
9
10
Crossover 11
12
13
14

Random reinsertion
Swap of two jobs
Randomization
Partial randomization
The NEH heuristic (Nawaz et al., 1983)

X
X
X

X

Partial reconstruction by the NEH
Partial reconstruction by a greedy stepper NEH

X
X

Iterated steepest descent 1-job interchange
Iterated first improvement 1-job interchange
Random steepest descent 1-job interchange
Random first improvement 1-job interchange
Order crossover
Precedence preservative crossover
Partially mapped crossover
One point crossover

solution. Other algorithms started from one tour and returned a new (or
same) one. Many LLHs were based on the effective NEH heuristic which
involved iterations of assignment trials (limited backtracking).

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

6.3

133

The Development of a SPOT for the FSP

This section presents an application of the SPOT hyper-heuristic for the
domain FSP, in the form of the three development phases mentioned in
Section 3.3:
P1: transformations for standards and sampling,
P2: determination of parameters, and
P3: generation of new LLHs (by modification).
All the domain-specific operations developed in phases P1 and P3 were
encapsulated in the interface functions within the class FSP UEA2, as shown
in Figure 4.1. The determined parameters were stored and used by the class
SPOT.

6.3.1

P1: Transformations and sampling

The designs of the U/EA and the U/EA2 standards are described in this
section. The first standard focuses on the input problem. Transformations
between permutations of FSPs and 0-1 matrices are introduced to meet the
U/EA standard. The latter standard aims at delivering meaningful and
sufficient information in terms of attributes, which can be partially borrowed
from existing algorithms.

134

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

6.3.1.1

Design and validation of transformations

Every solution of an FSP is a unique permutation (or sequence) of n jobs.
The size of the search space is n!. When n is a large number, the size can be
approximated by Stirling’s formula49 :

n! =

√

2πn

 n n
e

e λn ,

(6.1)

1
where 12n1+1 < λn < 12n
. Given an FSP, a Boolean function of the relationship

“following” between two jobs J i and J j can be defined as follows:

f l (i, j) =




0

iff. index of J i ≤ index of J j ,



1

otherwise.

(6.2)

Any permutation of n jobs can be transformed to an n × n 0-1 (or Boolean)
matrix, where each element is the value of function f l on its row index and
column index. For example, a permutation ( J 3 , J 2 , J 1 ) can be written as the
matrix





 0 1 1 


 0 0 1 




0 0 0

(6.3)

The conversion from a matrix to a permutation can be stated as follows. Let
the summation ∑1≤ j≤n f l (i, j) of the i-th row be the weight of the i-th job J i .
The permutation refers to the ordered jobs which are sorted by the weights
in the ascending order.
49 See http://en.wikipedia.org/wiki/Stirling’s approximation.

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

135

It is clear that f l (i, i ) = 0 must be a constant for any i. In a designed transformed FSP, each f l (i, j) is considered as a variable, where i , j. Therefore,
there are n2 − n Boolean variables. The new transformed problem can be
described as follows.
In a transformed FSP, n jobs are given to be processed on m
machines in the same order, the process time of job i on machine
j being ti,j (i = 1, 2, . . . , n; j = 1, 2, . . . , m). A 0 or 1 value of
“following” needs to be determined for each pair of jobs such that
the total elapsed time (makespan) is minimized.
The variables are unconstrained, and each variable is Boolean. Henceforth,
the transformed FSP meet the U/EA standard. The cardinal number of the
2

new search space is 2n −n . The new search space is, actually, much larger
than the original search space when n is a large number, for example,
n2 − n

limx→+∞ 2 n!

= limx→+∞ √ 2

n2 − n

2πn( ne )
2

n

= limx→+∞ eln 2(n −n)−
= limx→+∞

2
eln 2n

ln 2+ln π +ln n
−n(ln n−1)
2

(6.4)

= +∞.
Therefore, the transformation from permutations to matrices is a one-toone mapping and the transformation from matrices to permutations is a
many-to-one mapping.
Four instances were selected from HyFlex to test the compatibility of
the transformed FSP in the U/EA standard, as listed in Table 6.3. Even
though the instances seemed to be among small-scale instances in HyFlex,

136

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.3: Four benchmark FSP instances used in the development of the
SPOT algorithm
No.

0
1
2
3

Id

TA081
TA082
TA083
TA084

Size (n × m)

100 × 20
100 × 20
100 × 20
100 × 20

% excess the best-known

Resemblance r (%)

Suboptimum 1

Suboptimum 2

1.0158
0.8895
0.7176
0.8614

1.0803
1.0189
0.7973
0.9890

79.76
75.50
87.72
75.33

Approximability r̄ (%)

79.58

they were actually complicated enough50 to test and to train the SPOT hyperheuristic. In each instance, two arbitrary suboptima were selected to check
the resemblance between them. The approximability r̄ was 79.58%, as shown
in Table 6.3. Both approximability and generality were greater than those
in Chapter 5, respectively. Therefore, the transformed FSP was regarded as
compatible to the SPOT hyper-heuristic.

6.3.1.2

Design of decision attributes

Designing the attributes to meet the U/EA2 standard was the last task in
P1. The design involved the definitions of the groups of competitors, the raw
attributes and the label(s), similarly to that in Chapter 5.3.1. Each variable of
a “following” was considered as a group consisting of two competitor assignments: 0 and 1. Therefore, there were n2 − n groups and p = 2 assignments
in each group. The function getCompetitiveGroups() returned an array
of 2n2 − 2n elements of assignments, with n2 − n unique identities of groups.
Many existing heuristics, such as Palmer (1965), Campbell et al. (1970)’s
CDS, Gupta (1975) and Dannenbring (1977)’s RA, have proposed many
50 Evidence shows that they are still unsolved.

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

137

Table 6.4: The 5m + 3 raw attributes designed for a job J i against another job
J j in the FSP domain (i , j)
Raw att

Cols

Mean t
Std dev
Min t
Max t

1
1
1
1

Dist

1

Palmer
Gupta

1
1

Real time
CDS part1
CDS part2
RA part1
RA part2

m
m−1
m−1
m−1
m−1

∑

5m + 3

Definition

Reference

avg (ti,1 , ti,2 , . . . , ti,m )
stdev (ti,1 , ti,2 , . . . , ti,m )
min (ti,1 , ti,2 , . . . , ti,m )
max (ti,1 , ti,2
h , . . . , ti,m )
i
∑2≤k≤m−1 (m − k ) × |ti,k − t j,k−1 |
+ti,1 + ti,m
∑1≤k≤m [−(m − 2k) × ti,k ]
sgn† (ti,m − ti,1 )/ min(ti,1 + ti,2 , ti,2 + ti,3 ,
. . . , ti,m−1 + ti,m )
{ti,k |1 ≤ k ≤ m}
{∑1≤l ≤k ti,k |2 ≤ k ≤ m}
{∑k≤l ≤m ti,k |1 ≤ k ≤ m − 1}
{∑1≤l ≤k (m + 1 − k)ti,k |1 ≤ k ≤ m − 1}
{∑k≤l ≤m (m + 1 − k)ti,k |2 ≤ k ≤ m}



1
†: the sign function sgn( x ) = 0


−1,

Palmer (1965)
Gupta (1975)

Campbell et al. (1970)
Campbell et al. (1970)
Dannenbring (1977)
Dannenbring (1977)

x > 0,
x = 0,
x < 0.

decision attributes. Given a variable f l (i, j)51 , there are ra = 5m + 3 raw
attributes defined, as shown in Table 6.4. The first column in Table 6.4
shows the name of a raw attribute. The second column is the number of
raw attributes for each category. The third column presents the equations to
calculate the attributes. Some references are given in the last column. As a
result, the function FSP UEA2.getRawAttributes() returned a 2D array
(in 2n2 − 2n rows and 5m + 3 columns) of Boolean values. The class SPOT
then generated 2 × ra columns of normalized attributes, because there were
only two competitors in each group and some redundant attributes could be
removed. A generated data table had 10m + 6 columns and 2n2 − 2n rows.
The data table can be a subset in the U/EA2 attribute space.
51 A job J

i against a competitor J j .

138

CHAPTER 6. APPLICATION II: THE FSP DOMAIN
Because the group was defined with the Boolean values of the f l function,

the label was set to the f l function and its possible values were 0 and 1. As a
result, the objective of finding a permutation was transformed into finding a
matrix of “following” relations. Examples of generated data tables of labeled
training data and unlabeled test data can be found in Table B.5 and Table B.7
in Appendix B.

6.3.2

P2: Parameter determination

There are two important parameters to be determined one by one in
this phase, as described in Section 3.3.1. The first parameter is the size of
subproblems n0 , which is numeric. The other is the method of sampling for
generating subproblems, and it is non-numeric.

6.3.2.1

The size n0 of subproblems

The tests of n0 were set up as follows. The targets of sampling in the
FSP domain were the jobs, excluding the machines. The reason was that the
number of columns for training was a (linear) function against the number
m of machines. If the machines were sampled, the attributes in the training
data and the attributes in the test data must be inconsistent. Random selection
sampling on jobs was employed to determine the size n0 . The function
SPOT.run() was executed without the final step of generating LLHs. The
solving time of a subproblem was set to 6 seconds for the PHunter algorithm.
The algorithm selection was not enabled for two reasons. The main reason
was that the algorithm selection could noticeably reduce the precision of the
machine learning. The second reason was the intent to test the performance

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

139

(49,48,68,74,42,9,99,67,75,46,4,61,89,58,30,88,18,78,13,97,34,64,71,35,1,69,26,14,32,51,
25,33,0,98,31,47,81,8,12,70,52,94,29,83,96,10,2,63,7,92,65,66,86,45,22,59,19,3,90,21,60,
95,77,39,79,53,6,82,41,43,72,50,84,57,27,62,37,24,44,91,36,85,73,28,16,76,5,40,15,80,55,
23,87,17,11,93,20,38,54,56)
(i) A predicted permutation by the classifiers in the SPOT algorithm
(48,49,75,4,64,30,95,99,97,67,80,26,68,89,42,57,65,54,94,23,31,45,58,74,32,9,71,46,90,72,
78,62,34,92,70,13,18,47,15,96,29,60,76,37,25,59,33,14,22,6,77,69,41,83,87,12,61,21,82,51,
84,27,36,79,2,7,24,8,66,0,39,35,1,17,93,50,11,88,63,86,55,3,43,73,53,16,91,5,52,44,85,40,
10,81,98,19,20,38,28,56)
(ii) An NEH permutation by the NEH heuristic
(49,48,75,4,99,68,67,42,30,64,89,74,97,9,58,46,26,78,71,18,34,32,13,31,94,65,61,95,47,70,
14,25,45,69,33,29,92,96,90,51,57,88,12,35,1,83,72,80,59,0,60,22,8,62,2,23,77,7,6,54,21,37,
66,41,63,82,15,76,79,98,52,81,86,84,39,27,10,3,36,24,50,87,53,19,43,73,91,17,44,16,93,11,
55,85,5,40,28,20,38,56)
(iii) A weighted permutation by adding up the weights in (i) and (ii)

Figure 6.2: Examples of a predicted permutation, an NEH permutation and
a weighted permutation for the FSP instance TA082
of the supervised learning on a relatively large number (tens to a couple
of hundreds) of columns of data. The computation time of the embedded
subproblem solver PHunter was also set to six seconds. After training
classifiers and predicting labels, the generated matrix was transformed to the
weights of summations of “following” back to the given variables for sorting
a permutation of jobs. Similar to the weighted distance which combined
the general results of learning and variable-specific data of lengths of edges
(Xue, Chan, et al., 2011), the weighting in the FSP domain also considered a
random permutation by NEH to calculate the weighted summations, with
each permutation weighted as a half. An example is shown in Figure 6.2.

140

CHAPTER 6. APPLICATION II: THE FSP DOMAIN
The indicator used in the tests was r, which represented the average

resemblance between the “following” matrix of a given permutation to those
of the two suboptima listed in Table 6.3. The meaning of r could also be
regarded as the average coverage of suboptimum assignments of “following”.
Similarly to the development for the TSP, it is clear that 0 ≤ r ≤ 1 and a
greater value of r usually stood for a better result of learning and a higher
quality of including more promising assignments. It should be noted that
the labels generated by classifiers could not be directly used to measure
the coverage r before converting to a permutation. There were possibly
a lot of inconsistencies in a matrix of predicted Boolean values, such as
f l (i, j) = f l ( j, i ) = 0. When the labels were converted to weights and a new
“following” matrix was generated, the indicator r can be used to measure the
average coverage.
Five values of n0 were tested, as shown in Figure 6.3. One hundred
independent tests were carried out for each value for each instance. In the
“SPOT (direct)” entry, the jobs were sorted by the predicted weights in a
descending order. In the “SPOT (weighted permutation)” entry, the jobs
were sorted in a combined weight, as shown in Figure 6.2. The “longesttotal-time-first” presents a referential r of the permutation in the descending
order of the summation ∑1≤ j≤m ti,j of the time of a job J i . The “NEH” shows
r of the results of Nawaz et al. (1983)’s NEH heuristic which initialized with
the longest-total-time-first sequence. The “average of random NEHs” refers
to the average value of 100 NEH heuristics initialized with a randomized
permutation.
It can be observed from Figure 6.3 that the indicator r on the results of

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

Mean of r (%)

75
70
65

ation)

ermut
hted p ×
g
i
e
w
(
+
SPOT ×
t)
c
e
ir
+ OT (d
SP

×
+

+
×

141

+
×
NEH

Avg of random NEHs

60
55

longest-total-time-first

50
10

20

30

40

50

Standard deviation of r (%)

n0
5

+

4

SPO
+ T (direc
t)
× S
3
+
POT (weig
hted perm
utation) +
×
×
×
2 Avg of random NEHs

+
×

1
0
10

20

30

40

50

n0

×
+

Time cost (s)

30

20

10

×
+
utation) ×
+
SPOT (weighted perm
×
+
×
+
SPOT (direct)

0
10

20

30

40

50

n0

Figure 6.3: Average test results for determining n’ in the development of the
SPOT algorithm, on the four FSP training instances (100 runs)

142

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

the SPOT hyper-heuristic were noticeably greater than for the longest-totaltime-first sorting and greater than the results of the NEHs when n0 ≥ 20.
When the size n0 increased, both r and the time spent on learning gradually
and steadily increased. However, the two entries of SPOT methods were too
close on both r and time when n0 ≥ 30, therefore, the standard deviation of r
is included in Figure 6.3 to check the variances. The size n0 was finally set
to 30 after a careful balance between r (effect), time cost and, especially, the
possibility of training the classifiers from parallel subproblems, like those
described in Section 5.3.2.The weighted permutation is selected, because it
led to less variance and there was no significant differences on the time cost
and the means when n0 = 30.

6.3.2.2

The methods of sampling for subproblems

Segment selection is another sampling method, which chooses a number of
consecutive jobs from a high-qualified permutation. Because all the instances
were randomly generated, a random order or a noncompetitive permutation,
such as the conventional longest-total-time-first sorting in Figure 6.3, might
not generate samples that were much different from those by random selection. The segment selection in this chapter was based on the effective results
from the NEH heuristic.
Two groups of tests were carried out on the sampling methods, as shown
in Table 6.5 and Table 6.6. In the first group in Table 6.5, the segment selection
was compared with the random selection. It was found that the segment
selection seemed non-competitive against the random selection. In the other
group of tests in Table 6.6, two subproblems were selected in parallel. The

6.3. THE DEVELOPMENT OF A SPOT FOR THE FSP

143

Table 6.5: Average results and significants of r in tests on different subproblems sampled, where n0 = 30 (100 runs, significant values of p in bold)
Instance

Segment

Random

ANOVA
F

p

72.853
73.054
72.174
68.733

0.00
1.43
15.29
34.61

0.96
0.23
0.00
0.00

71.703

—

0.30

selection (%)

selection (%)

0
1
2
3

72.845
72.888
71.451
67.726

Average

71.227

Table 6.6: Average results and significants of r in tests on different combination of parallel subproblems sampled, where n0 = 30 (100 runs, significant
values of p in bold)
Instance

Segment + random

Random ×2

ANOVA

selections (%)

selections (%)

F

p

0
1
2
3

72.569
74.131
72.494
70.563

74.785
74.962
73.421
71.566

77.31
4.86
53.46
68.03

0.00
0.03
0.00
0.00

Average

72.439

73.183

—

0.01

“ Segment + random selections” column in Table 6.6 is the “Random ×2
selections” column and denotes two parallel random selections. Given the
methods of sampling as categories, the F values and the significance p are
shown in the last two columns in both tables, respectively.
It can be observed from Table 6.5 that the random selection resulted
in significant larger values of r in two instances and the average r of the
random selection was +0.476% more. According to Table 6.6, the parallel
selections with two random selections were significantly better than those
with a segment selection and a random selection in three instances, and
the average result was +0.744% more on r and the improvements were
very significant (p = 0.00). The final sampling method was set to two

144

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.7: Average results and significants of time of generation (timeoverall )
in tests on different combination of parallel subproblems sampled, where
n0 = 30 (100 runs, significant values of p in bold)
Instance

Segment + random

Random ×2

ANOVA

selections (s)

selections (s)

F

p

0
1
2
3

25.87
25.37
25.70
25.77

26.12
25.59
25.71
25.59

1.33
0.97
0.00
0.78

0.25
0.33
0.97
0.38

Average

25.68

25.75

—

0.48

parallel random selections. In fact, Kadioglu et al. (2010) also noticed that
“tuning on a collection of instances helps prevent over-tuning and allows
parameters to generalize to similar instances”. The difference between the
selected method and the longest-total-time-first in Figure 6.3 on r was a
considerable +20.976%. The noticeable difference indicated that some LLHs
generated by the SPOT algorithm should be capable of outperforming their
base algorithms significantly.
In addition, the time for generating the predicted permutations is shown
in Table 6.7. It can be observed that there was no significant difference in
any of the instances (as well as the average value) on time of the SPOT
hyper-heuristics for the two parallel methods. In summary, the two parallel
random selections were able to sample expected subproblems for training
the classifiers in about the same time.

6.3.3

P3: The generation of new LLHs

In the following, the labeled values of the following relationship are
converted to permutation to make the results more consistent. As designed

6.4. EXPERIMENTS AND OBSERVATIONS

145

and tested, the weighted permutation was used to modify some mutation
LLHs provided in HyFlex in the function FSP UEA2.setIntLabels(). The
LLHs that were modified, and their modifications are listed in Table 6.8.
The basic idea was in changing a mutation heuristic to a guided mutation
and changing a local search to a guided local search. The two LLHs in the
ruin-and-recreate category are also considered as multi-variable mutations.
No. 9 and No. 10 LLHs were modified to greedy random local search LLHs by
a greedy initialization and a greedy randomization. A rejection function was
employed in the greedy randomization for modifying the search sequences in
No. 9 and No. 10 LLHs. Given a swap of two jobs (i, j), there was probability
to reject the swap:

1−

min(| predicted(i ) − location(i )|, | predicted( j) − location( j)|)
.
n

(6.5)

The rejection function denoted that the jobs closer to the predicted locations
rejected to be shuffled in a higher probability. No. 2 LLH was a complete
randomization for all the jobs. No. 3 LLH was a complete randomization of a
part of the permutation. No. 4 LLH was a NEH using the input permutation
as the initial sequence. They were not included in the modification. The local
search heuristics No. 7 and No.8 were not included in the modification. The
main reason was the scopes of the two local search LLHs were not modified
by the new permutations, Section 6.5 gives more details. Similarly to those
LLHs in the TSP domain, six new LLHs were generated as Nos. 0’, 1’, 5’, 6’, 9’
and 10’. The original LLHs are called the base algorithms.

146

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.8: The modified LLHs for FSPs and modifications by the SPOT
heuristic generation
Type

Id

Original behavior

Modified behavior

Mutation

0

Reinserting a random job to
a random index.
Swapping two random jobs.

Choosing 3† random indices, doing
the most favorite reinsertion (1 in 6).
Choosing 3† random indices, doing
the most favorite swap (1 in 3).

Removing some random
indices and inserting back
by calling iterated NEHs.
Removing some random
indices and inserting back
by calling stepper NEHs.

On each removal, choosing 3† random
indices, always removing the least
favorite index.

9

Randomizing a sequence,
trying width-first steepest
descent 1-interchange
one by one.

Building a greedy sequence in the
descending order of the distant to
the predicted index, randomizing the
sequence by swaps with a rejection
function.

10

Randomizing a sequence,
trying width-first first
improvement 1-interchange
one by one.

1
Ruin-andrecreate

5

6

Local
search

Same as above.

Same as above.

†: Effective parameters according to tests, not rigorously examined in thesis.

6.4

Experiments and Observations

This section describes two groups of tests. In the first group of experiments, the individual performance of each new LLHs was examined, with
comparisons to their base algorithms. In the other group of tests, the five
new LLHs replaced their base algorithms. The heuristic selection method
PHunter solved the updated domain consisting of new LLHs. The results
were compared to the solving of PHunter with the original LLHs and the
winners in the FSP domain in CHeSC 2011.
All the tests were carried out on the same instances which had been
selected for the CHeSC 2011 competition, i.e. the numbers 1, 3, 8, 10 and 11

6.4. EXPERIMENTS AND OBSERVATIONS

147

shown in Table 6.1. The size n0 of the subproblems was set to 30 and two
subproblems were generated in two parallel random selections to provide
training data, as described in Section 6.3.2. The overall computation time
was set to the same as the normalized time in CHeSC 2011, which was
equivalent to 10 CPU minutes for a Pentium 4 3.0 GHz computer. All the
time costs of the processes in the SPOT hyper-heuristic, such as sampling
subproblems, solving the subproblems, learning instance-specific knowledge
and modifying LLHs, were fully included in the overall time in order to
make the results comparable. For the embedded subproblem solver PHunter,
the time limit for solving each subproblem was 1% of the overall time, i.e.
equivalent to 6 CPU seconds on a Pentium 4 3.0 GHz machine. The time cost
of rest processes the in SPOT was not directly restricted.

6.4.1

On the individual LLHs

Table 6.9 shows the results on the final makespan of the newly generated
LLHs in terms of percentage excess over the best-known solutions. There
were 100 runs for each configuration and each problem instance. In each
run, a random initial tour was generated by the NEH with a random seed.
All the LLHs started from the same initial tour independently. One-way
ANOVA tests were also included to check the significance of the changes.
The third column in Table 6.9 is the test instance, and the fourth and the fifth
columns show the excesses of the best and the average makespans found
by the base algorithms in HyFlex, respectively. The sixth and the seventh
columns indicate the best and the average makespans found by the new
LLHs, respectively. The eighth column denotes the change in the average

148

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

excess, where improvements (negative numbers) are written in bold. The
last two columns show the F values and the confidence p examined by the
one-way ANOVA tests. The significant values (p ≤ 0.05) are in bold.
It can be observed from the eighth column that all the newly generated
LLHs in the ruin-and-recreate and local search classes generally returned
better (negative in % excess) permutations than their base algorithms. The
average improvements in the ruin-and-recreate class were statistically significant and was more than those in the local search class. The differences
between the new LLHs and their base algorithms in the local search class
were almost significant, though the null hypophyses could not be fully objected (p = 0.10, 0.08). In comparison, the new LLHs in the mutation class
were neither significant nor robust. The observations on the LLHs in the
latter two classes were consistent with the interpretation and prediction
according to the indicator r in Section 6.3.2. The insignificant and unstable
improvements in the mutation category were because of the main objective
of mutation operations, which aim at diversifying a given solution.
The time cost of generating the LLHs is shown in Table 5.7. The third to the
sixth columns show the minimum, the average, the maximum values and
the standard deviation of the time cost of the whole SPOT hyper-heuristic for
each instance. It should be noted that a limit of 12 equivalent CPU seconds
for solving two subproblems was included in the data. It can be observed
that the maximum time cost in the large-scale instance with 500 jobs and
20 machines was about 60 seconds. Therefore, the SPOT could carry out
instance-specific heuristic generation in an “on-the-fly” fashion for practical
problems, for example, in a 10 minute time.

6.4. EXPERIMENTS AND OBSERVATIONS

149

Table 6.9: Test results on the makespan of the generated LLHs and their base
algorithms (10 runs, 100 iterations in each run, Depth of search = 1, Intensity
of mutation = 1, improvements on average performance in bold, significant
values of p in bold)
Category

Mutation

Group

No. 0
versus
No. 0’

No. 1
versus
No. 1’

Ruin-andrecreate

No. 5
versus
No. 5’

No. 6
versus
No. 6’

Local
search

No. 9
versus
No. 9’

No. 10
versus
No. 10’

Inst

1
3
8
10
11

1
3
8
10
11

1
3
8
10
11

1
3
8
10
11

1
3
8
10
11

1
3
8
10
11

% excess by
the base LLH

% excess by
the new LLH

Min

Avg

Min

Avg

4.642
4.100
2.342
3.734
2.154

7.004
6.435
3.069
5.488
2.827

4.432
4.291
2.206
3.698
2.063

7.053
6.378
3.081
5.484
2.837

4.351
4.371
2.247
3.975
2.154

4.739
3.892
2.391
3.805
1.843

3.396
2.983
1.836
3.037
1.710

3.510
2.999
1.821
2.930
1.638

3.574
2.903
1.829
2.966
1.566

∆ avg

ANOVA
F

p

0.049
-0.057
0.012
-0.004
0.010

1.41
1.93
1.00
0.02
0.71

0.23
0.17
0.32
0.90
0.40

Avg improvement (%)

-0.096

Avg

0.40

7.485
6.865
3.175
5.760
2.919

7.447
6.834
3.167
5.770
2.924

-0.038
-0.031
-0.008
0.009
0.005

0.59
0.40
0.41
0.05
0.15

0.44
0.52
0.52
0.83
0.70

Avg improvement (%)

0.177

Avg

0.60

6.510
5.742
3.103
5.415
2.827

6.188
5.513
3.012
5.381
2.766

-0.321
-0.230
-0.092
-0.033
-0.061

115.81
61.40
65.62
1.58
31.23

0.00
0.00
0.00
0.21
0.00

Avg improvement (%)

2.931

Avg

0.04

5.127
4.633
2.510
4.451
2.283

4.867
4.468
2.426
4.390
2.225

-0.260
-0.165
-0.084
-0.061
-0.057

95.50
37.20
70.67
6.18
36.93

0.00
0.00
0.00
0.01
0.00

Avg improvement (%)

3.171

Avg

0.00

5.097
4.565
2.424
4.240
2.165

5.016
4.507
2.399
4.223
2.147

-0.081
-0.058
-0.024
-0.018
-0.018

10.24
5.20
5.71
0.75
3.72

0.00
0.02
0.02
0.39
0.05

Avg improvement (%)

1.023

Avg

0.10

5.170
4.621
2.441
4.300
2.185

5.130
4.577
2.421
4.267
2.166

-0.041
-0.044
-0.020
-0.033
-0.019

2.65
3.23
3.98
2.51
4.00

0.10
0.07
0.05
0.11
0.05

Avg improvement (%)

0.839

Avg

0.08

4.561
3.876
2.244
3.957
2.108

4.221
2.951
2.195
3.868
2.097

3.105
2.616
1.791
2.814
1.627

3.299
2.903
1.761
2.975
1.551

3.607
3.063
1.701
2.930
1.612

150

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.10: Average overall time (timeoverall ) of generating new LLHs (10
runs)
Instance

1
3
8
10
11

Size

Time (s)

(n × m)

Min

Avg

Max

Std dev

100 × 20
100 × 20
500 × 20
200 × 20
500 × 20

11.19
12.11
59.25
19.35
60.19

13.51
14.26
62.77
21.65
62.41

16.77
17.24
65.38
24.87
64.80

1.61
1.59
2.06
1.87
1.50

Table 6.11 shows the computation time of the new LLHs and their base
algorithms for solving problems. No. 1 and No. 2 LLHs only changed the
indices of some jobs without any iteration or backtracking. Therefore, the
computation time and the differences in computation time in the first two
groups was not noticeable. The time was also hardly noticeable in the tests
for instances No. 1 and No. 3. Slight increments on time could be seen in
the results of the LLHs No. 5’ and No. 6’, while slight reductions could be
observed in those of the LLHs No. 9’ and No. 10’. Both the percentages
of the increments and the reductions were no more than 5%. In summary,
the newly generated LLHs spent almost the same amount of time as their
base algorithms, respectively. The new LLHs are eligible to replace their base
algorithms in HyFlex in practice.
Figure 6.4 represents the non-dominance of four new LLHs. The tests
were carried out on all the five 100 × 20 instances implemented in HyFlex. It
can be observed from Figure 6.4 that the two LLHs in the ruin-and-recreate
category were considerably improved in solution quality and the two in
the local search class were slightly improved in solution quality. All the
differences on time were very limited. The relevant significances are included
in Table C.6 in Appendix C. In summary, it is confident to have four new and

6.4. EXPERIMENTS AND OBSERVATIONS

151

Table 6.11: Average run time of each LLHs (10 runs, 100 iterations in each
run, instrument error < 0.16 ms)
Inst

Average run time of LLHs in the four group (ms)
0 0’ ∆

1† 0 0
3† 0 0
8 0 0
10 0 0
11 0 0

1 1’ ∆

0 0 0
0 0 0
0 0 0
0 0 0
0 0.1 0

5

5’ ∆

6

6’

∆

9’ ∆

9

10

10’

∆

0
0
0 0
1
1
0
0
0 0
0
0 0
0
0
0 0
1
1
0
0
0 0
0
0 0
0 40.7 40.7 0.0 13089.5 13051.7 -37.8 120.8 121.0 0.2 120.9 120.9 0.0
0 6.7 6.6 -0.1 831.9 831.3 -0.6 19.1 19.3 0.2 19.0 19.1 0.1
0 40.8 40.8 0.0 13067.3 12947.3 -120.0 120.9 121.1 0.2 120.7 120.4 -0.4

Avg imp (%)

0.3

0.4

-0.5

-0.1

†: Not included in the average improvements due to the level of instrument error.

Average excess over the best-known (%)

6.5
No. 5

No. 10

No. 5’

6

No. 10’
No. 9

5.5

No. 9’

No. 6

5

No. 6’
4.5

1

10
Average computation time (ms)

100

Figure 6.4: Comparisons between the new LLHs and their base algorithms for
solving the 100 × 20 FSP instances TA81, TA82, TA83, TA84 and TA85
non-dominated LLHs in the ruin-and-recreate and the local search categories
in the results of the SPOT heuristic generation.

152

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Table 6.12: The top three hyper-heuristics in the FSP domain in CHeSC 2011

6.4.2

No

Algorithm

Score

Author

1
2
3

ML
AdaptHH
VNS-TW

39
37
34

Mathieu Larose (Université de Montréa)
Mısır et al. (2012)
Hsiao, Chiang, and Fu (2011)

Comparisons with other hyper-heuristics

Based on the four new LLHs generated by the SPOT algorithm, many
existing hyper-heuristics, especially heuristic selection methods, can benefit.
A typical and straightforward way is to replace the base algorithms by the
generated LLHs. After replacement of their base algorithms, the set of fifteen
LLHs of the FSP domain in HyFlex consisted of the six new LLHs. The
heuristic selection method PHunter was also employed to run on the basis
of an updated LLH set. The combination of the SPOT preprocessor and the
body of the PHunter is noted as the SPOT-PHunter in the following tests.
The HyFlex framework and the competition CHeSC 2011 also provided
reliable results for twenty hyper-heuristics in the FSP domain52 . The top
three hyper-heuristics are listed in Table 6.12. There were five instances
selected, as shown in Table 6.1, so a maximum possible score was 50 for each
entry.
The SPOT-PHunter was tested as a normal algorithm player in CHeSC
2011. The parameters of the SPOT part were the same as determined in this
chapter, and the parameters of the PHunter part were the same as determined
in Chan et al. (2012). All the six new LLHs were used to replace their base
algorithms. The tests were conducted for the five instances selected in CHeSC
52 See http://www.asap.cs.nott.ac.uk/external/chesc2011/raw-results.html.

6.4. EXPERIMENTS AND OBSERVATIONS

153

Table 6.13: Median makespan of the PHunter, the SPOT-PHunter and three
other hyper-heuristics in the FSP domain in HyFlex (31 runs, best results in
bold for each instance)
Inst

.

1
3
8
10
11

Best results† in CHeSC 2011

Test results in thesis

ML

AdaptHH

VNS-TW

PHunter64bit

SPOT-PHunter

ANOVA p

6245
6323
26800
11384
26610

6240
6326
26814
11359
26643

6251
6328
26803
11376
26602

6246
6350
26822
11375
26609

6242
6323
26790
11359
26604

0.55
0.00
0.00
0.01
0.28

†: Results on a 32-bit Java.

2011, and each was tested 31 times. The time cost in each test was equivalent
to 10 minutes on a Pentium 4 3.0 GHz CPU. The median values were used
for scoring, as shown in Table 6.13. The second to the fourth columns show
the median solutions found by the top three hyper-heuristic approaches in
CHeSC 2011. The fifth and the sixth columns show the median results of the
PHunter and the SPOT-PHunter. The last column shows the significance
of one-way ANOVA test. The best results for each instance are written in
bold. Two entries of the PHunter64bit and the SPOT-PHunter were added
to the set of twenty hyper-heuristics. The updated scores, according to the
scoring system, of the top five hyper-heuristics are shown in Figure 6.5. It
should be noted that the tests in CHeSC 2011 were carried out on a 32-bit
Java environment.
It can be seen from Table 6.13 that the SPOT-PHunter found the three
best results in the five instances. The final score of the SPOT-PHunter was
43 out of 50, as shown in Figure 6.5. The runner-up was the AdaptHH
with a score of 31. The PHunter64bit solved instances with base algorithms
instead of newly generated LLHs. The PHunter64bit gained a score of 20.5,

154

CHAPTER 6. APPLICATION II: THE FSP DOMAIN
43

30

32

By
t

he

SP
OT

40

29.5
26

20

20.5

10

PHunter64bit SPOT-PHunter

AdaptHH†

ML†

VNS-TW†

(†: Results on a 32-bit Java)

Figure 6.5: Scores of the PHunter, the SPOT-PHunter and three other hyperheuristics in the FSP domain in HyFlex
which was the lowest in the five hyper-heuristics. The SPOT-PHunter won
the PHunter64bit in every instance according to Table 6.13. Furthermore,
the difference between the result of the PHunter64bit and that of the SPOTPHunter was significant in three instances out of the five, as shown in Table
C.7 in Appendix C. The observations on the scores and the improvements
meant that the SPOT heuristic generation generally improved the heuristic
selection of the PHunter64bit in solving in every FSP instance, where the
improvements were significant in three instances.
The main reason for the significant improvements is the observations
and interpretation of the significantly non-dominated LLHs, as presented in
Section 6.4.1. Another reason was No. 7 and No. 8 local search LLHs did not

6.5. DISCUSSION

155

Table 6.14: A flawed transformation of FSPs that resulting in a very low
approximability
No.

0
1
2
3

Id (n × m)

TA081
TA082
TA083
TA084

% excess the best-known

Resemblance r (%)

Suboptimum 1

Suboptimum 2

1.0158
0.8895
0.7176
0.8614

1.0803
1.0189
0.7973
0.9890

7.00
2.00
18.00
6.00

Approximability r̄ (%)

8.33

“homogenize” the differences brought by the new LLHs. That was probably
because the depth of search53 was only 1 in those local search LLHs. If a
3-job interchange local search or a 4-job interchange was implemented, the
new LLHs might not return significantly different results.

6.5

Discussion

The transformations in this chapter and Chapter 5 met the requirements
of the U/EA standard and had acceptable approximability, although, there
were some flawed transformations that failed to do so. For example, given
an FSP problem, let the position (index) of each job in the permutation be a
variable. The transformed problem seemed to meet the standard of U/EA,
however, the approximability was very low, as shown in Table 6.14. The
approximability of the transformed model was 8.33% which was too low for
the SPOT algorithm.
The reason why the main model in this chapter was converting the relationship of “following” into the variable, is based on the following observa53 It means the true integer depth of local search (range: 1, 2, . . . , n), not the decimal “Depth

of search” parameter (range: 0 to 1) in HyFlex.

156

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

tion. Taking the two suboptima for the instance No. 0 as examples, as shown
in Table A.1 in Appendix A, one can easily find that both permutations end
with the substring “{. . . , 48, 12, 6, 40}”. The first suboptimal permutation
started with a substring “{53, 77, 58, 21, . . .}”, and the other started with a
substring “{53, 21, 46, 0, . . .}”. In the four extreme positions in the heads and
the four in the tails, the two suboptima had six of the same jobs and five of
them were exactly in the same positions. After further investigation, it was
not difficult to discover that the relationship of “following” was much more
meaningful and robust for suboptima.
It could be concluded that the parameter n0 is more important than the
method of sampling, according to the values of resemblance r in Figure 6.3,
Table 6.5 and Table 6.6. It is clear that the differences of r varied much more
in Figure 6.3 than in the other two tables. The observation, in fact, verified
the development process of the determination n0 first, with the method set to
random selection.
The LLHs generated by the SPOT in this chapter were mutations and
some local search heuristics with a slight changing in their behavior. No. 7
and No. 8 local search LLHs did not benefit from the generated permutations.
The main reason was that the depths of search of the four local search LLHs
were, in fact, equal to 1. Both were variants of the 1-interchange local search.
The differences were mostly on the criterion of acceptance (steepest descent
or first improvement). The scopes of search of the two local search methods
did not change noticeably by reordering their search routes according to
a generated permutation. The capability of the 1-interchange local search
actually was also limited. As a result, the differences on the makespan by

6.5. DISCUSSION

157

the newly generated LLHs, as shown in Table 6.9, seemed passed onto the
results of the SPOT-PHunter, as shown in Table 6.13. In fact, the predicted
permutation by the SPOT hyper-heuristic could possibly benefit the local
search LLHs by extending the scope of search to a higher level. For example,
if a 1-interchange was very favorite according to learning results but failed
to improve the permutation, a conditional 2-interchange could be enabled.

158

CHAPTER 6. APPLICATION II: THE FSP DOMAIN

Chapter 7
Discussion and Conclusions

In the confrontation between the stream
and the rock, the stream always wins, not
through strength but by perseverance.
Brown Jr., H. Jackson

Heuristic generation is a subclass of hyper-heuristics. Instead of solving a
given problem directly, a heuristic generation approach generates new heuristics. This thesis presents a heuristic generation methodology called SPOT
(Suboptimum- and Proportion-based On-the-fly Training). The proposed
method was implemented and verified in two well-known NP-Complete
combinatorial optimization problem domains. Section 7.1 gives some discussion on the findings, implications and limitations while concluding remarks
are given in Section 7.2.

159

160

CHAPTER 7. DISCUSSION AND CONCLUSIONS

7.1

Discussion on the SPOT Methodology and Findings

In summary, the SPOT hyper-heuristic is a novel methodology aiming
at improving existing heuristics and data structures with instance-specific
information. Furthermore, there are new standard and new indicators developed for SPOT so that the instance-specific information can be obtained
from optima, suboptima or suboptima of subproblems for a given problem.
The limited scale of learning enables the capability of on-the-fly execution
of SPOT. Applications in the well-known NP-Complete problem domains
such as the TSP domain and the FSP domain were successful and verified
the SPOT hyper-heuristic generation in general.

7.1.1

Principal findings

Referring to the development procedures and the experimental results,
the presented SPOT hyper-heuristic was shown to be capable of conducting an on-the-fly supervised learning for finding instance-specific information and generating new heuristics according to the obtained instancespecific information. Particularly, it was proven to generate significantly
non-dominated new heuristics for large-scale FSP benchmark instances in a
few minutes. In terms of operations, SPOT contains three phases: (i) input to
the problem with certain regulations, (ii) suboptimum- and proportion-based
supervised learning for the instance-specific information and (iii) modify (or
create) of heuristics.
The input problems regulations included the U/EA standard and the
indicator of “compatibility”. As a requirement of SPOT, the U/EA standard

7.1. DISCUSSION ON THE SPOT METHODOLOGY AND FINDINGS 161
of the input combinatorial optimization problem was applied in the two
domains. Three transformations of combinatorial optimization problems,
one for the TSP and two for the FSP, were designed to remove the intervariable hard constraints so that a transformed problem can meet the U/EA
standard. An indicator of “resemblance” r was employed to measure the
compatibility of each transformation. The generality of applying SPOT to
transformed problems was measured by the similarities between the average
resemblance values of the instances. The approximability of transformed
problems was measured by the overall average resemblance value of all the
instances. Later tests and discussion verified the U/EA standard and the
compatibility. According to experience, some problem domains are natural
for the U/EA standard and the compatibility, such as the TSP domain. Some
others with strong inter-variable hard constraints or strong inter-variable
correlations are hard to meet. An example is the personnel scheduling
problem involving complex contracts and licenses (staff skills) and a 24-hour
job demand.
A distinct feature of the SPOT hyper-heuristic is the capability of learning
approximate information on the basis of suboptima and sampled proportions
instead of optima and the original problem. This is the key to the success
of SPOT. As stated in the Chapter 3, the capability came from the U/EA
standard and the compatibility of problems. Attributes were developed to
represent information from the suboptima of subproblems. In this thesis,
an ordinary measurement was considered as a “raw attribute”. An automated attribute generation procedure helps populating many “normalized
attributes” and assists the data in the learning process to meet the U/EA2

162

CHAPTER 7. DISCUSSION AND CONCLUSIONS

standard. Many attributes were about comparisons and increments, which
have no unit but represent certain concerns of competition in the possible
assignments. The instance-specific information about suboptimality was
in the form of representing competition in the opponent assignments. In
fact, besides the stochastic machine learning, there could also be many other
candidates for explicit or implicit learning, such as the backbone (W. Zhang &
Looks, 2005) and tour-merging (Cook & Seymour, 2003) in the TSP domain.
The generation of new heuristics, in fact, would be much more often in the
form of modifying existing heuristics rather than creating new ones completely.
The key process in generating new heuristics is the transformation from
instance-specific information to numeric or non-numeric objects that directly
interacted with the search heuristics. This thesis presents some examples
on how to utilize the obtained instance-specific information to modify (or
guide) the search in existing heuristics. For example, in FSP, the predicted
permutation based on the “following” relationship can advise the promising
mutation operation from a number of random mutation operations. This
representation (or interpretation) might not be necessary if the learning
involved was not stochastic learning, such as the backbone (W. Zhang &
Looks, 2005) in the TSP domain.
In the practical development of SPOT, there were two main parameters
to be determined in the SPOT heuristic generation. One was the size of
the subproblems, and the other was the method(s) to sample subproblems.
The first one was more important than the second one, according to the
U/EA standard. The indicators, one or more, represent the resemblance r
and help identify better configurations. Other numeric and non-numeric

7.1. DISCUSSION ON THE SPOT METHODOLOGY AND FINDINGS 163
parameters are considered as secondary and adopted from the literature.
These parameters may, somehow, affect the whole algorithm, though the
suggested values (or methods) in the literature would be of low-risk and be
more robust in general.
The main theoretical findings were the formal definitions on hyper-heuristics,
heuristic section and heuristic generation. As far as is known, the definitions
are the first exact formulations, especially for the latter two. In brief, if a
hyper-heuristic can only return finite or countably infinite different heuristics, it is a heuristic selection approach; if a hyper-heuristic can return some
heuristics in uncountable possible algorithms, it is a heuristic generation. The
presented formal definitions can distinguish heuristic selection approaches
and heuristic selection methods precisely, and most of the results are consistent with previous work. Furthermore, the scopes of the notions were
considerably extended. For example, the backbone (W. Zhang & Looks, 2005)
and the tour-merging (Cook & Seymour, 2003) in the TSP domain could be
regarded as heuristic generation methods by the definition.

7.1.2

Interpretation and implications of findings

First, all the findings verified the feasibility, the effectiveness and the efficiency of the SPOT heuristic generation. In fact, the verification did not only
come from positive supporting results, such as the significantly improved
ruin-and-recreate heuristics in the FSP domain, but also from the neutral and
even negative results in terms of solution quality or applicability, such as
the insignificant improvements in the TSP domain and the flawed transformation of FSPs. All the positive, neutral and negative cases supported the

164

CHAPTER 7. DISCUSSION AND CONCLUSIONS

values and the prediction of the indicator(s) of the SPOT development.
The two applications of the SPOT methodology in the two NP-Complete
problem domains showed that the presented algorithm was capable for
cross-domain applications. Particularly, every problem domain should be
transformed to meet the U/EA standard for a possible effective adaptation
of the SPOT hyper-heuristic. The capability of the SPOT would, in fact,
depend more on the capability of the developer in regard to transformation
than the specific given problem domain or the particular set of given problem
instances. It could probably also give a possible solution to algorithms, which
are encountering NFL theories. The proposed algorithm aimed at being able
to handle cross-domain problems in certain conditions, such as the U/EA
standard. However, the U/EA standard is closely related to perturbative
meta-heuristics, as mentioned in Section 3.3.1. The SPOT, therefore, could
not become a universal solver.
Computer power has been increasing continuously. However, the complexity, such as incomputability and NP-hardness, limited practical polynomial time programs in problem solving. It is very common to see an
up-to-date effective meta-heuristic or exact algorithm spending a long time,
up to weeks and months, on a large-scale combinatorial optimization problem. Therefore, it is not easy to find optima to discover instance-specific
information. The capability of learning instance-specific information from
suboptima of subproblems was is most interesting, and is theoretically enabled by the U/EA standard and a high-level of compatibility. The two
regulations made the learning and generation feasible and on-the-fly.
In a “compatible” domain, perturbative heuristics such as local search

7.1. DISCUSSION ON THE SPOT METHODOLOGY AND FINDINGS 165
should be relatively more effective than constructive heuristics. Hence the
two applications also mainly focused on perturbative heuristics. New LLHs
were generated by modifying existing LLHs. Therefore, it seems the SPOT
applications in thesis should be called heuristic “modification” instead of
generation. Nevertheless, new constructive heuristics can also be generated.
For example, the generating permutations by “direct” weighting in Section
6.3.2 was actually a constructive solution — though it was easily dominated.
It would also be interesting to investigate whether the instance-specific
information can be applied in exact algorithms. Xue, Chan, et al. (2011) gave
a preliminary trial on this topic.
The SPOT heuristic generation presented in this thesis was based on
stochastic supervised machine learning. However, it would probably be
possible to extract instance-specific information through non-stochastic machine learning methods, such as the backbone (W. Zhang & Looks, 2005)
and tour-merging (Cook & Seymour, 2003) in the TSP domain. It is because
the essence of the learning phase is to distinguish “promising” assignments
from a group of competitors. For the applications without sampling for
subproblems, the intersection or the union of assignments from different
suboptima, such as the backbone and tour-merging, could provide strong
information about suboptima assignments, too. Therefore, it would be interesting to investigate a non-stochastic learning version which continues to use
the “sample-learn-generate” framework of the SPOT algorithm on-the-fly.

166

CHAPTER 7. DISCUSSION AND CONCLUSIONS

7.1.3

Interpretation in the context of the literature

Generally, the findings in this thesis were consistent with previous research described in the literature. For example, the application in the TSP
domain re-implemented the sampling and heuristic modifications of Xue,
Chan, et al. (2011) and partially replicated the decision attributes. The development and results found some trends that were the same as those described
in Xue, Chan, et al. (2011). One main difference was that the average coverage
was noticeably improved in Xue, Chan, et al. (2011) when it was slightly and
insignificantly improved, as described in Chapter 5. Explanations are given
in Section 5.5. Another example was the redefinition of the three notions of
hyper-heuristics. Although the three definitions in this thesis changed all the
notions by increasing their extensions, their intensions remained. As a result, most of the clearly categorized hyper-heuristics, such as those in Burke,
Hyde, Kendall, Ochoa, Özcan, and Woodward (2010), could be classified to
the same category according to the new definitions.
The sampling for subproblems is the first main step on the SPOT hyperheuristic, if the transformations are considered as preparations. In fact,
the sampling in this thesis was more similar to the sampling in stochastic
learning, particularly the Naı̈ve Bayes, and some meta-heuristics, especially
the EDAs. There were normal distributions and the kernel method for a
Naı̈ve Bayes classifier to use (John & Langley, 1995). There were standard
probability-based (Mühlenbein & Paaß, 1996; Mühlenbein, Bendisch, & Voigt,
1996) and factorized distribution-based (Mühlenbein & Mahnig, 1999) estimations in EDAs, too. In this research, both random selection and some

7.1. DISCUSSION ON THE SPOT METHODOLOGY AND FINDINGS 167
non-random selections were also designed and tested. In some domains,
such as the TSP domain, it was better to use subproblems generated by
different methods of sampling.
The final results of the experiments showed that the SPOT methodology
was very competitive against the winners of CHeSC 2011 on the HyFlex
framework. However, it should be noted that the opponents’ results were
tested on a 32-bit Java, and the results in this thesis were tested on a 64-bit
Java. Although there was a performance factor measured by an official program to covert the local CPU time to a Pentium IV 3.0 GHz, the conversion
seemed not so precise when a lot of floating computations were involved.
One proof was that the PHunter got a slight but steady performance improvement on every single test instance in the TSP domain. Similar observation can
also be obtained in the FSP domain, see Table C.8 in Appendix C. However,
due to the demanded memory in machine learning, a 64-bit Java had to be
chosen for conducting all tests.

7.1.4

Limitations

The formal definitions contributed a few formulations of some important
concepts in hyper-heuristics. However, the definitions were, in fact, no more
than a few manipulations, which tied the distinguishability of the hyperheuristic notions to well-known concepts. In addition, the distinguishability
of heuristic generation, for example, was only applicable when there was an
infinitely large problem. Therefore, the given definitions are preliminary and
would benefit from modifications and refinements in future.
In this thesis, the methodology of examining the performance of a new

168

CHAPTER 7. DISCUSSION AND CONCLUSIONS

algorithm was based on experiments on benchmark instances and in comparison with well-known effective algorithms in an (hopefully) equivalent
environment. There were a number of factors that might cause some errors in
the final results of the two domains of applications, as discusses as follows.
Number of test instances: Five test instances, as many as tests in CHeSC
2011, were used in each domain. If there were more instances, the
results should become theoretically slightly less biased, if any.
Source of test instances: Although the test instances were well-known benchmark instances, more application domains would avoid biased results,
if any. For example, the instance selection in the TSP domain seemed
better than that in the FSP domain.
Unstable solutions: The stop criterion was a time limit, so the operations or
flops done by a CPU were not stable. A solution was changing the stop
criterion to iterations or operations.
Instrument error in measuring time in Java: In the implementations in this
study, the time cost was measured by a class ThreadMXBean provided
by Java. However, the precision was not very satisfactory. In certain
cases, the resolution on time provided by the class was about 160
milliseconds. That was too much in comparison with fast heuristics of
one millisecond.
32/64-bit environment of Java: It should be appreciated that there was an
official program in CHeSC 2011 to measure a CPU factor for the participant’s machine. However, the measurement seemed imbalanced
according to the test results in Chapter 5 and Chapter 6. A probable

7.1. DISCUSSION ON THE SPOT METHODOLOGY AND FINDINGS 169
reason was the float/integer ratio considered in the program might
vary in different applications of domains.
CPU and memory performance adjustment: A state-of-the-art CPU and memory are often capable of adjusting their voltages, clocks and performances to save power. Some models of CPUs also mapped one hard
core into two virtual cores. This would be an issue in measuring the
performance of an algorithm.
LLH implementation: The quality of LLHs also affected the results. If there
was some unnoticed error in the LLHs in HyFlex, there would be
inherent errors.
One answer to resolve the random errors, such as the unstable solutions
and instrument error, was to use the statistical techniques. For example, the
one-way ANOVA tests were generally employed in this thesis to compare
the differences in average values between two sets of values. Besides the
possible sources of errors listed above, the limited sources and scale of benchmark instances in HyFlex and the Formula 1-like scoring system in CHeSC
2011 might also lower the precision and the confidence of the experimental
results. Nevertheless, Xue, Chan, et al. (2011)’s results of a similar heuristic
generation method showed a good scalability for Euclidean TSP instances,
which covered three different sources and from 3,000 to 1,000,000 cities.
There were also some difficulties in the development phase of the SPOT
heuristic generation, as follows.
Transformation for the U/EA standard: Although it is not difficult to find
a U/EA transformation, transformations cannot be easily borrowed

170

CHAPTER 7. DISCUSSION AND CONCLUSIONS
from one domain to another. For example, in the personnel scheduling
domain in HyFlex, some perturbative techniques worked very well.
But both transformations in the two applications failed in practice.
Constructing a new high-level of compatible transformation may spend
a lot of time.

Memory usage of stochastic learning: The demand for memory could be
billions of bytes, when there are millions of rows of data records. This
problem is particularly serious when there is no attribute selection.
Interpretation of instance-specific information into heuristics: The instancespecific information obtained in this research was in the form of stochastic classifiers. The method of conversion of the classifiers would be
important for the final generated heuristics. It is a relative difficult task
to find a highly satisfactory conversion.
Although the main framework of the SPOT methodology is presented in
thesis, there are a few issues not mentioned:
Parameter analysis of the size of initial raw attributes: In thesis, the size,
as a parameter of attribute selection method, was referred to some
default value. Further works can be done in examining the size of raw
attributes in some given domains.
Parameter analysis of the size of sampled subproblems: In the two applications, the sizes were set to two, which was the minimal number
fulfilling the requirements of Definition 12. Further analysis can show
the difference of the measurements of compatibility by different sizes.

7.2. CONCLUSIONS

171

Descriptive compatibility: One issue is to determine a descriptive scale
to interpret the two numbers “generality” and “approximability” in
compatibility, e.g., “good” for 0.8 to 1.0 generality and “incompatible”
for 0 to 0.05 generality.
Transformation between domains: In the literature, Ruiz-Vanoye et al. (2011)
summarized a transformation map between hundreds of NP-Complete
problems. Maybe in the future an approximability tree or table can
be organized, so that a given domain A can take advantage of a welldefined transformation in another known domain B by mapping itself
to B first. A given practical domain A can also take advantage of
another well-tackled domain B by mapping, too.

7.2

Conclusions

This thesis presents the SPOT heuristic generation approach for solving
combinatorial optimization problems. With the proposed “sample-learngenerate” framework, SPOT samples small-scale subproblems, obtains instancespecific information from the suboptima of the subproblems by machine learning, then generates new heuristics by modifying existing heuristics and data
structures accordingly. All the operations are conducted on-the-fly, and it is
capable of generating non-dominated new heuristics for large-scale instances
very efficiently, such as in the FSP domain. In the development of SPOT
algorithm, two standards were incorporated to regulate the problem input
and the machine learning data. Besides, an automated attribute generation
procedure was introduced for populating more “normalized attributes” for

172

CHAPTER 7. DISCUSSION AND CONCLUSIONS

learning. An indicator of “resemblance” of suboptima is also introduced
to measure the compatibility of the input problem. Generally, if a problem
domain meets the standard of input and its compatibility is high, the SPOT
algorithm is expected to be a competitive method in the domain.
Additionally, formal definitions of the hyper-heuristic and its subclasses,
heuristic selection and heuristic generation, were proposed in this study.
Based on the new definitions, SPOT is a heuristic generation algorithm.
These definitions are mostly consistent with existing taxonomies on hyperheuristics, such as in Burke, Hyde, Kendall, Ochoa, Özcan, and Woodward
(2010).
SPOT was implemented in two well-known NP-Complete problem domains, the TSP and the FSP. Experiments were carried out in the two domains
for benchmarking purpose. The generated heuristics were non-dominated
and competitive in the FSP domain in HyFlex, while the improvements in the
TSP domain were limited. The results in both domains successfully validated
the compatibility predicted. To verify the long-time result of using SPOT,
one of the winners of the international hyper-heuristic competition CHeSC
2011, named PHunter, was tested with the generated heuristics. In TSP, the
heuristics generated by the SPOT method gave unstable improvements in
terms of solution quality. In FSP, the improvements were significant, and
the score of PHunter was approximately doubled (from 20.5 to 43 with 50
maximum) and became the highest among all the leading hyper-heuristics.
This is evidence on the success of SPOT.
There are two main contributions from this research. The first is the
SPOT heuristic generation methodology, and the other is the enhancement

7.2. CONCLUSIONS

173

of definitions of the hyper-heuristics and its two subclasses. In addition, the
capability of discovering instance-specific information from suboptima of
subproblems is an interesting property. It was also the key to the success
of the SPOT heuristic generation. SPOT can be applied in any combinatorial optimization domain if the problem instances in the domain meet the
standard requirements and the compatibility measures. Therefore, SPOT is a
conditionally cross-domain algorithm and its new property might probably
be an answer to the issues raised by the NFL theorems.
Future works on SPOT and related concepts include continued research
on the non-stochastic machine learning techniques and the refinements of
the definitions. It would also be meaningful to modify the-state-of-the-art
LLHs by employing SPOT to push forward the Pareto frontier for general
hyper-heuristics in terms of performance. Another interesting research could
be making a transformation map. On the map, the distances and paths of
transformations from well-known combinatorial optimization problems to
the required models of SPOT can be drawn, so that the adaption of SPOT
will be convenient in practice.

174

CHAPTER 7. DISCUSSION AND CONCLUSIONS

Appendix A

Suboptima of Training Instances

(i) Inst 4, Cost = 8873.002687969361

(ii) Inst 4, Cost = 8877.2061295038

Figure A.1: Suboptima of the training instance rat783 in the TSP

175

176

SUBOPTIMA OF TRAINING INSTANCES

(i) Inst 5, Cost = 57210.04615291555

(ii) Inst 5, Cost = 57142.007438059256

Figure A.2: Suboptima of the training instance pcb1173 in the TSP

SUBOPTIMA OF TRAINING INSTANCES

(i) Inst 6, Cost = 52288.38373274874

(ii) Inst 6, Cost = 52446.80109412457

Figure A.3: Suboptima of the training instance d1291 in the TSP

177

178

SUBOPTIMA OF TRAINING INSTANCES

Table A.1: Suboptima of the training instances in the FSP
Instance
(Cmax )

Solution

Inst 0 53, 77, 58, 21, 81, 0, 2, 39, 78, 75, 50, 34, 73, 3, 82, 32, 80, 61, 4, 46, 20, 5, 27,
(6265) 62, 88, 55, 30, 98, 59, 24, 85, 45, 15, 69, 11, 84, 57, 49, 1, 43, 65, 60, 52, 79, 10,
99, 8, 54, 90, 9, 93, 29, 38, 19, 71, 42, 92, 31, 66, 47, 41, 76, 36, 95, 14, 68, 33,
16, 44, 64, 94, 28, 25, 70, 63, 86, 56, 17, 23, 67, 89, 87, 51, 22, 37, 13, 72, 97, 18,
74, 7, 96, 91, 35, 83, 26, 48, 12, 6, 40
Inst 0 53, 21, 46, 0, 34, 82, 54, 4, 79, 81, 73, 58, 50, 78, 30, 39, 55, 80, 20, 38, 88, 11,
(6269) 99, 2, 49, 64, 75, 32, 52, 47, 60, 15, 69, 70, 8, 9, 10, 42, 43, 3, 59, 65, 44, 5, 84,
94, 92, 31, 27, 61, 29, 90, 14, 36, 17, 85, 77, 33, 1, 93, 35, 72, 98, 24, 41, 19, 86,
22, 16, 95, 28, 23, 37, 66, 76, 62, 45, 97, 71, 67, 87, 63, 74, 91, 7, 96, 89, 51, 18,
56, 13, 25, 57, 83, 26, 68, 48, 12, 6, 40
Inst 1 49, 48, 4, 94, 67, 30, 75, 51, 0, 65, 96, 83, 9, 89, 34, 68, 64, 35, 63, 99, 74, 61, 81,
(6238) 53, 26, 22, 2, 13, 12, 95, 18, 77, 98, 84, 8, 39, 59, 10, 71, 32, 19, 7, 46, 88, 15, 50,
21, 58, 41, 6, 82, 42, 47, 76, 3, 80, 78, 87, 31, 43, 1, 52, 40, 90, 62, 86, 24, 70, 25,
92, 29, 27, 36, 33, 85, 97, 69, 60, 72, 44, 57, 16, 5, 55, 73, 79, 37, 45, 14, 93, 11,
20, 66, 28, 91, 23, 17, 38, 56, 54
Inst 1 48, 94, 64, 49, 65, 68, 18, 61, 30, 96, 77, 13, 32, 2, 99, 34, 95, 88, 67, 51, 53, 89,
(6246) 42, 8, 81, 3, 1, 78, 9, 59, 52, 39, 75, 82, 19, 46, 90, 4, 6, 23, 15, 21, 27, 37, 31,
74, 76, 58, 83, 29, 73, 10, 22, 33, 97, 47, 92, 50, 45, 70, 0, 7, 35, 84, 14, 12, 79,
60, 25, 63, 26, 57, 69, 72, 44, 87, 43, 91, 85, 24, 55, 17, 40, 62, 86, 16, 41, 71, 36,
20, 66, 28, 5, 11, 54, 98, 93, 80, 38, 56
Inst 2 86, 9, 91, 60, 65, 10, 63, 66, 75, 28, 50, 23, 41, 46, 85, 22, 47, 57, 53, 4, 55, 97,
(6316) 11, 33, 21, 56, 19, 80, 68, 45, 32, 79, 69, 96, 12, 59, 74, 3, 52, 39, 84, 87, 40, 98,
77, 36, 24, 5, 93, 67, 14, 31, 15, 29, 76, 2, 27, 18, 48, 34, 95, 58, 16, 71, 20, 1,
26, 43, 30, 8, 42, 94, 89, 90, 7, 78, 62, 54, 88, 38, 92, 72, 83, 35, 25, 6, 70, 49, 0,
73, 51, 37, 61, 64, 44, 81, 17, 99, 13, 82
Inst 2 86, 9, 40, 60, 10, 96, 63, 91, 66, 41, 57, 75, 85, 50, 28, 23, 46, 47, 62, 22, 53, 59,
(6321) 55, 80, 45, 32, 97, 11, 21, 33, 79, 69, 4, 56, 38, 39, 92, 24, 3, 19, 68, 52, 65, 77,
84, 87, 5, 93, 29, 76, 14, 31, 67, 2, 27, 98, 18, 16, 48, 34, 95, 58, 71, 1, 20, 54, 26,
43, 30, 89, 42, 7, 49, 74, 8, 37, 78, 44, 36, 88, 72, 0, 83, 35, 73, 90, 25, 6, 70, 12,
64, 51, 61, 15, 81, 13, 82, 17, 94, 99
Inst 3 0, 67, 44, 4, 43, 59, 78, 88, 73, 68, 21, 47, 45, 34, 87, 28, 37, 64, 2, 38, 85, 32,
(6323) 75, 53, 95, 66, 54, 60, 15, 55, 70, 50, 84, 83, 25, 27, 91, 6, 90, 94, 76, 1, 72, 24,
8, 71, 57, 30, 18, 97, 14, 92, 62, 16, 58, 48, 99, 3, 10, 98, 7, 42, 89, 22, 69, 31, 23,
77, 17, 93, 49, 12, 86, 63, 11, 41, 82, 5, 80, 36, 56, 65, 61, 46, 51, 20, 33, 74, 40,
9, 81, 96, 29, 19, 13, 26, 52, 79, 35, 39
Inst 3 0, 59, 28, 44, 4, 50, 68, 45, 43, 73, 15, 54, 87, 49, 56, 2, 67, 64, 10, 53, 95, 94, 32,
(6331) 11, 88, 55, 71, 83, 1, 92, 78, 20, 24, 7, 37, 85, 57, 27, 72, 76, 98, 66, 91, 82, 69,
70, 77, 99, 97, 9, 62, 23, 38, 40, 34, 60, 12, 30, 86, 36, 17, 14, 31, 21, 22, 42, 18,
75, 25, 93, 89, 46, 84, 51, 3, 8, 90, 61, 29, 6, 16, 33, 5, 74, 19, 80, 65, 41, 52, 13,
58, 63, 96, 47, 48, 81, 26, 79, 35, 39

Appendix B
of Learning

Examples of Data and Results

179

180

EXAMPLES OF DATA AND RESULTS OF LEARNING

Table B.1: The group and raw attributes of assignments in the 20-nearestneighbor-first candidate set of two subproblems of the TSP instance rat783
(15840 rows × (1 column of group + 2 columns of raw attributes), n0 = 400,
sampled by a random selection and a rectangular selection, respectively)
Group
Id
316

..
.
167

A1
(distance)

A2
(angle)

13.04
17.20
23.09
23.77
27.02
27.80
29.00
30.41
34.13
37.16
38.12
40.20
40.79
42.20
43.46
45.18
45.28
45.49
46.53
48.51

0.08
0.62
1.26
-1.32
-0.89
-0.66
0.81
0.48
-1.02
-0.42
1.49
-1.47
0.20
0.63
0.40
-0.09
1.04
0.99
1.08
-1.32

..
.

..
.

16.40
17.20
26.08
26.93
28.32
30.23
30.87
36.24
36.40
36.67
37.36
40.45
42.15
43.05
43.86
45.61
49.19
49.98
50.64
51.09

-0.66
-2.19
-2.14
0.38
2.41
-0.60
1.14
2.06
-0.37
-2.95
0.27
2.99
-2.75
0.05
1.15
-2.48
-0.46
1.92
-1.41
2.27

EXAMPLES OF DATA AND RESULTS OF LEARNING

181

Table B.2: The full training table consisting of normalized attributes of assignments and the label of suboptima of two subproblems of the TSP instance
rat783 (15840 rows × (86 columns of attributes + 1 column of labels), attributes normalized from Table B.1)
A1 A2 A3 A4

A5

A6

A7

A8

A9 A10

···

A85 A86

Label

13.04
17.20
23.09
23.77
27.02
27.80
29.00
30.41
34.13
37.16
38.12
40.20
40.79
42.20
43.46
45.18
45.28
45.49
46.53
48.51
..
.

20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
..
.

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

0
1 0.76 0.56 0.55 0.48 0.47
0 1.32
1 0.75 0.72 0.64 0.62
0 1.77 1.34
1 0.97 0.85 0.83
0 1.82 1.38 1.03
1 0.88 0.85
0 2.07 1.57 1.17 1.14
1 0.97
0 2.13 1.62 1.20 1.17 1.03
1
0 2.22 1.69 1.26 1.22 1.07 1.04
0 2.33 1.77 1.32 1.28 1.13 1.09
0 2.62 1.98 1.48 1.44 1.26 1.23
0 2.85 2.16 1.61 1.56 1.38 1.34
0 2.92 2.22 1.65 1.60 1.41 1.37
0 3.08 2.34 1.74 1.69 1.49 1.45
0 3.13 2.37 1.77 1.72 1.51 1.47
0 3.24 2.45 1.83 1.78 1.56 1.52
0 3.33 2.53 1.88 1.83 1.61 1.56
0 3.46 2.63 1.96 1.90 1.67 1.62
0 3.47 2.63 1.96 1.90 1.68 1.63
0 3.49 2.64 1.97 1.91 1.68 1.64
0 3.57 2.70 2.02 1.96 1.72 1.67
0 3.72 2.82 2.10 2.04 1.80 1.74
..
..
..
..
..
..
..
.
.
.
.
.
.
.

···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
..
.

0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
0.85 0.85
..
..
.
.

1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

16.40
17.20
26.08
26.93
28.32
30.23
30.87
36.24
36.40
36.67
37.36
40.45
42.15
43.05
43.86
45.61
49.19
49.98
50.64
51.09

20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0
1 0.95 0.63 0.61 0.58 0.54
0 1.05
1 0.66 0.64 0.61 0.57
0 1.59 1.52
1 0.97 0.92 0.86
0 1.64 1.57 1.03
1 0.95 0.89
0 1.73 1.65 1.09 1.05
1 0.94
0 1.84 1.76 1.16 1.12 1.07
1
0 1.88 1.79 1.18 1.15 1.09 1.02
0 2.21 2.11 1.39 1.35 1.28 1.20
0 2.22 2.12 1.40 1.35 1.29 1.20
0 2.24 2.13 1.41 1.36 1.30 1.21
0 2.28 2.17 1.43 1.39 1.32 1.24
0 2.47 2.35 1.55 1.50 1.43 1.34
0 2.57 2.45 1.62 1.57 1.49 1.39
0 2.62 2.50 1.65 1.60 1.52 1.42
0 2.67 2.55 1.68 1.63 1.55 1.45
0 2.78 2.65 1.75 1.69 1.61 1.51
0 3.00 2.86 1.89 1.83 1.74 1.63
0 3.05 2.91 1.92 1.86 1.76 1.65
0 3.09 2.94 1.94 1.88 1.79 1.67
0 3.11 2.97 1.96 1.90 1.80 1.69

···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···
···

0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80
0.94 0.80

1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

182

EXAMPLES OF DATA AND RESULTS OF LEARNING

Table B.3: Results of the attribute selection for the full training table of Table
B.2 (15840 rows × (8 columns of attributes + 1 column of labels))
A1

A3

A5

A6

A7

A8

A14

A20

Label

13.04
17.20
23.09
23.77
27.02
27.80
29.00
30.41
34.13
37.16
38.12
40.20
40.79
42.20
43.46
45.18
45.28
45.49
46.53
48.51
..
.

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

1
1.32
1.77
1.82
2.07
2.13
2.22
2.33
2.62
2.85
2.92
3.08
3.13
3.24
3.33
3.46
3.47
3.49
3.57
3.72
..
.

0.76
1
1.34
1.38
1.57
1.62
1.69
1.77
1.98
2.16
2.22
2.34
2.37
2.45
2.53
2.63
2.63
2.64
2.70
2.82
..
.

0.56
0.75
1
1.03
1.17
1.20
1.26
1.32
1.48
1.61
1.65
1.74
1.77
1.83
1.88
1.96
1.96
1.97
2.02
2.10
..
.

0.55
0.72
0.97
1
1.14
1.17
1.22
1.28
1.44
1.56
1.60
1.69
1.72
1.78
1.83
1.90
1.90
1.91
1.96
2.04
..
.

0.35
0.46
0.62
0.64
0.73
0.75
0.78
0.82
0.92
1
1.03
1.08
1.10
1.14
1.17
1.22
1.22
1.22
1.25
1.31
..
.

0.29
0.38
0.51
0.53
0.60
0.62
0.64
0.67
0.76
0.82
0.84
0.89
0.90
0.93
0.96
1
1.00
1.01
1.03
1.07
..
.

1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

16.40
17.20
26.08
26.93
28.32
30.23
30.87
36.24
36.40
36.67
37.36
40.45
42.15
43.05
43.86
45.61
49.19
49.98
50.64
51.09

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

1
1.05
1.59
1.64
1.73
1.84
1.88
2.21
2.22
2.24
2.28
2.47
2.57
2.62
2.67
2.78
3.00
3.05
3.09
3.11

0.95
1
1.52
1.57
1.65
1.76
1.79
2.11
2.12
2.13
2.17
2.35
2.45
2.50
2.55
2.65
2.86
2.91
2.94
2.97

0.63
0.66
1
1.03
1.09
1.16
1.18
1.39
1.40
1.41
1.43
1.55
1.62
1.65
1.68
1.75
1.89
1.92
1.94
1.96

0.61
0.64
0.97
1
1.05
1.12
1.15
1.35
1.35
1.36
1.39
1.50
1.57
1.60
1.63
1.69
1.83
1.86
1.88
1.90

0.45
0.47
0.71
0.73
0.77
0.82
0.84
0.99
0.99
1
1.02
1.10
1.15
1.17
1.20
1.24
1.34
1.36
1.38
1.39

0.36
0.38
0.57
0.59
0.62
0.66
0.68
0.79
0.80
0.80
0.82
0.89
0.92
0.94
0.96
1
1.08
1.10
1.11
1.12

1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

EXAMPLES OF DATA AND RESULTS OF LEARNING

183

J48 pruned tree
-----------------A8 <= 0.986899
|
A6 <= 0.936329: 1 (711.0/69.0)
|
A6 > 0.936329
|
|
A7 <= 0.907959
|
|
|
A5 <= 1.425106
|
|
|
|
A8 <= 0.795966: 1 (308.0/86.0)
|
|
|
|
A8 > 0.795966: 0 (60.0/27.0)
|
|
|
A5 > 1.425106: 0 (229.0/106.0)
|
|
A7 > 0.907959: 0 (1007.0/299.0)
A8 > 0.986899: 0 (13525.0/288.0)
Number of Leaves :
Size of the tree :

6
11

Figure B.1: Instance-specific result of the J48 classifier for the TSP instance
rat783, training data shown in Table B.3

JRIP rules:
===========
1. (A7 <= 1.01303) and (A7 <= 0.823592) and (A6 <= 0.800735)
=> Label=1 (510.0/25.0)
2. (A7 <= 1.061977) and (A7 <= 0.906932) and (A5 <= 1.433721)
and (A8 <= 0.720805) => Label=1 (389.0/90.0)
3. (A7 <= 1.126515) and (A8 <= 0.794645) and (A5 <= 1.355669)
and (A6 <= 0.917267) => Label=1 (33.0/10.0)
4. (A8 <= 1.049303) and (A6 <= 1.059753) and (A8 <= 0.866758)
and (A5 <= 1.566151) => Label=1 (319.0/154.0)
5. (A14 <= 0.692356) and (A6 <= 1.194349) and (A14 <= 0.504353)
and (A7 <= 0.709541) => Label=1 (51.0/22.0)
6. (A8 <= 1.051403) and (A6 <= 1.155428) and (A14 <= 0.493171)
and (A14 >= 0.479119) => Label=1 (38.0/16.0)
7. (A8 <= 1.051403) and (A7 <= 1.01716) and (A20 <= 0.385248)
and (A5 <= 3.508232) and (A8 >= 0.938898) => Label=1
(37.0/15.0)
8. => Label=0 (14463.0/539.0)
Number of Rules : 8

Figure B.2: Instance-specific result of the JRip classifier for the TSP instance
rat783, training data shown in Table B.3

184

EXAMPLES OF DATA AND RESULTS OF LEARNING

Naive Bayes Classifier
Class 0: Prior probability = 0.9
A1: Normal Distribution. Mean = 34.1648 StandardDev = 12.2188
WeightSum = 14256 Precision = 0.07893093700787401
A3: Normal Distribution. Mean = 0 StandardDev = 0.1667
WeightSum = 14256 Precision = 1.0
A5: Normal Distribution. Mean = 4.6907 StandardDev = 5.3678
WeightSum = 14256 Precision = 0.005596767589330649
A6: Normal Distribution. Mean = 2.4956 StandardDev = 1.2026
WeightSum = 14256 Precision = 0.0020412012114637015
A7: Normal Distribution. Mean = 1.9844 StandardDev = 0.7277
WeightSum = 14256 Precision = 7.003549220879403E-4
A8: Normal Distribution. Mean = 1.6846 StandardDev = 0.5389
WeightSum = 14256 Precision = 3.5252928054789944E-4
A14: Normal Distribution. Mean = 1.0454 StandardDev = 0.2923
WeightSum = 14256 Precision = 1.501225443814196E-4
A20: Normal Distribution. Mean = 0.8162 StandardDev = 0.2232
WeightSum = 14256 Precision = 1.0161365097366262E-4
Class 1: Prior probability = 0.1
A1: Normal Distribution. Mean = 13.9825 StandardDev = 6.5628
WeightSum = 1584 Precision = 0.07893093700787401
A3: Normal Distribution. Mean = 0.0076 StandardDev = 0.1667
WeightSum = 1584 Precision = 1.0
A5: Normal Distribution. Mean = 1.675 StandardDev = 1.9162
WeightSum = 1584 Precision = 0.005596767589330649
A6: Normal Distribution. Mean = 0.9672 StandardDev = 0.4053
WeightSum = 1584 Precision = 0.0020412012114637015
A7: Normal Distribution. Mean = 0.7923 StandardDev = 0.3266
WeightSum = 1584 Precision = 7.003549220879403E-4
A8: Normal Distribution. Mean = 0.6861 StandardDev = 0.2872
WeightSum = 1584 Precision = 3.5252928054789944E-4
A14: Normal Distribution. Mean = 0.4315 StandardDev = 0.1895
WeightSum = 1584 Precision = 1.501225443814196E-4
A20: Normal Distribution. Mean = 0.3378 StandardDev = 0.149
WeightSum = 1584 Precision = 1.0161365097366262E-4

Figure B.3: Instance-specific result of the NaiveBayes classifier for the TSP
instance rat783, training data shown in Table B.3

EXAMPLES OF DATA AND RESULTS OF LEARNING

185

Table B.4: Eight selected columns of test data and the summation of predicted
labels of the assignments in the 20-nearest-neighbor-first candidate set of the
TSP instance rat783 (15660 rows × (8 columns of selected attributes + 1
column of summations of labels))
A1

A3

A5

A6

A7

A8

A14

A20

Σ label

7.62
13.04
17.89
25.55
28.64
33.54
34.00
36.00
36.24
36.69
37.00
40.05
42.45
49.65
52.01
55.01
55.80
57.22
58.18
..
.

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

1
1.71
2.35
3.36
3.76
4.40
4.46
4.73
4.76
4.82
4.86
5.26
5.57
6.52
6.83
7.22
7.33
7.51
7.64
..
.

0.58
1
1.37
1.96
2.20
2.57
2.61
2.76
2.78
2.81
2.84
3.07
3.26
3.81
3.99
4.22
4.28
4.39
4.46
..
.

0.43
0.73
1
1.43
1.60
1.88
1.90
2.01
2.03
2.05
2.07
2.24
2.37
2.78
2.91
3.08
3.12
3.20
3.25
..
.

0.30
0.51
0.70
1
1.12
1.31
1.33
1.41
1.42
1.44
1.45
1.57
1.66
1.94
2.04
2.15
2.18
2.24
2.28
..
.

0.21
0.36
0.49
0.71
0.79
0.93
0.94
0.99
1
1.01
1.02
1.11
1.17
1.37
1.44
1.52
1.54
1.58
1.61
..
.

0.14
0.24
0.33
0.46
0.52
0.61
0.62
0.65
0.66
0.67
0.67
0.73
0.77
0.90
0.95
1
1.01
1.04
1.06
..
.

2
2
2
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
..
.

2.00
9.22
20.88
21.93
27.17
30.81
31.02
31.14
31.89
32.00
34.89
40.22
41.05
46.49
47.00
50.99
52.01
53.01
53.37
55.57

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

1
4.61
10.44
10.97
13.58
15.40
15.51
15.57
15.95
16.00
17.44
20.11
20.52
23.24
23.50
25.50
26.00
26.50
26.68
27.78

0.22
1
2.26
2.38
2.95
3.34
3.36
3.38
3.46
3.47
3.78
4.36
4.45
5.04
5.10
5.53
5.64
5.75
5.79
6.03

0.10
0.44
1
1.05
1.30
1.48
1.49
1.49
1.53
1.53
1.67
1.93
1.97
2.23
2.25
2.44
2.49
2.54
2.56
2.66

0.09
0.42
0.95
1
1.24
1.40
1.41
1.42
1.45
1.46
1.59
1.83
1.87
2.12
2.14
2.32
2.37
2.42
2.43
2.53

0.06
0.29
0.65
0.69
0.85
0.97
0.97
0.98
1
1.00
1.09
1.26
1.29
1.46
1.47
1.60
1.63
1.66
1.67
1.74

0.04
0.18
0.41
0.43
0.53
0.60
0.61
0.61
0.63
0.63
0.68
0.79
0.81
0.91
0.92
1
1.02
1.04
1.05
1.09

2
2
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

186

EXAMPLES OF DATA AND RESULTS OF LEARNING

(i) The 20-nearest-neighbor-first
candidate set

(ii) Promising assignments classified by
the SPOT heuristic generation

Figure B.4: A set of promising edges selected from the 20-nearest-neighborfirst candidate set of the training instance rat783 in the TSP, according to
the Σ label column in Table B.4 (where black edges in the second figure were
the assignments with Σ label ≥ 2 and the gray edges were those with Σ label
= 1)

EXAMPLES OF DATA AND RESULTS OF LEARNING

187

Table B.5: The group and raw attributes of assignments of the “following”
relationship in two subproblems of the FSP instance TA082 (3480 rows × (1
column of group + 103 columns of raw attributes), n0 = 30, sampled by two
random selections, respectively)
A1

541

56.2 515.64 4555 -633 -0.02 96 41 64 82
47.2 759.82 4414 145 0.04 35 81 10 54

59
38

30
51

64
26

45
83

44 · · ·
87 · · ·

96
89

567

56.2 515.64 5403 -633 -0.02 96 41 64 82
49.6 991.10 5502
49 -0.09 72 99 62 10

59
1

30
64

64
16

45
2

44 · · ·
73 · · ·

96
99

562

56.2 515.64 5211 -633 -0.02 96 41 64 82
43.5 637.00 5415 637 0.06 40 40 36 22

59
57

30
87

64
8

45
23

44 · · ·
5 ···

96
87

565

56.2 515.64 4075 -633 -0.02 96 41 64 82
44.5 602.47 4156 -1173 -0.08 25 56 49 75

59
59

30
21

64
76

45
36

44 · · ·
53 · · ·

96
83

545

56.2 515.64 5635 -633 -0.02 96 41 64 82
42.2 694.03 5830 297 -0.03 31 19 50 3

59
75

30
31

64
40

45
8

44 · · ·
77 · · ·

96
96

560

56.2 515.64 4808 -633 -0.02 96 41 64 82
50.8 754.91 5450
93 0.02 54 58 34 53

59
38

30
18

64
78

45
28

44 · · ·
71 · · ·

96
95

566

56.2 515.64 5361 -633 -0.02 96 41 64 82
62.1 664.89 4824 342 0.02 59 29 32 87

59
67

30
93

64
21

45
85

44 · · ·
84 · · ·

96
93

553

56.2 515.64 6629 -633 -0.02 96 41 64 82
51.8 933.36 6948 1401 -0.17 78 28 23 8

59
4

30
2

64
58

45
93

44 · · ·
58 · · ·

96
97

544

56.2 515.64 5320 -633 -0.02 96 41 64 82
47.5 1043.95 6236 2466 0.09 5 6 25 62

59
7

30
80

64
57

45
25

44 · · ·
10 · · ·

96
99

557

56.2 515.64 4437 -633 -0.02 96 41 64 82
52.1 727.67 5087 191 0.02 41 46 40 38

59
8

30
38

64
77

45
97

44 · · ·
37 · · ·

96
97

551

56.2 515.64 4682 -633 -0.02 96 41 64 82
50.9 886.62 4498 -1528 -0.14 99 65 71 31

59
74

30
56

64
78

45
45

44 · · ·
74 · · ·

96
99

547

56.2 515.64 2599 -633 -0.02 96 41 64 82
54.1 654.62 4189 -1031 -0.03 39 88 73 60

59
22

30
61

64
59

45
33

44 · · ·
76 · · ·

96
95

559

56.2 515.64 4930 -633 -0.02 96 41 64 82
44.7 851.50 4997 -424 -0.02 89 71 13 97

59
14

30
35

64
40

45
26

44 · · ·
54 · · ·

96
97

555

56.2 515.64 3974 -633 -0.02 96 41 64 82
48.5 955.21 4141 -752 -0.09 41 88 82 40

59
78

30
79

64
2

45
9

44 · · ·
32 · · ·

96
97

554

56.2 515.64 5747 -633 -0.02 96 41 64 82
57.2 757.33 5611 176 -0.03 96 9 27 88

59
39

30
37

64
64

45
72

44 · · ·
91 · · ·

96
96

540

56.2 515.64 4357 -633 -0.02 96 41 64 82
44.5 579.53 4315 507 -0.02 44 25 35 73

59
33

30
37

64
25

45
38

44 · · ·
31 · · ·

96
96

..
.

..
.

..
.

..
.

..
.

.. . .
. .

..
.

45.4 888.34 4707 -369 -0.03 19 84 84 25
56.2 515.64 3098 -633 -0.02 96 41 64 82

11
59

58
30

33
64

35
45

55 · · ·
44 · · ·

97
96

..
.
331

..
.

A2 A3

..
.

..
.

A4

A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 · · · A103

Group

..
.

..
.

..
.

..
.

..
.

188

EXAMPLES OF DATA AND RESULTS OF LEARNING

Table B.6: The full training table from two subproblems of the FSP instance
TA082 (3480 rows × (206 columns of attributes + 1 column of labels), attributes normalized from Table B.5)
A1 A2

A3 A4

A5 A6

A7

A8

A9 A10 A11 A12 · · · A206

Label

56.2 1.19 515.64 0.68 4555 1.03 -633 -4.37 -0.02 -0.44
47.2 0.84 759.82 1.47 4414 0.97 145 -0.23 0.04 -2.28
56.2 1.13 515.64 0.52 5403 0.98 -633 -12.92 -0.02 0.19
49.6 0.88 991.10 1.92 5502 1.02
49 -0.08 -0.09 5.18
56.2 1.29 515.64 0.81 5211 0.96 -633 -0.99 -0.02 -0.30
43.5 0.77 637.00 1.24 5415 1.04 637 -1.01 0.06 -3.35
56.2 1.26 515.64 0.86 4075 0.98 -633 0.54 -0.02 0.23
44.5 0.79 602.47 1.17 4156 1.02 -1173 1.85 -0.08 4.38
56.2 1.33 515.64 0.74 5635 0.97 -633 -2.13 -0.02 0.68
42.2 0.75 694.03 1.35 5830 1.03 297 -0.47 -0.03 1.46
56.2 1.11 515.64 0.68 4808 0.88 -633 -6.81 -0.02 -0.98
50.8 0.90 754.91 1.46 5450 1.13
93 -0.15 0.02 -1.02
56.2 0.91 515.64 0.78 5361 1.11 -633 -1.85 -0.02 -0.72
62.1 1.10 664.89 1.29 4824 0.90 342 -0.54 0.02 -1.39
56.2 1.09 515.64 0.55 6629 0.95 -633 -0.45 -0.02 0.11
51.8 0.92 933.36 1.81 6948 1.05 1401 -2.21 -0.17 9.50
56.2 1.18 515.64 0.49 5320 0.85 -633 -0.26 -0.02 -0.19
47.5 0.85 1043.95 2.02 6236 1.17 2466 -3.90 0.09 -5.18
56.2 1.08 515.64 0.71 4437 0.87 -633 -3.31 -0.02 -0.81
52.1 0.93 727.67 1.41 5087 1.15 191 -0.30 0.02 -1.24
56.2 1.10 515.64 0.58 4682 1.04 -633 0.41 -0.02 0.12
50.9 0.91 886.62 1.72 4498 0.96 -1528 2.41 -0.14 8.14
56.2 1.04 515.64 0.79 2599 0.62 -633 0.61 -0.02 0.56
54.1 0.96 654.62 1.27 4189 1.61 -1031 1.63 -0.03 1.78
56.2 1.26 515.64 0.61 4930 0.99 -633 1.49 -0.02 0.86
44.7 0.79 851.50 1.65 4997 1.01 -424 0.67 -0.02 1.16
56.2 1.16 515.64 0.54 3974 0.96 -633 0.84 -0.02 0.19
48.5 0.86 955.21 1.85 4141 1.04 -752 1.19 -0.09 5.18
56.2 0.98 515.64 0.68 5747 1.02 -633 -3.60 -0.02 0.63
57.2 1.02 757.33 1.47 5611 0.98 176 -0.28 -0.03 1.58
56.2 1.26 515.64 0.89 4357 1.01 -633 -1.25 -0.02 0.84
44.5 0.79 579.53 1.12 4315 0.99 507 -0.80 -0.02 1.19
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
.
.

96 2.74 · · ·
35 0.36 · · ·
96 1.33 · · ·
72 0.75 · · ·
96 2.40 · · ·
40 0.42 · · ·
96 3.84 · · ·
25 0.26 · · ·
96 3.10 · · ·
31 0.32 · · ·
96 1.78 · · ·
54 0.56 · · ·
96 1.63 · · ·
59 0.61 · · ·
96 1.23 · · ·
78 0.81 · · ·
96 19.20 · · ·
5 0.05 · · ·
96 2.34 · · ·
41 0.43 · · ·
96 0.97 · · ·
99 1.03 · · ·
96 2.46 · · ·
39 0.41 · · ·
96 1.08 · · ·
89 0.93 · · ·
96 2.34 · · ·
41 0.43 · · ·
96 1.00 · · ·
96 1.00 · · ·
96 2.18 · · ·
44 0.46 · · ·
..
.. . .
.
. .

1.08
0.93
0.97
1.03
1.10
0.91
1.16
0.86
1.00
1.00
1.01
0.99
1.03
0.97
0.99
1.01
0.97
1.03
0.99
1.01
0.97
1.03
1.01
0.99
0.99
1.01
0.99
1.01
1.00
1.00
1.00
1.00
..
.

1
0
1
0
1
0
0
1
1
0
1
0
1
0
1
0
1
0
1
0
1
0
0
1
1
0
1
0
1
0
1
0
..
.

45.4 0.81 888.34 1.72 4707 1.52 -369
56.2 1.24 515.64 0.58 3098 0.66 -633

19 0.20 · · · 1.01
96 5.05 · · · 0.99

1
0

0.58 -0.03 1.58
1.72 -0.02 0.63

EXAMPLES OF DATA AND RESULTS OF LEARNING

189

J48 pruned tree
-----------------A89 <= 32
|
A90 <= 0.685864: 1 (274.0)
|
A90 > 0.685864
|
|
A92 <= 0.876712: 1 (8.0)
|
|
A92 > 0.876712: 0 (8.0)
A89 > 32
|
A159 <= 8239
|
|
A160 <= 0.975358: 0 (326.0)
|
|
A160 > 0.975358
|
|
|
A30 <= 1.054054: 1 (14.0)
|
|
|
A30 > 1.054054: 0 (8.0)
|
A159 > 8239
|
|
A90 <= 3.648649
|
|
|
A166 <= 4.821429
|
|
|
|
A14 <= 7.333334
|
|
|
|
|
A189 <= 6255
|
|
|
|
|
|
A40 <= 19.4
|
|
|
|
|
|
|
A16 <= 6.857143
|
|
|
|
|
|
|
|
A13 <= 9
|
|
|
|
|
|
|
|
|
A94 <= 1.321951
|
|
|
|
|
|
|
|
|
|
A10 <= 1.282051: 1 (78.0)
|
|
|
|
|
|
|
|
|
|
A10 > 1.282051
|
|
|
|
|
|
|
|
|
|
|
A2 <=1.148707: 0(2.0)
|
|
|
|
|
|
|
|
|
|
|
A2 > 1.148707: 1(4.0)
|
|
|
|
|
|
|
|
|
A94 > 1.321951
|
|
|
|
|
|
|
|
|
|
A2 <= 1.174407: 0 (4.0)
|
|
|
|
|
|
|
|
|
|
A2 > 1.174407: 1 (4.0)
|
|
|
|
|
|
|
|
A13 > 9
|
|
|
|
|
|
|
|
|
A13 <= 20
|
|
|
|
|
|
|
|
|
|
A160 <= 1.206111
...
|
|
A90 > 3.648649
|
|
|
A34 <= 1.36: 0 (166.0)
|
|
|
A34 > 1.36
|
|
|
|
A16 <= 18: 0 (4.0)
|
|
|
|
A16 > 18: 1 (6.0)
Number of Leaves :
Size of the tree :

157
313

Figure B.5: Instance-specific result of the J48 classifier for the FSP instance
TA082, training data shown in Table B.6

190

EXAMPLES OF DATA AND RESULTS OF LEARNING

JRIP rules:
===========
1. (A92 <= 0.882979) and (A169 >= 439) and (A21 <= 64) and
(A68 >= 0.910891) => Label=1 (404.0/0.0)
2. (A134 >= 1.336439) and (A92 <= 1.11976) and (A185 >= 3581)
and (A20 >= 0.946429) and (A68 <= 1.568389) and (A182 >=
0.954958) => Label=1 (108.0/2.0)
3. (A134 >= 1.061898) and (A134 >= 1.47425) and (A46 <=
2.333333) and (A182 >= 1.016753) and (A40 >= 0.9125) =>
Label=1 (124.0/10.0)
4. (A94 <= 1.030769) and (A59 >= 234) and (A100 <= 0.862408)
and (A90 <= 0.874074) and (A3 <= 754.905263) => Label=1
(134.0/2.0)
5. (A98 <= 1.091153) and (A134 >= 0.941898) and (A168 >=
1.716763) and (A26 >= 0.428571) and (A206 <= 1.021277) =>
Label=1 (106.0/8.0)
6. (A134 >= 0.950803) and (A94 <= 1.27027) and (A181 >= 2618)
and (A178 <= 1.158451) and (A132 >= 1.167996) and (A53 <=
174) => Label=1 (68.0/8.0)
7. (A92 <= 1.140625) and (A135 >= 3308) and (A69 <= 469) and (A80
>= 0.840686) and (A168 <= 1.143382) => Label=1 (100.0/2.0)
8. (A131 >= 2506) and (A92 <= 1.310606) and (A24 >= 1.24359)
and (A129 >= 2091) and (A14 <= 1.107143) and (A26 >= 0.857143)
=> Label=1 (72.0/0.0)
9. (A54 >= 0.843931) and (A96 <= 1.398104) and (A173 >= 942) and
(A166 <= 0.887446) and (A137 <= 5502) and (A42 >= 0.545455)
=> Label=1 (52.0/0.0)
10. (A136 >= 1.147994) and (A98 <= 1.246201) and (A46 <= 1.769231)
and (A7 >= -53) and (A108 <= 1.046472) and (A70 >= 0.778243)
=> Label=1 (66.0/2.0)
11. (A55 >= 192) and (A102 <= 1.219638) and (A23 >= 33) and (A4
<= 1.288403) and (A186 >= 0.99223) and (A186 <= 1.252018) and
(A3 >= 754.905263) and (A14 <= 4.4) => Label=1 (64.0/4.0)
12. (A96 <= 1.180995) and (A37 <= 56) and (A96 <= 0.805825) and
(A10 >= 0.21875) and (A50 <= 1.264706) and (A60 >= 0.621212)
=> Label=1 (52.0/2.0)
13. (A134 >= 1.069912) and (A10 <= -0.368421) and (A16 >= 1.8)
and (A182 <= 0.923543) => Label=1 (26.0/0.0)
14. (A181 >= 2642) and (A181 <= 2957) and (A92 <= 1.126984) and
(A3 <= 914.681579) and (A31 >= 79) and (A194 >= 0.960687) =>
Label=1 (54.0/0.0)
...
41. => Label=0 (1690.0/0.0)
Number of Rules : 41

Figure B.6: Instance-specific result of the JRip classifier for the FSP instance
TA082, training data shown in Table B.6

EXAMPLES OF DATA AND RESULTS OF LEARNING

191

Naive Bayes Classifier
Class 0: Prior probability = 0.5
A1: Normal Distribution. Mean = 50.4556 StandardDev = 5.2841
WeightSum = 1740 Precision = 0.4265306122448979
A2: Normal Distribution. Mean = 1.0188 StandardDev = 0.1557
WeightSum = 1740 Precision = 4.8687212103746397E-4
A3: Normal Distribution. Mean = 845.4133 StandardDev = 165.8667
WeightSum = 1740 Precision = 14.698139655172413
A4: Normal Distribution. Mean = 1.1252 StandardDev = 0.341
WeightSum = 1740 Precision = 0.0013090101322599194
A5: Normal Distribution. Mean = 5581.2359 StandardDev = 1062.31
WeightSum = 1740 Precision = 5.098496240601504
A6: Normal Distribution. Mean = 1.0177 StandardDev = 0.2476
WeightSum = 1740 Precision = 9.092975431530495E-4
A7: Normal Distribution. Mean = 545.3688 StandardDev = 803.2476
WeightSum = 1740 Precision = 81.51020408163265
A8: Normal Distribution. Mean = 0.2234 StandardDev = 5.6679
WeightSum = 1740 Precision = 0.05582411059757479
...
A206: Normal Distribution. Mean = 1.0171 StandardDev = 0.0649
WeightSum = 1740 Precision = 0.0015540015384615388
Class 1: Prior probability = 0.5
A1: Normal Distribution. Mean = 50.1061 StandardDev = 5.4521
WeightSum = 1740 Precision = 0.4265306122448979
A2: Normal Distribution. Mean = 1.0045 StandardDev = 0.1532
WeightSum = 1740 Precision = 4.8687212103746397E-4
A3: Normal Distribution. Mean = 792.9393 StandardDev = 190.5492
WeightSum = 1740 Precision = 14.698139655172413
A4: Normal Distribution. Mean = 0.9749 StandardDev = 0.3073
WeightSum = 1740 Precision = 0.0013090101322599194
A5: Normal Distribution. Mean = 5633.6625 StandardDev =1024.5525
WeightSum = 1740 Precision = 5.098496240601504
A6: Normal Distribution. Mean = 1.0401 StandardDev = 0.2506
WeightSum = 1740 Precision = 9.092975431530495E-4
A7: Normal Distribution. Mean = -31.8546 StandardDev = 677.081
WeightSum = 1740 Precision = 81.51020408163265
A8: Normal Distribution. Mean = 0.0055 StandardDev = 4.1143
WeightSum = 1740 Precision = 0.05582411059757479
...
A206: Normal Distribution. Mean = 0.9873 StandardDev = 0.061
WeightSum = 1740 Precision = 0.0015540015384615388

Figure B.7: Instance-specific result of the NaiveBayes classifier for the FSP
instance TA082, training data shown in Table B.6

192

EXAMPLES OF DATA AND RESULTS OF LEARNING

Table B.7: The test data and the summation of predicted labels of the assignments of the “following” relationship in the FSP instance TA082 (19800
rows × (206 columns of normalized attributes + 1 column of summations of
labels))
A1 A2

A3 A4

A5 A6

A7

56.9 1.07 1065.25 1.08 5215 0.74 766
53.3 0.94 984.54 0.92 7033 1.35 541
56.9 1.04 1065.25 1.79 6440 1.04 766
54.6 0.96 596.05 0.56 6187 0.96 494
56.9 1.21 1065.25 1.23 5728 0.81 766
47.0 0.83 865.00 0.81 7071 1.23 -293
56.9 1.35 1065.25 1.53 6575 1.28 766
42.2 0.74 694.03 0.65 5154 0.78 297
56.9 1.38 1065.25 1.67 5751 0.96 766
41.2 0.72 638.56 0.60 6012 1.05 -616
56.9 0.92 1065.25 1.60 3089 0.57 766
62.1 1.09 664.89 0.62 5442 1.76 342
56.9 0.93 1065.25 1.89 6380 1.08 766
61.3 1.08 562.85 0.53 5896 0.92 384
56.9 1.16 1065.25 1.28 7915 1.35 766
49.1 0.86 832.68 0.78 5850 0.74 -127
56.9 1.04 1065.25 1.69 7147 1.42 766
54.9 0.96 630.62 0.59 5037 0.70 1609
56.9 0.96 1065.25 1.75 3966 0.78 766
59.2 1.04 609.12 0.57 5096 1.28 -503
56.9 1.31 1065.25 1.29 6154 0.76 766
43.6 0.77 828.46 0.78 8064 1.31 -1787
56.9 1.16 1065.25 1.78 6223 1.16 766
49.3 0.87 596.93 0.56 5386 0.87 294
56.9 1.31 1065.25 1.67 5461 1.12 766
43.5 0.76 637.00 0.60 4873 0.89 637
56.9 1.23 1065.25 1.06 7010 1.37 766
46.4 0.82 1006.88 0.95 5100 0.73
85
56.9 1.17 1065.25 1.45 6639 0.88 766
48.5 0.85 733.52 0.69 7518 1.13 -917
56.9 1.06 1065.25 1.19 6758 0.89 766
53.8 0.95 896.48 0.84 7622 1.13 -861
56.9 1.05 1065.25 1.63 6415 1.07 766
54.1 0.95 654.62 0.61 6015 0.94 -1031
56.9 1.01 1065.25 1.41 7126 1.40 766
56.1 0.99 755.67 0.71 5082 0.71 1396
56.9 1.08 1065.25 0.79 6098 1.17 766
52.7 0.93 1345.08 1.26 5217 0.86 123
56.9 1.01 1065.25 1.23 5634 0.79 766
56.3 0.99 867.25 0.81 7151 1.27 -2264
..
..
..
..
..
..
..
.
.
.
.
.
.
.

A8

A9 A10 A11 A12 · · · A206

1.42 0.03 1.74
0.71 0.02 0.57
1.55 0.03 1.19
0.64 0.03 0.84
-2.61 0.03 -1.29
-0.38 -0.03 -0.78
2.58 0.03 -1.26
0.39 -0.03 -0.79
-1.24 0.03 -0.42
-0.80 -0.08 -2.38
2.24 0.03 1.32
0.45 0.02 0.76
1.99 0.03 1.90
0.50 0.02 0.53
-6.03 0.03 -0.42
-0.17 -0.08 -2.38
0.48 0.03 1.68
2.10 0.02 0.60
-1.52 0.03 -1.45
-0.66 -0.02 -0.69
-0.43 0.03 -0.19
-2.33 -0.17 -5.17
2.61 0.03 1.19
0.38 0.03 0.84
1.20 0.03 0.55
0.83 0.06 1.82
9.01 0.03 -1.10
0.11 -0.03 -0.91
-0.84 0.03 -1.45
-1.20 -0.02 -0.69
-0.89 0.03 -1.29
-1.12 -0.03 -0.78
-0.74 0.03 -1.03
-1.35 -0.03 -0.97
0.55 0.03 1.00
1.82 0.03 1.00
6.23 0.03 -1.23
0.16 -0.03 -0.82
-0.34 0.03 -0.68
-2.96 -0.05 -1.48
..
..
..
.
.
.

50.6 1.20 784.36 0.97 6265 1.06 1551 -21.85 0.03 0.58
42.3 0.84 812.22 1.04 5933 0.95 -71 -0.05 0.05 1.71

20 0.83 · · ·
24 1.20 · · ·
20 0.50 · · ·
40 2.00 · · ·
20 0.56 · · ·
36 1.80 · · ·
20 0.65 · · ·
31 1.55 · · ·
20 0.43 · · ·
46 2.30 · · ·
20 0.34 · · ·
59 2.95 · · ·
20 0.71 · · ·
28 1.40 · · ·
20 0.20 · · ·
99 4.95 · · ·
20 2.86 · · ·
7 0.35 · · ·
20 0.27 · · ·
74 3.70 · · ·
20 0.24 · · ·
84 4.20 · · ·
20 1.54 · · ·
13 0.65 · · ·
20 0.50 · · ·
40 2.00 · · ·
20 0.29 · · ·
68 3.40 · · ·
20 0.24 · · ·
85 4.25 · · ·
20 0.57 · · ·
35 1.75 · · ·
20 0.51 · · ·
39 1.95 · · ·
20 1.82 · · ·
11 0.55 · · ·
20 0.25 · · ·
79 3.95 · · ·
20 0.34 · · ·
58 2.90 · · ·
..
.. . .
.
. .

Σ label

1.02
0.98
1.11
0.90
1.03
0.97
1.03
0.97
1.22
0.82
1.06
0.94
1.08
0.93
1.00
1.00
1.06
0.94
1.01
0.99
1.03
0.97
1.06
0.94
1.14
0.88
1.04
0.96
1.00
1.00
1.01
0.99
1.04
0.96
1.08
0.93
1.01
0.99
1.00
1.00
..
.

2
2
0
3
0
3
3
0
0
3
0
3
0
2
1
2
3
1
2
2
0
3
1
2
1
2
2
1
0
3
0
3
0
3
2
1
1
3
1
3
..
.

51 4.64 · · · 1.03
11 0.22 · · · 0.97

0
3

EXAMPLES OF DATA AND RESULTS OF LEARNING

193

Table B.8: The weights (or summations of Σ label) according to Table B.7
(which generated the figure (i) in Figure 6.2)
Job
Id

Weight by
the SPOT

Job
Id

Weight by
the SPOT

Job
Id

Weight by
the SPOT

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

104
85
143
187
48
252
199
155
118
28
138
273
119
76
100
257
251
270
70
184
281
192
179
264
232
102
87
224
249
129
58
109
100
103
81

35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

84
244
227
291
197
255
206
27
207
236
174
42
110
6
3
215
100
122
198
292
262
294
222
54
181
192
48
225
148
81
165
166
40
11
85

70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99

120
83
210
247
18
40
251
196
71
197
259
110
202
134
221
244
171
265
66
48
190
241
160
277
123
192
136
77
106
39

194

EXAMPLES OF DATA AND RESULTS OF LEARNING

Table B.9: An example of adding the weights of a predicted permutation
and an NEH permutation (where the three columns of weights present the
figures (i), (ii) and (iii) in Figure 6.2, respectively)
Job

Weight by

Job

Weight by

Job

Weight by

Id SPOT NEH Sum

Id SPOT NEH Sum

Id SPOT NEH Sum

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99

32
24
46
57
10
86
66
48
37
5
45
94
38
18
27
88
84
93
16
56
96
59
54
91
77
30
26
74
83
42
14
34
28
31
20

69
72
64
81
3
87
49
65
67
25
92
76
55
35
47
38
85
73
36
95
96
57
48
19
66
44
11
61
98
40
5
20
24
46
32

101
96
110
138
13
173
115
113
104
30
137
170
93
53
74
126
169
166
52
151
192
116
102
110
143
74
37
135
181
82
19
54
52
77
52

23
80
76
97
63
87
68
4
69
78
53
9
35
1
0
71
29
40
65
98
90
99
73
13
55
60
11
75
47
21
50
51
7
2
25

71
62
43
97
70
91
52
14
82
89
21
27
37
0
1
75
59
88
84
17
80
99
15
22
45
41
56
31
78
4
16
68
9
12
51

94
142
119
194
133
178
120
18
151
167
74
36
72
1
1
146
88
128
149
115
170
198
88
35
100
101
67
106
125
25
66
119
16
14
76

39
22
70
82
3
8
85
62
17
64
89
36
67
43
72
81
52
92
15
12
58
79
49
95
41
61
44
19
33
6

34
26
29
83
23
2
42
50
30
63
10
93
58
53
60
90
79
54
77
13
28
86
33
74
18
6
39
8
94
7

73
48
99
165
26
10
127
112
47
127
99
129
125
96
132
171
131
146
92
25
86
165
82
169
59
67
83
27
127
13

Appendix C
Tables

Original Data of Figures and

Table C.1: Original data of Figure 1.4 (D. S. Johnson & McGeoch, 2002, p. 376)
Heuristic

% excess over the HK bound

Normalized time (s)

34.56
30.75
29.34
26.50
24.79
20.73
16.42
13.35
12.03
11.73
11.05
9.81
5.89
4.70
2.88
2.00
1.48
1.18
0.90
0.69

0.02
0.01
0.85
1.71
0.28
0.83
0.20
2.59
0.24
1129
0.44
1.04
823
1.41
1.50
2.06
18830
12.75
63.91
1840

Spacefill
Strip
Karp-20
NI
NN
CHCI
Greedy
FI
Savings
CCA
AppChristo
Christo-S
GENI-10
2-Opt-JM
3-Opt-JM
LK-JM
Tabu-SC-SC
MLLK-.1N
CLK-ABCC-N
Helsgaun.1N

195

196

ORIGINAL DATA OF FIGURES

Table C.2: Original data of Figure 5.1
n’

cov (%)

d¯

25
50
100
200
400
800

98.462
98.725
98.646
98.612
98.616
98.619

8-NN†

98.477

Timeoverall (s)

TimeL (s)

2.006
1.963
1.908
1.876
1.867
1.859

6.10
6.15
6.23
6.42
6.92
8.51

0.09
0.14
0.22
0.40
0.85
2.34

1.842

—

—

†: The 8-nearest-neighbor-first candidate set in HyFlex,
with no modification by SPOT algorithms.

ORIGINAL DATA OF FIGURES

197

Table C.3: Original data of Table 5.10
PHunter64bit
No. 0 No. 2 No. 6

No. 7

SPOT-PHunter
No. 8

No. 0 No. 2 No. 6

No. 7

No. 8

31 48194.9 6796.5 52510.6 66916.3 21221065.6 48194.9 6804.4 52950.5 67107.7 21429281.3
runs 48194.9 6813.4 52771.2 66689.6 21221065.6 48194.9 6801.5 52809.9 66790.8 21429281.3
48194.9 6825.0 52590.9 66686.0 21221065.6 48194.9 6802.1 52950.5 66606.4 21419868.0
48194.9 6799.4 52839.1 66841.1 21221065.6 48194.9 6807.9 52950.5 66790.8 21429281.3
48194.9 6820.1 52904.3 66693.0 21221065.6 48194.9 6805.7 52827.3 66606.4 21419868.0
48194.9 6808.4 52923.5 67025.8 21221065.6 48194.9 6802.3 53047.6 66606.4 21429281.3
48194.9 6812.4 52538.9 66722.6 21221065.6 48194.9 6802.1 52950.5 66606.4 21419868.0
48194.9 6824.0 53200.3 67073.3 21221065.6 48194.9 6803.4 52552.2 66606.4 21419868.0
48194.9 6805.0 52777.0 66887.8 21221065.6 48194.9 6803.7 52613.4 66606.4 21419868.0
48194.9 6813.3 52844.5 66704.0 21221065.6 48194.9 6804.6 52732.5 66728.0 21419868.0
48194.9 6820.2 53000.0 67272.9 21221065.6 48194.9 6804.8 52950.5 66894.9 21402829.4
48194.9 6817.2 53004.3 66700.4 21221065.6 48194.9 6803.4 52552.2 66790.8 21419868.0
48194.9 6819.6 52760.6 66744.5 21221065.6 48194.9 6801.5 52950.5 66606.4 21429281.3
48194.9 6797.5 52717.4 66749.6 21221065.6 48194.9 6801.5 53140.4 66790.8 20957289.4
48194.9 6811.6 52544.4 66487.7 21221065.6 48194.9 6801.5 52706.7 66790.8 21419868.0
48194.9 6810.3 52593.0 66699.4 21221065.6 48194.9 6804.6 52872.4 66606.4 21377588.8
48194.9 6812.6 52830.1 66490.6 21221065.6 48194.9 6802.8 52950.5 66606.4 21429281.3
48194.9 6809.9 52973.0 66655.8 21221065.6 48194.9 6801.5 52872.7 66790.8 21429281.3
48194.9 6815.3 52877.4 66960.9 21221065.6 48194.9 6804.7 52641.9 66790.8 21419868.0
48194.9 6805.0 52835.2 66910.3 21221065.6 48194.9 6802.1 52950.5 66393.5 21429281.3
48194.9 6812.4 52636.8 66801.1 21221065.6 48194.9 6802.1 52970.9 66606.4 21419868.0
48194.9 6801.6 52952.6 66930.1 21221065.6 48194.9 6803.7 53146.2 66606.4 21402829.4
48194.9 6810.1 52874.6 66635.1 21221065.6 48194.9 6802.4 52732.5 66614.9 21377588.8
48194.9 6829.8 52716.8 66978.5 21221065.6 48194.9 6802.4 52976.0 66614.9 21419868.0
48194.9 6824.8 53130.0 67164.0 21221065.6 48194.9 6802.4 52872.7 66606.4 21419868.0
48194.9 6814.8 52672.1 67072.4 21221065.6 48194.9 6803.7 52552.2 66606.4 21419868.0
48194.9 6799.2 52693.7 67305.3 21221065.6 48194.9 6802.4 53152.4 66606.4 21429281.3
48194.9 6818.3 52613.4 67100.0 21221065.6 48194.9 6802.1 52950.5 66790.8 21419868.0
48194.9 6807.2 52919.8 66987.9 21221065.6 48194.9 6807.9 52796.1 66606.4 21429281.3
48194.9 6823.2 52873.4 66760.7 21221065.6 48194.9 6804.9 52834.1 66606.4 21377588.8
48194.9 6820.4 52976.5 66826.0 21221065.6 48194.9 6802.1 52634.3 66790.8 21402829.4
Min 48194.9 6796.5 52510.6 66487.7 21221065.6 48194.9 6801.5 52552.2 66393.5 20957289.4
Med 48194.9 6812.6 52835.2 66826.0 21221065.6 48194.9 6802.4 52872.7 66606.4 21419868.0
Max 48194.9 6829.8 53200.3 67305.3 21221065.6 48194.9 6807.9 53152.4 67107.7 21429281.3
One-way ANOVA†
F
p
†: Factor: PHunter64bit = 1, SPOT-PHunter = 2.
‡: Exactly the same.

No. 0 No. 2 No. 6

No. 7

—‡
—

15.14
0.00

35.95
0.00

1.21
0.28

No. 8
210.25
0.00

198

ORIGINAL DATA OF FIGURES

Table C.4: Original data of Figure 5.3
Source

Hyper-heuristic

Score on each test instance

Overall

(alphabetical)

No. 0

No. 2

No. 6

No. 7

No. 8

score

Results† of
CHeSC 2011
(on a 32-bit
Java)

ACO-HH
AdaptHH
Ant-Q
AVEG-Nep
DynILS
EPH
GenHive
GISS
HAEA
HAHA
ISEA
KSATS-HH
MCHH-S
ML
NAHH
PHunter32bit
SA-ILS
SelfSearch
VNS-TW
XCJ

0
6
0
0
1
6
0
0
1
0
0
0
0
1
0
6
0
0
6
0

0
8
0
0
1
6
0
0
0
0
0
0
0
2
0
4
0
0
3
0

2
3
0
0
0
6
0
0
0
0
0
0
0
0
4
5
0
0
1
0

2
5
0
0
0
8
1
0
0
0
0
0
0
4
0
3
0
0
0
0

0
10
0
0
4
1
0
0
6
0
8
0
0
0
5
0
0
3
2
0

4
32
0
0
6
27
1
0
7
0
8
0
0
7
9
18
0
3
12
0

Experiments
in thesis

PHunter64bit
SPOT-PHunter

6
6

5
10

10
8

6
10

0
0

27
34

(†: Solutions were available at: http://www.asap.cs.nott.ac.uk/external/chesc2011/
raw-results.html.)

ORIGINAL DATA OF FIGURES

199

Table C.5: Original data of Figure 6.3
Algorithm

n’

r (%)
Average

Std dev

Timeoverall

TimeL

(s)

(s)

SPOT
(direct)

10
20
30
40
50

65.759
69.711
71.514
72.409
72.970

5.088
3.387
2.822
2.497
2.542

8.77
9.28
11.43
17.53
31.87

2.76
3.27
5.42
11.52
25.85

SPOT
(weighted
permutation)

10
20
30
40
50

68.951
71.052
71.703
72.151
72.367

3.022
2.283
2.045
2.020
2.102

8.80
9.36
11.45
17.61
31.89

2.79
3.35
5.44
11.60
25.88

longest-total-time-first

52.207

—

< 0.01

NEH

66.636

—

< 0.01

Avg of random NEHs

66.145

2.181

< 0.01

200

ORIGINAL DATA OF FIGURES

Table C.6: Original data of Figure 6.4 (Significant values of p in bold)
Group

No. 5
versus
No. 5’

No. 6
versus
No. 6’

No. 9
versus
No. 9’

No. 10
versus
No. 10’

Instance

% excess over best-known
Base LLH

New LLH

0
1
2
3
4

7.043
6.510
6.126
5.742
6.149

6.656
6.188
5.701
5.513
5.800

Avg

6.314

5.972

0
1
2
3
4

5.589
5.127
4.824
4.633
4.810

5.250
4.867
4.475
4.468
4.438

Avg

4.997

4.700

0
1
2
3
4

5.651
5.097
4.807
4.565
5.043

5.586
5.016
4.750
4.507
4.986

Avg

5.033

4.969

0
1
2
3
4

5.730
5.170
4.878
4.621
5.113

5.679
5.130
4.843
4.577
5.053

Avg

5.102

5.056

Run time (ms)

F

p

Base LLH

New LLH

170.90
115.81
215.11
61.40
151.12

0.00
0.00
0.00
0.00
0.00

1.76
1.70
1.67
1.62
1.51

1.64
1.87
1.59
1.79
1.75

0.00

1.65

1.73

0.00
0.00
0.00
0.00
0.00

102.95
108.27
101.42
102.77
98.36

103.49
104.44
101.35
103.04
99.17

0.00

102.75

102.30

0.01
0.00
0.03
0.02
0.03

4.82
4.73
4.71
4.82
4.65

5.15
4.93
4.71
4.99
4.63

0.02

4.75

4.88

0.05
0.10
0.17
0.07
0.02

5.12
5.17
4.56
5.12
4.46

4.90
4.76
4.88
4.82
4.56

0.08

4.88

4.78

176.63
95.50
164.49
37.20
226.30

6.48
10.24
4.99
5.20
4.80

3.96
2.65
1.87
3.23
5.12

ORIGINAL DATA OF FIGURES

201

Table C.7: Original data of Table 6.13 (Significant values of p in bold)
PHunter64bit

SPOT-PHunter

No. 1

No. 3

No. 8 No. 10 No. 11 No. 1

No. 3

No. 8 No. 10 No. 11

31
runs

6238
6264
6217
6248
6211
6248
6226
6239
6247
6229
6241
6252
6265
6258
6238
6256
6241
6258
6248
6261
6234
6250
6246
6262
6226
6246
6276
6238
6254
6235
6215

26794
26847
26806
26822
26800
26842
26827
26818
26859
26783
26868
26841
26770
26843
26815
26760
26824
26776
26778
26792
26820
26858
26843
26855
26868
26770
26849
26852
26810
26836
26811

6326
6353
6353
6350
6344
6353
6353
6350
6350
6366
6350
6350
6335
6353
6353
6363
6350
6350
6350
6344
6363
6325
6350
6347
6353
6323
6366
6353
6350
6353
6353

11419
11381
11346
11358
11397
11359
11350
11319
11389
11375
11391
11360
11365
11355
11396
11402
11379
11382
11415
11346
11359
11359
11330
11351
11362
11392
11389
11389
11400
11421
11370

26603
26626
26615
26602
26638
26593
26600
26629
26609
26603
26587
26608
26616
26621
26640
26683
26603
26625
26610
26626
26629
26622
26603
26603
26601
26643
26593
26608
26603
26603
26643

6218
6242
6249
6226
6248
6242
6242
6277
6238
6247
6252
6234
6259
6267
6231
6228
6247
6241
6213
6256
6261
6233
6238
6238
6248
6251
6239
6238
6221
6232
6242

26792
26752
26810
26755
26815
26801
26776
26783
26814
26811
26784
26768
26783
26773
26773
26784
26784
26774
26790
26820
26814
26830
26789
26850
26817
26778
26772
26814
26814
26840
26825

6303
6326
6323
6326
6325
6323
6350
6353
6331
6350
6350
6303
6329
6314
6323
6310
6323
6303
6303
6325
6353
6323
6303
6350
6309
6333
6326
6318
6350
6303
6303

11355
11371
11343
11333
11360
11365
11359
11359
11359
11368
11375
11378
11340
11323
11352
11393
11364
11373
11359
11359
11359
11359
11359
11359
11373
11319
11355
11381
11395
11359
11329

26591
26659
26614
26616
26560
26579
26659
26640
26572
26596
26586
26659
26585
26635
26584
26659
26628
26611
26586
26544
26677
26567
26601
26624
26634
26611
26566
26604
26636
26586
26577

Min
Med
Max

6211
6246
6276

26760
26822
26868

6323
6350
6366

11319
11375
11421

26587
26609
26683

6213
6242
6277

26752
26790
26850

6303
6323
6353

11319
11359
11395

26544
26604
26677

One-way ANOVA†

No. 1

No. 3

No. 8 No. 10 No. 11

F
p

0.36
0.55

11.47
0.00

47.44
0.00

†: Factor: PHunter64bit = 1, SPOT-PHunter = 2.

7.51
0.01

1.21
0.28

202

ORIGINAL DATA OF FIGURES

Table C.8: Original data of Figure 6.5
Source

Hyper-heuristic

Score on each test instance

Overall

(alphabetical)

No. 1

No. 3

No. 8

No. 10

No. 11

score

Results† of
CHeSC 2011
(on a 32-bit
Java)

ACO-HH
AdaptHH
Ant-Q
AVEG-Nep
DynILS
EPH
GenHive
GISS
HAEA
HAHA
ISEA
KSATS-HH
MCHH-S
ML
NAHH
PHunter32bit
SA-ILS
SelfSearch
VNS-TW
XCJ

3
10
0
0
0
2
0
0
0
0
0
0
0
5.5
5.5
0
0
0
1
0

0
5
0
0
0
3
0
0
0
0
0
0
0
8
8
1.5
0
0
4
0

0
5
0
0
0
4
1
0
2
0
0
0
0
8
0
0
0
0
6
0

1
9
0
0
0
0
0
0
0
0
0
0
0
3
4
2
0
0
5
0

0
3
0
0
0
4
2
0
1
0
0
0
0
5
0
0
0
0
10
0

4
32
0
0
0
13
3
0
3
0
0
0
0
29.5
17.5
3.5
0
0
26
0

Experiments
in thesis

PHunter64bit
SPOT-PHunter

4
8

1.5
8

3
10

6
9

6
8

20.5
43

(†: Solutions were available at: http://www.asap.cs.nott.ac.uk/external/chesc2011/
raw-results.html.)

References

Aaronson, S. (2005, March). Guest column: NP-complete problems and physical reality. SIGACT News, 36(1), 30–52. doi:10.1145/1052796.1052804
Aha, D. W., Kibler, D., & Albert, M. K. (1991, January). Instance-based
learning algorithms. Machine Learning, 6(1), 37–66. doi:10 . 1023 / A :
1022689900470
Ahuja, R. K., Ergun, O., Orlin, J. B., & Punnen, A. P. (2002). A survey of very
large-scale neighborhood search techniques. Discrete Applied Mathematics, 123(1-3), 75–102. doi:10.1016/S0166-218X(01)00338-9
Aickelin, U., Burke, E. K., & Li, J. (2009, April). An evolutionary squeaky
wheel optimization approach to personnel scheduling. IEEE Transactions on Evolutionary Computation, 13(2), 433–443. doi:10.1109/TEVC.
2008.2004262
Allahverdi, A., & Al-Anzi, F. S. (2002, July). Using two-machine flowshop
with maximum lateness objective to model multimedia data objects
scheduling problem for WWW applications. Comput. Oper. Res. 29(8),
971–994. doi:10.1016/S0305-0548(00)00097-6
Applegate, D. L., Bixby, R. E., Chvátal, V., & Cook, W. J. (2006). The traveling
salesman problem: a computational study. Princeton Series in Applied
Mathematics. Princeton, NJ, USA: Princeton University Press.
203

204

REFERENCES

Applegate, D. L., Bixby, R. E., Chvátal, V., Cook, W. J., Espinoza, D. G.,
Goycoolea, M., & Helsgaun, K. (2009). Certification of an optimal
TSP tour through 85,900 cities. Operations Research Letters, 37(1), 11–
15. doi:10.1016/j.orl.2008.09.006
Applegate, D. L., Bixby, R. E., Chvátal, V., & Cook, W. J. (2001). TSP cuts
which do not conform to the template paradigm. In M. Jünger & D. Naddef (Eds.), Computational combinatorial optimization (Vol. 2241, pp. 261–
304). Lecture Notes in Computer Science. London, UK: Springer-Verlag.
doi:10.1007/3-540-45586-8 7
Ausiello, G., Protasi, M., Marchetti-Spaccamela, A., Gambosi, G., Crescenzi,
P., & Kann, V. (1999). Complexity and approximation: combinatorial optimization problems and their approximability properties (1st). Secaucus, NJ,
USA: Springer-Verlag New York, Inc. doi:10.1007/978-3-642-58412-1
Babb, J., & Currie, J. (2008, July). The brachistochrone problem: mathematics
for a broad audience via a large context problem. The Montana Mathematics Enthusiast, 5(3), 169–184. Retrieved April 4, 2012, from http://www.
math . umt . edu / tmme / vol5no2and3 / TMME vol5nos2and3 a1 pp .
169 184.pdf
Bader El Den, M., & Poli, R. (2009, May). Grammar-based genetic programming for timetabling. In 2009 IEEE congress on evolutionary computation
(cec’09) (pp. 2532–2539). doi:10.1109/CEC.2009.4983259
Barnhart, C., Johnson, E. L., Nemhauser, G. L., Savelsbergh, M. W. P., &
Vance, P. H. (1998, March). Branch-and-price: column generation for
solving huge integer programs. Operations Research, 46(3), 316–329.
doi:10.1287/opre.46.3.316

REFERENCES

205

Bentley, J. J. (1992). Fast algorithms for geometric traveling salesman problems. OSAR Journal on Computing, 4, 387–411. doi:10.1287/ijoc.4.4.387
Bernoulli, J. (1696, June). Problema novum ad cuius solutionem Mathematici
invitantur. Acta Eruditorum, 15, 264–269.
Bernstein, M. (1987, January). Finding heuristics for flowshop scheduling.
SIGART Newsletter, (99), 32–33. doi:10.1145/24667.24670
Birattari, M. (2009). Tuning metaheuristics: a machine learning perspective (J.
Kacprzyk, Ed.). Studies in Computational Intelligence. Berlin, Germany:
Springer-Verlag. doi:10.1007/978-3-642-00483-4
Blum, A. L., & Rivest, R. L. (1992). Training a 3-node neural network is NPcomplete. Neural Networks, 5(1), 117–127. doi:10.1016/S0893-6080(05)
80010-3
Blum, C., & Roli, A. (2003, September). Metaheuristics in combinatorial
optimization: overview and conceptual comparison. ACM Computer
Survey, 35(3), 268–308. doi:10.1145/937503.937505
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm intelligence: from natural to artificial systems. Santa Fe Institute Studies in the Sciences of Complexity. New York, NY, USA: Oxford University Press. Retrieved April
4, 2012, from http://books.google.com.hk/books?id=PvTDhzqMr7cC
Bose, I., & Mahapatra, R. K. (2001). Business data mining — a machine
learning perspective. Information & Management, 39(3), 211–225. doi:10.
1016/S0378-7206(01)00091-X
Boyan, J., & Moore, A. (2000). ‘‘STAGE” learning for local search. Neural
computing surveys, 3, 35–38. Retrieved April 4, 2012, from http://ftp.
icsi.berkeley.edu/ftp/pub/ai/jagota/vol3 1.pdf

206

REFERENCES

Boyan, J. (1998, August). Learning evaluation functions for global optimization.
(Doctoral dissertation, Robotics Institute, Carnegie Mellon University,
Pittsburgh, PA, USA). Retrieved April 4, 2012, from http://www.ri.
cmu.edu/publication view.html?pub id=2954
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1983). Classification
and regression trees. Wadsworth Statistics / Probability series. Belmont,
CA, USA: Wadsworth Publishing Co. Inc.
Brown, D. E., Pittard, C. L., & Park, H. (1996). Classification trees with optimal
multivariate decision nodes. Pattern Recognition Letters, 17(7), 699–703.
doi:10.1016/0167-8655(96)00033-5
Burke, E. K., Hyde, M., & Kendall, G. (2006). Evolving bin packing heuristics
with genetic programming. In Proceedings of the 9th international conference on parallel problem solving from nature (ppsn’06) (Vol. 4193, pp. 860–
869). Lecture Notes in Computer Science. Germany: Springer-Verlag.
doi:10.1007/11844297 87
Burke, E. K., Hyde, M., Kendall, G., Ochoa, G., Ozcan, E., & Qu, R. (2010).
Hyper-heuristics: a survey of the state of the art (tech. rep. No. NOTTCS-TRSUB-0906241418-2747). School of Computer Science and Information
Technology, University of Nottingham. Nottingham NG8 1BB, UK.
Retrieved April 4, 2012, from http://www.cs.nott.ac.uk/∼gxo/papers/
hhsurvey.pdf
Burke, E. K., Hyde, M., Kendall, G., Ochoa, G., Özcan, E., & Woodward, J. R.
(2010). A classification of hyper-heuristic approaches. In M. Gendreau
& J.-Y. Potvin (Eds.), Handbook of metaheuristics (Vol. 146, pp. 449–468).
International Series in Operations Research & Management Science.
Springer US. doi:10.1007/978-1-4419-1665-5 15

REFERENCES

207

Burke, E. K., Hyde, M., Kendall, G., & Woodward, J. (2007). Automatic
heuristic generation with genetic programming: evolving a jack-of-alltrades or a master of one. In Proceedings of the 9th annual conference
on genetic and evolutionary computation (gecco’07) (pp. 1559–1565). New
York, NY, USA: ACM. doi:10.1145/1276958.1277273
Burke, E. K., Kendall, G., Newall, J., Hart, E., Ross, P., & Schulenburg, S. (2003).
Hyper-heuristics: an emerging direction in modern search technology.
In F. Glover & G. Kochenberger (Eds.), Handbook of metaheuristics (Vol.
57, pp. 457–474). International Series in Operations Research & Management Science. Springer New York. doi:10.1007/0-306-48056-5 16
Burke, E. K., Petrovic, S., & Qu, R. (2006, April). Case-based heuristic selection
for timetabling problems. J. of Scheduling, 9(2), 115–132. doi:10.1007/
s10951-006-6775-y
Campbell, H. G., Dudek, R. A., & Smith, M. L. (1970). A heuristic algorithm
for the n job, m machine sequencing problem. Management Science,
16(10), B630–B637. doi:doi:10.1287/mnsc.16.10.B630
Ceberio, J., Irurozki, E., Mendiburu, A., & Lozano, J. A. (2012). A review
on estimation of distribution algorithms in permutation-based combinatorial optimization problems. Progress in Artificial Intelligence, 1(1),
103–117. doi:10.1007/s13748-011-0005-3
Černý, V. (1985, January). Thermodynamical approach to the traveling salesman problem: an efficient simulation algorithm. Journal of Optimization
Theory and Applications, 45(1), 41–51. doi:10.1007/BF00940812
Chan, C. Y., Xue, F., Ip, W. H., & Cheung, C. F. (2012, January). A hyperheuristic inspired by pearl hunting. In Y. Hamadi & M. Schoenauer
(Eds.), Learning and intelligent optimization (pp. 349–353). Lecture Notes

208

REFERENCES
in Computer Science. Springer-Verlag. doi:10.1007/978-3-642-344138 26

Clark, P., & Niblett, T. (1989). The CN2 induction algorithm. Machine Learning,
3, 261–283. doi:10.1007/BF00116835
Cohen, W. W. (1995). Fast effective rule induction. In Proceedings of the twelfth
international conference on machine learning (ml95) (pp. 115–123). Tahoe
City, CA, USA. Retrieved April 4, 2012, from http://www.cs.cmu.edu/
∼wcohen/postscript/ml-95-ripper.ps

Colorni, A., Dorigo, M., & Maniezzo, V. (1992). Distributed optimization by
ant colonies. In F. J. Varela & P. Bourgine (Eds.), Towards a practice of
autonomous systems: proceedings of the first european conference on artificial
life (pp. 134–142). Cambridge, MA, USA: MIT Press.
Cook, W. J., & Seymour, P. (2003, July). Tour merging via branch-decomposition.
INFORMS Journal on Computing, 15(3), 233–248. doi:10.1287/ijoc.15.3.
233.16078
Cotta, C., & Troya, J. M. (2003, April). Embedding branch and bound within
evolutionary algorithms. Applied Intelligence, 18(2), 137–153. doi:10 .
1023/A:1021934325079
Croes, G. A. (1958, December). A method for solving traveling-salesman
problems. Operations Research, 6(6), 791–812. doi:10.1287/opre.6.6.791
Crowston, W. B., Glover, F., Thompson, G. L., & Trawick, J. D. (1963). Probabilistic and parametric learning combinations of local job shop scheduling rules.
ONR Research Memorandum. Pittsburgh, PA, USA: Defense Technical
Information Center.
Cruz-Reyes, L., Gómez-Santillán, C., Pérez-Ortega, J., Landero, V., Quiroz, M.,
& Ochoa, A. (2012, March). Algorithm selection: from meta-learning to

REFERENCES

209

hyper-heuristics. In V. M. Koleshko (Ed.), Intelligent systems (Chap. 4,
pp. 77–102). Rijeka, Croatia: InTech. doi:10.5772/2350
Culberson, J. C. (1998, June). On the futility of blind search: an algorithmic
view of “no free lunch”. Evolutionary Computation, 6(2), 109–127. doi:10.
1162/evco.1998.6.2.109
Dannenbring, D. G. (1977, July). An evaluation of flow shop sequencing
heuristics. Management Science, 23(11), 1174–1182. Retrieved from http:
//www.jstor.org/stable/2630656
Davis, M. (1965). The undecidable: basic papers on undecidable propositions, unsolvable problems and computable functions. Hewlett, New York, NY, USA:
Raven Press.
Denzinger, J., Fuchs, M., & Fuchs, M. (1997). High performance ATP systems
by combining several AI methods. In Proceedings of the 15th international
joint conference on artifical intelligence — volume 1 (pp. 102–107). IJCAI9́7.
San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Retrieved
April 4, 2012, from http://pages.cpsc.ucalgary.ca/∼denzinge/papers/
Denzinger.SR-96-09.ps.gz
Denzinger, J., & Offermann, T. (1999). On cooperation between evolutionary algorithms and other search paradigms. In Proceedings of the 1999
congress on evolutionary computation (cec’99) (pp. 2317–2324). doi:10 .
1109/CEC.1999.785563
de Werra, D., & Hertz, A. (1989). Tabu-search techniques: a tutorial and an
application to neural networks. OR Spektrum, 11(3), 131–141. doi:10.
1007/BF01720782
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs.
Numerische Mathematik, 1, 269–271. doi:10.1007/BF01386390

210

REFERENCES

Dorigo, M., & Gambardella, L. M. (1997a). Ant colonies for the travelling
salesman problem. Biosystems, 43(2), 73–81. doi:10.1016/S0303-2647(97)
01708-5
Dorigo, M., & Gambardella, L. M. (1997b). Ant colony system: a cooperative
learning approach to the traveling salesman problem. IEEE Transactions
on Evolutionary Computation, 1(1), 53–66. doi:10.1109/4235.585892
Dorigo, M., Maniezzo, V., & Colorni, A. (1991). Positive feedback as a search
strategy (tech. rep. No. 91-016). Dipartimento di Elettronica, Politecnico
di Milano. Milan, Italy. Retrieved April 4, 2012, from http://citeseerx.
ist.psu.edu/viewdoc/summary?doi=10.1.1.52.6342
Eiben, A. E., Hinterding, R., & Michalewicz, Z. (1999, July). Parameter control
in evolutionary algorithms. IEEE Transactions on Evolutionary Computation, 3(2), 124–141. doi:10.1109/4235.771166
Eiben, A. E., & Smith, J. E. (2003). Introduction to evolutionary computing.
Natural Computing Series. Berlin, Germany: Springer-Verlag. Retrieved
April 4, 2012, from http://www.cs.vu.nl/∼gusz/ecbook/ecbook.html
Fan, W., & Xue, F. (2006). Optimize cooperative agents with organization
in distributed scheduling system. In K. L. De-Shuang Huang & G. W.
Irwin (Eds.), Computational intelligence (Vol. 4114, pp. 502–509). Lecture
Notes in Artificial Intelligence. Germany: Springer-Verlag. doi:10.1007/
978-3-540-37275-2 61
Feo, T. A., & Resende, M. G. C. (1989). A probabilistic heuristic for a computationally difficult set covering problem. Operations Research Letters,
8(2), 67–71. doi:10.1016/0167-6377(89)90002-3

REFERENCES

211

Feo, T. A., & Resende, M. G. C. (1995). Greedy randomized adaptive search
procedures. Journal of Global Optimization, 6, 109–133. doi:10 . 1007 /
BF01096763
Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7(2), 179–188. doi:10.1111/j.1469-1809.1936.
tb02137.x
French, A. P., Robinson, A. C., & Wilson, J. M. (2001, November). Using a hybrid genetic-algorithm/branch and bound approach to solve feasibility
and optimization integer programming problems. Journal of Heuristics,
7(6), 551–564. doi:10.1023/A:1011921025322
Fukunaga, A. S. (2008, March). Automated discovery of local search heuristics
for satisfiability testing. Evolutionary Computation, 16(1), 31–61. doi:10.
1162/evco.2008.16.1.31
Gama, J. (1999). Discriminant trees. In Proceedings of the sixteenth international
conference on machine learning (icml’99) (pp. 134–142). San Francisco, CA,
USA: Morgan Kaufmann Publishers Inc. Retrieved April 4, 2012, from
http://www.liaad.up.pt/∼jgama/ml99.ps.gz
Gambardella, L. M., & Dorigo, M. (1995). Ant-Q: a reinforcement learning
approach to the traveling salesman problem. In A. Prieditis & S. Russell
(Eds.), Machine learning: proceedings of the twelfth international conference
on machine learning (pp. 252–260). San Francisco, CA: Morgan Kaufmann
Publishers. Retrieved April 4, 2012, from http://www.idsia.ch/∼luca/
ml95.pdf
Garey, M. R., & Johnson, D. S. (1979). Computers and intractability; a guide to
the theory of NP-completeness. New York, NY, USA: W. H. Freeman & Co.
doi:10.1137/1024022

212

REFERENCES

Garey, M. R., Johnson, D. S., & Sethi, R. (1976, May). The complexity of
flowshop and jobshop scheduling. Mathematics of operations research,
1(2), 117–129. doi:10.1287/moor.1.2.117
Gendreau, M., & Potvin, J.-Y. (2010). Handbook of metaheuristics (2nd) (F. S.
Hillie, Ed.). International Series in Operations Research & Management
Science. New York, NY, USA: Springer Science+Business Media. doi:10.
1007/978-1-4419-1665-5
Gerdts, M. (2012). Optimal control of ODEs and DAEs. Göttingen, Germany:
Walter De Gruyter Inc. doi:10.1515/9783110249996
Glover, F. (1986). Future paths for integer programming and links to artificial
intelligence. Computers & Operations Research, 13(5), 533–549. doi:10.
1016/0305-0548(86)90048-1
Glover, F., & Laguna, M. (1997). Tabu search. Kluwer Academic Publishers.
doi:10.1007/978-1-4615-6089-0
Gomes, C. P., & Selman, B. (1997). Algorithm portfolio design: theory vs.
practice. In Proceedings of the thirteenth conference on uncertainty in artificial intelligence (pp. 190–197). UAI9́7. San Francisco, CA, USA: Morgan
Kaufmann Publishers Inc. Retrieved April 4, 2012, from http://dl.acm.
org/citation.cfm?id=2074226.2074249
Grefenstette, J. J. (1986, January). Optimization of control parameters for
genetic algorithms. IEEE Transactions on Systems, Man and Cybernetics,
16(1), 122–128. doi:10.1109/TSMC.1986.289288
Gupta, J. N. D. (1975). A search algorithm for the generalized flowshop
scheduling problem. Computers & Operations Research, 2(2), 83–90. doi:10.
1016/0305-0548(75)90011-8

REFERENCES

213

Gutin, G., & Glover, F. (2005). Further extension of the TSP assign neighborhood. Journal of Heuristics, 11(5-6), 501–505. doi:10.1007/s10732-0053487-y
Hackley, C. W. (1847). Elementary course of geometry. New York, NY, USA:
Harper & Brothers, Publishers. Retrieved April 4, 2012, from http :
//www.google.com.hk/books?id=jWILAAAAYAAJ
Hall, M. A. (1999, April). Correlation-based feature selection for machine learning.
(Doctoral dissertation, Department of Computer Science, University
of Waikato, Hamilton, NewZealand). Retrieved April 4, 2012, from
http://www.cs.waikato.ac.nz/∼mhall/thesis.pdf
Held, M., & Karp, R. M. (1970, December). The traveling-salesman problem
and minimum spanning trees. Operations Research, 18(6), 1138–1162.
doi:10.1287/opre.18.6.1138
Heller, J. (1960, April). Some numerical experiments for an M x J flow shop
and its decision-theoretical aspects. Operations Research, 8(2), 178–184.
doi:10.1287/opre.8.2.178
Helsgaun, K. (2000). An effective implementation of the Lin-Kernighan traveling salesman heuristic. European Journal of Operational Research, 126(1),
106–130. doi:10.1016/S0377-2217(99)00284-2
Helsgaun, K. (2009). General k-opt submoves for the Lin-Kernighan TSP
heuristic. Mathematical Programming Computation, 1, 119–163. doi:10.
1007/s12532-009-0004-6
Hilario, M., Kalousis, A., Nguyen, P., & Woznica, A. (2009, September). A
data mining ontology for algorithm selection and meta-mining. In V.
Podpečan, N. Lavrač, J. N. Kok & J. de Bruin (Eds.), Proceedings of the
2009 european conference on machine learning and principles and practice

214

REFERENCES
of knowledge discovery in databases (ecml pkdd 2009) (pp. 76–87). Bled,
Slovenia. Retrieved April 4, 2012, from http://www.ecmlpkdd2009.
net/wp- content/uploads/2008/09/service- oriented- knowledgediscovery 2.pdf#page=82

Holland, J. H. (1992). Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control and artificial intelligence.
Cambridge, MA, USA: MIT Press.
Ho, N. B., & Tay, J. C. (2005, September). Evolving dispatching rules for
solving the flexible job-shop problem. In The 2005 IEEE congress on
evolutionary computation (Vol. 3, pp. 2848–2855). doi:10.1109/CEC.2005.
1555052
Hsiao, P.-C., Chiang, T.-C., & Fu, C. (2011, September). A variable neighborhood search-based hyperheuristic for corss-domain optimization
problems in CHeSC 2011 competition. (p. 89). The OR Society. Nottingham, UK.
Huberman, B. A., Lukose, R. M., & Hogg, T. (1997, January). An economics
approach to hard computational problems. Science, 275(5296), 51–54.
doi:10.1126/science.275.5296.51
Hunt, J. E., & Cooke, D. E. (1996, April). Learning using an artificial immune
system. Journal of Network and Computer Applications, 19(2), 189–212.
doi:10.1006/jnca.1996.0014
Hutchinson, J. E. (1981). Fractals and self-similarity. Indiana University Mathematics Journal, 30(5), 713–747. Retrieved April 4, 2012, from http://
maths.anu.edu.au/∼john/Assets/Research%20Papers/fractals selfsimilarity.pdf

REFERENCES

215

Hutter, F., Hoos, H. H., Leyton-Brown, K., & Stützle, T. (2009, September).
ParamILS: an automatic algorithm configuration framework. Journal
of Artificial Intelligence Research, 36(1), 267–306. Retrieved April 4, 2012,
from http://www.jair.org/media/2861/live-2861-4703-jair.pdf
Ibaraki, T., & Muroga, S. (1970, January). Adaptive linear classifier by linear
programming. IEEE Transactions on Systems Science and Cybernetics, 6(1),
53–62. doi:10.1109/TSSC.1970.300329
Ignall, E., & Schrage, L. (1965, June). Application of the Branch and Bound
technique to some flow-shop scheduling problems. Operations Research,
13(3), 400–412. doi:10.1287/opre.13.3.400
Jakobović, D., Jelenković, L., & Budin, L. (2007). Genetic programming heuristics for multiple machine scheduling. In Proceedings of the 10th european conference on genetic programming (eurogp’07) (Vol. 4445, pp. 321–
330). Lecture Notes in Computer Science. Germany: Springer-Verlag.
doi:10.1007/978-3-540-71605-1 30
John, G. H., & Langley, P. (1995). Estimating continuous distributions in
Bayesian classifiers. In Proceedings of the eleventh conference on uncertainty
in artificial intelligence (uai’95) (pp. 338–345). San Francisco, CA, USA:
Morgan Kaufmann Publishers Inc. Retrieved April 4, 2012, from http:
//dl.acm.org/citation.cfm?id=2074158.2074196
Johnson, D. S., & McGeoch, L. A. (1997). Traveling salesman problem: a
case study in local optimization. In E. H. L. Aarts & J. K. Lenstra
(Eds.), Local search in combinatorial optimization (Chap. 8, pp. 215–310).
London, UK: John Wiley and Sons. Retrieved April 4, 2012, from http:
//www2.research.att.com/∼dsj/papers/TSPchapter.pdf

216

REFERENCES

Johnson, D. S., & McGeoch, L. A. (2002). Experimental analysis of heuristics
for STSP. In G. Gutin & A. P. Punnen (Eds.), The traveling salesman
problem and its variations (Chap. 9, pp. 369–443). New York, NY, USA:
Kluwer Academic Publishers. doi:10.1007/b101971
Johnson, S. M. (1954). Optimal two- and three-stage production schedules
with setup times included. Naval Research Logistics Quarterly, 1(1), 61–68.
doi:10.1002/nav.3800010110
Joslin, D. E., & Clements, D. P. (1999, May). ‘‘squeaky wheel” optimization.
Journal of Artificial Intelligence Research, 10(1), 353–373. Retrieved April
4, 2012, from http://www.jair.org/media/561/live-561-1759-jair.pdf
Kadioglu, S., Malitsky, Y., Sellmann, M., & Tierney, K. (2010). Isac – instancespecific algorithm configuration. In Proceedings of 19th European conference on artificial intelligence (ecai 2010) (pp. 751–756). Amsterdam,
Netherlands: IOS Press. Retrieved April 4, 2012, from http://www.itu.
dk/people/kevt/papers/isac-ecai2010.pdf
Kanda, J., Carvalho, A., Hruschka, E., & Soares, C. (2011, August). Selection of
algorithms to solve traveling salesman problems using meta-learning.
International Journal of Hybrid Intelligent Systems, 8(3), 117–128. doi:10.
3233/HIS-2011-0133
Karp, R. M. (1972). Reducibility among combinatorial problems. In R. E.
Miller & J. W. Thatcher (Eds.), Complexity of computer computations
(pp. 85–103). New York, NY, USA: Plenum Press. doi:10.1007/9781-4684-2001-2 9
Kassim, H. O., & Cadbury, R. G. (1996). The place of the computer in chemical engineering education. Computers & Chemical Engineering, 20, Sup-

REFERENCES

217

plement 2, S1341–S1346. European Symposium on Computer Aided
Process Engineering 6. doi:10.1016/0098-1354(96)00230-X
Keller, R. E., & Poli, R. (2007, September). Linear genetic programming
of parsimonious metaheuristics. In 2007 IEEE congress on evolutionary
computation (cec 2007) (pp. 4508–4515). doi:10.1109/CEC.2007.4425062
Keller, R. E., & Poli, R. (2008). Cost-benefit investigation of a genetic-programming
hyperheuristic. In Proceedings of the evolution artificielle, 8th international
conference on artificial evolution (ea’07) (Vol. 4926, pp. 13–24). Lecture
Notes in Computer Science. Germany: Springer-Verlag. doi:10.1007/
978-3-540-79305-2 2
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. In Proceedings
of IEEE international conference on neural networks (Vol. 4, pp. 1942–1948).
IEEE Press. doi:10.1109/ICNN.1995.488968
Kennedy, J., Eberhart, R., & Shi, Y. (2001). Swarm intelligence. San Diego, CA:
Morgan Kaufmann Publishers.
Kim, B.-I., & Wy, J. (2010). Last two fit augmentation to the well-known
construction heuristics for one-dimensional bin-packing problem: an
empirical study. The International Journal of Advanced Manufacturing
Technology, 50, 1145–1152. doi:10.1007/s00170-010-2572-z
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983, May). Optimization by
simulated annealing. Science, 220(4598), 671–680. doi:10.1126/science.
220.4598.671
Kleunen, J. P. C. (1976). Computers and operations research: a survey. Computers & Operations Research, 3(4), 327–335. doi:10.1016/0305-0548(76)
90015-0

218

REFERENCES

Knuth, D. E., & Moore, R. W. (1975). An analysis of alpha-beta pruning.
Artificial Intelligence, 6(4), 293–326. doi:10.1016/0004-3702(75)90019-3
Ko, K.-I., & Friedman, H. (1982). Computational complexity of real functions.
Theoretical Computer Science, 20(3), 323–352. doi:10.1016/S0304-3975(82)
80003-0
Koonce, D. A., & Tsai, S.-C. (2000). Using data mining to find patterns in
genetic algorithm solutions to a job shop schedule. Comput. Ind. Eng.
38(3), 361–374. doi:10.1016/S0360-8352(00)00050-4
Kotsiantis, S. B. (2007, October). Supervised machine learning: a review of
classification techniques. Informatica, 31(3), 249–268. Retrieved April 4,
2012, from http://www.informatica.si/vol31.htm#No3
Kotsiantis, S. B., Zaharakis, I. D., & Pintelas, P. E. (2006, November). Machine
learning: a review of classification and combining techniques. Artificial
Intelligence Review, 26(3), 159–190. doi:10.1007/s10462-007-9052-3
Koza, J. R. (1992). Genetic programming: on the programming of computers by
means of natural selection. Cambridge, MA, USA: MIT Press.
Kriegel, H.-P., Kröger, P., Schubert, E., & Zimek, A. (2008). A general framework for increasing the robustness of PCA-based correlation clustering
algorithms. In Proceedings of the 20th international conference on scientific and statistical database management (ssdbm ’08) (Vol. 5069, pp. 418–
435). Lecture Notes in Computer Science. Germany: Springer-Verlag.
doi:10.1007/978-3-540-69497-7 27
Kwak, C., & Yih, Y. (2004, February). Data-mining approach to production
control in the computer-integrated testing cell. IEEE Transactions on
Robotics And Automation, 20(1), 107–116. doi:10.1109/TRA.2003.819595

REFERENCES

219

Lachenbruch, P. A., & Goldstein, M. (1979). Discriminant analysis. Biometrics,
35(1), 69–85. doi:10.2307/2529937
Lawler, E. L., & Wood, D. E. (1966, August). Branch-and-bound methods: a
survey. Operations Research, 14(4), 699–719. doi:10.1287/opre.14.4.699
Le Cessie, S., & Van Houwelinge, J. C. (1992). Ridge estimators in logistic
regression. Applied Statistics, 41(1), 191–201. doi:10.2307/2347628
Lee, C.-Y., Piramuthu, S., & Tsai, Y.-K. (1997). Job shop scheduling with a genetic algorithm and machine learning. International Journal of Production
Research, 35(4), 1171–1191. doi:10.1080/002075497195605
Leyton-Brown, K., Nudelman, E., Andrew, G., McFadden, J., & Shoham,
Y. (2003). A portfolio approach to algorithm select. In Proceedings of
the 18th international joint conference on artificial intelligence (pp. 1542–
1543). IJCAI’03. Acapulco, Mexico: Morgan Kaufmann Publishers Inc.
Retrieved April 4, 2012, from http://robotics.stanford.edu/∼eugnud/
papers/portfolio-IJCAI.pdf
Li, L., Ju, S., & Zhang, Y. (2008, October). Improved ant colony optimization
for the traveling salesman problem. In 2008 international conference
on intelligent computation technology and automation (icicta) (pp. 76–80).
doi:10.1109/ICICTA.2008.265
Lindner, G., & Studer, R. (1999). AST: support for algorithm selection with a
CBR approach. In J. Zytkow & J. Rauch (Eds.), Principles of data mining
and knowledge discovery (Vol. 1704, pp. 418–423). Lecture Notes in Computer Science. Germany: Springer. doi:10.1007/978-3-540-48247-5 52
Lin, S. (1965, December). Computer solutions of the traveling salesman
problem. Bell System Technical Journal, 44(10), 2245–2269. Retrieved

220

REFERENCES
April 4, 2012, from http : / / www. alcatel - lucent . com / bstj / vol44 1965/articles/bstj44-10-2245.pdf

Lin, S., & Kernighan, B. W. (1973). An effective heuristic algorithm for the
travelling-salesman problem. Operations Research, 21, 498–516. doi:10.
1287/opre.21.2.498
Liu, H., & Motoda, H. (2001). Instance selection and construction for data mining.
Norwell, MA, USA: Kluwer Academic Publishers.
Li, X., & Olafsson, S. (2005). Discovering dispatching rules using data mining.
Journal of Scheduling, 8(6), 515–527. doi:10.1007/s10951-005-4781-0
Lobjois, L., & Lemaı̂tre, M. (1998). Branch and bound algorithm selection
by performance prediction. In Proceedings of the fifteenth national/tenth
conference on artificial intelligence/innovative applications of artificial intelligence (pp. 353–358). AAAI ’98/IAAI ’98. Menlo Park, CA, USA:
American Association for Artificial Intelligence. Retrieved April 4, 2012,
from http://aaaipress.org/Papers/AAAI/1998/AAAI98-050.pdf
Lobo, F. G., & Goldberg, D. E. (2004, December). The parameter-less genetic
algorithm in practice. Inf. Sci. Inf. Comput. Sci. 167(1-4), 217–232. doi:10.
1016/j.ins.2003.03.029
Lourenço, H., Martin, O., & Stützle, T. (2003). Iterated local search. In F.
Glover & G. Kochenberger (Eds.), Handbook of metaheuristics (Chap.
11, Vol. 57, pp. 320–353). International Series in Operations Research
& Management Science. New York, NY, USA: Springer New York.
doi:10.1007/0-306-48056-5 11
Mandelbrot, B. (1967, May). How long is the coast of Britain? statistical
self-similarity and fractional dimension. Science, 156(3775), 636–638.
doi:10.1126/science.156.3775.636

REFERENCES

221

Marinakis, Y., Migdalas, A., & Pardalos, P. M. (2005a). A hybrid geneticGRASP algorithm using Lagrangean relaxation for the traveling salesman problem. Journal of Combinatorial Optimization, 10, 311–326. doi:10.
1007/s10878-005-4921-7
Marinakis, Y., Migdalas, A., & Pardalos, P. M. (2005b, December). Expanding
neighborhood GRASP for the traveling salesman problem. Computational Optimization and Applications, 32(3), 231–257. doi:10.1007/s10589005-4798-5
Markovitch, S., & Rosenstein, D. (2002, October). Feature generation using
general constructor functions. Machine Learning, 49(1), 59–98. doi:10.
1023/A:1014046307775
McCallum, A., & Nigam, K. (1998). A comparison of event models for Naive
Bayes text classification. In Proceedings of aaai-98 workshop on learning
for text categorization (pp. 41–48). Madison, WI, USA: AAAI Press. Retrieved April 4, 2012, from http://www.aaai.org/Papers/Workshops/
1998/WS-98-05/WS98-05-007.pdf
Merrill, A. S. (1919, January). An isoperimetric problem with variable endpoints. American Journal of Mathematics, 41(1), 60–78. doi:10 . 2307 /
2370478
Miller, D. L., & Pekny, J. F. (1995). A staged primal-dual algorithm for perfect
b-matching with edge capacities. ORSA Journal on Computing, 7(3), 298–
320. doi:10.1287/ijoc.7.3.298
Milor, L., & Sangiovanni-Vincentelli, A. L. (1994, June). Minimizing production test time to detect faults in analog circuits. Computer-Aided Design
of Integrated Circuits and Systems, IEEE Transactions on, 13(6), 796–813.
doi:10.1109/43.285252

222

REFERENCES

Mısır, M., Verbeeck, K., Causmaecker, P. D., & Vanden Berghe, G. (2012).
An intelligent hyper-heuristic framework for CHeSC 2011. In Sixth
international conference on learning and intelligent optimization (lion 6).
Lecture Notes in Computer Science. Springer.
Mladenović, N., & Hansen, P. (1997). Variable neighborhood search. Computers & Operations Research, 24(11), 1097–1100. doi:10 . 1016 / S0305 0548(97)00031-2
Moscato, P., & Cotta, C. (2003). A gentle introduction to memetic algorithms.
In F. Glover & G. Kochenberger (Eds.), Handbook of metaheuristics (Vol.
57, pp. 105–144). International Series in Operations Research & Management Science. New York, NY, USA: Springer New York. doi:10.1007/0306-48056-5 5
Mühlenbein, H., Bendisch, J., & Voigt, H.-M. (1996). From recombination
of genes to the estimation of distributions II. continuous parameters.
In PPSN IV: proceedings of the 4th international conference on parallel
problem solving from nature (pp. 188–197). London, UK: Springer-Verlag.
doi:10.1007/3-540-61723-X 983
Mühlenbein, H., & Mahnig, T. (1999, December). Fda — a scalable evolutionary algorithm for the optimization of additively decomposed functions.
Evolutionary Computation, 7(4), 353–376. doi:10.1162/evco.1999.7.4.353
Mühlenbein, H., & Paaß, G. (1996). From recombination of genes to the
estimation of distributions I. binary parameters. In PPSN IV: proceedings
of the 4th international conference on parallel problem solving from nature
(pp. 178–187). London, UK: Springer-Verlag. doi:10.1007/3-540-61723X 982

REFERENCES

223

Murthy, S. K. (1998). Automatic construction of decision trees from data:
a multi-disciplinary survey. Data Mining and Knowledge Discovery, 2,
345–389. doi:10.1023/A:1009744630224
Nawaz, M., Enscore, E. E., & Ham, I. (1983). A heuristic algorithm for the
m-machine, n-job flow-shop sequencing problem. Omega, 11(1), 91–95.
doi:10.1016/0305-0483(83)90088-9
Nemhauser, G. L., & Wolsey, L. A. (1988). Integer and combinatorial optimization.
New York, NY, USA: Wiley-Interscience.
Ochoa, G., Hyde, M., Curtois, T., Vázquez-Rodrı́guez, J. A., Walker, J., Gendreau, M., . . . Burke, E. K. (2012). HyFlex: a benchmark framework for
cross-domain heuristic search. LNCS, 7245, 136–147. doi:10.1007/9783-642-29124-1 12
Ochoa, G., Vázquez-Rodrı́guez, J. A., Petrovic, S., & Burke, E. K. (2009). Dispatching rules for production scheduling: a hyper-heuristic landscape
analysis. In Proceedings of the eleventh conference on congress on evolutionary computation (cec’09) (pp. 1873–1880). Piscataway, NJ, USA: IEEE
Press. doi:10.1109/CEC.2009.4983169
Olafsson, S., & Li, X. (2010). Learning effective new single machine dispatching rules from optimal scheduling data. International Journal of
Production Economics, 128(1), 118–126. doi:10.1016/j.ijpe.2010.06.004
Palmer, D. S. (1965). Sequencing jobs through a multi-stage process in the
minimum total time – a quick method of obtaining a near optimum.
OR, 16(1), 101–107. doi:doi:10.2307/3006688
Papadimitriou, C. H. (1977). The Euclidean travelling salesman problem is
NP-complete. Theoretical Computer Science, 4(3), 237–244. doi:10.1016/
0304-3975(77)90012-3

224

REFERENCES

Papadimitriou, C. H., & Steiglitz, K. (1982). Combinatorial optimization: algorithms and complexity. Upper Saddle River, NJ, USA: Prentice-Hall,
Inc.
Papadimitriou, C. H., & Vempala, S. (2006, February). On the approximability
of the traveling salesman problem. Combinatorica, 26(1), 101–120. doi:10.
1007/s00493-006-0008-z
Pelikan, M., Goldberg, D. E., & Cantú-Paz, E. (1999). BOA: the Bayesian optimization algorithm. In W. Banzhaf, J. Daida, A. E. Eiben, M. H. Garzon,
V. Honavar, M. Jakiela & R. E. Smith (Eds.), Proceedings of the genetic and
evolutionary computation conference (gecco-99) (Vol. 1, pp. 525–532). Orlando, FL, USA: Morgan Kaufmann Publishers. Retrieved April 4, 2012,
from http://archive.org/details/BoaTheBayesianOptimizationAlgorithm
Petersen, L., Minkkinen, P., & Esbensen, K. H. (2005). Representative sampling for reliable data analysis: theory of sampling. Chemometrics and
Intelligent Laboratory Systems, 77(1-2), 261–277. doi:10.1016/j.chemolab.
2004.09.013
Pillay, N., & Banzhaf, W. (2007). A genetic programming approach to the generation of hyper-heuristics for the uncapacitated examination timetabling
problem. In Proceedings of the aritficial intelligence 13th portuguese conference on progress in artificial intelligence (epia’07) (Vol. 4874, pp. 223–
234). Lecture Notes in Computer Science. Germany: Springer-Verlag.
doi:10.1007/978-3-540-77002-2 19
Poli, R., & Graff, M. (2009). There is a free lunch for hyper-heuristics, genetic
programming and computer scientists. In Proceedings of the 12th European conference on genetic programming (eurogp ’09) (Vol. 5481, pp. 195–

REFERENCES

225

207). Lecture Notes in Computer Science. Germany: Springer. doi:10.
1007/978-3-642-01181-8 17
Puchinger, J., & Raidl, G. (2005). Combining metaheuristics and exact algorithms in combinatorial optimization: a survey and classification. In
J. Mira & J. Álvarez (Eds.), Artificial intelligence and knowledge engineering
applications: a bioinspired approach (Vol. 3562, pp. 113–124). Lecture Notes
in Computer Science. Germany: Springer. doi:10.1007/11499305 5
Punnen, A. P. (2004). The traveling salesman problem: applications, formulations and variations. In G. Gutin & A. P. Punnen (Eds.), The traveling
salesman problem and its variations (Chap. 1, Vol. 12, pp. 1–28). Combinatorial Optimization. Netherland: Springer. doi:10.1007/0-306-48213-4 1
Quinlan, J. R. (1993). C4.5: programs for machine learning. San Francisco, CA,
USA: Morgan Kaufmann Publishers Inc.
Raschke, R. L., Krishen, A. S., Kachroo, P., & Maheshwari, P. (2012). A combinatorial optimization based sample identification method for group
comparisons. Journal of Business Research. (to appear). doi:10.1016/j.
jbusres.2012.02.024
Reeves, C. R. (1995). A genetic algorithm for flowshop sequencing. Computers
& Operations Research, 22(1), 5–13. doi:10.1016/0305-0548(93)E0014-K
Reinartz, T. (2002, April). A unifying view on instance selection. Data Mining
and Knowledge Discovery, 6(2), 191–210. doi:10.1023/A:1014047731786
Reinelt, G. (1994). Candidate sets. In The traveling salesman: computational
solutions for TSP applications (Chap. 5, Vol. 840, pp. 381–412). Lecture
Notes in Computer Science. Berlin, Heidelberg: Springer-Verlag. doi:10.
1007/3-540-48661-5 5

226

REFERENCES

Rennie, J. D., Shih, L., Teevan, J., & Karger, D. R. (2003). Tackling the poor
assumptions of Naive Bayes text classifiers. In Proceedings of the twentieth
international conference on machine learning (icml-2003) (pp. 616–623).
Washington DC, USA: AAAI Press. Retrieved April 4, 2012, from http:
//aaaipress.org/Papers/ICML/2003/ICML03-081.pdf
Reza Hejazi, S., & Saghafian, S. (2005). Flowshop-scheduling problems with
makespan criterion: a review. International Journal of Production Research,
43(14), 2895–2929. doi:10.1080/0020754050056417
Rice, J. R. (1976). The algorithm selection problem. In M. Rubinoff & M. C.
Yovits (Eds.), (Vol. 15, pp. 65–118). Advances in Computers. Elsevier.
doi:10.1016/S0065-2458(08)60520-3
Richardson, L. F. (1961). The problem of contiguity: an appendix to statistic of deadly quarrels. General systems: Yearbook of the Society for the
Advancement of General Systems Theory, 6, 139–187.
Rish, I. (2001, November). An empirical study of the naive bayes classifier (tech.
rep. No. RC 22230 (W0111-014)). IBM Research Division. Yorktown
Heights, NY, USA. Retrieved April 4, 2012, from http://www.research.
ibm.com/people/r/rish/papers/RC22230.pdf
Rubinstein, R. Y. (1997). Optimization of computer simulation models with
rare events. European Journal of Operational Research, 99(1), 89–112. doi:10.
1016/S0377-2217(96)00385-2
Rubinstein, R. Y. (1999). The cross-entropy metod for combinatorial and continuous optimization. Methodology and Computing in Applied Probability,
1(2), 127–190. doi:10.1023/A:1010091220143

REFERENCES

227

Rudolph, G. (1994, January). Convergence analysis of canonical genetic algorithms. IEEE Transactions on Neural Networks, 5(1), 96–101. doi:10.1109/
72.265964
Ruiz-Vanoye, J. A., Pérez-Ortega, J., R., R. A. P., Dı́az-Parra, O., Frausto-Solı́s,
J., Huacuja, H. J. F., . . . F., J. A. M. (2011). Survey of polynomial transformations between NP-complete problems. Journal of Computational and
Applied Mathematics, 235(16), 4851–4865. doi:10.1016/j.cam.2011.02.018
Sahni, S. (1974, December). Computationally related problems. SIAM Journal
on Computing, 3(4), 262–279. doi:10.1137/0203021
Sha, D. Y., & Liu, C.-H. (2005, June). Using data mining for due date assignment in a dynamic job shop environment. The International Journal of Advanced Manufacturing Technology, 25(11), 1164–1174. doi:10.1007/s00170003-1937-y
Sima, Q. (2010). Records of the grand historian (2nd) (Y.-W. Wang, Ed.). The
Twenty-four Histories of Baina Edition. Taipei, Taiwan: Commercial
Press (Taiwan), Ltd. (Original work published 91B.C.).
Smith, D. E. (1929). A source book of mathematics (G. D. Walcott, Ed.). Source
Books in ths Histrory of the Sciences. New York, NY, USA: McGrawHill Book Company, Inc. Retrieved April 4, 2012, from http://archive.
org/details/sourcebookinmath00smit
Smith-Miles, K. A. (2008a, December). Cross-disciplinary perspectives on
meta-learning for algorithm selection. ACM Computing Surveys, 41(1),
6:1–6:25. doi:10.1145/1456650.1456656
Smith-Miles, K. A. (2008b, June). Towards insightful algorithm selection for
optimisation using meta-learning concepts. In 2008 IEEE international

228

REFERENCES
joint conference on neural networks (ijcnn 2008) (pp. 4118–4124). doi:10.
1109/IJCNN.2008.4634391

Smith, R. B. (1913). Carthage and the Carthaginians. London, UK: Longmans,
Green, and Co. Retrieved April 4, 2012, from http://www.archive.org/
details/CarthageAndTheCarthaginians
Soare, R. I. (1996, September). Computability and recursion. The Bulletin of
Symbolic Logic, 2(3), 284–321. doi:10.2307/420992
Ssu-ma, C. (1994). Sun tzu and Wu Ch’i (T.-f. Cheng, Z. Lu, W. H. Nienhauser
& R. Reynolds, Trans.). In W. H. Nienhauser (Ed.), The memoirs of prehan china (Chap. 5, Vol. 7, pp. 39–40). The Grand Scribe’s Records.
Bloomington, Indiana, USA: Indiana University Press. (Original work
published 91B.C.).
Stillwell, J. (2010). Mathematics and its history (3rd). Undergraduate Texts in
Mathematics. New York, NY, USA: Springer-Verlag. doi:10.1007/978-14419-6053-5
Stützle, T. (1998). Local search algorithms for combinatorial problems — analysis,
improvements, and new applications. (Doctoral dissertation, Department
of Computer Science, University of Essex, Colchester, UK).
Stützle, T., & Hoos, H. H. (1997, April). MAX-MIN ant system and local
search for the traveling salesman problem. In Proceedings of the IEEE
international conference on evolutionary computation (icec’97) (pp. 309–314).
Indianapolis, IN, USA. doi:10.1109/ICEC.1997.592327
Suykens, J., & Vandewalle, J. (1999). Least squares support vector machine
classifiers. Neural Processing Letters, 9, 293–300. doi:10.1023/A:1018628609742

REFERENCES

229

Taillard, E. (1990). Some efficient heuristic methods for the flow shop sequencing problem. European Journal of Operational Research, 47(1), 65–74.
doi:10.1016/0377-2217(90)90090-X
Taillard, E. (1993). Benchmarks for basic scheduling problems. European
Journal of Operational Research, 64(2), 278–285. doi:10.1016/0377-2217(93)
90182-M
Talukdar, S., Baerentzen, L., Gove, A., & De Souza, P. (1998, December).
Asynchronous teams: cooperation schemes for autonomous agents.
Journal of Heuristics, 4(4), 295–321. doi:10.1023/A:1009669824615
Tay, J. C., & Ho, N. B. (2008). Evolving dispatching rules using genetic programming for solving multi-objective flexible job-shop problems. Computers & Industrial Engineering, 54(3), 453–473. doi:10.1016/j.cie.2007.08.
008
Thompson, S. K. (2002). Sampling (2nd). Wiley Series in Probability and Statistics. New York, NY, USA: Wiley-Interscience. doi:10.1002/9781118162934
Tsutsui, S., Pelikan, M., & Goldberg, D. E. (2003, August). Using edge histogram
models to solve permutation problems with probabilistic model-building genetic algorithms (tech. rep. No. 2003022). Illinois Genetic Algorithms Lab,
Department of General Engineering, University of Illinois at UrbanaChampaign. Urbana, Illinois, USA. Retrieved April 4, 2012, from http:
//illigal.org/2003/04/20/using-edge-histogram-models-to-solvepermutation- problems- with- probabilistic- model- building- geneticalgorithms/
Turing, A. M. (1936). On computable numbers, with an application to the
Entscheidungsproblem. In Proceedings london mathematical society (Vol.

230

REFERENCES
42, 4, pp. 230–265). 2. Retrieved April 4, 2012, from http://www.cs.
virginia.edu/∼robins/Turing Paper 1936.pdf

van der Bruggen, L. J. J., Lenstra, J. K., & Schuur, P. C. (1993, August). Variabledepth search for the single-vehicle pickup and delivery problem with
time windows. Transportation Science, 27(3), 298–311. doi:10.1287/trsc.
27.3.298
Vázquez-Rodrı́guez, J. A., & Ochoa, G. (2011). On the automatic discovery of
variants of the NEH procedure for flow shop scheduling using genetic
programming. 62(2), 381–396. doi:10.1057/jors.2010.132
Vázquez-Rodrı́guez, J. A., & Petrovic, S. (2010). A new dispatching rule based
genetic algorithm for the multi-objective job shop problem. Journal of
Heuristics, 16, 771–793. doi:10.1007/s10732-009-9120-8
Vilalta, R., & Drissi, Y. (2002). A perspective view and survey of meta-learning.
Artificial Intelligence Review, 18, 77–95. doi:10.1023/A:1019956318069
Virgil. (1893). The æneid of Vergil (J. Rhoades, Trans.). London, UK: Longmans, Green, and Co. (Original work published 29-19B.C.). Retrieved
April 4, 2012, from http : / / scans . library. utoronto . ca / pdf / 1 / 5 /
aeneidbooks16tra00virguoft/aeneidbooks16tra00virguoft.pdf
Virgil. (1904). Virgil’s Aeneid (C. E. Bennett, Ed.). Bennett’s Latin series. Boston,
MA: Allyn and Bacon. (Original work published 29-19B.C.). Retrieved
April 4, 2012, from http://archive.org/details/aeneid00virggoog
Wikipedia. (2012a). How long is the coast of britain? statistical self-similarity
and fractional dimension — wikipedia, the free encyclopedia. (page version ID: 482513336). Retrieved April 4, 2012, from http://en.wikipedia.
org / w / index . php ? title = How Long Is the Coast of Britain % 3F
Statistical Self-Similarity and Fractional Dimension&oldid=482513336

REFERENCES

231

Wikipedia. (2012b). NP (complexity) — wikipedia, the free encyclopedia.
(page version ID: 484061972). Retrieved April 4, 2012, from http://
en . wikipedia . org / w / index . php ? title = NP (complexity ) &oldid =
484061972
Williamson, D. P., Hall, L. A., Hoogeveen, J. A., Hurkens, C. A. J., Lenstra, J. K.,
Sevast’janov, S. V., & Shmoys, D. B. (1997, April). Short shop schedules.
Operations Research, 45(2), 288–294. doi:10.1287/opre.45.2.288
Witten, I. H., & Frank, E. (2005). Data mining: practical machine learning tools
and techniques (2nd). Morgan Kaufmann Series in Data Management
Systems. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.
Woeginger, G. J. (2003). Combinatorial optimization — eureka, you shrink!
In M. Jünger, G. Reinelt & G. Rinaldi (Eds.), (Chap. Exact algorithms
for NP-hard problems: a survey, Vol. 2570, pp. 185–207). Lecture Notes
in Computer Science. New York, NY, USA: Springer-Verlag New York,
Inc. doi:10.1007/3-540-36478-1 17
Wolpert, D. H., & Macready, W. G. (1997, April). No free lunch theorems for
optimization. IEEE Transactions on Evolutionary Computation, 1(1), 67–82.
doi:10.1109/4235.585893
Woodhouse, R. (1964). A treatise on isoperimetrical problems and the calculus
of variations [A history of the calculus of variations in the eighteenth
century]. Bronx, NY, USA: AMS Chelsea Publishing. (Original work
published 1810).
Xue, F., Chan, C. Y., Ip, W. H., & Cheung, C. F. (2011). A learning-based
variable assignment weighting scheme for heuristic and exact searching
in Euclidean traveling salesman problems. NETNOMICS: Economic

232

REFERENCES
Research and Electronic Networking, 12, 183–207. doi:10.1007/s11066-0119064-7

Xue, F., & Fan, W. (2007). Multi-agent optimization design for multi-resource
job shop scheduling problems. In S. Huang, L. Heutte & M. Loog (Eds.),
Advanced intelligent computing theories and applications: with aspects of artificial intelligence (Vol. 4682, pp. 1193–1204). Lecture Notes in Computer
Science. Germany: Springer-Verlag. doi:10.1007/978-3-540-74205-0 123
Xu, L., Hutter, F., Hoos, H. H., & Leyton-Brown, K. (2008, June). SATzilla:
portfolio-based algorithm selection for SAT. Journal of Artificial Intelligence Research, 32(1), 565–606. Retrieved April 4, 2012, from http :
//www.aaai.org/Papers/JAIR/Vol32/JAIR-3214.pdf
Xu, L., Krzyżak, A., & Suen, C. Y. (1992, June). Methods of combining multiple classifiers and their applications to handwriting recognition. IEEE
Transactions on Systems, Man and Cybernetics, 22(3), 418–435. doi:10.1109/
21.155943
Yousef, M., Nebozhyn, M., Shatkay, H., Kanterakis, S., Showe, L. C., & Showe,
M. K. (2006). Combining multi-species genomic data for microRNA
identification using a naı̈ve bayes classifier. Bioinformatics, 22(11), 1325–
1334. doi:10.1093/bioinformatics/btl094
Zaki, M. J., Parthasarathy, S., Li, W., & Ogihara, M. (1997). Evaluation of
sampling for data mining of association rules. In Proceedings of the
7th international workshop on research issues in data engineering (ride ’97)
(pp. 42–50). Washington, DC, USA: IEEE Computer Society. doi:10 .
1109/RIDE.1997.583696

REFERENCES

233

Zhang, G. P. (2000, November). Neural networks for classification: a survey.
IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications
and Reviews, 30(4), 451–462. doi:10.1109/5326.897072
Zhang, W., & Looks, M. (2005). A novel local search algorithm for the traveling salesman problem that exploits backbones. In Proceedings of the
19th international joint conference on artificial intelligence (pp. 343–348).
IJCAI’05. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.
Retrieved April 4, 2012, from http://www.cs.wustl.edu/ ∼zhang/
publications/bgtsp.pdf
Zobolas, G. I., Tarantilis, C. D., & Ioannou, G. (2009). Minimizing makespan
in permutation flow shop scheduling problems using a hybrid metaheuristic algorithm. Computers & Operations Research, 36(4), 1249–1267.
doi:10.1016/j.cor.2008.01.007