Artificial Intelligence

Problems
AND THEIR SOLUTIONS
License, Disclaimer Of Liability, And Limited Warranty
By purchasing or using this book and disc (the “Work”), you agree that
this license grants permission to use the contents contained herein,
including the disc, but does not give you the right of ownership to any
of the textual content in the book / disc or ownership to any of the
information or products contained in it. This license does not permit
uploading of the Work onto the Internet or on a network (of any kind)
without the written consent of the Publisher. Duplication or dissemi-
nation of any text, code, simulations, images, etc. contained herein is
limited to and subject to licensing terms for the respective products,
and permission must be obtained from the Publisher or the owner of
the content, etc., in order to reproduce or network any portion of the
textual material (in any media) that is contained in the Work.

MERCURY LEARNING AND INFORMATION (“MLI” or “the Publisher”) and any-

one involved in the creation, writing, or production of the companion
disc, accompanying algorithms, code, or computer programs (“the
software”), and any accompanying Web site or software of the Work,
cannot and do not warrant the performance or results that might be
obtained by using the contents of the Work. The author, developers,
and the Publisher have used their best efforts to insure the accuracy
and functionality of the textual material and/or programs contained in
this package; we, however, make no warranty of any kind, express or
implied, regarding the performance of these contents or programs. The
Work is sold “as is” without warranty (except for defective materials
used in manufacturing the book or due to faulty workmanship).

The author, developers, and the publisher of any accompanying content,

and anyone involved in the composition, production, and manufacturing
of this work will not be liable for damages of any kind arising out of the
use of (or the inability to use) the algorithms, source code, computer
programs, or textual material contained in this publication. This includes,
but is not limited to, loss of revenue or profit, or other incidental, physi-
cal, or consequential damages arising out of the use of this Work.

The sole remedy in the event of a claim of any kind is expressly limited
to replacement of the book and/or disc, and only at the discretion of the
Publisher. The use of “implied warranty” and certain “exclusions” vary
from state to state, and might not apply to the purchaser of this product.
Artificial Intelligence
Problems
AND THEIR SOLUTIONS

Dr. Danny Kopec

Shweta Shetty
Christopher Pileggi

MERCURY LEARNING AND INFORMATION

Dulles, Virginia
Boston, Massachusetts
New Delhi
Copyright ©2014 by Mercury Learning and Information LLC. All rights reserved.

This publication, portions of it, or any accompanying software may not be reproduced in any way,
stored in a retrieval system of any type, or transmitted by any means, media, electronic display or
mechanical display, including, but not limited to, photocopy, recording, Internet postings, or scanning,
without prior permission in writing from the publisher.

Publisher: David Pallai

Mercury Learning and Information
22841 Quicksilver Drive
Dulles, VA 20166
[email protected]
www.merclearning.com
1-800-758-3756

D. Kopec, S. Shetty, & C. Pileggi. Artificial Intelligence Problems and Their Solutions.
ISBN: 978-1-938549-83-0

The publisher recognizes and respects all marks used by companies, manufacturers, and developers
as a means to distinguish their products. All brand names and product names mentioned in this book
are trademarks or service marks of their respective companies. Any omission or misuse (of any kind) of
service marks or trademarks, etc. is not an attempt to infringe on the property of others.

Library of Congress Control Number: 2013957999

141516321

Printed in the United States of America

This book is printed on acid-free paper.

Our titles are available for adoption, license, or bulk purchase by institutions, corporations, etc.
For additional information, please contact the Customer Service Dept. at 1-800-758-3756 (toll free).
Digital versions of our titles are available at: www.authorcloudware.com

The sole obligation of Mercury Learning and Information to the purchaser is to replace the
book and/or disc, based on defective materials or faulty workmanship, but not based on the operation
or functionality of the product.
 CONTENTS

Preface ..............................................................................................................xi
List of Figures and Tables ...................................................................................xiii

Chapter 1: Introduction .................................................................. 1

1.1 Goals and Purpose of This Book...............................................................1
1.2 Background and Previous Work ...............................................................3
1.3 Contributions of This Book.......................................................................5
1.4 References .................................................................................................5

Chapter 2: Problem Solving ............................................................ 7

2.1 Polya’s Five Steps for Problem Solving ....................................................7
2.2 Problem-Solving Techniques ..................................................................11
2.2.1 Heuristics .......................................................................................18
2.2.2 Additional Heuristics for Problem Solving ...................................21
2.3 The Human Window...............................................................................22
2.4 Human Window Criteria and Ranking of Solutions ..............................24
2.5 Classification ............................................................................................27
2.6 References ...............................................................................................28

Chapter 3: The Missionaries and Cannibals Problem ..................... 31

3.1 Background..............................................................................................31
3.2 Choosing an Appropriate Representation ..............................................32
3.3 Solution ................................................................................................34
3.4 Human Problem Solving.........................................................................38
3.5 Human Window Analysis of Solutions ...................................................42
3.5.1 The Most Human Window–Compatible Solution ........................43
3.5.2 The Least Human Window–Compatible Solution .......................45
vi • Artificial Intelligence Problems and Their Solutions

3.6 Best Machine Solution ...........................................................................48

3.7 Playable Program ....................................................................................49
3.8 References ...............................................................................................49

Chapter 4: The 12 Coins Problem ................................................. 51

4.1 Background..............................................................................................51
4.2 Solving a Smaller Problem ......................................................................52
4.3 Solution ................................................................................................56
4.4 Human Problem Solving.........................................................................62
4.5 Human Window Analysis of Solutions ...................................................63
4.5.1 The Most Human Window–Compatible Solution ........................64
4.5.2 The Least Human Window–Compatible Solution .......................66
4.6 Best Machine Solution ...........................................................................69
4.7 Playable Program ....................................................................................69
4.8 References ...............................................................................................70

Chapter 5: Cryptarithms ............................................................... 71

5.1 Background..............................................................................................71
5.2 Problem-Solving Techniques ..................................................................72
5.3 Solution ................................................................................................73
5.4 Human Problem Solving.........................................................................84
5.5 Human Window Analysis of Solutions ...................................................84
5.5.1 The Most Human Window–Compatible Solution ........................85
5.5.2 The Least Human Window–Compatible Solution .......................87
5.6 Best Machine Solution ...........................................................................89
5.7 Playable Program ....................................................................................90
5.8 References ...............................................................................................91

Chapter 6: The Red Donkey Puzzle ............................................... 93

6.1 Background..............................................................................................93
6.2 Solution ................................................................................................94
6.2.1 Bidirectional Search .......................................................................98
6.3 Human Problem Solving.........................................................................99
6.4 Human Window Analysis of Solutions .................................................100
6.4.1 The Most Human Window–Compatible Solution ......................100
6.4.2 The Least Human Window–Compatible Solution .....................104
6.5 Best Machine Solution .........................................................................106
6.6 Playable Program ..................................................................................107
6.7 References .............................................................................................107
 Contents • vii

Chapter 7: The 15 Puzzle ............................................................ 109

7.1 Background............................................................................................109
7.2 Problem-Solving Techniques ................................................................110
7.3 Solution ..............................................................................................111
7.3.1 Solving the First Row...................................................................111
7.3.2 The Corner Technique ................................................................116
7.3.3 Solving the Second Row ..............................................................118
7.3.4 Solving the Third Row .................................................................119
7.3.5 Solving the Fourth Row ...............................................................122
7.4 Human Window Analysis of Solutions .................................................124
7.4.1 The Most Human Window–Compatible Solution ......................126
7.4.2 The Least Human Window–Compatible Solution .....................126
7.5 Best Machine Solution .........................................................................128
7.6 Playable Program ..................................................................................130
7.7 References .............................................................................................130

Chapter 8: The Knight’s Tour Problem ........................................ 133

8.1 Background............................................................................................133
8.2 Problem-Solving Techniques ................................................................134
8.3 Solution ..............................................................................................135
8.4 Human Problem Solving.......................................................................142
8.5 Human Window Analysis of Solutions .................................................142
8.5.1 The Most Human Window–Compatible Solution ......................144
8.5.2 The Least Human Window–Compatible Solution .....................145
8.6 Best Machine Solution .........................................................................148
8.7 Playable Program ..................................................................................149
8.8 References .............................................................................................150

Chapter 9: Mastermind ............................................................... 151

9.1 Background............................................................................................151
9.2 Problem-Solving Techniques ................................................................152
9.3 Solution ..............................................................................................153
9.3.1 Example 1 .....................................................................................154
9.3.2 Example 2 .....................................................................................156
9.3.3 Example 3 .....................................................................................159
9.3.4 Example 4 .....................................................................................161
9.4 Human Problem Solving.......................................................................164
9.5 Human Window Analysis of Solutions .................................................164
9.6 Best Machine Solution .........................................................................165
viii • Artificial Intelligence Problems and Their Solutions

9.7 Playable Program ..................................................................................166

9.8 References .............................................................................................166

Chapter 10: The Monty Hall Problem ......................................... 169

10.1 Background..........................................................................................169
10.2 Problem-Solving Techniques ..............................................................170
10.3 Solution ..............................................................................................172
10.4 Human Problem Solving.....................................................................172
10.5 References ...........................................................................................173

Chapter 11: Rubik’s Cube ........................................................... 175

11.1 Background..........................................................................................175
11.2 Problem-Solving Techniques ..............................................................176
11.3 Solution ..............................................................................................177
11.3.1 Cube Components .....................................................................177
11.3.2 Subgoal 1: The Top Cross ..........................................................182
11.3.3 Subgoal 2: The Top Corners ......................................................184
11.3.4 Subgoal 3: The Middle Layer ....................................................185
11.3.5 Subgoal 4: The New Top Cross .................................................187
11.3.6 Subgoal 5: The New Top’s Corners ...........................................189
11.3.7 Subgoal 6: Correct Corner Positioning .....................................190
11.3.8 Subgoal 7: The Top Edges .........................................................192
11.4 Human Problem Solving.....................................................................194
11.5 Human Window Analysis of Solutions ...............................................196
11.5.1 The Most Human Window–Compatible Solution ....................197
11.5.2 The Least Human Window–Compatible Solution ...................198
11.6 Best Machine Solution .......................................................................201
11.7 Playable Program ................................................................................203
11.8 References ...........................................................................................204

Chapter 12: The Prisoner’s Dilemma .......................................... 207

12.1 The Traditional Problem .....................................................................207
12.2 The Iterated Prisoner’s Dilemma .......................................................209
12.3 Applications in Diverse Areas .............................................................210
12.4 References ...........................................................................................212

Chapter 13: Miscellaneous Problems .......................................... 215

13.1 Cards/Coins in the Dark .....................................................................215
13.1.1 Cards in the Dark .......................................................................215
 Contents • ix

13.1.2 Coins Version of the Problem ....................................................218

13.1.3 References ..................................................................................218
13.2 Ten Pirates and Their Gold.................................................................218
13.3 Halmos’s Handshake Problem ............................................................222
13.4 Random Airline Seats Problem ..........................................................228
13.5 The Birthday Problem ........................................................................230

Chapter 14: Conclusion: Toward a Theory for Problem Solving ... 233
14.1 Human Window Study ........................................................................233
14.2 Lessons Learned .................................................................................233
14.2.1 The Missionaries and Cannibals Problem.................................235
14.2.2 The 12 Coins Problem ...............................................................235
14.2.3 Cryptarithms ..............................................................................236
14.2.4 The Red Donkey Puzzle ............................................................237
14.2.5 The 15 Puzzle .............................................................................238
14.2.6 The Knight’s Tour Problem .......................................................238
14.2.7 Mastermind ................................................................................238
14.2.8 The Monty Hall Problem...........................................................239
14.2.9 Rubik’s Cube ..............................................................................239
14.2.10 The Prisoner’s Dilemma ..........................................................239
14.2.11 Miscellaneous Problems ..........................................................239
14.3 Retrospective, Conclusions, and Future Work ..................................240
14.4 Supplemental References ...................................................................241
14.5 Composite Problem Playability Sites..................................................243

On The CD-ROM ....................................................................... 245

Appendix A—Student Solutions
Appendix B—12 Coins Algorithm
Appendix C—Human Window Study
Appendix D—Program Solutions to Problems

Index .............................................................................................................247
 PREFACE
We live in times that are transitional and fast-paced. Just about every-
one seems to own at least one electronic device, such as iPhones, smart
phones, notebooks, and so on, that they consider vital to their existence.
Patience is short, and everyone wants quick results (e.g., Google) without
having to do a lot of difficult research. Students come to classes unabash-
edly deploying these devices, and instructors don’t know if they’re text mes-
saging/social networking, or addressing the topic at hand. Attention spans
seem to be very short. We are consistently distracted by the small things that
need to, or can, be done. In the end, our lives seem to be made up of small,
choppy episodes as we go about conducting our daily tasks. The notion
that one sits still at a desk uninterrupted for many hours seems foreign to
our society today. Persistence, trial and error, search, inquisitiveness, and
simply “thinking” appear to be activities of the past. And it is clear that
there will be no turning back from this technological age and our depen-
dencies on technology. Yet, over and over again, we hear that our children’s
test scores are declining.
One perspective on intelligence measures the ability to solve problems.
Another considers how an individual or a society is able to survive under
adverse conditions. So the natural question is: what will happen to our
society and our children if someone “turns off the power”?
Therefore, a book about solving problems that uses nontrivial and some-
times difficult problems is especially timely, because it addresses the needs
of a number of possible constituent audiences. This book assembles in one
place a set of interesting and challenging artificial intelligence (AI)–type
problems that students regularly encounter in computer science, mathe-
matics, and AI courses. These problems are not new, and students from
xii • Artificial Intelligence Problems and Their Solutions

all backgrounds can benefit from the kind of deductive thinking that goes
into solving them. Another constituent audience is the computer science,
mathematics, or AI instructor who is looking for answers to these problems.
This book explores the kinds of thinking, methods, and applications used
in computer science and AI to solve these problems. Furthermore, a solid
understanding of how amenable a solution is to humans (hence the notion
of the “Human Window”) adds more depth to how the problems and their
solutions can and should be presented. This book is also helpful to those
who would like to improve their problem-solving skills. No doubt today’s
young people will soon discover that they are lacking in these basic skills,
and this book will be one resource they can turn to for help practicing and
enhancing their problem-solving abilities. Furthermore, there will be ben-
efits from developing and maintaining life-long learning skills.
It must be emphasized that this book is not about solving problems
quickly. Skillful problem solving—whether learning to play chess, becom-
ing a competent violinist, or reading this book—requires time. So get ready
to enjoy, learn, benefit, and grow!
Danny Kopec
Merrick, NY
March, 2014
 LIST OF FIGURES AND TABLES

List of Figures

Figure 1: Human Window Solutions ................................................................22

Figure 2: The Missionaries and Cannibals Problem
[Graphic was in Color in AI / New Millennium] .............................31
Figure 3: Search Tree Diagram ........................................................................35
Figure 4: State Transition Diagram ..................................................................36
Figure 5: Problem State Space Matrix .............................................................36
Figure 6: Solution Using Problem State Space Matrix ....................................37
Figure 7: State Transition in Problem State Space ..........................................38
Figure 8: Instructor’s Diagram for The Missionaries
and Cannibals Problem ....................................................................39
Figure 9: Student 1’s Partial Solution ...............................................................39
Figure 10: Student 1’s Solution.........................................................................40
Figure 11: Student 2’s Solution.........................................................................40
Figure 12: Student 3’s Solution.........................................................................41
Figure 13: The “Symmetric Solution” ..............................................................46
Figure 14: The “Constraint Rules Solution” ....................................................48
Figure 15: The 12 Coins Problem ....................................................................51
Figure 16: Graphical Depiction of Some of the n Coins Problems ................59
Figure 17: The “Generic Table Solution”.........................................................65
Figure 18: The “Logically Enumerated Solution” ...........................................68
Figure 19: The 12 Coins Problem ....................................................................69
Figure 20: The Send More Money Problem....................................................71
Figure 21: Step 1 ...............................................................................................74
Figure 22: Step 2 ...............................................................................................75
Figure 23: Step 3 ...............................................................................................75
Figure 24: Step 4 ...............................................................................................76
Figure 25: Step 5 ...............................................................................................77
Figure 26: Step 1 ...............................................................................................78
Figure 27: Step 2 ...............................................................................................79
Figure 28: Step 3 ...............................................................................................79
Figure 29: Step 4 ...............................................................................................80
Figure 30: Step 5 ...............................................................................................81
Figure 31: Step 6 ...............................................................................................81
Figure 32: The “Donald Solution” (Shortened) ...............................................88
xiv • Artificial Intelligence Problems and Their Solutions

Figure 33: Alphabetic App ................................................................................90

Figure 34: The Red Donkey Puzzle .................................................................93
Figure 35: The Red Donkey Puzzle .................................................................94
Figure 36: Step 9 ...............................................................................................96
Figure 37: Step 14 .............................................................................................96
Figure 38: Step 21 .............................................................................................96
Figure 39: Step 26 .............................................................................................96
Figure 40: Step 31 .............................................................................................96
Figure 41: Step 41 .............................................................................................96
Figure 42: Step 41 .............................................................................................97
Figure 43: Step 48 .............................................................................................97
Figure 44: Step 52 .............................................................................................97
Figure 45: Step 60 .............................................................................................97
Figure 46: Step 67 .............................................................................................97
Figure 47: Step 71 .............................................................................................97
Figure 48: Step 76 .............................................................................................98
Figure 49: Step 81 .............................................................................................98
Figure 50: The Initial State...............................................................................98
Figure 51: The Goal State.................................................................................98
Figure 52: Subgoal 1 While Searching Backward from the Goal State ..........99
Figure 53: The “Donkey Solution” (Shortened) ............................................103
Figure 54: The “Hollow Solution” (Shortened) .............................................106
Figure 55: The 15 Puzzle ................................................................................109
Figure 56: Initial State ....................................................................................112
Figure 57: Move the Empty Space to the 1 Tile ............................................112
Figure 58: Move the 1 Tile to the Upper Left Corner ..................................113
Figure 59: The 4 Tile Positioned ....................................................................114
Figure 60: The Center “Square”.....................................................................114
Figure 61: Positioning the 2 Tile ....................................................................115
Figure 62: 3 Tile Preferred Position ...............................................................115
Figure 63: Positioning the 3 Tile ....................................................................116
Figure 64: 3 Tile Special Case ........................................................................116
Figure 65: Positioning the 3 Tile (Special Case) ............................................117
Figure 66: 1 and 2 Tiles Over the Corner ......................................................117
Figure 67: Reposition the 1 and 2 Tiles .........................................................118
Figure 68: Row 2 Solved .................................................................................118
Figure 69: Position the 9 and 12 Tiles ............................................................119
Figure 70: Acceptable Center “Square” .........................................................119
Figure 71: 10 Tile Positioned Badly ...............................................................120
 List of Figures and Tables • xv

Figure 72: Moving the 9 Tile ..........................................................................120

Figure 73: Rotate the Center “Square” ..........................................................121
Figure 74: Obtaining the Sufficient Center “Square” ...................................121
Figure 75: Third Row Complete ....................................................................122
Figure 76: Incorrect Last Row .......................................................................122
Figure 77: Shifted 9 and 10 Tiles....................................................................123
Figure 78: Shifting the 11 and 12 Tiles ..........................................................123
Figure 79: Rotate the Center “Square” Tiles .................................................123
Figure 80: Solving the Last Row ....................................................................124
Figure 81: The “Optimal Solution” ................................................................127
Figure 82: The Knight’s Tour Problem ..........................................................133
Figure 83: Initial State (Visiting Square 1).....................................................136
Figure 84: Visiting Squares 2–4 ......................................................................136
Figure 85: Visiting Squares 5 and 6 ................................................................136
Figure 86: Upper Right Corner (Square 7) ...................................................137
Figure 87: Visiting Squares 8–11 ....................................................................137
Figure 88: Break the Rules—First Time (Visiting Squares 12–14) ..............137
Figure 89: Visiting Squares 15 and 16 ............................................................138
Figure 90: Visiting Squares 17–19 ..................................................................138
Figure 91: Visiting Squares 20–22 ..................................................................138
Figure 92: Visiting Squares 23–25 ..................................................................138
Figure 93: Lower Right Corner (Square 26) .................................................138
Figure 94: Visiting Squares 27–29 ..................................................................139
Figure 95: Visiting Squares 30 and 31 ............................................................139
Figure 96: Upper Left Corner (Square 32)....................................................139
Figure 97: Visiting Squares 33–36 ..................................................................139
Figure 98: Breaking the Rules—Second Time
(Visiting Squares 37 and 38) .........................................................140
Figure 99: Visiting Squares 39–41 ..................................................................140
Figure 100: Visiting Squares 42–44 ................................................................140
Figure 101: Visiting Squares 45–47 ................................................................140
Figure 102: Visiting Squares 48–50 ................................................................140
Figure 103: Visiting Center Squares 51–54 ...................................................141
Figure 104: Visiting Center Squares 55–58 ...................................................141
Figure 105: Visiting Center Squares 59–61 ...................................................141
Figure 106: Goal State (Visiting Squares 62–64) ...........................................141
Figure 107: The “Semi-Magic Solution” (Shortened) ..................................147
Figure 108: Mastermind .................................................................................151
Figure 109: Sample Guess ..............................................................................153
xvi • Artificial Intelligence Problems and Their Solutions

Figure 110: Ex. 1—Guess 1 ............................................................................154

Figure 111: Ex. 1—Guess 2 ............................................................................154
Figure 112: Ex. 1—Guess 3 ............................................................................155
Figure 113: Ex. 1—Guess 4 ............................................................................155
Figure 114: Ex. 1—Alternate Guess 4............................................................156
Figure 115: Ex. 2—Guess 1 ............................................................................156
Figure 116: Ex. 2—Guess 2 ............................................................................157
Figure 117: Ex. 2—Guess 3 ............................................................................157
Figure 118: Ex. 2—Guess 4 ............................................................................158
Figure 119: Ex. 2—Guess 5 ............................................................................159
Figure 120: Ex. 3—Guess 1 ............................................................................159
Figure 121: Ex. 3—Guess 2 ............................................................................159
Figure 122: Ex. 3—Guess 3 ............................................................................160
Figure 123: Ex. 3—Guess 4 ............................................................................161
Figure 124: Ex. 4—Guess 1 ............................................................................162
Figure 125: Ex. 4—Guess 2 ............................................................................162
Figure 126: Ex. 4—Guess 3 ............................................................................162
Figure 127: Ex. 4—Guess 4 ............................................................................163
Figure 128: Ex. 4—Guess 5 ............................................................................163
Figure 129: The Monty Hall Problem ...........................................................169
Figure 130: Rubik’s Cube ...............................................................................175
Figure 131: Corner Pieces ..............................................................................178
Figure 132: Edge Pieces .................................................................................178
Figure 133: Center Pieces ..............................................................................179
Figure 134: Cube Faces ..................................................................................179
Figure 135: Top Layer.....................................................................................180
Figure 136: Bottom Layer ..............................................................................180
Figure 137: Left Layer ....................................................................................180
Figure 138: Right Layer ..................................................................................181
Figure 139: Front Layer .................................................................................181
Figure 140: Back Layer ...................................................................................181
Figure 141: Subgoal 1 .....................................................................................182
Figure 142: Inverted Edge .............................................................................183
Figure 143: Subgoal 2 .....................................................................................184
Figure 144: State 1 ..........................................................................................185
Figure 145: State 2 ..........................................................................................185
Figure 146: State 3 ..........................................................................................185
Figure 147: Subgoal 3 .....................................................................................186
Figure 148: Pre-Algorithm State ....................................................................186
 List of Figures and Tables • xvii

Figure 149: Incorrect Orientation ..................................................................187

Figure 150: Subgoal 4 .....................................................................................188
Figure 151: State 1 ..........................................................................................188
Figure 152: State 2 ..........................................................................................188
Figure 153: State 3 ..........................................................................................188
Figure 154: State 4 ..........................................................................................188
Figure 155: Alternate State .............................................................................189
Figure 156: Subgoal 5 .....................................................................................189
Figure 157: State 1 ..........................................................................................190
Figure 158: State 2 ..........................................................................................190
Figure 159: State 3 ..........................................................................................190
Figure 160: Subgoal 6 .....................................................................................191
Figure 161: State 1 (Adjacent) ........................................................................191
Figure 162: State 2 (Diagonal) .......................................................................191
Figure 163: Subgoal 7 (Goal State) ................................................................192
Figure 164: Orange-Blue Edge to the Right..................................................193
Figure 165: Orange-Green Edge to the Right ...............................................193
Figure 166: The “Short Solution” ...................................................................199
Figure 167: The Prisoner’s Dilemma (Figure used in AI in
New Millennium / Shutterstock #) .............................................207
Figure 168: Payoff Matrix for The Prisoner’s Dilemma ................................208
Figure 169: Cards Laid Out on a Table ..........................................................215
Figure 170: The Deck of Cards Laid Out on a Table ....................................216
Figure 171: The Two Piles of Cards ...............................................................217
Figure 172: After Flipping All the Cards in Pile 1.........................................217
Figure 173: Pirate and Gold ...........................................................................218
Figure 174: Ten Pirates and Their Gold ........................................................219
Figure 175: Two Pirates ..................................................................................220
Figure 176: Three Pirates ...............................................................................220
Figure 177: Four Pirates .................................................................................221
Figure 178: Halmos’s Handshake Problem ....................................................222
Figure 179: Ten People and Handshakes.......................................................223
Figure 180: How Many Shook his Wife’s Hand? ...........................................224
Figure 181: Person A—8 Shakes ....................................................................224
Figure 182: Person A’s Wife—0 Shakes .........................................................225
Figure 183: Person B—7 Shakes ....................................................................225
Figure 184: Person B’s Wife—1 Shake...........................................................225
Figure 185: Person C—6 Shakes ....................................................................226
Figure 186: Person C’s Wife—2 Shakes .........................................................226
xviii • Artificial Intelligence Problems and Their Solutions

Figure 187: Person D—5 Shakes ...................................................................226

Figure 188: Person D’s Wife—3 Shakes ........................................................226
Figure 189: Halmos—4 Shakes ......................................................................227
Figure 190: Halmos’s Wife—4 Shakes ...........................................................227
Figure 191: Airline Seats.................................................................................228
Figure 192: The Birthday Problem ................................................................230

List of Tables
Table 1: Classification of Problems ...................................................................28
Table 2: Ranking The Missionaries and Cannibals Problem Solutions
According to the Human Window......................................................44
Table 3: Solution to the 3 Coins Problem ........................................................53
Table 4: Solution to the 6 Coins Problem ........................................................55
Table 5: Solution to The 12 Coins Problem .....................................................57
Table 6: Number of Weighings Required to Find an Odd Coin
Among n Coins in Different Cases .....................................................59
Table 7: Configuration of Coins in the Two Weighings ...................................60
Table 8: Possible Outcomes After the Two Weighings ....................................61
Table 9: Ranking The 12 Coins Problem Solutions According to
the Human Window ............................................................................64
Table 10: Ranking Cryptarithmetic Solutions According to the Human
Window..............................................................................................86
Table 11: Ranking The Red Donkey Puzzle Solutions According to the
Human Window ..............................................................................101
Table 12: Ranking The 15 Puzzle Solutions According to the
Human Window ..............................................................................125
Table 13: Ranking The Knight’s Tour Problem Solutions According to the
Human Window ..............................................................................143
Table 14: The Monty Hall Problem................................................................171
Table 15: Ranking Rubik’s Cube Solutions According to the
Human Window ..............................................................................195
Table 16: Example 1 ........................................................................................211
Table 17: Example 2 ........................................................................................211
Table 18: Problem-Solving Strategies ............................................................234
 CHAPTER

1
INTRODUCTION

T
hese days, students have an abundance of technologies available to
help them solve problems —be it in school or in life. Unfortunately,
most students use technology primarily for entertainment and social
networking. As far as using technology for educational purposes is con-
cerned, applications, for the most part, are limited to finding quick answers
or solutions to assignments. In fact, information about almost every subject
is so readily available and accessible that during the past two decades, evi-
dence of students’ efforts to solve problems that require critical thinking
has become somewhat rare.
Problem solving, issue analysis, and decision making (once considered
standard for college graduates) are increasingly becoming the most sought-
after skills in the employment market [1]. Thus, developing these skills will
be helpful not only in education but in employment and other areas of
life as well. However, people are often overwhelmed by complex problems
and tend to get lost in them [2]. A systematic way to teach students how to
understand a problem, using the appropriate techniques according to the
specific problem types, is clearly needed.

1.1 GOALS AND PURPOSE OF THIS BOOK

The goal of this book is to collect, classify, and study human and machine
solutions to some classic problems that have frequently been posed in
mathematics, computer science, and artificial intelligence (AI). Overall, we
classify and view these problems as AI problems because of their particular
2 • Artificial Intelligence Problems and Their Solutions

characteristics: they are nontrivial, at least for humans; they involve logic;
they may involve search and memory; solutions to these problems involve
certain techniques, such as deduction, inference, solving subproblems, and
pattern recognition, which are applicable and extendable to other prob-
lems; they are by and large well-known and have survived the test of time;
some of these problems have been used as a testbed for AI problem-solving
techniques—for example, The 15 Puzzle, Mastermind, The Knight’s Tour
Problem, The Red Donkey Puzzle, and Cryptarithms; and they are ame-
nable for studying human problem-solving activity. The long-term, higher
goal of our work is the development of a theory for problem solving. The
problems we have chosen for investigation include the following:

1. The Missionaries and Cannibals Problem: In this problem, a group

of three missionaries and three cannibals are to be transported safely
from the East bank of a river to the West bank, using a boat that can
carry only two passengers at a time. The condition is that in any state
the cannibals can never outnumber the missionaries. The goal is to get
everyone to the other side of the river in the fewest number of boat
trips.
2. The 12 Coins Problem: In this problem, there are 12 coins, and we
must determine the 1 coin that weighs differently from the other 11
with only three weighings, using a balance scale.
3. Cryptarithms: Some examples are SEND + MORE = MONEY or
DONALD + GERALD = ROBERT. The goal of these problems is to
find the unique numbers (from 0 to 9) representing the letters of the
alphabet that satisfy the equations.
4. The Red Donkey Puzzle: This is a sliding block puzzle in which a
2” × 2” square block must be moved from the top of a 4” × 5” board to
the bottom center. The other pieces in the puzzle include four 1” × 2”
vertical blocks, four 1” × 1” square blocks, and one 2” × 1” horizontal
block.
5. The 15 Puzzle: This is another sliding block puzzle in which 15 unit
square blocks (numbered 1 to 15) must be rearranged in ascending
order in a clockwise direction on a 4” × 4” grid.
6. The Knight’s Tour Problem: This puzzle involves finding a path on
a chessboard such that a chess knight visits every square on the board
exactly once. A more strict variation of the problem requires the knight to
 Introduction • 3

return to the square of origin on its last move. It is common for the knight
to start on the square in the lower left corner of the board (known as a1).
7. Mastermind: This is a two-player code-breaking board game where
one of the players (called the code-breaker) has to decode or guess the
code created by the other player (called the code-maker) in at most
10 guesses. This code is represented by a sequence of colored pegs
arranged in a particular order.
8. The Monty Hall Problem: This problem involves guessing which
door out of three has a prize (e.g., a car) behind it. The dilemma lies in
whether to stick with your first choice or switch doors after the game
host tells you which of the three doors does not have the prize behind it.
9. Rubik’s Cube: This popular sliding block puzzle is a three-dimensional
cube composed of small cubes with different colors on each surface.
The goal is to arrange these small cubes so the cube overall has only one
color on each surface.
10. The Prisoner’s Dilemma: Here we consider a set of problems arising
from classical game theory that involves an optimal scheme for how pris-
oners can get a minimal sentence, depending on whether they cooper-
ate or defect in their private confessions.
11. Miscellaneous Problems: We have also chosen five smaller problems
which are in the same genre as the ten problems presented in this book,
but on a smaller scale. The five problems primarily involve developing a
strong graphical abstraction and understanding of the problem, devising
schema, using probability, and logic.
This book focuses on solutions to these classic problems using specific
problem-solving strategies such as search, problem reduction, deduction,
and mathematical formulas. It also examines how people typically try to
solve problems and presents ideas for how to analyze problems and break
them down to make them more manageable and easier to solve.

1.2 BACKGROUND AND PREVIOUS WORK

Human problem solving is an important topic that has been widely

studied and is the basis for both artificial intelligence and its “cousin” from
psychology: cognitive science. From our perspective, before we develop
4 • Artificial Intelligence Problems and Their Solutions

and use machines to solve problems for us, we must first have a thorough
understanding of how humans think and solve problems.
Much of the inspiration for our work comes from George Polya’s book
How to Solve It. This little book proved groundbreaking and survived the
test of time in helping problem solvers to become more cognizant of their
activities and how they can be more successful. (See Chapter 2 for more
about Polya’s work.)
Allen Newell and Herbert Simon have contributed significantly to
the topic of problem solving and their research is summarized in the
book Human Problem Solving. According to Newell, Shaw, and Simon,
a problem has a set of paths, only one (or several) of which lead(s) to
a desired goal. Finding that correct path(s) is solving the problem. A
problem is also defined as the difference between an existing state and
a desired state [3]. Newell and Simon give several examples of problem
solving, such as solving a maze or crossword puzzle, finding the com-
bination of a safe, and translating one language into another [4]. Their
research includes topics such as representation of an external environ-
ment in an internal form, strategies such as the creation of subgoals and
subproblems to reach a solution, working backward from a goal, and
using heuristics to find the best path to a solution. They also stress the
process of relying on simple abstractions of a problem space as a way to
approach more complex problems.
Cerveny, Garrity, and Sanders discuss two types of problem solving.
One is linear problem solving, which involves performing a sequence of
steps to reach a goal, and the other is concurrent problem solving, which is
iterative in nature and involves comparing the current state with an ideal
state and performing the steps required to reach that ideal state [5].
Gunzelmann and Anderson state that the two important initial steps
to solving a problem are understanding the problem and representing the
problem in an internal form. They also stress the importance of planning
in problem solving. Humans tend to increase the degree of planning if it is
rewarded with faster solutions [6].
Finding the path leading to a goal is sometimes not enough. Due to
constraints such as time, computing power, and other resources, solving a
problem efficiently with optimum resource utilization is necessary. Accord-
ing to Anzai and Simon, several strategies are available for solving a prob-
lem, some more efficient than others. The efficiency of a strategy depends
 Introduction • 5

on various factors such as speed, load on memory, and ease of retention of

information in memory, among others [7].

1.3 CONTRIBUTIONS OF THIS BOOK

1. Classification of the problems under consideration.

2. Discussion of general problem-solving techniques as applicable to the
problems under consideration.
3. Application of appropriate problem-solving strategies to the selected
problems as required by the respective problem domains.
4. Study of people’s problem-solving behaviors in a problem-solving set-
ting.
5. Recommendations of appropriate strategies and approaches in proper
order according to problem type.
6. Exposition of Donald Michie’s concept of the “Human Window” for
solving AI-type problems [8].

1.4 REFERENCES

1. Ruggiero, V.R. (1998). The Art of Thinking. Boston, MA: Addison-Wesley.

2. Coleman, P.T. (2011). The Five Percent: Finding Solutions to Seemingly
Impossible Conflicts. New York: Public Affairs, Perseus.
3. Newell, A., Shaw, J.C., and Simon, H.A. (1958). The Process of Creative
Thinking. Pittsburgh, PA: Carnegie-Mellon University, The Rand Corporation.
4. Newell, A., and Simon, H.A. (1972). Human Problem Solving. Englewood
Cliffs, NJ: Prentice-Hall.
5. Cerveny R.P., Garrity, E.J., and Sanders, G.L. (1990). “A Problem-Solving
Perspective on Systems Development.” Journal of Management Information
Systems 6 (4): 103–122. Published by: M.E. Sharpe, Inc. Stable.
6 • Artificial Intelligence Problems and Their Solutions

Available at http://www.jstor.org/stable/40398777. Accessed on December 23.

Source: For the Learning of Mathematics (1) 2: 35–42. Published by: FLM
Publishing Association. Stable
Available at http://www.jstor.org/stable/40247714. Accessed on December 23,
2012.
6. Gunzelmann, G., and Anderson, J.R. (2003). Problem Solving: Increased
Planning with Practice. Cognitive Systems Research 4 (1): 57–76.
7. Anzai, Y., and Simon, H.A. (1979). The Theory of Learning by Doing.
Psychological Review, 86 (2): 124-140.
8. Michie, D. (1982). Experiments on the Mechanization of Game-Learning: 2—
Rule-Based Learning and the Human Window. The Computer Journal 25 (1):
105–113.
 CHAPTER

2
PROBLEM SOLVING

P
roblem solving is an important skill that helps us in all aspects of life.
Therefore, it is essential that we follow a systematic approach in solv-
ing problems. Polya’s work [1] was groundbreaking because it led to
the development of a five-step sequence for problem solving, which essen-
tially can be universally applied, independent of domain.

2.1 POLYA’S FIVE STEPS FOR PROBLEM SOLVING

1. Understand the problem.

2. Develop an algorithm (plan).
3. Implement the algorithm (coding in computer science).
4. Test the algorithm (debugging).
5. Revise the algorithm (possibly returning to Steps 1 to 4).
The wonderful thing about Polya’s five steps [1] is that they can be
applied with only slight revision to mathematical problems, computer pro-
gramming, engineering, management science, business, and more!

1. Understand the Problem

What Polya means by this is don’t even bother trying to solve a prob-
lem before you have spent some time acclimating yourself to it. Learn the
essence of the problem. Try to classify the problem. What kind of problem
is it? Can the problem be broken down into parts? What is a good represen-
tation for the problem? Is the problem related to other problems that you
have encountered or solved?
8 • Artificial Intelligence Problems and Their Solutions

Even as we write this section we have run into a PROBLEM.

The laptop keeps dying. Certain computers have been known to
have problems with their power supply. So we wonder, what could
it be? The power supply (one of the two power supplies available
is not the right one for the machine)? So we try both power sup-
plies. Nothing changes. It became fairly clear that the power supply
was not the problem. Every few minutes the laptop dies, and that
does not portend well for successfully completing this book by our
deadline. One solution we tried was to simply stop work early and
let the machine “cool off” overnight. It seemed to be better in the
morning, but after about an hour of operation, it died again.
What other unique symptoms are there to the problem?
We have noted that in the past few weeks the laptop seems to
feel pretty hot (on the top left, above the keyboard) where there is
a fan vent.
Postmortem to the Problem: A week or so later, the computer’s
motherboard and fan were replaced under warranty. The machine
is now working fine.

2. Develop an Algorithm
Clearly, problem solving without thinking is not very effective. For
example, in chess, we say it is better to play with a plan, even a bad plan,
than no plan at all. We work by the following definition for an algorithm:
“A well-defined, possibly repeating, finite series of steps to
solve a problem.”

Having noted the problem described in the box above, and cou-
pled with a previous experience (also described above), and recalling
that we had seen a youngster once solve a problem with his computer
by simply opening it up and replacing a fan, an idea emerges:
There is a very small (5-inch fan) available that is sometimes
used during hot summer days for quick local cooling. Perhaps that
fan could be used to cool the laptop.
 Problem Solving • 9

3. Implement the Algorithm

This is when you must take action. All the time we have dozens of
ideas and plans in our heads. Plans without implementation become
something like “little bad memories” or reminders that we should keep
lists, notes, and so on, and not just leave everything in our head. Still,
you cannot do everything at once, and people are not machines. We need
some kind of order in our lives and in our activities—some directive—
and thus a PLAN that enables us to assess the effectiveness of a proposed
problem solution.

The little 5-inch fan has been blowing for about 45 minutes, and
the laptop is still working fine. The area on the top left is still hot,
but not as hot as it was before (just using the sense of touch). Does
this mean that the problem is solved?

4. Test the Algorithm

Testing is a very important step in problem solving. All of us want
to be able to say, “The problem has been solved!” How much of that
decision is based on desire, emotion, ego, or just the desire to move on?
For people who are goal oriented, ambitious, conscientious, obsessive,
or just plain determined, making the decision whether or not to accept
a solution can be overwhelming. Hence we have to answer the following
questions:

Does the solution work? Does it work fully? Does it work par-
tially? Now how confident are you that you have solved the problem?
How much more work is necessary to be certain that the problem
has been solved? More than an hour and a half later, the fan solution
seems to be working.

5. Revise the Algorithm

Based on the outcomes above, can we make a decision as to how to
proceed? Do we return to Stage 1, in which we reconsider our under-
10 • Artificial Intelligence Problems and Their Solutions

standing of the problem? Or do we consider alternative solutions/plans/

algorithms (Stage 2)? Did we correctly implement the planned solution
or algorithm?

Alongside the “fan” solution, alternate solutions have been con-

sidered, such as using another laptop or backing up all the key files
in question. One may resort to these alternative plans, depending
on the result of the current algorithm/testing.

A problem has provided some given data, some unknowns, some condi-
tions, an initial state, and a goal state. The aim is to use the given data to
find the unknowns while moving from the initial to the goal state. During
the initial stages of problem solving, the solver’s goals are to understand the
problem and represent it in an internal form [2].
One of the key steps in facilitating problem understanding is to rep-
resent the problem in a format that clearly presents the various problem
states (stages) of a proposed solution and the transitions between them.
This often reduces the complexity of the problem. Some problems may best
be expressed in a graphical way, such as a diagram or a graphical sketch,
whereas others may be better represented in a search tree format.
An important aspect of knowledge representation is the level of
detail. A knowledge representation can be extensional or intensional. Exten-
sional representations show every step, case, or example, whereas inten-
sional representations are short and implicit—for example, a formula [3].
We need to choose the representation that best represents the problem and
helps us understand and solve it easily.
According to Polya, an important step in understanding a problem is
isolation of the principal parts—that is, the data, the conditions, and the
unknowns in a problem. These are examined individually, in relation to one
another and in relation to the whole problem [1]. Each of these approaches
helps us understand the problem better.
Once we understand the problem, we go on to planning to solve it. It is
a way to manage the complexity of the problem. It is observed that if plan-
ning leads to faster solutions, the degree of planning increases [4].
 Problem Solving • 11

One of the ways to deal with a complex problem is to solve a simpler

version of the problem. The solutions and methods used for the simpler
problem can then be used to plan for the harder problems [5].
Solving a problem involves searching through a large number of pos-
sibilities consisting of a set of paths leading to a goal state. We may also
create subgoals and solve them to reach the final goal [6]. Algorithms and
heuristics are two of the basic ways to solve a problem.
Algorithms in computer science will typically be computationally inten-
sive and involve an exhaustive search in the state space. Therefore, they
will always lead to the goal state and are 100% correct [7]. The heuristic
approach to problem solving is more suitable for humans with their limita-
tions in terms of memory capacity and processing speed.
A heuristic is a rule of thumb that allows one to select one or a few of
the possible paths through the state space that are more likely to lead to the
goal state. Though heuristics usually lead to solutions with relatively less
processing [4], this does not always guarantee a solution and sometimes
may even lead us in the wrong direction [7].

2.2 Problem-Solving Techniques

We will now examine some problem-solving techniques. Whole books

can be written about many of these techniques, but our purpose here is
simply to introduce and distinguish these techniques from one another in
the hope that the problem solver will find them useful.
■ Induction
In mathematics, induction is often used as a soundproof technique.
Unfortunately, when used in the real world, it is not necessarily sound.
Example 1: I find a black raven. I find 100 black ravens. Every raven
I find is black. Therefore, the next raven I find will be black. (This is not
necessarily true.)
Example 2: Every time a team gets a 3 to 0 lead in a playoff series in
baseball, it wins. This has happened 39 times in history, but in 2004, the
Yankees had a 3 to 0 lead against the Boston Red Sox—and they lost!
12 • Artificial Intelligence Problems and Their Solutions

■ Deduction
Deduction involves the attempt to reach a conclusion from some facts
that logically follow one another. Often it may involve general information
from which we try to reach a specific conclusion [8].*
Example 1: The famous modus ponens is a common example:
All men are mortal. Socrates was a man. Therefore, Socrates was mortal.
Example 2: There was a burglary at 365 85th Street, Apt. 4B, between
4:00 and 6:00 pm. John Stealer was the only person seen (by security cam-
era) to enter and leave that apartment between those hours. Thus, it is logi-
cal to assume (by deduction) that John Stealer was the burglar.
Example 3: In medicine, we use deductions to determine what disease
is afflicting a patient. For example, if a patient has a fever, headache, and
sore throat, it would probably be determined that she has the flu because
she has “flulike” symptoms.
■ Abduction
Abduction is an attempt to make a deduction that is not sound. This is
discussed in John Sowa’s book Conceptual Structures. He gives this example:
Example 1: All Quakers are pacifists. Richard Nixon was a Quaker.
Modus ponens would suggest that Nixon was a pacifist. (This is not neces-
sarily true.)
Example 2: Chocolate is the food of love. I am in love. Hence, the
chocolate I ate made me fall in love.
Example 3: Smoking causes cancer. Tom has cancer. Hence, Tom was
a smoker.
■ Empirical
The term empirical is often used with the terms data and evidence—
for example, “empirical data suggest” or “empirical evidence shows.” This
refers to the data or evidence that is available or has been observed.

* An excellent book on deduction is Human Reasoning: The Psychology of Deduction, by.

Ruth M.J. Byrne, Jonathan B.T. St. Evans, Stephen E. Newstead. For example, they point
out that in the case of modus ponens, the inference about Socrates being mortal is rela-
tively easy. If people see the series 1,3, they would likely conclude the next number is 5.
However, if told the next number is not 5 such a conclusion is wrong. This is called “the
contrapositive” or modus tollens. e.g., Given p => q and not(q), means not(p)
 Problem Solving • 13

Example 1: Empirical evidence suggests that smoking causes cancer.

Example 2: Empirical data suggests that distracted drivers cause a large
proportion of motor vehicle accidents. This would suggests that text mes-
saging while driving—something that distracts drivers—should be illegal
(modus ponens).
■ Exhaustive Enumeration
This is an attempt to solve a problem by considering every possibility in
a search space. It is certainly not efficient.
Example 1: In a class, you could find out who an unsigned exam paper
belonged to by exhaustive enumeration—for example, by checking all the
exam papers that did have signatures.
Example 2: The possibilities for “win,” “draw,” and “loss” in tic-tac-toe
can fairly easily be determined by exhaustive enumeration.
Example 3: All the possibilities for playing chess could theoretically
be identified by exhaustive enumeration, but since there are 1042 estimated
reasonable games, it is not possible in practice.
■ Contradiction
Contradiction is a problem-solving technique often used to prove some-
thing is true or is impossible. It is based on trying to demonstrate that the
opposite of an argument, cannot coexist with the argument itself.
Example 1: Given some premise like A > B, then A ≤ B cannot be true
at the same time.
Example 2: If a person is suspected of committing a crime but has an
alibi (e.g., proof that she was out of the country at the time of the crime),
then that would be proof by contradiction.
■ Recursive
Recursive problems or solutions involve describing a problem (or func-
tion) in terms of itself. Recursive problem-solving techniques are subtle and
elegant. They involve applying the same steps over and over but to different
“arguments.” For these reasons, recursive programs are much shorter than
their iterative counterparts.
Example 1: Fact(N) = N * Fact(N – 1); Fact(1) = 1
Here, the function Factorial is defined in terms of itself but applied to
a different argument.
14 • Artificial Intelligence Problems and Their Solutions

Example 2: To solve the Towers of Hanoi problem with seven discs,

one need only solve the problem for six discs. To solve the Towers of Hanoi,
for six discs all that needs to be done is to solve the problem for five discs.
The process repeats itself until the exit condition is met. This is recursion
at work. It is a kind of divide-and-conquer technique. Remember, however,
that there must be an “exit” condition—for example, when there is only one
disc left (the biggest disc), it must be moved to the goal peg.
■ Divide and Conquer
Divide and conquer is a problem-solving technique based on the notion
of breaking up a problem into smaller parts. Few problems can be solved
“at once,” so it is better if they are broken up into logical components.
Example 1: One cannot just sit down and write an entire book, but
one can begin with an outline, a few words, a phrase, a complete sentence,
and finally a paragraph. Almost all problem solving involves some form of
divide and conquer.
Example 2: Driving 20 hours straight to Miami is virtually impossible.
But breaking the trip up into four segments of 5 hours each is much more
manageable.
■ Solving a Subproblem
Similar to divide and conquer, this involves identifying specific compo-
nents of a problem and solving them. The components may be independent
of one another (hence no dependency), but each component is a part of the
same whole problem.
Example 1: If we need to identify all the factors of a number, finding
first the odd factors and then the even factors is a way of solving the inde-
pendent subproblems of the problem.
Example 2: If a solution can lie in any of four quadrants, solving one
quadrant is solving a subproblem.
■ Solving a Subgoal
This is distinct from solving a subproblem because to solve the main
problem, you must be able to solve smaller problems that are parts of the
larger problem.
Example 1: A building cannot be constructed without preparing its
foundation.
 Problem Solving • 15

Example 2: In the case of The 12 Coins Problem, you cannot solve

12 coins until you can solve 4 coins.
Example 3: In the case of The Red Donkey Puzzle, you cannot get the
donkey piece to its goal until it can pass the horizontal, which obstructs it.
■ Problem Reduction
This involves turning a problem into a simpler problem or simpler form.
Example 1: We use algebra to simplify an equation.
Example 2: In a cryptarithm, if we solve for the leftmost characters, we
will have fewer possibilities for the other characters.
■ Solving a Related Problem
When we are “in tune” with a problem, we are able to identify and solve
related problems.
■ Using Prior Knowledge
When we guessed that the problem with the laptop was that it was get-
ting too hot (it continues to work “cooled down”), that guess was based on
prior knowledge and experience. As it is said, “experience counts.” Expe-
rience should not be underrated. The more you can employ your knowl-
edge and experience to support your problem solving, the shorter and more
effective the process is likely to be.
Example 1: While making a decision as to what route to take to a cer-
tain destination, a certain road is avoided because it is known there is often
heavy traffic or roadwork.
Example 2: Knowing the hypotenuse of a triangle helps solve the
problem of finding the area of a parallelogram.
Example 3: Having taken a professor’s course before, you are more
familiar with his or her teaching and examination question styles.
■ Generate and Test
This problem-solving method literally entails how it is named. Solu-
tions are produced according to the constraints on a problem. A com-
mon example is The Eight Queens Problem (see Lucci and Kopec,
2013, p. 72), where queens are “placed” on a chessboard (the “genera-
tion” component) and then “tested” as to whether they meet the problem
constraints.
16 • Artificial Intelligence Problems and Their Solutions

Example 1: In The Eight Queens Problem, a new queen is placed in a

new row and column at each “generation.”
Example 2: A man notes that he left his car near the elevator on the
fifth floor of the parking garage. This information helps him generate
potential solutions to the problem of finding his car in the parking garage
when he returns.
■ Imposing a Structure: Hierarchy
Being able to structure a problem is always better than having to
tackle the problem cold without any prior knowledge. Structure can be
conducive to identifying subgoals and subproblems that can and should
be solved.
Example 1: By having a corporate structure, it is clear who is respon-
sible for what.
Example 2: By developing a problem tree, it is clear how to proceed
from one component of a problem to another.
Example 3: A map helps to identify the points on a trip with the goal
of reaching a destination.
■ Solving an Analogous Problem (Analogy)
Problem solving is thinking. Thinking should involve looking for pat-
terns and relationships among problems. For example, if we realize that
taking a certain approach to a situation has worked before, we are more
likely to use the same (or a similar) approach again in the future.
In his biography, the great Chess World Champion Mikhail Botvinnik
(1948–1963) related the story of how the Russian government would only
support his computer chess research (with powerful computers) when he
could solve some power station problems (he was an electrical engineer),
using more powerful computers.
Example 1: Outsourcing for help. If outsourcing has worked before in
a problem situation (e.g., you need a book cover designed), you will likely
outsource again.
Example 2: Eating pizza before going to bed causes acid reflux. Hence,
you should avoid eating anything with tomato sauce before going to bed
(i.e., tomato sauce contributes to acid reflux).
 Problem Solving • 17

■ Come Up With a “Bright Idea”

This is akin to the expression “Think outside the box.” Do not think in a
conventional way. Sometimes this can be aided by getting away from the
problem—akin to “incubation” below. Bright ideas are not cheap, and it
is hard to figure out how an individual comes up with a bright idea. For
some people, their brightest ideas occur in the shower. The key is being
relaxed so you can think freely without pressure. The idea you come up
with must in some way support your problem-solving goal.
Example 1: Your car is getting old, and you would like to buy a new one.
It makes more sense to buy a used car that is about two to five years old and
has low mileage. Doing this saves money but still produces a viable solution.
Example 2: Often in life, we find that we make snap judgments that
do not work out. While in certain situations (see below—intuition), it is
good to go with your instincts, it is sometimes better to hold off making any
decision at all (i.e., “incubation” or “sleep on it”). Or as we sometimes say in
chess instruction, sometimes the best move is not to make any move at all.
This may aid in “coming up with a bright idea.”
■ Intuition
Intuition can be a very powerful problem-solving method. It is not acci-
dental, and it is certainly formed with the aid of prior experience. Intu-
ition may come from instincts, which may come from having confronted
analogous problems or situations. Intuition must be distinguished from
unfounded guessing. There is usually a basis for intuition that may be hard
to identify or express.
Example 1: “My intuition is that it would be a good idea for you not to
deal with. . . .” This suggestion is usually based on facts, wisdom, and prior
experience.
Example 2: “Try substituting 1 for L in the right side of the equation. . . .”
Advice like this does not come from lack of knowledge but more often from
prior knowledge or experience.
■ Consider Boundary Conditions
In computer science, it is important to solve problems for all values
of X. Therefore, testing for cases when X is either very large or very small is
important for the purpose of having a general and complete solution.
18 • Artificial Intelligence Problems and Their Solutions

Example 1: Consider when X = 0.

Example 2: Consider when X is very large.
■ Working Backward (Retrograde Analysis): Backtracking
Sometimes it is important to be able to reconstruct the steps that led to
a problem situation (e.g., in the case of having lost something or in accident
analysis). Lucci and Kopec (2013, p. 235) discuss this topic.
Example 1: What chess piece fell off that square?
Example 2: What were the causes of a car accident?
■ Incubation
This is a highly recommended approach to problem solving. It essen-
tially suggests “sleeping on a problem.” Sometimes getting away from a
problem is the best thing you can do. Recognize when you have worked
long enough on a problem, and then just get away from it—for a while.
While you are mentally “away” from the problem, your subconscious may
still be working on it.
Example 1: If you have worked a few hours on a problem and are still
stuck, it may be most effective to give the problem a rest, have a cup of
coffee (maybe even take a few hours off), and come back to it later or even
the next day.
■ Decomposition (Equivalence Classes): Cases
Occasionally, it is possible to solve a problem by breaking it down into
distinct cases. If this can be done, a large problem is broken into smaller
problems that are more manageable. This is akin to solving subproblems.
Example 1: A large manufacturing problem is broken down into stages
that are ordered and can be achieved.

2.2.1 Heuristics

One of the most important ideas that has come out of Polya’s work is
the notion of heuristics or “rules of thumb.” In effect, heuristics are a very
powerful problem-solving technique that are used universally. That’s why
we think it deserves its own section. Judea Pearl considered this topic so
important that he wrote an entire book about it [10]. When one consid-
ers how we go through our daily lives, it is apparent that we are indeed a
 Problem Solving • 19

heuristically controlled species. Heuristics also play an important role in the

decision-making process and in the lives of other species.
Heuristics may be viewed as “attributes,” “features,” “factors,” “charac-
teristics,” or even “hunches” that contribute to the decision-making process
in problem solving, but they are not simply idiosyncrasies or superstitions.
Let us consider some examples:
Example 1: In baseball, when a player at bat is physically small and
light in stature, the outfield will tend to play in, although he may have con-
siderable power. The outfielders are applying a natural heuristic that usu-
ally works—but sometimes it simply doesn’t apply.
Example 2: The heuristic for driving on highways says if traffic is heavy
and things are moving very slowly, you should stay in the right lane. This
usually helps. Also, if you are talking on the phone while driving (hands-
free), you should stay in the right lane.
Example 3: Before teaching a class, it is a good idea to get to your
office at least two hours in advance to check for any materials you may need
and review them.
Example 4: Avoid rush hour! Try to avoid driving into New York City
between 7:30 am and 9:00 am, and avoid driving out of New York
City between 4:00 pm and 6:00 pm.
Example 5: Brook’s Rule: Whatever the amount of time a programmer
tells you is necessary to complete a job, multiply it by three. This is because
people are not good estimators of the time and effort needed to complete a
job—what Brooks calls “gutless estimating.”
Example 6: When developing a computer program, consider the gen-
eral case data as well as the boundary cases—for example, when N = 0 and
N is a very large number.
Example 7: For The Knight’s Tour Problem, when given a choice of
moves, always move to the edge. This heuristic will work most of the time.
Example 8: In baseball, it is easier for the outfielder coming in to catch
a fly ball than for the infielder going out. Hence, the infielder should usu-
ally defer to the outfielder.
Example 9: There is the heuristic “Be aggressive but not stupid.”
Those five words represent a heuristic that means a lot and can cover
many domains. For example, in football, it would mean aggression without
20 • Artificial Intelligence Problems and Their Solutions

committing penalties. In chess, it means playing actively with your pieces

but not giving your pieces away. In baseball, it could mean trying to steal
bases without getting thrown out.
The same concept could be generalized to many situations in life. There
are times when you must be firm or “aggressive,” but not so aggressive that
your action becomes “stupid.” That would apply to social situations or even
academic ones. Let us say you are listening to a lecture in class and disagree
strongly with the professor. If you can make your point in a strong, articu-
late, intellectual manner, without being abusive or overly loud, and you have
valid supporting evidence for your view, then you are more likely to be taken
seriously than someone who is just very aggressive, outspoken, and impolite.
Example 10: If you have something troubling you and it is late at night,
it is often better to “sleep on it” than try to resolve it when you are tired.
This is a “common sense” heuristic. It is a fact that when you are tired, you
are apt to make bad decisions. Sometimes you may have done or said things
when tired that you have regretted for a long time. After a good night’s
sleep, your frame of mind is usually more organized, optimistic, and ready
to take on the challenges of a new day.
Example 11: “Don’t do today what you can put off to tomorrow.” What
a bad heuristic! It’s pretty much saying, “Put off until tomorrow what you
need to do today.” That, however, is obviously a bad idea and a familiar one
to all of those who are certified procrastinators.
The concept could easily be an extension of Example 10, but we do not
advocate it, nor is it what we are trying to say here. Do not confuse work
with “a problem.” If there is work that you are already doing, it is not a bad
idea to do as much of that work as you can while your mind is working well
and you are not overly tired. In other words, get done what you can in one
sitting. There is always some overhead involved in leaving a project. The
longer you leave a project, the more time you have to spend trying to regain
your final status, ideas, and plan. Yes, if you can accomplish small parts of
a project in some systematic, regular work schedule, that is excellent (incu-
bation may even be occurring), but it is not better than actually getting
something done. Nonetheless, it is always a good idea to review work that
you have done—whether under time pressure or when you were tired. The
Institute of Medicine’s landmark report To Err Is Human attributes many
human errors to sheer tiredness. One must be able to recognize tiredness
and the need to stop working when necessary.
 Problem Solving • 21

2.2.2 Additional Heuristics for Problem Solving

In this section, we present some additional heuristics that have been

effectively employed by problem solvers.

1. Look for a simpler, related problem. This will no doubt lead to a bet-
ter understanding of the problem itself.
2. Work backward from specific to general. That is, take the facts
to try to deduce what the nature of the problem is. In AI, this is called
backward chaining.
3. Word forward from general to specific. In essence, we are trying
to use generalizations to reach a specific conclusion. For example,
say we encounter one bad student in a school. Then we encounter
several bad students in the school. Hence, all the students we have
encountered are bad. So when we encounter a new student in that
school, we will probably assume that he or she is a bad student. Of
course, this reasoning is unsound, and it can lead to uncalled-for
prejudices.
4. Narrow the condition. See if the problem can be understood better
and solved if the conditions are more restrictive.
5. Widen the conditions to see if that will make the problem easier
to understand and solve. Methods #4 and # 5 may help in problem
comprehension.
6. See if there is a counterexample that will disprove your hypoth-
esis for a solution.
7. Change the conceptual mode. This is akin to incubation. See if you
can look at the problem from a different perspective. This may include
restating the problem, drawing diagrams, or choosing alternative repre-
sentations.
8. Try to approximate a proposed answer. It is always a good idea to
determine if you are in the “right ballpark.”
9. Consider if all the data have been used. It is usually a good idea to
see how different data affects a solution.
10. Look for patterns. Patterns can lend insight to problem
solutions.
22 • Artificial Intelligence Problems and Their Solutions

2.3 THE HUMAN WINDOW

The solutions to problems must fall inside the Human Window

(Figure 1). The Human Window is a region constrained by human mem-
ory capabilities and computational limitations [3]. Thus, solutions must
be executable within the limitations of human memory and must be com-
prehensible. Furthermore, the solutions should be 100% correct and be
of manageable grain size (i.e., the solutions must be neither too memory
intensive nor too computation intensive) [3].

Too memory intensive Too computation intensive

The Human Window

FIGURE 1 Human Window Solutions.

The Human Window illustrates the limitations of the human brain’s

ability to process information and the need for artificial intelligence solu-
tions to fall within its boundaries. The late Donald Michie is credited with
this idea. For a solution to fall within the boundaries of the Human Win-
dow, that solution must be ideal in terms of, correctness, grain size, execut-
ability, and comprehensibility [8].
The correctness of a solution is determining whether or not it is 100%
correct. A solution to a complex problem that is incorrect will not fit well
within the Human Window.
The grain size of a solution refers to the computational and mem-
ory constraints of humans and the Human Window. The key idea is that
the solutions to problems of sufficient complexity (i.e., AI-type problems)
are limited in terms of the amount of detail with respect to computation
and memory needed by people to execute the solutions and comprehend
them. The solutions to complex problems should have a manageable grain
size—that is, not too computationally intensive (large grain size) or memory
intensive (i.e., a database, small grain size).
The executability of a solution is measured by how easily one can find
the solution, and the comprehensibility of a solution is measured by how
 Problem Solving • 23

well one can understand it. Although these criteria assess how well a solu-
tion fits within the Human Window, executability and comprehensibility
are not necessarily dependent on each other. For example, some solutions
to problems can be executed but not necessarily understood. Other types
of solutions may be easy for people to understand, but they may be unable
to execute them.
Along with the above criteria, it is also important to take into consider-
ation the level of detail of a solution to a complex problem in terms of its
knowledge representation and determine whether or not it is extensional
(explicit, detailed, and long) or intensional (implicit, short, and compact).
Extensional representations will usually show every case, every example,
of some information, while intensional representations will often be short
(e.g., a formula or an expression that represents some information).
This book presents solutions to the chosen problems that include sec-
tions on the Human Window for each. In this section, various human solu-
tions will be compared based on a set of Human Window criteria.
Kopec has addressed several of the selected problems in his courses and
has accumulated a number of solutions and resources for their study and
solution. For each of the above complex problems, we have found and ana-
lyzed solutions that have already been developed. Among those solutions,
we have ranked each of them based on our assessment of how they meet the
above criteria (see Section 2). Using these justifications and our opinion, we
have evaluated solutions that may be considered the “Most Human Win-
dow Compatible (MHWC)” and the “Least Human Window Compatible
(LHWC)” for each complex problem. We have also attempted to find the
solutions to the problems that can be deemed the “best” for a machine by
evaluating the runtime and space efficiency of specific algorithms. Finally,
Christopher Pileggi has also attempted to determine the best way to repre-
sent these solutions.
We look at the problem solutions from these points of view:
1. Intensional versus extensional solutions
2. Choice of knowledge representation
3. Suitability in terms of meeting the constraints of the Human Window
(i.e., correctness, grain size, executability, and comprehensibility)
4. Problem-solving method
5. Flexibility
24 • Artificial Intelligence Problems and Their Solutions

6. Mode of conveyance
7. Possible optimality
Finding the most and least Human Window–compatible solutions for
complex problems based on these factors can be beneficial in a number of
ways. For example, finding the best and most suitable solution for humans
can be very beneficial to the study and understanding of problem solving.
The same can be said for finding the optimal solution for a machine prob-
lem. Not only will these solutions apply to the problems they are meant for,
but they may also provide a gateway to solving other human and computa-
tional problems. They can also be used as effective teaching tools.

2.4 HUMAN WINDOW CRITERIA AND RANKING OF

SOLUTIONS

For each complex problem in the book, we ranked the solutions under
consideration based on the following criteria:
Intensional versus Extensional: The solution to a problem will fall
under the category of being either intensional or extensional. An exten-
sional solution is one that can be explicit, detailed, and usually very long,
while an intensional solution is one that is usually implicit, short, and com-
pact. We indicate whether a solution is extensional or intensional.
Intensionality/Extensionality: Although determining whether a
solution is intensional or extensional is imperative, that doesn’t make the
category black and white. An intensional solution can be measured on
how intensional it is, and an extensional solution can be measured on how
extensional it is. For example, consider two extensional solutions to The
Missionaries and Cannibals Problem: a graph with images and a table with
directions. The graph with images can be considered as more extensional
than the table. As a matter of fact, we can say a solution is even more exten-
sional if it were, for example, to involve watching the missionaries and can-
nibals crossing the riverbank in person. We will give each solution a rank
between 1 and 10 to rate how intensional or extensional a solution is, with
10 being the highest level of intensionality or extensionality.
Choice of Knowledge Representation: When analyzing and ranking
solutions to identify which is considered “most suitable,” it is important to
 Problem Solving • 25

take note of how each solution is represented. By analyzing the rankings

and taking note of how a problem is represented, a conclusion can be drawn
as to which representation (e.g., table, tree, pseudocode, etc.) may be con-
sidered ideal for humans. For each solution, we will indicate what choice of
representation was used and analyze its affects and attributes.
Meeting the Constraints of the Human Window: This is important
as to whether or not a solution falls within the Human Window. That is,
does it meet each of the following constraints? For each solution, we will
answer this question with a “yes” or “no.” However, to fully answer this
question, we must first analyze each of the following criteria:
• Correctness: To understand a solution, the solution must be correct.
Otherwise, the solution is essentially meaningless. For each solution,
we indicate if it is correct, incorrect, or correct but with cycles. That
is, the solution provides the correct answer but with some extra steps
involved (i.e., backtracking).
• Grain Size: The tradeoff between the amount of memory and
computation needed to carry out a solution is essential for a solution
to fall within the constraints of the Human Window. Too much or
too little of either may make a solution so unclear that it is either not
comprehensible (too small a grain size) or not executable (too large
a grain size). For each solution, we rank this aspect of it in terms of
its grain size. Very Small means the level of memory needed to carry
out the solution is too high, while Very Large means the level of
computation needed to carry out the solution is too high. Ideal means
there is a balance between how much a person needs to memorize
and calculate to be able to understand (comprehend) and carry out
(execute) a solution.
• Executability: The executability of a solution is the measurement
of how well a person can actually carry it out. The number of steps
needed, the resources required (i.e., tables), and the background
required (i.e., mathematics) are only some of the factors that
determine whether or not a person can execute a particular solution.
Each solution is ranked on its executability from 1 to 10, 10 meaning
it is very executable.
• Comprehensibility: Although a solution may be executable,
that does not necessary mean a person can comprehend it. One’s
26 • Artificial Intelligence Problems and Their Solutions

comprehensibility of a solution is how well he or she can understand

it. This can be determined by factors such as images, colors, the size
of the solution, the organization of data, and so on. Since we are
not experimenting with human subjects for the Human Window
analysis, the ratings are based on our experience with people and
their attitudes toward complex problems. Our basis will range
between adept problem solvers and non–problem solvers, and we
will determine a possible average between the two. Each solution is
ranked from 1 to 10, 10 meaning it is very comprehensible.
Problem-Solving Method: Just as we must take note of a solution’s
choice of representation, it is also important to note the problem-solving
method(s) used. Although there are a wide variety of methods to choose
from, the average person may prefer one method over another. For exam-
ple, one may find it significantly easier to solve a problem by breaking it up
into subgoals instead of using a top-down approach. With that, comparing
a solution’s ratings to the method(s) used to solve it may determine which
method or methods a person prefers. For each solution, the method(s) used
to carry it out is indicated.
Flexibility of Solutions: It is possible that a particular solution can
be represented differently. The flexibility of a solution is how easily it
can be represented in some other way, without changing the problem-solv-
ing method used or detracting from its suitability in terms of meeting the
constraints of the Human Window. Solutions can be changed, however, to
make them more suitable, especially in terms of their comprehensibility.
For example, a black-and-white image may be easier to comprehend if it
were in color. Each solution is ranked on its flexibility from 1 to 10, with 10
meaning it is very flexible.
Mode of Conveyance: Another important attribute of a solution is
how “portable” it is. For example, consider two fully comprehensible solu-
tions—one that can be presented by hand and one that can only be carried
out through an image maker on a computer. The one that can be hand
drawn is more portable because it can be used anywhere to solve a prob-
lem. On the other hand, this may affect a solution’s ability to meet the con-
straints of the Human Window. Each solution is rated on the possible ways
it can be conveyed.
Possible Optimality: A property that makes a solution commend-
able overall is its optimality. That is, the solution is carried out in the least
 Problem Solving • 27

number of possible steps. Not only does this reduce the problem size, but it
also tells us what the best possible solution may entail. We indicate whether
or not a solution is optimal.
Along with these rankings, each of these 10 problems is ranked on two
additional factors:
Difficulty Level: This measures how difficult we perceive each prob-
lem to be based on our study of its solutions. The rating is based on how
difficult we find the problems after attempting to solve them and how dif-
ficult we feel others would find them. Each solution is ranked from 1 to 10,
with 10 meaning it is very difficult.
Complexity: Not only is it important to measure how easy or hard it
is for a person to understand a solution, but it also important to measure
how easy or hard it is for a machine. Each problem’s complexity of the best
machine solution is specified in terms of Big-Oh notation.

2.5 CLASSIFICATION

Table 1 classifies the problems according to the type of problem-solving

method used. Each of the problems is discussed in terms of background
historically, choice of suitable representation, problem-solving techniques
used, its solution(s), human problem-solving approaches, Human Window
analysis, and the best machine solutions.

No. Problem Problem Type

The Missionaries and Can-
1 Logic, constraint satisfaction, search
nibals Problem
2 The 12 Coins Problem Logic, optimizing information
Combinatorics, constraint satisfaction, solving
3 Cryptarithmetic
subproblems
Sliding block puzzle, search, constraint
4 The Red Donkey Puzzle
satisfaction, solving subproblems
Sliding block puzzle, search, heuristics,
5 The 15 Puzzle
constraint satisfaction
6 The Knight’s Tour Problem Search, heuristics
28 • Artificial Intelligence Problems and Their Solutions

7 Mastermind Logic, optimizing information

8 The Monty Hall Problem Logic, probability

9 Rubik’s Cube Subgoals, patterns, subproblems

Game theory, Nash Equilibrium, Pareto
10 The Prisoner’s Dilemma
Optimality, schema
11 a. Cards/Coins in the Dark Probability, logic
b. Ten Pirates and Their
Gold Schema, logic
c. Halmos’s Handshake
Problem Permutations, logic
d. Airline Seats Problem Probability, logic
e. The Birthday Problem Probability, logic

TABLE 1 Classification of Problems.

2.6 REFERENCES

1. Polya, G. (1945). How to Solve It, 2nd ed. Princeton, NJ: Princeton University
Press.
2. Burton, L., and Burton, M. (1980). “Problems and Puzzles.” For the Learning
of Mathematics (1) 2: 20–23. Published by: FLM Publishing Association. Stable
http://www.jstor.org/stable/40247710. Accessed on December 23, 2012.
3. Lucci, S., and Kopec, D. (2013). Artificial Intelligence in the 21st Century.
Dulles, Virginia: Mercury Learning, Inc.
4. Gunzelmann, G., and Anderson, J.R. (2002). Problem Solving: Increased Plan-
ning with Practice. Cognitive Systems Research 4 (2003) 57–76.
5. Minsky, M. (1960). Steps Toward Artificial Intelligence. In Computers and
Thought, (Eds. E. Feigenbaum and J. Feldman), McGraw-Hill, New York,
pp. 406–450.
6. Newell, A., and Simon, H.A. (1972). Human Problem Solving. Englewood
Cliffs, NJ: Prentice-Hall.
7. Unterrainer, J.M., and Owen, A.M. (2006). “Planning and Problem Solving:
From Neuropsychology to Functional Neuroimaging.” Journal of Physiology—
Paris 99: 308–317.
 Problem Solving • 29

8. Jonathan St. Evans, B.T., Newstead, S.E., and Byrne, Ruth M.J. (1993). Human
Reasoning: The Psychology of Deduction, Hove, England: Lawrence Erlbaum
Associates.
9. Kopec, D. (1983). Human and Machine Representations of Knowledge. (PhD
thesis) Edinburgh, Scotland: University of Edinburgh.
10. Pearl, J. (1984). Heuristics: Intelligent Search Strategies for Computer Problem
Solving. Reading, MA: Addison-Wesley.
 CHAPTER

3
THE MISSIONARIES AND
CANNIBALS PROBLEM

FIGURE 2 The Missionaries and Cannibals Problem.

3.1 BACKGROUND

The Missionaries and Cannibals Problem (Figure 2) is a river-

crossing puzzle and is a variant of the Jealous Husbands Problem (another
variant is The Brothers and Sisters Problem), which first appeared in the
ninth-century manuscript originally entitled Propositiones ad Acuendos
Juvenes (Problems to Sharpen the Young) [1]. River-crossing problems usu-
ally involve transporting objects from one riverbank to the other, using a
boat that can carry a certain number of objects. The goal is to transport the
32 • Artificial Intelligence Problems and Their Solutions

objects optimally (in the least number of trips) while satisfying some con-
straints.
Around the thirteenth century, The Jealous Husbands Problem became
popular throughout northern Europe. The problem involved husbands and
wives as the entities, with the restriction being that the husbands were not
allowed to outnumber the wives on either bank [2]. By the end of the nine-
teenth century, the problem had a revival and became known as The Mis-
sionaries and Cannibals Problem. The rules were exactly the same as The
Jealous Husbands Problem, but wives were replaced with missionaries and
the husbands were replaced with cannibals [3].
Today, The Missionaries and Cannibals Problem is widely used in the
field of artificial intelligence. It is considered an important problem for
several reasons. One is because it is an example of a toy problem, which is a
problem that seems insufficient and not worth studying or solving but helps
explain a more general problem-solving technique. Another is because of
its use in the study of problem representation. It is, as computer scien-
tist Saul Amarel pointed out, a prime example of the fact that “the choice
of appropriate representations is capable of having spectacular effects on
problem-solving efficiency” [4].
On the whole, The Missionaries and Cannibals Problem is a relatively
simple problem. Any solution is optimal if it can move all six missionaries
and cannibals to the opposite bank in 11 steps while adhering to the con-
straints of the problem. From analyzing the problem, it can be concluded
that there are four ways to achieve the goal state in 11 steps.
The problem is stated as follows: Given three missionaries and three
cannibals on the West bank of a river, find a way to safely transport every-
one to the other side of the river. The constraints on the problem are:

1. The boat can only carry two persons at a time.

2. At any point in time (state), the cannibals cannot outnumber the
missionaries.

3.2 CHOOSING AN APPROPRIATE REPRESENTATION

This problem is quite easy to understand and can be tackled by

adults as well as children. However, choosing the correct sequence of
transformations from the starting state to the goal state, while adhering to
 The Missionaries and Cannibals Problem • 33

the constraints of the problem, can be challenging. Experience has dem-

onstrated that the choice of representation is very closely related to suc-
cessfully solving the problem. The representation not only has to clearly
depict the various states in the problem but the moves and transitions from
one state to another as well. The following figure shows one way to repre-
sent the initial state:

MMM
CCC

In this figure, the three boxes represent the West bank, the river, and
the East bank, respectively. M stands for missionary, and C stands for
cannibal.
Since the boat is initially stationary, it is not shown. When the boat is in
motion, it can be represented by an arrow that shows in which direction the
boat is moving. Thus, the following figure represents two cannibals being
transported from the West bank of the river to the East bank:

MMM CC
C

The transition can be shown as follows:

MMM
CCC

MMM CC
C

The problem state can be represented in other ways. For example,

the initial state can be represented as a set of three numbers, with the
first number representing the number of missionaries, the second num-
ber the number of cannibals, and the third number the presence or
absence of a boat. However, this would only represent the state on one
side of the river.
34 • Artificial Intelligence Problems and Their Solutions

For example, 331 represents the initial state on the West bank with
three missionaries, three cannibals, and a boat. Similarly, the state at the
West bank after transporting two cannibals to the other side can be repre-
sented as 310.
The transition can be shown as:
331 310
It is clear that the first representation is more extensional when com-
pared to the second one. One could use a verbal description or use a formal
notation [6] to describe the states, the moves, and the transition between
the moves, but it is quite apparent that the graphical notation is more trans-
parent and facilitates visualization of states of the problem.
From the initial state, it may seem there are many possible moves.
However, when one delves deeper and considers the constraints of the
problem, it can be seen that two entities must be in the boat to get the
boat back to the West bank. In addition, to avoid the constraint of cannibals
outnumbering the missionaries, either two cannibals or one missionary and
one cannibal can be transported on the first move. Further moves that lead
to safe states are shown in Figure 3.

3.3 SOLUTION

One of the ways to represent this problem and its various moves is in
the search tree format, as shown in Figure 3. A minimum of eleven moves
are required to reach the goal state.
Yet another way of representing the solution is a state transition dia-
gram (Figure 4). The boxes on the left side depict all the safe states on the
West bank, and the boxes on the right side depict the corresponding states
on the East bank. Only states that are safe for both sides of the river are
shown. The arrows represent the transition from one state to another as
well as the direction of the boat.
The moves represented in Figure 4 can also be translated to a problem
state space matrix, as shown in Figure 5. The two matrices depict the set of
problem states on the two banks of the river. The darker-shaded cells corre-
spond to the unsafe states where the cannibals outnumber the missionaries,
 The Missionaries and Cannibals Problem • 35

W E Initial Goal
MMM MMM
CCC State CCC State

MMM CC MM MC
C CC

MMM C C MM M C
C CC

MMM CC C Sub-goal 1

MMM C CC

MC MM CC

MC MC MC

CC MM MC
Sub-goal 2

CC C MMM

C CC MMM

C C MMM C M MM
C CC

CC MMM MC MM
C CC

MMM
CCC Goal

FIGURE 3 Search Tree Diagram.

36 • Artificial Intelligence Problems and Their Solutions

West bank East bank

Initial
M M M C C C
State

1.
M M M C C CC C
State 2
2.C

3.C
State 4 M M M C C C C State 1
4.C

M M M C C C State 3

5.M
M
State 6 M M C C M C
6.M
C

M C M M C C State 5

7.M
M
State 8 C C C M M M
8.C

9.C
State 10 C C M M M C State 7
10.
CC

C 11 M M M C C State 9
.C
C

M M M C C C Goal State

FIGURE 4 State Transition Diagram.

WEST EAST
Initial Initial
Missionaries State Missionaries State

3 0 0 0

2 1

1 2

0 x 3 x
Goal 0 1 2 3 Goal 3 2 1 0
State State
Cannibals Cannibals

FIGURE 5 Problem State Space Matrix.

and the lighter-shaded cells depict the safe states. The striped cells depict
the states corresponding to the unsafe states on the other side of the river.
For example, having two missionaries and one cannibal on the East bank is
 The Missionaries and Cannibals Problem • 37

a safe state per se. However, when you consider the corresponding state on
the West bank, (i.e., one missionary and two cannibals), you know it is an
unsafe state for the West bank, and thus this state is invalid and should be
avoided. Therefore, the solution involves transitioning from the initial state
(cell marked 0) to the goal state (cell marked X).
The moves are numbered and are mentioned in the cell representing
the state after the move is made. The moves made on the West bank (i.e.,
the departing of the boat) are represented by odd numbers in the matrix
labeled WEST, and the moves made on the East bank are represented by
even numbers in the matrix labeled EAST. The solution in matrix form is
shown in Figure 6.

WEST EAST
Initial Initial
Missionaries State Missionaries State

3 3 1 0 0 4 2 0

2 1 6

1 5 2

0 11 9 7 3 x 10 8
Goal 0 1 2 3 Goal 3 2 1 0
State State
Cannibals Cannibals

FIGURE 6 Solution Using Problem State Space Matrix.

Another way of representing the matrix is portrayed in Figure 7 [4].

This figure represents the problem state space of the West bank at all times.
The odd-numbered moves (from the West bank) are represented with the
arrows with black arrowheads, and the even-numbered moves (from the
East bank) are represented with the arrows with white arrowheads.
Figure 7 helps us understand a very important aspect of the prob-
lem: symmetry under time reversal [4]. This means that the moves from
the start and end of the solution are similar and opposite. That is, if we
work backwards from the solution, we are moving the cannibals and mis-
38 • Artificial Intelligence Problems and Their Solutions

Initial State
(Source)
3
M

1
Key
Legal State
Illegal State
0 West to East
C East to West
Goal State M No. of Missionaries
0 1 2 3
(Destination) C No. of Cannibals

FIGURE 7 State Transitions in Problem State Space.

sionaries from the East bank to the West bank. This property can help us
perform a bidirectional search of the problem and thus reduce the search
by a factor of two.

3.4 HUMAN PROBLEM SOLVING

As part of our research, we observed students trying to solve The Mis-

sionaries and Cannibals Problem. They spent about 20 minutes on it, and
some were able to come up with the solution. Most of them used the dia-
gram the instructor used when introducing the problem. Some tried other
representations, but the pictorial representation seemed to be the easiest
for them to understand.
Most students recognized that a key subgoal of the problem was to first
isolate the missionaries and cannibals. Second, due to the nature of the ini-
tial state, it was natural to first transport all the cannibals to the other side of
the river. But that is just a temporary solution. The next step is to get all the
missionaries to the other side of the river. However, this involves moving
the cannibals back to the original state to fetch the missionaries.
Figure 8 shows the diagram the instructor used when he presented the
problem.
 The Missionaries and Cannibals Problem • 39

West East

MMM

CCC

FIGURE 8 Instructor’s Diagram for The Missionaries and Cannibals

Problem.

Here is how three students tackled the problem. (The complete solu-
tions are available in Appendix A.)

Student 1:
The student uses a diagram similar to the instructor’s (Figure 9).
The diagram represents the states after every move.
A written explanation of
W E
each move accompanies the
diagram. MMM
1
CCC
The student starts with
two entities on the boat but 2
MM MC MC M and C in the
CC boat
realizes that one has to stay
back and the other has to re- MMM
3 M C Drop C, M leaves
C
turn to fetch other people.
The student starts with an
FIGURE 9 Student 1’s Partial Solution.
M and a C but abandons the
solution after two steps.
The next attempt is made by moving two Cs first but again abandoning
the attempt after three steps.
In the third attempt, two Ms are moved first, but this fails to satisfy the
constraint imposed by the problem in the second move.
The student goes back to the first attempt and tries again, this time
reaching the sixth step and successfully moving one C and one M to the
other side of the river.
The time allotted is over (Figure 10).
40 • Artificial Intelligence Problems and Their Solutions

Conclusion
W E
The student tries the three possible first
MMM
0
CCC
moves, and, after reaching dead ends, he
learns about good and bad moves. By the
MM M and C
1
CC
MC
leave
third attempt, the student has evaluated the
best of the three possible first moves. In the
MMM
2
CC
C M leaves fourth attempt, the student makes this move
and proceeds further, carefully avoiding the
3 MMM CCC CC leaves
moves that violate the problem constraint.
4
MMM
CC C leaves The student has gathered this knowledge
C
from the previous three trial-and-error at-
MM
5 MC
CC
MM leave tempts.
MM The student first tries to understand the
6 MC leave
CC
problem by using the trial-and-error tech-
nique and plans for the next attempt.
FIGURE 10 Student 1’s Solution.

Student 2:
The student uses characters M and C to represent
MMM
individual missionaries and cannibals respectively, and CCC
uses arrows to depict the direction of the boat. MMM
CC
C
No other explicit representation of the moves is MMM
C
CC
shown.
MMM CCC
First, all the cannibals are transferred to the East MMM
CC
bank of the river. C
MM
MC
Then all the cannibals are returned to the West CC
MM
bank, and all the missionaries are brought to the East CC
MC

bank, carefully avoiding violating the problem con- CC

MMM
C
straint.
CCC MMM
Finally, after all the missionaries are moved to the
MMM
C
East bank, the cannibals are transferred one by one to CC

the East bank (Figure 11). CC

MMM
C
MMM
Conclusion CCC

The student solves the problem in time. It seems FIGURE 11 Student 2’s
she recognizes that one of the subgoals is to separate Solution.
 The Missionaries and Cannibals Problem • 41

the Cs from the Ms by moving all the Ms or all the Cs first. However, while
transferring the Ms, some Ms and some Cs sometimes have to be trans-
ferred together to maintain balance to avoid having Cs outnumber Ms. The
student realizes that moving the Cs back to the West bank again is necessary
to get all the Ms on the East bank.

Student 3:
Student 3 uses a tabular format with three columns to represent the
problem and the solution. The first column depicts the state on the West
bank, the second column represents the state
on the East bank, and the third column repre- 3M MC
sents the move. 3C
2M M
The student starts by moving a missionary M
2C C
and a cannibal in the first move and proceeds 3M
to move all the cannibals to the East bank. C 2C
2C
After this, the student skips some steps, 3M 3C
reaches the second to last step, and then the M 2M
last step, and does not show the moves in the C 2C
table. 3M
3C
The last two steps are the exact opposite of
the first two steps (Figure 12). FIGURE 12 Student 3’s Solution.

Conclusion
It may be possible that the student was able to identify that this solution
has time reversal symmetry and thus jumped to the last few steps and was
either trying to figure out or gave up figuring out the steps in between. It
seems there was an attempt to use bidirectional search, as the end states were
mentioned with some steps in the middle missing. It may be possible that
apart from searching from the initial state toward the goal state, the student
was working backward toward the goal.

Analysis of the Solution

We realize it will be easier to solve the problem if we separate the mis-
sionaries and cannibals. This can be done by either moving all the mission-
aries or all the cannibals to the other side of the river first. Trying to move
all the missionaries first is impossible because after moving two of them,
we get into an unsafe state. Moving all the cannibals to the other side is
42 • Artificial Intelligence Problems and Their Solutions

possible. This is accomplished after the first four moves. However, at least
one cannibal has to return back to fetch the missionaries. Now we can start
moving the missionaries to the other side. The sixth move is extremely im-
portant. It maintains the equilibrium by moving a missionary and a cannibal
to the West bank to get the last of the missionaries from that side in the
seventh move, while also ensuring a safe state on the East bank.
Note that this step seems like it is leading away from the solution (i.e.,
this move causes the distance between the current state and the goal state
to increase) [5]. The next four steps get the remaining cannibals back on
the East bank.
By analysis, it was found that people divide the problem into three or
four major subproblems [5]. It is evident from the solution in Figure 10 that
the first four steps involve moving all the cannibals to the East bank. The
next three steps involve carefully move the cannibals back to retrieve all the
missionaries, while balancing the number of missionaries and cannibals on
either side of the river. Finally, the last four steps again move all the can-
nibals to the East bank.

3.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

After doing some research, we have collected a number of solutions

for The Missionaries and Cannibals Problem for use in our analysis of the
Human Window. In this section, we present our findings and an analysis of
those findings. We also present the solution we have determined to be the
most and the least Human Window compatible and an explanation as to
why it is considered as such.
When analyzing the solutions (Table 2), we can see there are a wide
variety of possible representations that can be used to describe and solve
this problem. One important detail to note is that every one of them is
extensional. This may be due to the problem’s relatively small state space,
which makes it possible to show the solution in a reasonably sized image,
graph, table, and so on. There are only 16 possible states in total, but only
10 which are legal are required to describe the solution. Another reason is
because the problem is not part of an even larger problem. That is, there
will always be three missionaries and three cannibals in the problem, and
these numbers do not vary.
 The Missionaries and Cannibals Problem • 43

We can see that the majority of solutions were either table or graphical
representations. This is most likely because this is a problem that involves
visualization, as there is an environment with two kinds of entities interact-
ing with it. Also, graphs and tables help provide visualization that minimizes
the need for descriptions.
Overall, the solutions had mixed rankings for each of their categories.
Some were more executable than comprehensible and some more compre-
hensible than executable. However, those that contained images or graph-
ics generally fit more within the Human Window. It can also be seen that
they have a higher ranking of extensionality as well, while those represented
by text or tables were ranked relatively low. Finally, these solutions used a
graphical/visual representation as their primary problem-solving technique,
while others, especially those represented by tables, used more of a top-
down approach.

3.5.1 The Most Human Window–Compatible Solution

The Most Human Window–Compatible (MHWC) solution is the “Sym-

metric Solution” (Figure 13). (A good example of this solution can be found
at [6].) This solution is a depiction of the state space of the problem as a
graphical image. The vertices of this graph are the problem’s state repre-
sentations. These representations show a visual of the environment, indicat-
ing the river, the boat, the bank that the boat is closest to, and letters that
represent the missionaries and cannibals. Each of these components is sep-
arate from the others, and the environment is completely redrawn in each
state. Finally, a key is provided to indicate the meaning of the initials. Each
of these details makes the overall representation highly comprehensible.
The edges of the graph indicate which states are adjacent to each other.
They also are given a weight that indicates how many missionaries and/or
cannibals are to be placed in the boat to move between each state. This is
done with the use of numbers and initials. This makes the solution very
executable.
This solution shows all of the legal states and how they are reached,
including the states that do not lead to a solution. This provides a depiction
of the entire state space, revealing every possible solution to the problem,
and allows a person to solve the problem in any way he or she sees fit. One
of the most intriguing details about the solution is that it reveals another
44 • Artificial Intelligence Problems and Their Solutions

Difficulty

2/10

Complexity

O(n)

Int or Intly/ Grn Prob Solv

Name Repr HW? Corr? Exec Compr Flex Mode of Conv Opt? Tot
Ext? Extly Sz Mthd
Constraint Table Generator;
Ext. 1/10 Table Y Y Small 6/10 2/10 Const. Satisf. 5/10 Y 14/40
Rules Hand
Copin Graphic Tp-Dwn; Graphic Img
Ext. 5/10 Y Y Small 9/10 6/10 8/10 Y 28/40
Diagram Diagrm Graphic Rep. Maker; Hand
Tp-Dwn;Vis.
Game Ext. 9/10 Imgs Y Y Ideal 10/10 9/10 10/10 Img Maker; Hand Y 38/40
Rep.
Logic State Graphic Img
Ext. 3/10 Y Y Small 7/10 5/10 Graphic Rep. 7/10 Y 22/40
Diagram Diagrm Maker; Hand
Tp-Dwn; Table Generator;
Ma’s Table Ext. 1/10 Table Y Y Ideal 9/10 4/10 8/10 Y 22/40
Table Repr. Hand
Descr. Tp-Dwn; Vis. Text Doc.; Img
O+G Ext. 6/10 Y Y Ideal 10/10 7/10 9/10 Y 32/40
w/ Img Rep. Maker; Hand
Prolog Tp-Dwn; Table Generator;
Ext. 2/10 Table Y Y Ideal 10/10 7/10 8/10 Y 27/40
Table Table Rep. Hand
Graphic Graphic Img
Symmetric Ext. 7/10 Y Y Ideal 10/10 8/10 Graphic Rep. 10/10 Y 35/40
Img Maker
Text Tp-Dwn; Vis.
Ext. 1/10 Text Imgs Y Y Ideal 10/10 7/10 9/10 Text Doc.; Hand Y 27/40
Symbol Rep.
Graph/ Graphic Img
Z Ext. 2/10 Y Y Ideal 7/10 3/10 Graphic Rep. 7/10 Y 29/40
Grid Maker; Hand

Key

Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?

Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?

Rep.: Representation—How is the solution represented?

HW?: Human Window—Does the solution exist in the Human Window?

Corr?: Correctness—Is the solution correct?

Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?

Exec: Executability—How executable is this solution?

Compr: Comprehensibility—How comprehensible is this solution?

Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?

Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?

Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 2 Ranking The Missionaries and Cannibals Problem Solutions According to the Human Window.
 The Missionaries and Cannibals Problem • 45

method of problem solving: the use of subgoals. When you examine this
solution further, you will notice that the first half of the solution is an exact
mirror image of the second half. This shows that when the first half of the
problem is solved, the second half can be solved by replicating the steps for
the first half in reverse and placing the missionaries and cannibals on re-
verse banks. The solution depicts this symmetry by presenting the graph as
a symmetrical image. Finally, this symmetry once again reveals the benefits
of using a bidirectional search to solve this problem. That’s why this is the
MHWC solution, despite the fact that it did not receive the highest overall
ranking on the table.

3.5.2 The Least Human Window–Compatible Solution

The solution that is the Least Human Window–Compatible (LHWC)

in our assessment is the “Constraint Rules Solution” (Figure 14). It is called
this because it uses constraint satisfaction as its problem-solving method.
For this solution, to find the goal state, one must follow a set of constraints,
or “Production Rules” (see Section 3.6). These rules correspond to the
constraints one must adhere to when attempting to solve The Missionaries
and Cannibals Problem.
The main idea behind using this method to find the solution is (1) for
each step, look at the current state; (2) look at the rules provided to see
which can be followed; (3) implement this rule; and (4) indicate the new
state as well as the rule used. This process is to be continued until a goal
state is reached.
Although these rules efficiently achieve the goal state in the least num-
ber of moves possible (an ideal approach for a computer program), a solu-
tion such as this may be too complicated for people, especially with regard
to a relatively small problem such as The Missionaries and Cannibals Prob-
lem. Humans prefer visual and intuitive representations, especially when
constraints are involved. Further analyzing Figure 14, we can see that the
solution is divided into two parts: a table of these Production Rules and a
table that depicts an example solution when following these rules. We will
analyze both.
The example solution table shows a possible transition of states if one
were to apply the Production Rules to The Missionaries and Cannibals
Problem. Each step is represented by a row, and it shows the current state
46 • Artificial Intelligence Problems and Their Solutions

Start Goal
C ×1 MC C ×2, W C ×2, W MC C ×1
MC MC
MC MC

M ×1, M ×1,
C ×1, W C ×1, W

MC MC C M M C MC MC
C MC MC MC C MC MC C MC MC MC C
M M M M

M ×1, E M ×1, E

MC C ×1, E C ×1, E MC
C MC MC C
M M

C ×2, W C ×2, W

C M M C
C M M C
C M M C

C ×1, E C ×1, E

M M
C MC C
C MC C
M M

M ×2, W M ×2, W
M x1,
C x1, E
MC MC
MC MC MC MC

Key:
M = Missionary W = West Bank
C = Cannibal E = East Bank
For example, M ×1, W
One missionary travels to the west bank
FIGURE 13 The “Symmetric Solution.”

and the rule applied to get to that state. When analyzing the table, a pos-
sible issue that arises with this method is tracing the solution in general. To
fully understand how the solution is carried out, one must either memorize
each rule, which is simply too difficult, or continuously refer back to the
table to find a corresponding rule, which can be very tedious.
 The Missionaries and Cannibals Problem • 47

Aside from the example table provided in the figure, comprehend-

ing the Production Rules may be hard in and of itself. (An example of
these rules can be found in [7].) This is because they appear similar to “if
statements” commonly used by computer scientists. The average person
may very well not be able to understand the real implications behind a
series of “if” statements. However, an implementation of the Production
Rules is a better solution for a machine rather than a human being (see
Section 3.6).
Additionally, for the solution in general, the missionaries and cannibals
can be seen to be represented by variables, and the banks and the boat are
represented by functions that take those variables as parameters. Again, a
computer scientist or mathematician can probably understand this better
than the average person.
Key
x = Missionaries (x = 0, 1, 2, or 3)
y = Cannibals (y = 0, 1, 2, or 3)
W(x, y) = Number of missionaries and/or cannibals on the West bank
E(x, y) = Number of missionaries and/or cannibals on the East bank
B(x, y) = Number of missionaries and/or cannibals on the boat and crossing
the river to get to the opposite bank

Constraint Rules
Rule
Constraint
No.
1 B(2, 0) iff : (x – 2 ≥ y || x – 2 = 0 in one bank) && (x + 2 ≥ y in the other)
2 B(1, 0) iff : (x – 1 ≥ y || x = 0 in one bank) && (x + 1 ≥ y in the other)
B(1, 1) iff : (x – 1 ≥ y – 1 || x = 0 in one bank) && (x + 1 ≥ y + 1 || x = 0 in
3
the other)
4 B(0, 1) iff : (y – 1 < x || x = 0 in one bank) && (y + 1 ≤ x || y = 0 in the other)
5 B(0, 2) iff : (y – 2 ≤ x || x = 0 in one bank) && (y + 2 ≤ x || x = 0 in the other)

Initial State: W(0, 0) E(3, 3)

Goal State: W(3, 3) E(0, 0)
48 • Artificial Intelligence Problems and Their Solutions

Example
Step State Rule Applied
0 W(0, 0); E(3, 3) (Init)
1 W(1, 1); E(2, 2); B(1, 1) 3
2 W(0, 1); E(3, 2); B(1, 0) 2
………
10 W(2, 2); E(1, 1); B(1, 0) 2
11 W(3, 3); E(0, 0); B(1, 1) 3

FIGURE 14 The “Constraint Rules Solution.”

3.6 BEST MACHINE SOLUTION

For a computer, The Missionaries and Cannibals Problem is gener-

ally easy to solve. The best solutions for the problem are those that find
a solution quickly and take very little space, usually representing a prob-
lem state space with a search tree, whereby the initial state is the root of
the tree. Along with the bidirectional search [8], search algorithms such as
the breadth-first search, depth-first search, and iterative-deepening
have proved to be very effective in solving this problem. This is because the
state space for the problem is so small, especially when the constraints are
met, thereby reducing the depth and the branching factor of the tree.
Additionally, a set of Production Rules are used to help guide these
searches to the goal state. The Production Rules are the constraints that
must be followed to solve the problem correctly (i.e., without breaking
the rules). These rules are essentially “if-then” statements. According to
Robertson, “Some researchers claim to be able to predict the extent to
which the rules that apply in one area can transfer [domain-specific knowl-
edge] to another” [9]. An example of these rules can be seen in Figure 14
(see Section 3.5.2) [7]. With the rules implemented and the constraints put
into place, the worst-case scenario for any search algorithm is a tree with
only 16 nodes (the legal states). Therefore, the overall complexity of The
Missionaries and Cannibals Problem is only in the order of O(n).
 The Missionaries and Cannibals Problem • 49

3.7 PLAYABLE PROGRAM

Some Websites have games developed in Flash for The Missionaries

and Cannibals Problem where one can try and use different problem-solv-
ing techniques. These games are highly extensional representations; the
problem solver can actually manipulate the objects and use the trial-and-
error technique to find the solution. Furthermore, one would have to draw
the representations multiple times before arriving at the solution if one
tries to solve this problem manually. Thus, such games get students inter-
ested in problems such as this. One such game is available at http://www.
smartestgames.com/game/missionaries-and-cannibals/.
Thus, choice of representation, the trial-and-error method, planning,
solving subgoals, and symmetry are some of the techniques that are helpful
in solving a problem like The Missionaries and Cannibals Problem.

3.8 REFERENCES

1. O’Connor, J.J., and Robertson, E.F. (2012). Propositiones ad acuendos iuvenes

by Alcuin. Accessed on December 4, 2013.
2. Dold-Samplonius, Y., Dauben, J.W., Folkerts, M., and Van Dalen, B. (2002).
“Jealous Husbands Crossing the River: A Problem from Alcuin to Tartaglia,
Raffaella Franci,” From China to Paris: 2000 Years Transmission of Mathemati-
cal Ideas. Stuttgart: Franz Steiner Verlag.
3. Pressman, I., and Singmaster, D. (1989). “‘The Jealous Husbands’ and ‘The
Missionaries and Cannibals.’” The Mathematical Gazette (73) 464: 73–81.
4. Amarel, S. (1968). “On Representations of Problems of Reasoning about Ac-
tions,” Machine Intelligence (3): 131–171. http://aitopics.org/sites/default/files/
classic/Machine%20Intelligence%203/MI3-Ch.10-Amarel.pdf. Accessed on
January 8, 2014.
5. Eysenck, M.W. (2002). Simply Psychology, 2nd ed. New York: Psychology
Press.
6. Wickler, G. (2013). “The Missionaries and Cannibals Problem.” Gerhard Wick-
ler’s Home Page. Edinburgh, Scotland: Artificial Intelligence Applications In-
50 • Artificial Intelligence Problems and Their Solutions

stitute School of Informatics. Available at http://www.aiai.ed.ac.uk/~gwickler/

missionaries.html. Accessed on January 8, 2014.
7. Mousavi, F. (2012). “Missionaries and Cannibals Problem in AI.” Programming
Notebook. Available at http://blog.mousavi.net/2013/07/29/missionaries-and-
cannibals-problem-in-ai-2/. Accessed on January 8, 2014.
8. Lucci, S., and Kopec, D. (2013). Artificial Intelligence in the 21st Century.
Dulles, Virginia: Mercury Learning Inc.
9. Robertson, S.I. (2001). Problem Solving. New York: Psychology Press.
 CHAPTER

4
THE 12 COINS PROBLEM
Score
11188

Click button to reset the scale

Image published with permission from Novel Games Limited.

http://www.novelgames.com/en/spgames/coins/
FIGURE 15 The 12 Coins Problem.

4.1 BACKGROUND

The 12 Coins Problem (Figure 15) belongs to a class of counter-

feit coin problems that involve finding a counterfeit coin among a group
of coins. The earliest known appearance of this kind of problem was in
E. D. Schell’s article in the January 1945 issue of American Mathematical
Monthly, which required finding an underweight coin among a group of
eight coins using a beam balance only twice [1]. There are many variations
of the counterfeit coin problems, with some of them providing an extra
standard coin for comparison and others revealing whether the odd coin is
heavier or lighter than the rest of the coins.
52 • Artificial Intelligence Problems and Their Solutions

For The 12 Coins Problem in this chapter, given that 1 of the 12 coins
has a different weight, the challenge is to find the odd coin in just three
weighings and also to determine if the coin is heavier or lighter than the
others.
The data given for this problem is:
There is only one odd coin.
The odd coin has a weight different from the others.
The unknowns are:
Which coin is the odd one?
Is the odd coin heavier or lighter than the others?
The constraint is:
Only three weighings are allowed.

4.2 SOLVING A SMALLER PROBLEM

We can convert this problem to a simpler one by reducing the number

of coins and the number of weighings. Heeding Polya’s advice, “If you can’t
solve a problem, then if there is an easier problem that you can solve, find
it,” we look for a problem with fewer coins.
Let us start using the problem reduction technique. If we start with only
2 coins, it is impossible to find which of the 2 coins is odd unless we know
if the odd coin is heavier or lighter than the other or if we have a standard
coin to compare with. Thus, the 2 coins problem is not the minimal prob-
lem (without an extra standard coin or the knowledge of the relative weight
of the odd coin with respect to the good one). Hence, we can check if the
3 coins problem is an easier version of The 12 Coins Problem. If we reduce
the number of coins to 3, we may find a method that can be applied for
solving The 12 Coins Problem.
In the 3 coins problem, if we know that the odd coin weighs less or
more, only one weighing is required to identify the odd coin. Otherwise,
we must first compare the first and second coins. If they are equal, we
know that the third coin is odd. One more weighing of the third coin with
any of the first two coins will tell us if the odd one is heavier or lighter than
the other two. However, if the scales are not equal in the first weighing, it
 The 12 Coins Problem • 53

means one of the first two coins is odd and the third coin is not odd. In any
case, after the results of first weighing, the 3 coins problem becomes similar
to the “2 coins problem with the standard coin,” and thus the odd coin and
its relative weight can be found in a total of two weighings. A solution to
the 3 coins problem with the coins named A, B, and C is shown in Table 3.

Weighing 1 Weighing 2
Result
A?B A?C
A=C Not possible because one of them is odd
A=B
(Thus, C is A>C C is odd and light
odd)
A<C C is odd and heavy
A=C B is odd and light
A>B
(Thus, A or B A>C A is odd and heavy
is odd)
A<C Not possible because ONLY one of them is odd
A=C B is odd and heavy
A<B
(Thus, A or B A>C Not possible because ONLY one of them is odd
is odd)
A<C A is odd and light

TABLE 3 Solution to the 3 Coins Problem.

Thus, we can deduce the following:

Coins with equal weights are good coins and can be separated from the
coins that have not been weighed yet, and this group of remaining coin(s)
contains the odd coin.
When the scales are unequal, one side of the scale contains the odd
coin.
Once the odd coin is detected, we need to know its weight relative to a
good coin to determine if it is heavier or lighter than the rest.
Moving to the slightly more complex 4 coins problem, if we divide the
coins into two groups of 2 coins and compare the weights of these groups,
we know which of the two groups is heavier (or lighter). However, we do
not know which group has the odd coin, nor do we know if the odd coin is
heavy or light. Thus, very little information is gained from the first weigh-
ing. Furthermore, two more weighings are needed to compare the coins in
each group individually to identify the odd coin.
54 • Artificial Intelligence Problems and Their Solutions

Alternatively, we could take the first 3 coins and solve them in two
weighings like we did for the 3 coins problem. If both groups weigh the
same, we know the fourth coin is odd, and one more weighing with one of
the known good coins can tell us if the fourth coin is overweight or under-
weight. Thus, the 4 coins problem can be solved in three weighings.
This problem helps us conclude that dividing coins into groups of three
gives us more information in the first weighing, as compared to dividing
into groups of two—the information about the group that has good coins
only. This is because after the first weighing, we know that either the first
two groups or the third one have all good coins, thus separating the good
coins from the groups containing the odd coin, and thereby reducing the
problem into one with fewer coins.
If we try to solve the 6 coins problem with the knowledge gained from
the previous two problems, we can proceed as follows: Divide the coins to
three groups of 2 coins. After comparing the first two groups, we know the
odd coin is in one of first two groups if they do not weigh the same or that
the odd coin is in the third group if the first two groups weigh the same in
the first weighing.
In either case, one more weighing of the odd (or possibly odd) group
with one of the good group(s) will tell if the odd coin is heavy or light, and
the third weighing will help identify the odd coin in the odd group (Table 4).

Weighing 1 Weighing 2
Weighing 3 Result
AB ? CD AB ? EF
Not possible
AB = EF because ONLY
one of them is odd
Not possible because ONLY
E=F
AB = CD one of them is odd
AB > EF
E>F F is odd and light
(E or F is
E<F E is odd and light
odd)
Not possible because ONLY
E=F
one of them is odd
AB < EF
E>F E is odd and heavy
E<F F is odd and heavy
 The 12 Coins Problem • 55

Not possible because ONLY

C=D
one of them is odd
AB = EF
C>D D is odd and light
C<D C is odd and light
Not possible because ONLY
A=B
AB > CD one of them is odd
AB > EF
A>B A is odd and heavy
A<B B is odd and heavy
Not possible
AB < EF because ONLY
one of them is odd
Not possible because ONLY
C=D
one of them is odd
AB = EF
C>D C is odd and heavy
C<D D is odd and heavy
Not possible be-
AB < CD AB > EF cause ONLY one
of them is odd
Not possible because ONLY
A=B
one of them is odd
AB < EF
A>B B is odd and light
A<B A is odd and light

TABLE 4 Solution to the 6 Coins Problem.

From the 6 coins problem, we can conclude the following:

When we divide coins into groups of 3 coins, we know by the second
weighing if the odd coin is heavy or light and which group contains the odd
coin.
This is because by the end of two weighings, we know the relative
weight of each group with respect to the other two groups, which separates
out the odd group and tells us if the odd coin is heavy or light.
The 9 coins problem is a special case, because we divide the coins into
three groups of 3 coins. Thus, after comparing one of the three groups
56 • Artificial Intelligence Problems and Their Solutions

with the other two groups in two weighings, we know the odd group and
the relative weight of the odd coin. Therefore, the problem is reduced to a
3 coins problem when the relative weight of the odd coin is known. Hence,
one more weighing gives us the odd coin.
Alternatively, we could consider the 3 coins in each of the three groups
as one big coin. The problem could then be reduced to a 3 coins problem,
and after two weighings, we know if the three groups or the big coin is odd
and if it is heavy or light. Thus, now we know the group of 3 coins contain-
ing the actual odd coin. This problem is thus reduced to a 3 coins problem
where we know that the odd coin is heavy or light, and thus one weigh-
ing is enough to find the odd coin. This technique is known as recursion,
where a problem is continuously broken down to a subproblem that, when
it reaches the elementary or base case, can be solved directly.
We can conclude that as the number of coins increases, we need to
compare as many coins as possible in the first two weighings to reduce the
problem to one that is as small as possible. We have just seen how a 9 coins
problem was reduced to a 3 coins problem after two weighings.

4.3 SOLUTION

In an observation conducted in an AI class, one of the students came up

with the solution to The 12 Coins Problem shown in Table 5.

Weighing 1
Weighing 2 Weighing 3 Result
ABCD ? EFGH
Not possible because one of
A=L
them is odd
ABC = IJK
A>L L is odd and light
A<L L is odd and heavy
I=J K is odd and light
ABCD = EFGH
ABC > IJK I>J J is odd and light
I<J I is odd and light
I=J K is odd and heavy
ABC < IJK I>J I is odd and heavy
I<J J is odd and heavy
 The 12 Coins Problem • 57

Not possible because one of

G=H
them is odd
ABE = CDF
G>H H is odd and light
G<H G is odd and light
A=B F is odd and light
ABCD > EFGH
ABE > CDF A>B A is odd and heavy
A<B B is odd and heavy
C=D E is odd and light
ABE < CDF C>D C is odd and heavy
C<D D is odd and heavy
Not possible because one of
G=H
them is odd
ABE = CDF
G>H G is odd and heavy
G<H H is odd and heavy
C=D E is odd and heavy
ABCD < EFGH
ABE > CDF C>D D is odd and light
C<D C is odd and light
A=B F is odd and heavy
ABE < CDF A>B B is odd and light
A<B A is odd and light

TABLE 5 Solution to The 12 Coins Problem.

In this solution, the 12 coins are divided into three groups of 4 coins
each. In the first weighing, we know either that the third group of coins,
which has not yet been weighed, has the odd coin or that one of the first
two groups has the odd coin. In the first case, the problem is reduced to a
4 coins problem. By removing one coin from this group, the problem can
be solved as a 3 coins problem in two more weighings. If that fourth coin
that was omitted is found to be odd (by the second weighing), it can be
weighed against a known good coin to check if it is heavy or light in the
fourth weighing.
In the second case, we know the relative weights between the first two
groups and can use this knowledge for the next weighing. We then divide
58 • Artificial Intelligence Problems and Their Solutions

the heavier group into 2 coins each and place it on either side of the scale.
We then take 2 coins from the lighter group and place one on either side
of the scale and conduct the second weighing. This would reduce the num-
ber suspect odd coins to 3 possible coins if the scales do not balance or 2
possible coins (the remaining 2 from the lighter group) if the scales do bal-
ance. In any case, one more weighing will suffice to detect the odd coin.
This solution, though correct, cannot be extended to any number of coins.
We can generalize The 12 Coins Problem to an n coins problem. It
was found that the number of weighings required to find an odd coin
among n coins can be deduced from the following formulas developed by
Bennet Manvel in his article “Counterfeit Coin Problems” in Mathematics
Magazine:
An underweight coin in a set of k coins can be found in [log 3 k] weighings.
If we know only that the counterfeit coin is a different weight, then
⎡log 3 (2k + 3)⎤ weighings are required.
⎡log 3 (2k + 1)⎤ weighings are required if a standard coin is provided.
In Table 6, we tabulated the number of weighings for several n coin
problems using the formulas for the three cases mentioned above.

Case 1: Finding an Case 2: Finding an odd Case 3: Finding an

underweight coin in coin with a different odd coin with a dif-
a set of k coins weight when a standard ferent weight
No. of coin is provided *
Coins = k
No. of weighings = No. of weighings = No. of weighings =
⎡log 3(k)⎤ ⎡log 3(2k + 1)⎤ ⎡log 3(2k + 3)⎤

3 1 2 2
4 2 2 3
6 2 3 3
9 2 3 3
11 3 3 3
12 3 3 3
 The 12 Coins Problem • 59

13 3 3 4
39 4 4 4
40 4 4 5
120 5 5 5
* Please note that [x] means the “ceiling” of x, or the smallest integer greater than or equal to
x, where x is a real number [1].
TABLE 6 Number of Weighings Required to Find an Odd Coin Among n Coins in Different Cases.

The graph in Figure 16 illustrates the application of the formulas in

Table 6.
Noteworthy is how the log3 growth rates seem to define a stepwise
function.

7.00
‘N’ coins problem
6.00

5.00
No. of weighings

4.00

3.00
No. of weighings = log3(k)
2.00
No. of weighings = log3(2k+1)
1.00
No. of weighings = log3(2k+3)
0.00
1 3 9 27 81 243
No. of coins

FIGURE 16 Graphical Depiction of Some of the n Coins Problems.

Another solution that Jack Wert created is available at http://www.cut-

the-knot.org/blue/OddCoinProblems.shtml and http://www.cut-the-knot.
org/blue/OddCoinProblemsShort.shtml. This is a solution for The 12 Coins
Problem that can be extended to any (3n – 3)/2 number of coins [2].
Divide the coins into three equal groups of four. Let us name the coins
A, B, C, D, E, F, G, H, I, J, K, and L.
60 • Artificial Intelligence Problems and Their Solutions

Divide each of the three groups into a main subgroup of 3 coins and
then 1 coin.

Place the first two groups of 4 coins on either side of the scale and note
the results. This is the first weighing.

Then, rotate the main subgroups—in other words, remove the main
subgroup of the first group from the left pan of the scale, move the main
subgroup of the second group from the right pan to the left, and keep the
main subgroup of the third group on the right pan of the scale. Keep the
single coins from the first two groups as they are on the scale. This is the
second weighing.

The weighings are described in Table 7.

Left Pan Right Pan

Weighing 1 (W1) A, B, C, D E, F, G, H

Weighing 2 (W2) E, F, G, D I, J, K, H

TABLE 7 Configuration of Coins in the Two Weighings.

 The 12 Coins Problem • 61

Each of the weighings has three possible outcomes:

= means the two scales are equal.
< means the right pan is heavier than the left one.
> means the left pan is heavier than the right one.
Table 8 shows the nine possible outcomes of the two weighings.

Is the How many

Probable Total
W1 W2 odd coin more
Case odd Next step number of
result result heavy or weighings
coin(s) weighings
light? needed?
Compare coin
L with any of
Can’t say
1 = = L the good coins 1 3
yet.
to know its
relative weight.
Can’t say Compare coins
2 < < D or H 1 3
yet. D or H with
any of the good
coins to know
Can’t say the relative
3 > > D or H 1 3
yet. weight of the
odd coin.
4 = < I, J, or K Heavy This problem 1 3
is reduced to
5 = > I, J, or K Light 1 3
the 3 coins
6 < = A,B, or C Light problem, 1 3
where the
7 > = A,B, or C Heavy 1 3
relative weight
8 < > D,E, or F Heavy of the odd coin 1 3
is known.
Thus, solve
9 > < D,E, or F Light the reduced 1 3
problem.

TABLE 8 Possible Outcomes After the Two Weighings.

The key aspect of this solution is that it tries to maximize the amount
of information gained in the least number of weighings. In the first two
weighings, 11 out of the 12 coins are weighed. The weighings are then con-
figured in such a way so that after two weighings, the problem is reduced to
62 • Artificial Intelligence Problems and Their Solutions

a smaller problem that can be solved in just one more weighing, depending
on the nine possible results. Also, we can leverage the information gained
from the previous two weighings in the last weighing.
According to this solution, to solve a (3n – 3)/2 coins problem in general,
the following steps must be followed [2]:
Divide the coins into three equal main groups. Each is to be divided
into (n – 1) subgroups of 3n-2, 3n-3, . . . , 30 coins.
Place the first two main groups on each side of the scale. This is the first
weighing.
Rotate only the largest subgroups, as described above in The 12 Coins
Problem solution. This is the second weighing.
If the condition of the balance changes, as seen in the last six cases
of The 12 Coins Problem solution above, the subgroup containing the
odd coin will be identified, and the problem will be reduced to a smaller
problem.
Continue the rotation of the next largest subgroups until there is a
change in the condition of the balance or until there is only one coin on
each side of the scale. Then rotate them and the odd coin, and its relative
weight will be found.

4.4 HUMAN PROBLEM SOLVING

This problem was presented to a class of around 20 graduate students

studying artificial intelligence. It was presented in two phases: as a 6 coins
problem and then as The 12 Coins Problem. The choice of representation
most students used was a search tree, with different branches of the tree
signifying the different outcomes of the weighings.
Most students were able to solve the 6 coins problem in three weighings,
with one of the students concluding that if the relative weight of the odd
coin is known, only two weighings were sufficient.
The class struggled with The 12 Coins Problem because it is harder than
the 6 coins problem. Most of them were able to solve the problem in four
weighings, and one student solved it in the minimum three weighings. (This
student’s solution is shown in Table 8.) It used the information gained from
 The 12 Coins Problem • 63

the relative weights of two groups of coins compared in the first weighing
to narrow down the number of possible odd coins.
From the analysis of The 12 Coins Problem, by reducing it into smaller
problems, we learn more about the problem and get clues about how to
solve the original problem. Also, decomposing a bigger problem into small
parts and solving the smaller problems recursively helps to solve the bigger
problem. We have seen that problem reduction and recursion techniques
were useful in solving The 12 Coins Problem.

4.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

When analyzing the solutions (Table 9), we can see some similarities in
their rankings. First, note that each of them is extensional. The solutions we
found are specific to The 12 Coins Problem. Since the solution is relatively
small (three weighings), any method can depict the problem in its entirety.
If the table contained a solution to the general n coins problem, however, it
would most likely be classified as intensional.
Overall, their rankings of the solutions’ extensionality are mixed. Those
that used actual images depicting coins and scales were generally ranked
higher, while those represented by text were ranked lower. However, not
many of the solutions found used images as a means of representation.
The choices of knowledge representation used for the solutions are a
mix between tables and text. The text solutions give a general description
about each possible outcome for all three weighings. The tables simplify
these explanations with the use and organization of symbols. In general,
each solution uses the three weighings method similarly and emphasizes
the subproblems generated by it.
The rankings pertaining to how the solutions fit into the Human Win-
dow are mixed as well. In general, each of the solutions was given a high
ranking for executability. Overall, it is easy to follow the steps of the three
weighings regardless of the representation used. The comprehensibility
of the solutions, however, was mixed. Those that used a text representa-
tion were generally ranked lower, while those represented by tables were
ranked higher. Finally, it can be noted that each of the solutions used the
same general problem-solving techniques, which entail looking for subgoals
by employing using visual representations.
64 • Artificial Intelligence Problems and Their Solutions

Difficulty
3/10

Complex
O(n(log 3n))

Int Prob.
Inlty/ Grn Mode of
Name or Rep. HW? Corr? Exec Compr Solv. Flex Opt? Tot
Exlty Sz Conv
Ext? Mthd
Txt Vis. Rep.;
B&Y Ext. 2/10 Y Y Ideal 8/10 8/10 8/10 Txt Editor Y 26/40
Desc Subgoals
Txt Vis. Rep.; Txt Editor;
Duncanson Ext. 4/10 Y Y Ideal 7/10 6/10 9/10 Y 26/40
Desc Subgoals Hand
Tble Mkr;
Vis. Rep.;
FA Ext. 6/10 Tble Y Y Ideal 10/10 7/10 7/10 Txt Editor; Y 30/40
Subgoals
Hand
Generic Tbl. Rep.; Tbl Mkr;
Ext. 7/10 Tble Y Y Ideal 10/10 8/10 8/10 Y 33/40
Table Subgoals Hand
Graphic
Logically Txt Vis. Rep.;
Ext. 3/10 Y Y Small 7/10 5/10 5/10 Img Mkr; Y 20/40
Enumerated Desc Subgoals
Hand
Logic Vis. Rep.;
LogicFlow Ext. 5/10 Y Y Small 9/10 6/10 7/10 Img Mkr Y 27/40
Dgrm Subgoals
Tbl Mkr;
Txt; Vis. Rep.;
Poskitt Ext. 7/10 Y Y Ideal 8/10 5/10 7/10 Txt Editor; Y 27/40
Tble Subgoals
Hand

Key

Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?

Intly/Extly: Intensionality/Extensionality—How intensional or extensional is it?

Rep?: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is the solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution? (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 9 Ranking The 12 Coins Problem Solutions According to the Human Window.

4.5.1 The Most Human Window–Compatible Solution

The solution assessed as MHWC is the “Generic Table Solution”

(Figure 17). This solution is the same as the one in Section 4.3. It is rep-
resented as a table containing four columns. The first three columns are
 The 12 Coins Problem • 65

Weighing 1
ABCD? Weighing 2 Weighing 3 Result
EFGH
A=L Not possible as one of them is odd
ABC = IJK A>L L is odd and light
A<L L is odd and heavy

ABCD = I=J K is odd and light

EFGH ABC > IJK I>J J is odd and light
I<J I is odd and light
I=J K is odd and heavy
ABC < IJK I>J I is odd and heavy
I<J J is odd and heavy
G=H Not possible as one of them is odd
ABE = CDF G>H H is odd and light
G<H G is odd and light

ABCD > A=B F is odd and light

EFGH ABE > CDF A>B A is odd and heavy
A<B B is odd and heavy
C=D E is odd and light
ABE < CDF C>D C is odd and heavy
C<D D is odd and heavy
G=H Not possible as one of them is odd
ABE = CDF G>H G is odd and heavy
G<H H is odd and heavy

ABCD < C=D E is odd and heavy

EFGH ABE > CDF C>D D is odd and light
C<D C is odd and light
A=B F is odd and heavy
ABE < CDF A>B B is odd and light
A<B A is odd and light

FIGURE 17 The “Generic Table Solution.”

66 • Artificial Intelligence Problems and Their Solutions

used to show the possible outcomes the scale may produce during any of
the three weighings, with each column representing a different weighing.
These outcomes contain letters that represent the coins and a conditional
operator that represents the result of the weighing. The coins (letters) that
are grouped together are those present on a specific side of the scale. The
last column indicates the result of the third weighing, thus concluding
whether a coin is heavy or light, or if the weighing is impossible.
This solution is comprehensible because it is concise and the columns
in the table organize the solution well. This solution is also highly execut-
able. When the result of the first weighing is found, the remaining possible
results for the second weighing are directly adjacent to that result in the
next column. This adjacency is also present for the second and third weigh-
ings. This makes it possible for the solution to be found easily and quickly,
and it reduces the information needed to be found in the table overall.
Although the solution is comprehensible, the overall comprehensibility
could be improved if a small picture of the actual scale were present next to
each possible outcome to show if the scale is balanced or not. Additionally,
it could be improved even further if the letters had circles around them,
which would better indicate that they are in fact representing coins.

4.5.2 The Least Human Window–Compatible Solution

The solution assessed as LHWC is the “Logically Enumerated Solu-

tion” (Figure 18). This solution gives a textual description as to what to do
at each weighing, given a particular outcome.
While this solution shows each possible outcome and demonstrates
how to solve the problem in the least number of weighings, a table repre-
sentation may be better than a descriptive representation. This is because
particular solutions to The 12 Coins Problem can be found instantly on a
table, while they may be hard to find in the midst of a lengthy description.
In particular, this solution may not be very comprehensible, since the
descriptions are grouped as small paragraphs. Within those paragraphs are
representations of the coins, the coin groupings, and scales, which may be
seen as inefficient. Each coin is given a name that corresponds to a particu-
lar group, such as X1, X2, Z3, and so on. To indicate that the coins belong
to a particular group on the scale, the coins are concatenated using the “–”
symbol.
 The 12 Coins Problem • 67

Finally, the groups are indicated as being on the left or right side of the
scale by separation using a conditional operator. Although this works for the
“Generic Table Solution,” it may not work in this type of representation.

• Coin Groups: {X1, X2, X3, X4}; {Y1, Y2, Y3, Y4}; {Z1, Z2, Z3, Z4}
• Weigh X1-X2-X3-X4 against Y1-Y2-Y3-Y4.
• Case 1
• 1st weighing: If X1-X2-X3-X4 = Y1-Y2-Y3-Y4
The counterfeit must be in group Z. Takeany of the X or Y coins as a control coin (i.e., X1).
Weigh Z1+Z2 against Z3+X1.

Case 1.1
2nd weighing: Z1+Z2 >Z3+X1.
Z1 or Z2 is heavy or Z3 is light. Weigh Z1 against Z2.
(or if Z1+Z2 < Z3+X1, substitute heavy for light and vice versa in the following.)

♦ Case 1.1.1
♦ 3rd weighing: Z1 = Z2
Z3 is light.

♦ Case 1.1.2
♦ 3rd weighing: Z1 > Z2
Z1 is heavy.

♦ Case 1.1.3
♦ 3rd weighing: Z1 < Z2
Z2 is heavy.

Case 1.2
2nd weighing: Z1+Z2 = Z3+X1
Z4 must be the counterfeit.
Weigh it against any of the control coins to determine whether it is heavy or
light.(i.e., X1)

♦ Case 1.2.1
♦ 3rd weighing: Z4 < X1
Z4 is light.

♦ Case 1.2.2
♦ 3rd weighing: Z4 > X1
Z4 is heavy.
68 • Artificial Intelligence Problems and Their Solutions

• Case 2
• 1st weighing: X1-X2-X3-X4 < Y1-Y2-Y3-Y4 (analogous holds if X1-X2-X3-X4 >
Y1-Y2-Y3-Y4; switch heavy for light and vice versa)
One of the Xs is lighter than the rest, or one of the Ys is heavier than the rest.
All the Zs can be taken as control coins
Weigh X1-X2-Y1-Y2 against X3-Y3-Z1-Z2.

Case 2.1
2nd weighing: X1-X2-Y1-Y2 > X3-Y3-Z1-Z2
Y1 or Y2 is heavy, or X3 is light.
Weigh Y1 against Y2.

♦ Case 2.1.1
♦ 3rd weighing: Y1 > Y2
Y1 is heavy.

♦ Case 2.1.2
♦ 3rd weighing: Y1 < Y2
Y2 is heavy.

♦ Case 2.1.3
♦ 3rd weighing: Y1 = Y2
X3 is light.

Case 2.2
2nd weighing: X1-X2-Y1-Y2 = X3-Y3-Z1-Z2
X4 is light or Y4 is heavy. Weigh X4 against Z1.

♦ Case 2.2.1
♦ 3rd weighing: X4 < Z1
X4 is light.

♦ Case 2.2.2
♦ 3rd weighing: X4 = Z1
Y4 is heavy.

FIGURE 18 The “Logically Enumerated Solution.”

 The 12 Coins Problem • 69

4.6 BEST MACHINE SOLUTION

In general, there are no particular categories of algorithms or heuris-

tics needed for a machine to solve The 12 Coins Problem. Any algorithm
that implements the general method of solving The 12 Coins Problem in
three weighings is guaranteed to find the optimal solution. If implemented
correctly, the same technique can be used to solve any n coins problem in
⎡log 3 (2k + 3)⎤ weighings. Correctly implementing an algorithm will find
a solution to the n coins problem in at most O(nlog3n) time. The n rep-
resents the number of coins in the problem, and the log3n represents the
number of weighings. It is not simply log 3n because the algorithm must find
a way to group every coin at each weighing. Although some coins are not
used in the second and third weighings and will inevitably deal with fewer
than n coins, they are still in the order of n.

4.7 PLAYABLE PROGRAM

You can find help playing The 12 Coins Problem at

http://www.mathplayground.com/coinweighing.html (Figure 19).

Score
11188

Image published with permission from Novel Games Limited.

HTTP://WWW.NOVELGAMES.COM/EN/SPGAMES/COINS/

Click button to reset the scale

FIGURE 19 The 12 Coins Problem.

70 • Artificial Intelligence Problems and Their Solutions

4.8 REFERENCES

1. Manvel, B. (1977). “Counterfeit Coin Problems.” Mathematics Magazine 50:

90–92.
2. Bogomolny, A. (2013). Odd Coin Problems: the 120 Marble Problem—Five
Weighings from Interactive Mathematics Miscellany and Puzzles. Available
at http://www.cut-the-knot.org/blue/OddCoinProblems.shtml. Accessed on
January 9, 2014.
 CHAPTER

5
CRYPTARITHMS

S E N D

+ M O R E

M O N E Y
FIGURE 20 The Send More Money Problem.

5.1 BACKGROUND

A cryptarithm is a type of mathematical puzzle in which the digits are

replaced by the letters of the alphabet or other symbols [1]. The first cryp-
tarithm appeared in 1864 in American Agriculturist magazine. In 1931,
Simon Vatriquant introduced the term crypt-arithmetic (cryptarithme in
French) under the pseudonym MINOS in Sphinx, a Belgian recreational
mathematics magazine [2, 3]. In 1955, J. A. H. Hunter coined the term
alphametic to denote cryptarithms whose letters form meaningful words
or phrases.
72 • Artificial Intelligence Problems and Their Solutions

In artificial intelligence, cryptarithms are studied as a class of con-

straint satisfaction problems, where the goal is to assign the correct
values (the digits) that belong to a domain (0 to 9 in the case of decimal
numbers) to a set of variables (the alphabetic characters), while satisfying
the constraints (the result of the arithmetic operations should be correct)
of the problem [4].
The constraints are as follows:
Each alphabetic character represents one digit only, and each digit is
represented by exactly one alphabetic character.
A numeric representation of a word does not start with the digit zero.
When alphabetic characters are replaced by digits, the resultant sum is
arithmetically correct.

5.2 PROBLEM-SOLVING TECHNIQUES

Deduction, generate and test, backtracking, solving subproblems, and

solving linear equations are some of the techniques that can be helpful for
solving cryptarithmetic puzzles. Usually the problem itself contains some
clues that provide more information and clues to solving other letters or
symbols. In doing so, we can build a database of information that can fur-
ther help in finding more clues and ultimately the solution. For example,
it is easy to spot 0 and 9 in an equation because 0 added to any number is
the same number, and 9 added to a number is also the same number with a
carryover if there is a carryover from the previous column. Also, when two
decimal digits are added, the carryover is a maximum of 1, and thus an extra
digit in the sum means that digit is 1. Furthermore, when digits represent-
ing one of the two operands being added are from the extreme end of the
domain of possible values (e.g., 0, 1, or 9), it is relatively easier to decipher
the other operand if we know whether or not the sum results in a carryover.
Thus, we are able to logically determine results from given information [5].
In this chapter, we consider two well-known problems.
Problem 1:
SEND + MORE = MONEY
This is perhaps the most classic cryptarithmetic puzzle. It was intro-
duced by Henry Dudeney in Strand Magazine in 1924 [1, 6]. This problem
 Cryptarithms • 73

is also an alphametic, since the words in the puzzle form a sensible phrase
SEND MORE MONEY (Figure 20). One of the notable characteristics of
this problem is the extra digit in the sum (i.e., the M), because it is an
important clue. Another is that some of the letters, such as M, E, and N, are
repeated, so the sooner they are deciphered, the more quickly the problem
can be solved.

S E N D

+ M O R E

M O N E Y

Problem 2:
DONALD + GERALD = ROBERT (Clue: D = 5)
This is another well-known problem, and it has been the subject of
Newell and Simon’s study of human problem solving using cryptarithmetic.

D O N A L D

+ G E R A L D

R O B E R T

Note the hint that D = 5. This very helpful in finding more information
about all the variables connected to D by the different constraints—namely,
the T, the carryover to the second column, G, and R.

5.3 SOLUTION

Problem 1:
SEND + MORE = MONEY
Here is a solution that is very similar to many others available online.
Since the sum has one digit more than the two numbers being added, we
choose to start solving the problem from the left side. We will refer to
the columns as first, second, third, and so on, starting from the right side. As
we begin to find clues to the solution, we keep adding them to a Knowledge
Table and refer to it to obtain additional clues.
74 • Artificial Intelligence Problems and Their Solutions

c5 c4 c3 c2

S E N D

+ M O R E

M O N E Y

It is evident from the problem that M = 1, since it is a carryover from

the previous column and also because the sum cannot start with a leading 0.
Figure 21 demonstrates this.

FIGURE 21 Step 1.

Assuming there is no carryover from the third column, S + 1 ≥ 10, and

thus S can only be 9, which proves O = 0.
However, if there is a carryover from the third column, 1 + S + 1 ≥
10—that is, S + 2 ≥ 10, so S = 8, which gives O = 0, or S = 9, which gives
O = 1, which is not possible because M is already 1. Thus, if there is a car-
ryover, S = 8.
In any case, O = 0. Figure 22 illustrates how we reach conclusions for
O = 0, and S = 8 or 9.
 Cryptarithms • 75

Deductions

c5 = 1 → c4 + S + 1 greater than 9 1 c4 c3 c2

S E N D
S + 1 greater than 9 → S = 8 if c4 = 1 + 1 O R E
or S = 9 if c4 = 0
1 O N E Y
In both cases above, S = 8 or S = 9 → O = 0

Knowledge Unknowns 1 c4 c3 c2
Table S, E, N, D, R, Y
c2, c3, c4 S E N D
c5 = 1
M=1 + 1 0 R E
Remaining Numbers
O=0 2, 3, 4, 5, 6, 7, 8, 9 1 0 N E Y

S = 8 if c4 = 1 or
S = 9 if c4 = 0

FIGURE 22 Step 2.

FIGURE 23 Step 3.
76 • Artificial Intelligence Problems and Their Solutions

We continue by moving to the third column from the right. Looking at

Figure 23, we see that E + 0 = N, which is not possible because anything
added to zero is itself. This means c3 = 1. Therefore, E + 1 = N. The maxi-
mum that E can be is 9, and if E = 9, N = 0, which is not possible because
O = 0. This means E is not equal to 9 and E < 9. Thus, N < 10, c4 = 0, and
S = 9.
Continuing to the fourth column, we can see in Figure 24 that N + R = E
and c3 = 1, so N + R = E + 10. This equation is valid if c2 = 0.
In addition, we know that E + 1 = N. Substituting this in the above
equation, we get E + 1 + R = E + 10, so R = 9. But S = 9, so our assumption
that c2 = 0 column is incorrect. Hence, the correct equation is
1 + N + R = E + 10
That is, 1 + E + 1 + R = E + 10, which gives R = 8.

Deductions
1 0 1 c2
c3 = 1 → c2 + N + R = E +10
9 E N D
E + 1 = N → c2 + (E + 1 ) + R = E + 10 + 1 0 R E
→ c2 + R = 9
1 0 N E Y
If c2 = 0, R = 9, which is not possible
because S = 0 → c2 = 1

c2 = 1 → R = 8

Knowledge Table

M=1 Unknowns 1 0 1 1
O=0 E, N, D, Y 9 E N D
S=9
c4 = 0 + 1 0 8 E
c3 = 1 1 0 N E Y
Remaining Numbers
E+1=N
E<9 2, 3, 4, 5, 6, 7
c2 = 1
R=8

FIGURE 24 Step 4.
 Cryptarithms • 77

Continuing to the last column, we have D + E = Y, which generates a

carryover to the next column (see Figure 24). Thus, D + E ≥ 10. Now the
numbers 0, 1, 8, and 9 are taken.
Hence, 12 ≤ D + E ≤ 17
and D and E can be 2, 3, 4, 5, 6, or 7.
The following combinations of the above numbers have the sum
between 12 and 17
5 + 7 = 12
6 + 6 = 12
6 + 7 = 13
7 + 7 = 14

Deductions
c1 = 1 → D + E = Y > 10

The numbers 0, 1, 8, and 9 are taken 1 0 1 1

→ 12 ≤ D + E ≤ 17 9 5 6 D
+ 1 0 8 E
D ! = E → the following possible equations give the
solution: 1 0 6 5 Y
5 + 7 = 12

If E = 5, D = 7, Y = 2, and N = 6
6 + 7 = 13
This is the solution.

Knowledge Table
M=1
O=0
Unknowns
S=9 1 0 1 1
c4 = 0 9 5 6 7
c3 = 1
Remaining Numbers
E+1=N + 1 0 8 5
E<9 3, 4
1 0 6 5 2
c2 = 1
R=8
E= 5
N=6
D= 7
Y =2

FIGURE 25 Step 5.
78 • Artificial Intelligence Problems and Their Solutions

Since D is not equal to E (a constraint of the problem), only the first

and third equations are valid. Also, since E occurs more frequently in this
problem, let us try the values 5, 6, and 7 for E first (see Figure 25)
If we assign E = 5, D = 7, Y = 2, and N = 6, we can see that each of these
assignments are valid, and thus, have reached the solution.
Trying E = 6, we get D = 7, Y = 3, and N = 7, which is not possible
because D = 7.
Also trying E = 7, we get D = 6, Y = 3, and N = 8, which is not possible
because we have proven R = 8.
Similarly, trying D = 5 leads to E = 7, so N = 8, which is again invalid
because R = 8.
Thus, we conclude the problem is solved and there is no other solution.
Problem 2:
We name the carryovers c2, c3, . . . , c6 for the columns 2, 3, . . . , 6.

c6 c5 c4 c3 c2

D O N A L D

+ G E R A L D

R O B E R T

Deductions

D = 5 → T = 0 and c2 = 1 c6 c5 c4 c3 1
5 O N A L D
c2 = 1 → 1 + 2L = R + c3
+ G E R A L D
1+ 2L = R + c3 → 1 + 2L = R R O B E R T
→ R is odd or 1+ 2L=R + 10
Knowledge Table Unknowns
D=5 O, N, A, L, G, E, R, B c6 c5 c4 c3 1
T=0
c2 = 1 c3, c4, c5, c6 5 O N A L 5
R is odd + G E R A L 5
2L + 1 = R or
Remaining Numbers R O B E R 0
2L + 1 = R + 10
1, 2, 3, 4, 6, 7, 8, 9

FIGURE 26 Step 1.
 Cryptarithms • 79

Substituting the clue, we get T = 0 with a carryover to the second

column. Thus, c2 = 1 and 1 + L + L = R, so R is odd (see Figure 26).
Moving to the fifth column, O + E = O, which means if there is no
carryover, E = 0, and if there is a carryover, E = 9 (see Figure 26).
Since T is already 0, E = 9 and thus confirming a carryover from the
fourth column and a carryover to the sixth column—that is, c5 = c6 = 1.

FIGURE 27 Step 2.

Deduction
1 1 c4 c3 1
c3 + A + A = 9 or c3 + A + A = 19 → c3 = 1
If c4 = 1 → A = 9 5 O N A L 5
If c4 = 0 → A = 4 + G 9 R A L 5
E = 9 → A = 4 and c4 = 0 R O B 9 R 0

Knowledge Table Unknowns

D=5 1 1 0 1 1
O, N, L, G, R, B
T=0
5 O N 4 L 5
c2 = 1
R is odd + G 9 R 4 L 5
2L + 1 = R or Remaining Number
2L + 1 = R + 10 R O B 9 R 0
E=9 1, 2, 3, 6, 7, 8
c5 = 1, c6 = 1
c3 = 1, c4 = 0
A=4

FIGURE 28 Step 3.
80 • Artificial Intelligence Problems and Their Solutions

Since A + A = 9 (which is an odd number), there is a carryover from the

second to the third column (see Figure 28). Also, there is no carryover from
the third to the fourth column because it would mean A + A + 1 = 9, thus
making A = 9. But 9 is already taken. So there is no carryover from the third
to the fourth column, and thus A = 4.

Deduction
1+5+G=R→6+G=R
1 1 0 1 1
6 + G = R and R is odd → R = 7 or 9
5 O N 4 L 5
E=9→R=7
R=7→G=1 + G 9 7 4 L 5
R O B 9 7 0

Knowledge Table Unknowns

D=5 O, N, L, B 1 1 0 1 1
T=0
5 O N 4 L 5
c2 = 1
R is odd + 1 9 7 4 L 5
2L + 1 = R or Remaining Numbers
2L + 1 = R + 10 7 O B 9 R 0
E=9 2, 3, 6, 8
c5 = 1, c6 = 1
c3 = 1, c4 = 0
A=4
R=7
G=1

FIGURE 29 Step 4.

Moving to the sixth column, we have 5 + G = R (see Figure 29).

Because R is odd, R = 7 or R = 9. But E = 9, so R = 7, which makes
G = 1 because there is a carryover from the fifth column.
In the fourth column, N + 7 = B with a carryover to the fifth column
(see Figure 30). Also, since there is no carry from the third column, N can
be 2, 3, 4, 6, or 8. To generate a carryover, N has to be greater than 2. N = 3
or 4 makes B = 0 or 1, both of which are already assigned. N = 6 makes
B = 3, and N = 8 makes B = 5, which is again invalid. Thus, N = 6 and B = 3.
 Cryptarithms • 81

Deduction
N + 7 = B and c5 = 1 → N + 7 = B + 10 1 1 0 1 1
N + 7 = B + 10 → N = B + 3 5 O N 4 L 5
N = B + 3 → Of the remaining numbers, only N=
6 and B = 3 satisfy the equation. + 1 9 7 4 L 5

7 O B 9 R 0
Knowledge Table Unknowns
D=5
O, L
T=0
c2 = 1 1 1 0 1 1
R is odd 5 O 6 4 L 5
2L + 1 = R or Remaining Numbers
2L + 1 = R + 10 + 1 9 7 4 L 5
2, 8
E=9
7 O 3 9 7 0
c5 = 1, c6 = 1
c3 = 1, c4 = 0
A=4
R=7
G=1
N=6
B=3

FIGURE 30 Step 5.

Now we also have some new data for the second column: 1 + L + L = 7
(see Figure 31). Because there is a carryover to the third column, L = 8
makes the sum = 1 + 8 + 8 = 17. The only remaining digit is 2, and the
remaining alphabet is O. Thus, O = 2.

Knowledge Table
D=5 Deduction
T=0 1 1 0 1 1
1 + L + L = 17 → L = 8
c2 = 1 5 O 6 4 L 5
R is odd L=8→O=2
2L + 1 = R or + 1 9 7 4 L 5
2L + 1 = R + 10
7 O 3 9 7 0
E=9 Unknowns
c5 = 1, c6= 1
c3 = 1, c4 = 0 The solution is as below:
A=4
R=7 Remaining Numbers 1 1 0 1 1
G=1 5 2 6 4 8 5
N=6
+ 1 9 7 4 8 5
B=3
L=8 7 2 3 9 7 0
O=2

FIGURE 31 Step 6.
82 • Artificial Intelligence Problems and Their Solutions

It is evident from the above two solutions that we also have computed
the value of the carryovers, which are also known as auxiliary variables.
Even though auxiliary variables are not directly a part of the solution, they
help in the problem-solving process.
Let us consider Problem 1: SEND + MORE = MONEY.
The variables are S, E, N, D, M, O, R, and Y.
The domain for these variables is 0 to 9.
The auxiliary variables are c1, c2, c3, and c4 (the carryovers).
The domain for auxiliary variables is 0, 1.
The constraints are as follows:
All the variables represent different numbers.
The numbers do not start with digit 0.
S = c4
c3 + S + M = O + 10*c4
c2 + E + O = N + 10*c3
c1 + N + R = E + 10*c2
D + M = Y + 10*c1
If we conduct an exhaustive search for the solution through the prob-
lem state space, we soon realize that it is a herculean task, because if a
problem consists of 10 different variables, there are 10! = 3,628,800 pos-
sible assignments [4]. However, as observed previously, if we can decipher
some of the variables using the clues inherent in the problem, such as an
extra digit in the sum or a 0 or 9, we can reduce the number of possible
combinations—but only to a certain extent.
Alternatively, we can formulate the problem as a search tree by consider-
ing one of the values found initially as the initial state and the goal state, being
the situation where the value of every variable in the solution is known [4].
We can also divide the problem into subproblems. In this case, we can think
of a column as a subproblem and consider solving one subproblem at a
time by computing the values of the variables in that column. However, it is
futile to consider a column in isolation because the carryovers from the sum
in one column affect the sum in the next column, and thus the columns are
connected to each other. This relationship between the columns—or, more
 Cryptarithms • 83

specifically, between the variables—is evident from the list of constraints and
can be more clearly seen via a constraint hypergraph, which George Luger
developed for this problem [4]. Therefore, we have to consider the con-
straints that connect one column to the other. If we look closely, it is the aux-
iliary variables (i.e., the carryovers) that connect two columns. Fortunately,
the domain of these auxiliary variables is very small, because the values are
either 1 or 0. Thus, if we correctly identify these auxiliary variables, we may
be able to prune the search tree and reduce the search effort and time. Addi-
tionally, we can assume one of the legal values for the auxiliary variable and
find the values for the remaining variables in a particular column.
This heuristic, which involves considering the variables’ fewest possible
values, is called the minimum remaining values (MRV) heuristic [4]. Thus,
after applying the “generate and test technique” to the variables with MRV,
the domain of the other variables in the column is also reduced. We also
consider the second possible value for the auxiliary values and compare the
resulting domains of the variables in the column.
For example, in Problem 1, consider column 4 after substituting M =
c4 = 1:
c3 + S + 1 = O + 10
If c3 = 1, S + 1 = O + 10
Thus, S + 1 ≥ 10, and the only possible legal value for S is 9 and for
O is 0.
Similarly, if c3 = 1, S + 2 ≥ 10, and the only possible legal value for S is
8 and for O is 0.
Now the question is whether or not the order in which the two possible
values of c3 needs to be considered. In this case, the order does not matter
because the domain of S is reduced to only one legal value (i.e., 8 or 9) in
both cases (for both possible values of c3)—that is, 1 or 0, respectively. In
cases where testing a probable value for a variable results in a domain that
is too small or null for its neighboring variable, such a variable is considered
last. This heuristic, choosing the value that rules out the least possible values
for the neighboring variables, is called the least constraining value (LCV) [4].
Thus, when using the MRV heuristic, we can find the value of O and
can reduce the domain of S to two values: 8 and 9. This prevents us from
having to test all other values from the domain for the variables, which
would later fail anyway.
84 • Artificial Intelligence Problems and Their Solutions

After choosing the LCV* for a variable, we can start deriving possi-
ble legal values for the remaining variables and proceed further down the
search tree toward the solution. This technique is called forward checking.
Continuing this way, if we were to reach a dead end or a state where no
legal values are left for a variable, we trace back to the last assignment. This
technique is called backtracking.

5.4 HUMAN PROBLEM SOLVING

If we analyze the solution to Problem 1, we realize that it starts by assign-

ing the value 1 to M, which is deduced from constraints 2 and 3. Then in
column 4, the auxiliary variable c3 is chosen as the MFV,** since its domain
is smaller compared to the other two variables. Thus, we are left with only
two possible values for the letter S and only one value for O. Following the
substitution of O = 0 in column 3, the relation between E and N is found,
which helps find the value of R. After deciphering all the remaining auxil-
iary variables, the domain of the variables in the first column is reduced to
just three numbers, due to an exhaustive search using a combination of the
three values, and the application of constraints leads to the solution.
The solution to Problem 2 starts with a similar path, substituting the clue
and deducing the values/parities of the related variables. The solution pro-
ceeds to finding the value 9 and the values of most of the auxiliary variables.
Also, the columns are observed one by one, substituting the knowledge gained.

5.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

Overall, each solution we found follows the same general concept: show
(at least) the completely encrypted and decrypted cryptarithm and give a
step-by-step description of how the solution was found.
When analyzing the solutions we found Table (10), we can see that they
are all represented using text. This is because a mathematical arithmetic
problem, along with any reasoning behind it, is easily understood by words

* least constraining value

** minimum remaining values
 Cryptarithms • 85

rather than pictures or diagrams. Therefore, the choice of representation is

less likely to play a role in determining the best solution. Rather, the organi-
zation, detail, and description of steps play a major factor in identifying the
solution that would be considered the most compatible.
The majority of the solutions are extensional, mainly consisting of meth-
ods to solve the SEND MORE MONEY problem and the DONALD
GERALD ROBERT problem. They are considered extensional because
they explain the step-by-step procedure for solving either one of these two
cryptarithms in their entirety. The intensional solutions found, such as the
“Primer Solution,” summarize the possible arithmetic constraints and tech-
niques for solving any cryptarithm, even for those that involve arithmetic
other than addition [7].
The rankings for both executability and comprehensibility are generally
mixed. Since each solution uses text as a means representation, they were
analyzed more profoundly, especially in terms of organization, detail, and
description of steps, as mentioned before. However, it can be seen that
the intensional solutions were generally ranked lower. This is most likely
because they use general arithmetic techniques that one must remember
to solve cryptarithms instead of the solution to a cryptarithm being readily
available to the problem solver, as in the case of extensional solutions.
Not surprisingly, each solution uses constraint satisfaction as a problem-
solving technique. This method, along with the help of solving for subgoals,
is almost crucial to cryptarithmetic. As for the intensional solutions, they
use constraint satisfaction accompanied with the use of brute force. This is
because, unlike the SEND MORE MONEY problem, cryptarithms are not
generally unique. Therefore, it is possible that a single letter or symbol can
represent multiple possible values, and not a specific digit. To solve these
types of cryptarithms, brute force must be used.
Finally, it is also important to note that the solutions’ flexibility is ranked
fairly high. Since they are represented using text, these solutions can be
presented in many different ways.

5.5.1 The Most Human Window–Compatible Solution

The solution assessed as the MHWC is the “Knowledge Table Solution.”

This solution is the same solution presented in Section 5.3. This solution is
a simple, yet detailed solution for the SEND MORE MONEY cryptarithm.
86 • Artificial Intelligence Problems and Their Solutions

Difficulty
8/10

Complex.
O((n – (constrs + N/A))!)

Int or Intly/ Prob. Solv. Mode of

Name Rep. HW? Corr? Grn Sz Exec Compr Flex Opt? Tot
Ext? Extly Mthd Conv
Txt
Constr.
Ariels Ext. 6/10 Txt Y Y Ideal 7/10 6/10 9/10 Editor; Y 28/40
Satisf.
Hand
Constr.
Code/ Txt
Code Int. 7/10 Y Y Ideal 6/10 4/10 Satisf.; Brute 8/10 N 25/40
Algrthm Editor
Force
Txt
Constr.
Donald Ext. 9/10 Txt Y Y Small 5/10 5/10 9/10 Editor; Y 28/40
Satisf.
Hand
Txt
Knowledge Txt w/ Constr.
Ext. 8/10 Y Y Ideal 10/10 10/10 9/10 Editor; Y 37/40
Table Pics Satisf.
Hand
Txt
Txt w/ Constr. Editor;
M and Cs Ext. 8/10 Y Y Ideal 9/10 8/10 8/10 Y 33/40
Tables Satisf. Tbl Mkr;
Hand
Txt
Pavalli Ext. 9/10 Txt Y Y Ideal 7/10 9/10 Constr. Satisf 9/10 Editor; Y 34/40
Hand
Constr. Txt
Y w/ Vry
Primer Int. 4/10 Txt Y 6/10 5/10 Satisf.; Brute 8/10 Editor; N 23/40
Cycles Large
Force Hand

Key
Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?
Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?
Rep.: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is this solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 10 Ranking Cryptarithmetic Solutions According to the Human Window.

 Cryptarithms • 87

We could consider it the MHWC for several reasons. One is because it

is divided into steps, making the solution highly comprehensible. Another
reason is because the solution divides the columns of the equation into sub-
problems, making it highly executable.
A third reason is because the equation is reiterated consistently at every
step, showing the current values at that particular step in the solution and
columns that are being analyzed. In addition, the symbols and values of
the carryovers are displayed as well. This prevents the need to memorize
or write down the values of the letters and carryovers and keeps the reader
engaged.
A fourth reason is the level of detail of its description. This solution has
just the right amount of detail in that it explains exactly why the letters are
assigned their particular values. Also, the language of the description is very
explicit, so even those without an advanced mathematical background can
properly understand the solution.
The final, and most substantial, reason as to why it is considered the
MHWC solution is because of the use of the Knowledge Tables. These
tables display the letter values, deductions, and unknowns that remain at
each particular step. As the solution proceeds further, more elements are
added or removed from these tables. These tables are significant because
it is a compact and concise way of organizing data. This makes the solution
very comprehensible and very executable.

5.5.2 The Least Human Window–Compatible Solution

The solution assessed as the LHWC solution is the “Donald Solution”

(Figure 32). This is a solution to the DONALD GERALD ROBERT cryp-
tarithm, with a representation similar to one presented in [8]. The solu-
tion consists of equivalence functions that implicitly and briefly show the
potential values of the letters and carryovers at each step. Although the
solution is organized as steps, it shows how those values are obtained with
only a brief mathematical explanation and a small amount of description.
These factors could potentially lower both the executability and compre-
hensibility of the solution, especially for those who are not familiar with
cryptarithms. Nonetheless, this solution may be suitable for those who are
highly mathematically and/or logically inclined.
88 • Artificial Intelligence Problems and Their Solutions

DONALD + GERALD = ROBERT

Solution:

c1 c2 c3 c4 c5
DONALD
GERALD
----------------
ROBERT
Steps:
1) T is even (D + D = an even number)

2) D 5 c5 = 1
D 4 c5 = 0
3) L 5 c4 = 1
L 4 c4 = 0

4) A 5 c3 = 1
A 4 c3 = 0

5) c5 = 1 R is odd
c5 = 0 R is even (L + L)
R is not 0 (R and L cannot both be 0)
6) c4 = 1 E is odd
c4 = 0 E is even (A + A = an even number)
E not 0 (E and A cannot both be 0)
7) c2 = 0 E=0
c2 = 1 E = 9 and c3 + N + R > 9

8) c1 + D + G 9 (no carryover)
R D
R G
...
...
...
...
...

FIGURE 32 The “Donald Solution” (Shortened).

Another feature that could be an issue with this solution is the fact that,
unlike the “Knowledge Table Solution,” the cryptarithm is not reiterated in
between steps. This further reduces the amount of guidance provided to the
reader and thus may further reduce the comprehensibility of the solution.
 Cryptarithms • 89

Finally, you may notice that the solution begins by analyzing the letters
in the rightmost column as opposed to the leftmost column. As seen in the
previous solutions in this chapter, we usually begin by analyzing the let-
ters in the leftmost column and then moving toward the rightmost column.
This approach is more efficient because it is likely to determine the values
of the carryover variable earlier in the process. However, beginning from
the rightmost column may result in more steps, which is the case in the
“Donald Solution.” Therefore, this may also reduce the executability of the
solution.

5.6 BEST MACHINE SOLUTION

Up to this point, we have only analyzed cryptarithms whose values con-

tained up to 10 possible values (i.e., digits 0 to 9). That is, we have been
solving for equations whose values are in base 10, which is essentially the
number system we all use every day. However, the possibility of using dif-
ferent bases has most certainly been considered.
For a machine, determining if a cryptarithm has a solution is NP-
complete when generalized to arbitrary bases [9]. This means there is no
known algorithm that can guarantee a solution to a cryptarithm in an accept-
able amount of time as the number of unknown digits (base) increases (a
solution can be verified in polynomial time, however). For any cryptarithm
whose values are of base n, the worst-case runtime for finding that solution
is O(n!). This is the case when all n digits are used in the equation and when
there are no constraints that can be concluded by the use of mathematical
arithmetic rules. However, it is more than likely that cryparithmetic prob-
lems will not contain every one of its n values and will contain mathematical
constraints of some kind. Thus, the average running time would be
O( (n – (constraints + unused digits))! ).
Fortunately, most cryptarithmetic problems that are analyzed tend to
contain letters with values in base 10 (e.g., the SEND MORE MONEY
problem), since this is the base we are used to. For cryptarithms in base
10, we replace n with 10 and thus find that the worst-case running time is
O(10!) and the average case is O( (10 – (constraints + unused digits)) ! ).
Although this greatly reduces the problem space, 10! is still relatively
large. For a machine, one of the most efficient algorithms for finding
90 • Artificial Intelligence Problems and Their Solutions

solutions to cryptarithms is a parallel genetic algorithm. This algorithm

is essentially the same as a typical genetic algorithm (i.e., the “Code
Solution” [10]) but is executed with the aid of multiple processors. In an
experiment done at Shahid Chamran University, the time required to solve
a cryptarithm with a parallel genetic algorithm increased proportionately as
the number of unknown variables increased. After doing this experiment
on a base 10 cryptarithm containing all 10 variables, it took 3.5 seconds to
find the solution using a parallel generic algorithm, whereas the depth-first
algorithm took a little more than an hour [11].

5.7 PLAYABLE PROGRAM

Today, there are many apps available on iPhone and Android phones
that have kept the interest in cryptarithmetic alive. With the advent of
modern computers a few decades ago, people started publishing computer-
generated cryptarithms that were too complex to be solved by humans,

Image published with permission from Thomas L. Peterson, Eximietate, Inc.

http://www.letsgetwordv.com/#am_app.

FIGURE 33 Alphabetic App.

 Cryptarithms • 91

leading to a decrease in interest in these games [12]. Cryptarithms that

fall inside the Human Window are fun to solve and help in improving our
problem-solving skills. However, keeping track of the knowledge table, the
latest state of the problem, and the different variable possibilities using
strictly a pen and paper can be a tedious job. This is where playing apps
and computer games containing cryptarithms can help. We can choose the
variables’ values based on logic and heuristics, while the computer can do
the complex calculations and memory management. One of the apps can be
found at http://www.letsgetwordy.com/#am_app (Figure 33).

5.8 REFERENCES

1. Logicville. (2012). Cryptarithm. Available at http://www.logicville.com/

cryptarithm.htm. Accessed on January 8, 2014.
2. Wikipedia. (2012). Simon Vatriquant. Available at http://fr.wikipedia.org/wiki/
Simon_Vatriquant. Accessed on January 8, 2014.
3. Soares, J. (2012). The Sphinx Memorial. http://cry ptarithms.awardspace.us/
sphinx.html. Accessed on January 8, 2014.
4. Luger, G.F. (2008). Artificial Intelligence: Structures and Strategies for Com-
plex Problem Solving 6th ed., Reading, MA: Addison-Wesley.
5. Polk, T.A., and Newell, A. (1995). “Deduction as Verbal Reasoning.” American
Psychological Association (102) 3: 533–536.
6. Elitmus Guide. (2012). Guide to Solve Cryptic Multiplication Question.
Available at http://elitmusguide.blogspot.in/2012/07/guide-to-solve-cryptic-
multiplication.html. Accessed on January 8, 2014.
7. Soares, J. (2002). A Primer on Cryptarithmetic. Available at http://cryptarithms.
awardspace.us/primer.html. Accessed on January 8, 2014.
8. Moran, D.B. (2013). Name Addition: DONALD + GERALD = ROBERT.
Available at http://www.dougmoran.com/examples/puzzle-name-addition.txt.
Accessed on January 8, 2014.
9. Eppstein, D. (1987). “On the NP-Completeness of Cryptarithms.” SIGACT
News 18 (3): 38–40. Available at http://www.ics.uci.edu/~eppstein/pubs/Epp-
SN-87.pdf.Accessed on January 8, 2014.
92 • Artificial Intelligence Problems and Their Solutions

10. Pachnekar, T (2011). Cryptarithmetic. Available at http://www.codeproject.

com/Articles/176768/Cryptarithmetic. Accessed on January 8, 2014.
11. Abbasian, R., and Mazloom, M. (2009). Solving Cryptarithmetic Problems
Using Parallel Genetic Algorithm. Available at Available at http://staff.science.
uva.nl/~masoudm/Papers/mazloom,milad.pdf. Accessed on January 8, 2014.
12. Soares, J. (2003). Will Cryptarithmetic Survive Innovation? http://cryptarithms.
awardspace.us/survival.htm. Accessed on April 22, 2013.
 CHAPTER

6
THE RED DONKEY PUZZLE

FIGURE 34 The Red Donkey Puzzle.

6.1 BACKGROUND

Apart from being known as a variation of Dad’s Puzzler, which is

perhaps the earliest-known puzzle of its kind [1], The Red Donkey
Puzzle (Figure 34) is credited as the third most sold sliding block puzzle
(after The 15 Puzzle and Dad’s Puzzler) in the book Sliding Piece Puz-
zles by L. Edward Hordern [2]. The earliest appearance of this puzzle was
recorded in France, where it was known as L’Âne Rouge (The Red Donkey)
and enjoyed great popularity. The English version of the puzzle, known
as The Red Donkey, had surfaced in the early 1930s [1], while in Poland
94 • Artificial Intelligence Problems and Their Solutions

it was known as Klocki (wooden blocks) and in Japan as Hakoiri musume

(Daughter in the Box) [3].
The Red Donkey Puzzle is a type of sliding block puzzle that consists
of several tiles (or blocks) bounded by a set area in which the tiles must
be organized into a specific arrangement. Specifically, The Red Donkey
Puzzle is bounded by a 4” × 5” rectangular area and consists of 10 blocks:
a 2” × 2” square block (known as the Donkey piece), four 1” × 2” vertical
blocks, a 2” × 1” horizontal block, and four 1” × 1” square blocks. The con-
straints on the puzzle are as follows:
The blocks can only slide within the bounded rectangular area.
The blocks cannot be lifted from the area to be replaced.
The arrangement of the blocks in the initial state of the puzzle is shown
in 35 [4]. In its initial state, the 2” × 2” block, which is the largest block, is at
the top center of the bounded area, and the goal is to move it to the bottom
center. We have also numbered the blocks so they can be easily referenced
and the solution is clearly understood.

FIGURE 35 The Red Donkey Puzzle.

6.2 SOLUTION

In this section, we study a solution consisting of 81 moves, similar to

the one provided by Martin Gardner in the March 1964 issue of Scientific
American. For The Red Donkey Puzzle, any solution that reaches the goal
 The Red Donkey Puzzle • 95

state in 81 moves is considered optimal, as this is known to be the least

number of possible moves required to solve the puzzle. An online game
(available at www.bsswebsite.me.uk) is used to play the puzzle and to show
some of the important stages in the solution. Here is the solution with num-
bers representing blocks:
10 (halfway), 6, 5, 7 (down), 4
9 (halfway), 7, 4, 5, 8 (up, left)
10, 4, 8 (right, down), 5, 10
8, 6, 4, 9, 7, 5, 8 (down, left), 4, 6, 3
2, 1, 10, 8, 4, 1, 2, 3, 6, 7, 8 (left, up), 5, 1, 4, 7
2, 10, 8 (up, right), 7, 4
1, 9 (left, down), 2, 10 (down, left), 3
6, 2, 10, 8 (halfway), 7, 4, 1, 9, 10 (down), 2
6, 3, 7, 8, 4, 1, 2, 8, 7, 3
6, 8 (right, up), 5, 10, 9
2, 7, 8, 5, 9 (up, right), 2
After some trial and error and after gaining some understanding of the
problem, we realize that the empty space created by moving the unit square
blocks must be shaped as either a horizontal or vertical rectangle to allow
the larger blocks to move. Thus, the small square blocks are more flexible
when compared with the larger blocks, since they can be used to create
space for the larger blocks, either in a horizontal or vertical direction. Fur-
thermore, one of the key observations is that the horizontal block in the
middle is obstructing the movement of the donkey block and the vertical
blocks. Hence, it would be a good idea to move this block away from the
middle to the side. Thus, one of the important subgoals of the problem
would be to move the horizontal block toward a corner or a side, away from
the middle, enabling free movement of the other blocks.
The following moves lead to the state shown [1] in Figure 36:
10 (halfway), 6, 5, 7 (down), 4, 9 (halfway), 7, 4 and 5
In this state, we are able to move block 5 to a side. After five more
moves, block 5 moves a row down with the help of the small blocks as shown
in Figure 37 and after move 21, block 5 is in the lowest row (Figure 38).
96 • Artificial Intelligence Problems and Their Solutions

FIGURE 36 Step 9. FIGURE 37 Step 14. FIGURE 38 Step 21.

We now have space to move the vertical blocks so block 2 can move lat-
erally, as seen in Figure 39 after the 26th move. However, the small square
blocks need to be adjacent to block 2 to help it move because the vertical
blocks still restrict block 2’s movement. The small square blocks are adja-
cent to the donkey block in Step 31 (Figure 40).

FIGURE 39 Step 26. FIGURE 40 Step 31. FIGURE 41 Step 41.

By Step 41 (Figures 41 and 42), the horizontal block has moved to the
bottom right corner, allowing vertical blocks 4 and 1 to move further down,
making space for one of the two remaining unit square blocks to move up.
It is very important to have one of the single unit square blocks remain at
the bottom to avoid creating a unit square space, which can hinder move-
ment of the larger blocks. With the three small square blocks on top, verti-
cal blocks 4 and 1 can be moved to the left side so the donkey block can
move a row down, as shown in Step 48 (Figure 43). An observation of the
current state reveals that the donkey block is still constrained by the vertical
 The Red Donkey Puzzle • 97

blocks on both sides, and it is necessary to move them to the top to enable
the donkey block to move freely farther down.

FIGURE 42 Step 41. FIGURE 43 Step 48. FIGURE 44 Step 52.

Steps 48 to 71 concentrate on moving all the vertical blocks to the top,

which is another important subgoal. This will clear the donkey block’s path
toward its goal at the bottom. After Step 48, the small blocks 8 and 10 can
be arranged to create vertical space, enabling block 6 to move to the top.
Then the donkey block can conveniently slide to its right to make space for
the vertical blocks 1 and 4 to move to the top, as seen in Step 52 (Figure
44). At this point, block 10 can be moved back to the bottom. As seen in
Step 60 onward (Figures 45 to 49), block 6 has to move down temporarily
to get unit square blocks 7 and 8 down while the donkey block slides left.

FIGURE 45 Step 60. FIGURE 46 Step 67. FIGURE 47 Step 71.

With all four of the small square blocks adjacent to block 2 in

Step 71, the donkey block can slide to the bottom row by displacing
98 • Artificial Intelligence Problems and Their Solutions

blocks 5, 9, and 10. The last step is achieved in the 81st move, when
the donkey block moves right to the center.

FIGURE 48 Step 76. FIGURE 49 Step 81.

6.2.1 Bidirectional Search

If we know the goal state, a bidirectional search is helpful by search-

ing from both the initial state and the goal state to meet halfway when a
solution path exists.

FIGURE 50 The Initial State. FIGURE 51 The Goal State.

The initial and goal states are shown in Figures 50 and 51, respectively.
A careful observation of the goal state suggests that a significant milestone
while working backward is when the donkey block is moved to the left and
one row up, while the horizontal block moves to the bottom right. This is
the state represented in Step 71 in the previous solution.
 The Red Donkey Puzzle • 99

FIGURE 52 Subgoal 1 While Searching Backward from the Goal State.

Moving back to the initial state, we can consider the state in Figure 52
as our intermediate goal state and work toward it. Thus, our new subgoal
in our forward search is to move the donkey block and the horizontal
block two rows below their current positions. The second of the two sub-
goals, if achieved optimally, can be done in 21 steps, as shown in 52.
Thus, the search technique of mean-ends analysis can be used to reduce
the difference between the current state and the goal state [4, 5]. Kose
(2012) solved the Klotski puzzle using the bidirectional search in 116
moves [6].

6.3 HUMAN PROBLEM SOLVING

This problem is too complex to be solved in 20 minutes, especially with-

out any simulation tool like an actual game or a playable program. There-
fore, some of the solutions available online were studied instead of the ones
done by a class of students. From our experience with this problem, the
puzzle is very difficult, and without a strategy, one gets stuck at dead ends,
resulting in a fair amount of backtracking.
The solution studied in Section 6.2 divides the problem into several
subgoals:
■ Moving the horizontal block to the bottom to make space for the don-
key block to move to the center
100 • Artificial Intelligence Problems and Their Solutions

■ Moving the small square blocks to the top to help move the vertical
blocks out of the donkey block’s pathway to the center
■ Moving all the vertical blocks to the top once the donkey block is in
the center

6.4 HUMAN WINDOW ANALYSIS OF SOLUTIONS

Although The Red Donkey Puzzle was once extremely popular, we

were unable to find many solutions for our analysis. The solutions we have
found, however, are very similar to one another.
When analyzing 11 that the solutions are all ranked similarly to one
another in almost all fields. This is because, as the rankings for flexibility
will show you, there are only so many ways The Red Donkey Puzzle can be
represented. This is because the puzzle has fixed block sizes and shapes,
which are key features that every representation must have.
From Table 11, it could be seen that each solution is represented exten-
sionally. This is because they each use images of the actual puzzle and show
every block movement made to achieve the end state. For a problem as
difficult as The Red Donkey Puzzle, it is better to show the actual puzzle
instead of describing the solution using words. As a result, the extensional-
ity and executability of each solution are relatively the same, and they each
use the same problem-solving technique of visual representation.
The attribute with the biggest difference in ranking is comprehensibility.
While each solution uses the same representation, they all contain additional
properties that set them apart from one another including block colors, num-
bering, and pictures. Therefore, the comprehensibility of a solution may be a
definitive factor as to whether or not it should be considered MHWC.

6.4.1 The Most Human Window–Compatible Solution

The solution we consider the MHWC is the “Donkey Solution” (53).

An example of this type of representation can be found in [7]. Instead of
showing each of the 81 steps, it has been shortened, displaying only the
important subgoals. This solution has many qualities that make it highly
comprehensible. Each tile is represented with a different shading and
number, making them easily distinguishable from one another. This applies
 The Red Donkey Puzzle • 101

Difficulty
9/10

Complex
O(bd/2)

Int or Intly/ Prob. Solv. Mode of

Name Rep. HW? Corr? Grn Sz Exec Compr Flex Opt? Total
Ext? Extly Mthd Conv
Vis. Rep.;
Donkey Ext. 6/10 Pics Y Y Ideal 9/10 9/10 7/10 Img Mkr Y 31/40
Subgoal
Vis. Rep.;
Gnome Ext. 7/10 Pics Y Y Ideal 9/10 5/10 5/10 Img Mkr Y 26/40
Subgoal
Vis. Rep.; Img Mkr;
Hollow Ext. 5/10 Pics Y Y Ideal 8/10 4/10 5/10 N 22/40
Subgoal Hand
Vis. Rep.;
Klajok Ext. 7/10 Pics Y Y Ideal 9/10 6/10 5/10 Img Mkr Y 27/40
Subgoal
Vis. Rep.;
Neon Ext. 9/10 Pics Y Y Ideal 9/10 8/10 5/10 Img Mkr Y 31/40
Subgoal

Key
Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?
Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?
Rep.: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is this solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 11 Ranking The Red Donkey Puzzle Solutions According to the Human Window.

especially to the donkey piece, which has an actual image of a donkey on it.
Additionally, above each image in the solution is a number that represents
its corresponding step number. Finally, the quality that greatly increases the
overall comprehensibility is the inclusion of the arrows. The arrows indicate
which tile was moved from the previous state to get to the current state and
in how many units it moved. This prevents the need to compare pictures for
two consecutive steps to determine which tile has been moved. The picture
of the current step can simply be viewed, and the arrangement of the previ-
ous step can be determined by it.
102 • Artificial Intelligence Problems and Their Solutions

Initial State

Subgoal 1: Move the horizontal piece to the bottom.

Subgoal 2: Move two 1" × 1" tiles adjacent to the donkey.

 The Red Donkey Puzzle • 103

Subgoal 3: Move one 1" × 1" tile to the bottom and the donkey two rows down.

Subgoal 4: Move the vertical tiles to the top.

Subgoal 5: Move the donkey to the goal position.

FIGURE 53 The “Donkey Solution” (Shortened).

104 • Artificial Intelligence Problems and Their Solutions

6.4.2 The Least Human Window–Compatible Solution

The solution that was assessed as LHWC is the “Hollow Solution”

(Figure 54). When analyzing this solution, it can be seen that the images
that represent the puzzle are simply thin black lines that are put together
in such a way as to make an image of the puzzle, including the tiles and the
border. Although the tiles and border are spaced out efficiently enough
to differentiate themselves from one other, the fact that all the tiles look
almost the same can make them indistinguishable from one another, pos-
sibly reducing the comprehensibility of the solution. This applies especially
when trying to identify the tiles that were moved from a previous step to
achieve the arrangement of the current step. Additionally, the background
color is the same as the tile color. This produces the effect that the tiles
are hollow and can make the image look like a cluster of thin black lines,
thus making the puzzle components even more indistinguishable from one
another.
Initial State

Subgoal 1: Move the horizontal piece to the bottom.

 The Red Donkey Puzzle • 105

Subgoal 2: Move two 1" × 1" tiles adjacent to the donkey.

Subgoal 3: Move one 1" × 1" tile to the bottom and the donkey two rows down.

Subgoal 4: Move the vertical tiles to the top.

106 • Artificial Intelligence Problems and Their Solutions

Subgoal 5: Move the donkey to the goal position.

Goal
FIGURE 54 The “Hollow Solution” (Shortened).

6.5 BEST MACHINE SOLUTION

Currently, no special heuristics can be used to solve The Red Donkey

Puzzle. Similarly, there are no known algorithms that can solve the puzzle
in the smallest number of possible steps. However, the two best-known
solutions to solving the puzzle are the bidirectional search [6] and the sim-
ple breadth-first search.
The Red Donkey Puzzle has a relatively large state space, consisting of
25,955 possible block arrangements, along with 47,151 possible moves [8].
When analyzing these values in terms of a tree structure, the possible tile
arrangements can be seen as the tree’s nodes, while the tile movements can
be seen as the arcs. Since The Red Donkey Puzzle can have a maximum
of three possible moves per tile arrangement, the branching factor of the
tree produced by a search is very small. However, given the large number
of possible moves and tile arrangements, the depth produced can become
quite large.
In his paper “Solving Klotski,” Karl Wiberg shows the benefits of using
the breadth-first search to solve The Red Donkey Puzzle. Just like any
implementation of the breadth-first search, it analyzes every node in a spe-
cific depth k of the tree before analyzing every node in depth k + 1 in the
tree. This search can find the solution in 116 steps [9].
Another, more sophisticated, search method that can be used to solve
The Red Donkey Puzzle is the bidirectional search, which Erdal Kose
developed. Interestingly enough, this search was able to find the solution
 The Red Donkey Puzzle • 107

in 116 steps, just like the breadth-first search [6]. However, this search is
more sophisticated because it greatly reduces the number of nodes needed
to be visited and expanded. This is because both ends of the search start
with a parent node and meet each other at a common middle step (see
Section 6.2).
Since solutions are found efficiently through the use of search algo-
rithms, the complexity of The Red Donkey Puzzle can be evaluated in
terms of a search tree. In the worst case, the runtime and space needed for
the breadth-first search is O(bd), where b is the branching factor and d is
the depth of the tree. For the bidirectional search, however, the runtime
and space needed is O(bd/2) [6]. This shows that the bidirectional search is
more efficient than the breadth-first search, even though they find the solu-
tion in the same number of steps.

6.6 PLAYABLE PROGRAM

Many playable programs for The Red Donkey Puzzle are available
online, where you can play the puzzle, become familiar with the problem,
and learn to differentiate between good and bad moves.
It would be extremely difficult to memorize all the movements, so you
would have to draw all the moves on a piece of paper. However, drawing
the search tree can be difficult and tedious and would require a lot of space.
This is why being able to solve the puzzle on a computer is preferable.
Physical puzzles, such as those made of cardboard or wooden blocks, and
virtual puzzles, such as games on a computer or a smart phone, are only
some examples. These are the most extensional forms of representation of
the problem, as the problem solver can physically or virtually manipulate
the problem states. An online game can be found at http://www.bsswebsite.
me.uk/Puzzlewebsite/Reddonkeypuzzle/Reddonkeypuzzle.htm.

6.7 REFERENCES

1. Gardener, M. (1964). “Mathematical Games.” Scientific American, March: 133.

2. Brandeis. (2013). Red Donkey. Available at http://www.cs.brandeis.
edu/~storer/JimPuzzles/ZPAGES/zzzRedDonkey.html. Accessed on January 8,
2014.
108 • Artificial Intelligence Problems and Their Solutions

3. Bsswebsite. (2013). Red Donkey Puzzle. Available at http://www.bsswebsite.

me.uk/Puzzlewebsite/Reddonkeypuzzle/Reddonkeypuzzle.htm. Accessed on
January 8, 2014.
4. Sweller, J. (1988). “Cognitive Load During Problem Solving: Effects on
Learning.” Cognitive Science 12: 257–285.
5. Newell, A., and Simon, H.A. (1972). Human Problem Solving. Englewood
Cliffs, NJ: Prentice-Hall.
6. Kose, E. (2012). Comparing AI Search Algorithms and Their Efficiency When
Applied to Path Finding Problems. Ph.D. Thesis. New York, NY: CUNY, The
Graduate Center, Department of Computer Science.
7. Whiting, R. (2009). Red Donkey. Available at http://home.comcast.net/~l-whiting/
attbi/Donkey_solution.html. Accessed on January 8, 2014.
8. Blazewicz, D. (2011). Red Donkey. Available at https://sites.google.com/site/
klajok/klotski/huarong-trail/red-donkey. Accessed on January 8, 2014.
9. Wiberg, K. (2009). Solving Klotski. Available at http://www.treskal.com/kalle/
klotski.pdf. Accessed on January 8, 2014.
 CHAPTER

7
THE 15 PUZZLE

FIGURE 55 The 15 Puzzle.

7.1 BACKGROUND

The 15 Puzzle (Figure 55) was invented by Noyes Chapman, a post-

master who resided in Canastota, New York [1]. It is another example of
a sliding block puzzle. However, unlike The Red Donkey Puzzle, which
consists of different-sized tiles, The 15 Puzzle’s tiles are all square and are
the same size. It is currently known as the earliest sliding puzzle created.
In 1874, Chapman developed a puzzle that involved “16 numbered
blocks that were to be put together in rows of four.” The goal of the puzzle
was to arrange these blocks so the numbers on a row of blocks added up
110 • Artificial Intelligence Problems and Their Solutions

to 34. In 1879, a new and improved version of The 15 Puzzle spread through-
out the United States after Chapman’s son introduced the puzzle [1]. Even-
tually, the puzzle was popular all over the United States, and by 1880, it
had spurred a puzzle craze. The puzzle eventually made its way to Canada
and then to Europe. By July 1880, however, its popularity had begun to
wane [2].
At one point Chapman unsuccessfully attempted to patent the puzzle.
This is most likely because it was too similar to Ernest Kinsey’s “Puzzle-
Blocks” which was patented in 1878 [1].
The invention of The 15 Puzzle is often incorrectly accredited to
American chess player, chess composer, and recreational mathematician
Samuel Lloyd. Until his death, Lloyd insisted that he had invented the
puzzle in 1891, and at one point, he revived its popularity when he intro-
duced The 14-15 Puzzle, which was very similar, except two of the tiles
were switched (see Section 7.2). However, despite these claims, it was in
fact Chapman who invented The 15 Puzzle [1].
The 15 Puzzle consists of 15 square tiles numbered 1 to 15, bounded in
a closed 4” × 4” area. The 15 tiles fill up 15 of the spaces in the area, while
the 16th space is left empty so each tile has limited space to move. The goal
is to reposition the squares so they are placed in numerical order by row.

7.2 PROBLEM-SOLVING TECHNIQUES

Unlike the problems we have analyzed so far, The 15 Puzzle does not
have a specific predefined initial setup. Therefore, any arrangement of the
tiles, including the goal arrangement, can be considered an initial state.
There are 16! * (2.09 × 1013) arrangements in which the tiles can be
placed. However, according to a study by Woosly Johnson and William Story,
there are only 16!/2 possible positions in which the goal setup can actually
be achieved. In 1879, Johnson and Story researched The 15 Puzzle and
discovered that all puzzles with arrangements of even parity are solvable,
whereas those of odd parity are not [3]. That is, if the number of tile inter-
changes (or permutations) needed to transform a particular arrangement
into a solvable arrangement is even, then that particular arrangement itself
is solvable. If it is odd, it is unsolvable. More than a decade after this dis-
covery, Sam Lloyd presented his variation of the problem to the public
 The 15 Puzzle • 111

that involved solving his version of the 15 Puzzle whereby titles numbered
14 and 15. The 14-15 Puzzle consists of one interchange, is of odd parity,
and therefore is unsolvable [4].
Although the number of possible positions is lower than expected, 16!/2
(1.05 × 1013) possible initial tile arrangements is still a very large problem!
However, finding the solution to The 15 Puzzle can be very straightforward.
The most valuable problem-solving technique to use is solving subgoals.
The subgoal is to achieve a completed row, and once that is accomplished,
one can go on to the next step. However, as you will see (in Section 7.3),
there are special situations, such as solving the fourth row, in which the
completed third row must be uprooted.
Generally speaking, The 15 Puzzle is a type of n puzzle, with the value
n representing an integer greater than 1. The rules and goal of any n puzzle
are the same as The 15 Puzzle. The only difference is that it consists of n
square tiles numbered from 1 to n placed in a closed sqrt(n +1) × sqrt(n + 1)
area, with the nth location left as an empty tile space. However, no matter
how big the problem, the same problem-solving technique can be used. It
may just take more moves to complete. In this book, we perform our analy-
sis solely on The 15 Puzzle.

7.3 SOLUTION

A very common solution to The 15 Puzzle involves focusing on and solv-

ing individual tiles one by one in numeric order. Doing so will not only solve
the puzzle in an orderly fashion, but it will also make it possible to solve for
subgoals, as mentioned in Section 7.2. Specifically, a subgoal is the comple-
tion of a row. With the exception of the final step, once a row is solved, none
of the tiles in the row need to be moved again.
Since there are so many possible initial tile arrangements, this solution
will show general ways to solve each tile and therefore will not be based on
an initial arrangement.

7.3.1 Solving the First Row

The first step is to move the 1 tile to the upper left corner of the board.
This is a relatively simple step to achieve. Since the other tiles have not
been properly positioned, it does not matter how they are moved as long
112 • Artificial Intelligence Problems and Their Solutions

as the 1 tile is placed in its correct position. Let’s say the puzzle is arranged
like in Figure 56.

FIGURE 56 Initial State.

The black square represents the empty tile space, and the gray squares
represent tiles that we will not consider at this point.
First, arrange the tiles so the empty space is near the 1 tile (Figure 57).

FIGURE 57 Move the Empty Space to the 1 Tile.

Now, move the 1 tile to the upper left corner (Figure 58).
 The 15 Puzzle • 113

FIGURE 58 Move the 1 Tile to the Upper Left Corner.

Basically, the general idea behind moving a target tile is to move all
temporarily disregarded tiles around it until the empty space is in the gen-
eral direction of the correct position relative to the target tile. Then use that
empty space to move the target tile toward its goal position. Continue this
until the target tile is in the correct position.
Figure 58 shows the general idea behind moving a tile throughout
the board. From now on, future diagrams will only show moves that are
114 • Artificial Intelligence Problems and Their Solutions

significant to positioning a specific tile. To refresh yourself on how to move

a tile around the board, refer back to Figure 58.
Next, we move the 4 tile to the upper right corner of the puzzle board.
Just as we did with the 1 tile, this tile can be easily positioned by simply
moving the temporarily disregarded tiles around. Position it as shown in
Figure 59.

FIGURE 59 The 4 Tile Positioned.

Now we position the 2 tile to the right of the 4 tile. Although the 4 and
2 tile positions are very close to each other, the 4 tile does not need to be
disturbed from its correct position in order to position the 2 tile next to it.
This can be done by moving the 2 tile and the empty spot to any position
in the center “square” located in between the 1 and 4 tiles. For example, a
sufficient tile arrangement would be as shown in Figure 60.

FIGURE 60 The Center “Square.”

 The 15 Puzzle • 115

The center “square” we are referring to is indicated by the bolded bor-

ders. Now simply perform the movements in Figure 61.

FIGURE 61 Positioning the 2 Tile.

At this point, you have probably noticed that we have not placed the
2 tile in its correct position. However, temporarily placing it here will allow
us to easily position the 3 tile.
Now, arrange the 3 tile and empty slot so they are in the positions shown
in Figure 62.

FIGURE 62 3 Tile Preferred Position.

This should be very easy, since the 1, 2, or 4 tiles do not have to be

moved to reach this state. Now perform the moves in Figure 63.
116 • Artificial Intelligence Problems and Their Solutions

FIGURE 63 Positioning the 3 Tile.

7.3.2 The Corner Technique

After positioning the 2 tile, the first row might look like Figure 64.

FIGURE 64 3 Tile Special Case.

In this arrangement, the technique used in Figure 63 cannot be used.

Instead, make sure the empty spot is located underneath the 2 tile and
perform the moves in Figure 65.
 The 15 Puzzle • 117

FIGURE 65 Positioning the 3 Tile (Special Case).

Now rotate the tiles in the “square” so they are arranged as in Figure 66.

FIGURE 66 1 and 2 Tiles Over the Corner.

118 • Artificial Intelligence Problems and Their Solutions

Then, reposition tiles 1 and 2 back in place, as in Figure 67.

FIGURE 67 Reposition the 1 and 2 Tiles.

We have now completed what we call the corner technique. It is named

this due to the fact that we moved the 1 and 2 tiles over the corner of the
board in order to create room for the 3 tile to move into position. This
technique will be used again to complete row 4.
So now we have completed the first row. Furthermore, none of the
tiles in the top row need to be moved again and will remain in place for the
remainder of the solution.

7.3.3 Solving the Second Row

The next goal is to solve the second row. To do so, follow the same tech-
niques that were used to solve the first row. Move the 5 tile to the upper
left, the 8 tile to the upper right, the 6 tile to the left of the 8 tile, and use
the same techniques that were used to solve the 3 tile to solve the 7 tile
(Section 7.31 and Section 7.3.2). If done correctly the puzzle should look
like Figure 68.

FIGURE 68 Row 2 Solved.

 The 15 Puzzle • 119

At this point, the tiles in the second row do not need to be repositioned
again. We are now halfway done, leaving us with two more rows to solve. As
you can see, the amount of space is being reduced, and there is less room to
move tiles. Positioning the last seven tiles will be a little trickier.

7.3.4 Solving the Third Row

Moving on, the next goal is to solve the third row. First, move the 9 tile
to the leftmost position, and then move the 12 tile to the rightmost position,
as in Figure 69.

FIGURE 69 Position the 9 and 12 Tiles.

Next, we position both the 10 and 11 tiles simultaneously. To do this, it

is important that the 10 and 11 tiles and the empty slot are located in the
center square. Figure 70 shows a possible arrangement.

FIGURE 70 Acceptable Center “Square.”

120 • Artificial Intelligence Problems and Their Solutions

If this were the case, simply rotating the tiles in the “square” twice
would complete the third row. However, this will certainly not always be
the case. For example, you may have a position like the one in Figure 71.

FIGURE 71 10 Tile Positioned Badly.

Notice how we have the empty space below the 9 tile. This would be
acceptable if one of the temporarily disregarded tiles was to the right of it,
but unfortunately that is not the case. With this arrangement, it would be
impossible to solve the puzzle without moving the 9 tile. Follow the steps
in Figure 72.

FIGURE 72 Moving the 9 Tile.

Now, rotate the tiles in the center square so one of the temporarily
disregarded tiles is next to the 9 tile and the empty slot is above it, as shown
in Figure 73.
 The 15 Puzzle • 121

FIGURE 73 Rotate the Center “Square.”

Then perform the moves in Figure 74.

FIGURE 74 Obtaining the Sufficient Center “Square.”

Now we have obtained our requirement. Similarly, if the empty slot

began under the 12 tile, perform a reflection of the above moves on
the right side of the puzzle. Basically, it is vital that the 9 and 12 tiles
have temporarily disregarded tiles underneath them to complete the
third row.
As you can see, it is also possible that the 10 and 11 tiles are inconve-
niently placed occupying each other’s target positions. Does this look famil-
iar? Simply use the “corner technique” to properly position them. If done
correctly the third row will be complete, as shown in Figure 75.
122 • Artificial Intelligence Problems and Their Solutions

FIGURE 75 Third Row Complete.

7.3.5 Solving the Fourth Row

We have only one row to deal with. At this point, the puzzle is either
solved or positioned incorrectly, as in Figure 76.

FIGURE 76 Incorrect Last Row.

Let us assume the tiles in the last row are positioned incorrectly. Until
this point, we have followed a basic strategy that prevented the need to
displace any tiles that belong to a completed row. To proceed in finding the
solution, that rule must be broken. To deal with this arrangement, we begin
by once again using the corner technique. First move the 9 and 10 tiles so
they are aligned vertically to each other. This involves moving the 9 tile into
the last row and moving the 10 tile into the position where the 9 tile should
be (Figure 77).
 The 15 Puzzle • 123

FIGURE 77 Shifted 9 and 10 Tiles.

Then do the same to the 11 and 12 tiles (Figure 78).

FIGURE 78 Shifting the 11 and 12 Tiles.

At this point, the tiles in the center “square” should only be those that
belong to the final row (i.e., 13, 14, and 15). Arrange them as in Figure 79.

FIGURE 79 Rotate the Center “Square” Tiles.

124 • Artificial Intelligence Problems and Their Solutions

Finally, perform the moves in Figure 80.

FIGURE 80 Solving the Last Row.

At last, the puzzle is solved!

7.4 HUMAN WINDOW ANALYSIS OF SOLUTIONS

Like The Red Donkey Puzzle, we have not found many solutions for
our Human Window analysis of The 15 Puzzle. Additionally, the majority
of them are identical to the solution discussed in Section 7.3 and to one
another. Therefore, they each use similar problem-solving methods and
representations. Each solution demonstrates a general method in which
descriptions are used, along with visual images as guidance, to solve The 15
Puzzle regardless of its initial arrangement. Therefore, the solutions can be
considered intensional because of this generality. However, they are ranked
relatively low on their intensionality because they are partially extensional
(i.e., diagrams).
 The 15 Puzzle • 125

Difficulty
4/10

Complex
O(bd/2)

Name Int or Intly/ Rep. HW? Corr? Grn Sz Exec Compr Prob. Solv. Flex Mode of Opt? Total
Ext? Extly Mthd Conv
Txt Editor;
Descr. Y w/
In-Place Int. 3/10 Y Small 5/10 7/10 Sub-goals 7/10 Img Maker; N 22/40
w/ Vis. Cycles
Hand
Txt Editor;
Descr. Vis. Repr;
Jurgen Int. 4/10 Y Y Ideal 4/10 6/10 8/10 Img Maker; N 22/40
w/ Vis. Sub-goals
Hand
Txt Editor;
Optimal Ext. 3/10 Text Y Y Ideal 9/10 5/10 Vis. Repr. 9/10 Y 26/40
Hand
Txt Editor;
Descr. Vis. Repr;
Real Int. 9/10 Y Y Ideal 9/10 7/10 7/10 Img Maker; N 32/40
w/ Vis. Sub-goals
Hand
Txt Editor;
Descr. Vis. Repr;
Tile Int. 5/10 Y Y Ideal 10/10 9/10 7/10 Img Maker; N 33/40
w/ Vis. Sub-goals
Hand
Descr. Vis. Repr; Txt Editor;
W-how Int. 5/10 Y Y Ideal 6/10 8/10 7/10 N 26/40
w/ Vis. Sub-goals Img Maker

Key
Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?
Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?
Rep.: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is this solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 12 Ranking The 15 Puzzle Solutions According to the Human Window.

When analyzing the solutions’ rankings for comprehensibility and exe-

cutability (Table 12), we can see that they are generally mixed. Therefore,
it is the features that affect these ranking categories that determine which
solution should be considered the MHWC.
126 • Artificial Intelligence Problems and Their Solutions

The only intensional solution that uses a somewhat different technique

than the others is the “In-Place Solution.” The subgoals required to solve it
are the same, except that it guarantees that no tile that reaches its correct
position will move again [5].
Finally, the only solution classified as extensional is the “Optimal Solu-
tion.” which is discussed further in Section 7.4.2.

7.4.1 The Most Human Window–Compatible Solution

The solution considered the MHWC is the “Tile Solution,” which is the
same as the one in Section 7.3. It provides a step-by-step method on how
to solve The 15 Puzzle with any initial arrangement. At each step, a simple
description is given on what tiles to move and how to do it. Provided at
each step is a picture of the puzzle itself, which contain tiles that are either
shaded or numbered. The tiles that are shaded are unimportant at a par-
ticular step and can be ignored, while the numbered tiles are important or
have been correctly placed. These images are used to show what the puzzle
would look like after the execution of each step. Finally, arrows are used to
show how certain tiles were moved at a step. These brief, but to the point,
directions, along with the images, make the solution very comprehensible
and executable.
To improve this solution even further, actual images of the puzzle
can be used instead of a grid or a table. This will make the solution even
more extensional. A solution that uses this approach is the “Real-Puzzle
Solution” [6].

7.4.2 The Least Human Window–Compatible Solution

The solution assessed as the LHWC is the “Optimal Solution” (Figure 81).
A representation similar to this was developed by B. MacKenzie [7]. It is
referred to as such because it demonstrates how to solve The 15 Puzzle in
the least number of possible steps. Optimal solutions may be feasible for
machines by clever or precise algorithms, but for people, finding optimal
solutions to nontrivial problems is usually very difficult, if not impossible.
Finding an optimal solution to The 15 Puzzle depends on the initial
arrangement of the tiles. Of all 16!/2 possible initial tile arrangements, only
 The 15 Puzzle • 127

The Solution

1 5 9 13
2 6 10 14
3 7 11 15
4 8 12 X
X = vacant position (empty tile space)
Moves (direction of X):
1. L1 11. R1 21. R3 31. R2 41. R2
2. U1 12. U1 22. D1 32. U1 42. D2
3. R1 13. R2 23. L3 33. R1 43. R1
4. U2 14. U1 24. D1 34. D2
5. L3 15. L3 25. R2 35. L3
6. D3 16. D1 26. U2 36. U1
7. R2 17. R3 27. L1 37. R3
8. U2 18. U1 28. D3 38. D2
9. L2 19. L3 29. L1 39. L3
10. D2 20. U1 30. U2 40. U2

Key
L = left
R = right
U = up
D = down

FIGURE 81 The “Optimal Solution.”

16 can be solved in 43 steps with multi-tile movements, which is considered

optimal. Otherwise, the minimal number of single-tile moves required to
solve the puzzle can average at around 80 steps [7]. However, as we have
seen, optimal solutions will generally have little to do with the Human Win-
dow. Although this solution demonstrates a very efficient method in solving
The 15 Puzzle in the least number of steps, we will only be analyzing the
solution’s representation.
128 • Artificial Intelligence Problems and Their Solutions

This section analyzes 1 of the 16 possible arrangements that can

be solved in 43 steps, which is optimal for The 15 Puzzle. This solution
can be considered the LHWC for many reasons. One is because of its rep-
resentation. Everything in this solution is in text, including the puzzle itself.
The puzzle is represented as a collection of numbers that indicate the tiles.
They are organized in such a way as to represent the actual puzzle (i.e.,
there are four rows and four numbers to a row). However, this may not be
as comprehensible as an actual image of the puzzle itself.
Another reason this can be the LHWC is because of the way the steps
are carried out. Each step consists of a letter and a number. The letter
represents the direction in which the tiles should be moved relative to the
empty space, and the number represents the number of tiles that must
move in that direction. For example, “U2” means the first two tiles that are
underneath the current empty space must move up into that space. This
representation can be confusing, especially since there are no representa-
tions of the puzzle that correspond to each step. As a result, a person would
most likely not be able to determine which step to backtrack to if he or she
made a mistake and would have to start over.
Finally, it can be considered LHWC because this solution is exten-
sional. For this solution, all the steps that solve this puzzle from beginning
to end are shown, so it is classified as extensional. While the solution is very
executable (since it provides the steps to solve an entire puzzle arrange-
ment), this is a solution to only one specific tile arrangement. If a person
were to attempt to solve The 15 Puzzle with a different arrangement, he or
she may not have the required knowledge to do so.

7.5 BEST MACHINE SOLUTION

In the field of artificial intelligence, solutions to The 15 Puzzle, and the

n puzzle in general, are usually modeled and studied with the application
of search algorithms aided by heuristics. Although simple searches such as
the breadth-first search would be able to find an optimal solution to The
15 Puzzle, the amount of space and time required to search every possible
arrangement is incredibly large, even with the bidirectional search. There-
fore, it is best to employ searches that use heuristics as a guide and to limit
bad choices.
 The 15 Puzzle • 129

Heuristics give the search algorithm an estimate as to where the goal

may be, greatly reducing the number of puzzle positions that need to be
searched. The most commonly used search to find an optimal solution to
The 15 Puzzle is the A* search equipped with an admissible heuristic.
The A* Search combines the uniform cost search, which measures the
number of steps it takes to reach a specific puzzle arrangement from the
initial arrangement, with an admissible heuristic, which prevents the search
from overestimating the goal [8]. These two properties, if practically opera-
tive, enable the A* search to find an optimal solution.
The A* search can use several admissible heuristics to solve The 15
Puzzle. One is the best-first heuristic, which maintains an open node
queue and pursues the node at the front of the queue first but retains in
memory the other nodes on the queue to pursue other possible paths [9].
Another is the Manhattan distance heuristic, which uses taxicab geom-
etry to measure how far a specific tile is from its goal position. These
heuristics, however, are slow and may still generate too many states, con-
suming a lot of memory [8, 9]. One solution to this problem is the use of the
iterative-deepening-A* search. This search reduces the number of nodes
normally searched by the A* search, thus reducing the need for space [8].
This search, however, is still slow.
Another, and better, heuristic that can be used is a disjoint pattern
database heuristic. This works by finding the sum of the “number of
needed moves to well order tiles” in a pattern of a given data structure. This
structure is a priority minimum queue combined with a hash table. With
this heuristic, fewer than 100,000 nodes are expanded, and any 15 Puzzle
arrangement can be solved in less than 10 seconds [8].
Like The Red Donkey Puzzle, the worst-case runtime and space
required to solve any arrangement of the puzzle is in the order of O(bd/2)
when using a bidirectional search. Given that the number of tiles that can
move is limited, the branching factor is small. However, given the high
number of possible arrangements, the depth can become very large. For-
tunately, since the A* search with the disjoint pattern database heuristic
greatly reduces the number of nodes expanded, a solution can be found
with a complexity less than that of O(bd/2).
For the general n puzzle, finding a solution to a legal tile arrangement
can be done in polynomial time. However, the problem of always finding an
optimal solution to a particular arrangement is NP-Hard [10].
130 • Artificial Intelligence Problems and Their Solutions

7.6 PLAYABLE PROGRAM

A playable 15 Puzzle game is available at http://migo.sixbit.org/puzzles/

fifteen/. With a simple click, the player can move the tiles around into the
empty space. Additionally, the player can perform multi-tile moves as well.
Finally, there is an option to attempt the 14-15 Puzzle if you so choose.
Good luck with that!

7.7 REFERENCES

1. Slocum J., and Sonneveld, D. (2006). The 15 Puzzle. Beverly Hills, CA:
Slocum Puzzle Foundation
2. Slocum, J., and Weisstein, E.W. (2013)15 Puzzle. Available at http://mathworld.
wolfram.com/15Puzzle.html. Accessed on January 9, 2014.
3. Johnson, W.W., and Story, W.E. (1879). “Notes on the ‘15’ Puzzle.” American
Journal of Mathematics 2 (4): 397–404.
4. Grant, H. (2002). The Fifteen Puzzle. Available at http://www.math.ubc.
ca/~cass/courses/m308-02b/projects/grant/fifteen.html. Accessed on January 9,
2014.
5. Yates, J. (2013). Fifteen Puzzle Solution: How to Solve the Famous 15 Puzzle.
Available at http://www.chessandpoker.com/fifteen-puzzle-solution.html.
Accessed on January 9, 2014.
6. Schmidt, W. (2013). Wayne Schmidt’s 15-Puzzle Solving Page. Available at
http://www.waynesthisandthat.com/15puzzle.htm. Accessed on January 9,
2014.
7. Norskog, B. (2013). The Fifteen Puzzle Can Be Solved in 43 ‘Moves.’”
Available at http://cubezzz.dyndns.org/drupal/?q=node/view/223. Accessed on
January 9, 2014.
8. Chiang, K-Y. (2013). Finding Optimal Solution of 15 Puzzle. Available at
http://www.cs.utexas.edu/~kychiang/files/Finding_Optimal_Solution_of_15_
puzzle.ppt. Accessed on January 9, 2014.
 The 15 Puzzle • 131

9. Lucci, S., and Kopec, D. (2013). Artificial Intelligence in the 21st Century.
Dulles, Virginia: Mercury Learning Inc.
10. Ratner, D., and Warmuth, M.K. (1986). Finding a Shortest Solution for the
N × N Extension of the 15-PUZZLE Is Intractable. National Conference on
Artificial Intelligence. Santa Cruz, CA: University of California Santa Cruz.
 CHAPTER

8
THE KNIGHT’S TOUR PROBLEM

FIGURE 82 The Knight’s Tour Problem.

8.1 BACKGROUND

A knight’s tour is a sequence of moves made by a chess knight, starting

from any initial square on a chessboard (usually on the lower left corner),
such that each square on the board is visited exactly once. While doing so,
it must be taken into consideration that the knight is limited to the moves it
would normally make on a chess board (i.e., the “L” shape).
The Knight’s Tour Problem (Figure 82) entails finding the correct
sequence of steps for the knight that will complete a knight’s tour. The
problem is normally solved on a standard 8” × 8” grid, similar to that of a
134 • Artificial Intelligence Problems and Their Solutions

standard chessboard. However, knight’s tours have been studied on general

m × n grids.
The concept of The Knight’s Tour Problem dates back to around
2200 BC, but the earliest-known reference is around AD 900 in India. It
was first introduced in a Sanskrit work on poetics titled Kavyalankara, by
the Kashmirian poet Rudrata. In Kavyalankara, the pattern of The Knight’s
Tour was presented as an elaborate poetic figure called the turagapada-
bandha, which translates to “arrangement in the steps of a horse.” The
knight’s tour was “apparently presented by a series of syllables on the squares
that make sense when read in the sequence of the tour.” The same verse in
four lines of eight syllables can be read by following the path of the knight
on tour, with each syllable being thought of as a square on a chessboard [1].
Over the years, The Knight’s Tour Problem evolved as mathemati-
cians began to take an interest. Leonhard Euler and H. C. Von Warnsdorff
helped to develop notable solutions [2].
In general, there are two kinds of knight’s tours. One is the open tour,
which simply involves the knight visiting each square only once. The other
is a closed (or reentrant) tour, which is like an open tour, except the knight
must be able to go back to the square it started on after visiting all of the
squares on the chessboard. This is a variation of a famous NP-Complete
problem called The Hamiltonian Cycle Problem. The point is that for a
given graph, it is impossible to determine apriori if a Hamiltonian Cycle of
that graph (e.g., a closed knight’s tour in this case) is possible.
Mathematician Allen Schwenk proposed a proven theorem that an
m × n chessboard with m ≤ n has a closed knight’s tour unless one or more
of the following three conditions hold: m and n are both odd; m = 1, 2, or 4;
or m = 3 and n = 4, 6, or 8 [3].

8.2 PROBLEM-SOLVING TECHNIQUES

Many solutions to The Knight’s Tour Problem have been discovered

on an 8” × 8” board with the use of several problem-solving techniques.
One of them is exhaustive enumeration, but that would be very inefficient
because an 8” × 8” board contains 4 × 1051 possible sequences. A more effi-
cient problem-solving technique is divide and conquer, in which the chess-
board is divided into smaller boards that are solved individually. The most
 The Knight’s Tour Problem • 135

efficient method is solving subgoals, and it is also the most commonly used
technique. The solutions we have found use some of the more popular sub-
goal methods.
One method is a relatively straightforward method discovered by
Abraham de Moivre. When starting at a corner square, the knight will first
visit all of the squares around the edges of the board and then visit all of the
remaining squares in the center. The heuristic is that at every choice point,
the knight visits the square closest to the edge (i.e., the square with the
smallest in-degree) first. A similar method is used in the solution presented
in this chapter (see Section 8.3).
A second popular method is square and diamonds. This solution essen-
tially divides the chessboard into four 4” × 4” subboards. In each of these
subboards, four patterns are formed with the knight during the tour: a
rhombus, a square, and a mirror image for each. This method is used in the
“D&S Solution” [4].
A third method is tessellation. The basis of this method is to move the
knight in a path that would trace the shapes of four polygons that, when
combined together, form a tessellation. A tessellation is a group of shapes
that can fit perfectly with one another on a plane without overlapping. With
this method, the knight can start at any square on the board and form the
polygon that the square is a part of. After completing that shape, the knight
can simply move to another square that is a part of one of the other three
polygons and travel along a path to trace its shape. The tour is complete
after the knight traces the shapes of the four polygons. If done correctly,
the tessellation method can create a closed tour. This method is used in the
“Tessellation Solution,” which was developed by Daniel E. Thomasson [5].

8.3 SOLUTION

This section presents a relatively simple solution, with steps and rules
that are easy to follow and remember. It is based on the method developed
by Abraham de Moivre, in which the knight travels on the outer edges of
the chessboard. Using this method, we can still achieve a closed tour. How-
ever, there are specific points in the solution where the heuristic must be
broken. This approach is similar to one developed by Andrew Joseph, a
former student of Dr. Kopec’s.
136 • Artificial Intelligence Problems and Their Solutions

A closed knight’s tour can be achieved by making the knight follow two
general rules:
Rule 1: Stick as close to the outer edges of the board as possible.
Rule 2: When the opportunity presents itself, corner squares must be
taken.
For this solution, our knight starts at the lower left corner of the board
(Figure 83).

FIGURE 83 Initial State (Visiting Square 1).

The first general idea is to go in a spiral motion around the board, which
is something the knight will be doing for most of the solution. We will start
by traversing the outer squares of the board in a clockwise motion (Figures
84 and 85).

FIGURE 84 Visiting Squares 2–4. FIGURE 85 Visiting Squares 5 and 6.

At this point, we have the opportunity to move into the upper right-
hand corner of the chessboard. We will use this opportunity and have the
knight move there (Figure 86).
 The Knight’s Tour Problem • 137

FIGURE 86 Upper Right Corner (Square 7).

Now proceed with Rule 1 as normal (Figure 87).

FIGURE 87 Visiting Squares 8–11.

Landing on square 11 is one of two very important visits to take note

of, as the rules will now be broken temporarily. Make the moves shown in
Figure 88.

FIGURE 88 Break the Rules—First Time (Visiting Squares 12–14).

138 • Artificial Intelligence Problems and Their Solutions

Although this set of moves does not help the knight visit the lower
right corner of the board, visiting squares 12 to 14 will help the knight with
future moves. For this “favored” solution, we recommend that the problem
solver simply memorize this exception, as it will be necessary when we make
a similar maneuver at square 36. Until then, continue to follow the rules
(Figures 89 to 92).

FIGURE 89 Visiting Squares 15 and 16. FIGURE 90 Visiting Squares 17–19.

FIGURE 91 Visiting Squares 20–22. FIGURE 92 Visiting Squares 23–25.

Again, we have an opportunity to visit a corner square. So we will take

it (Figure 93).

FIGURE 93 Lower Right Corner (Square 26).

 The Knight’s Tour Problem • 139

Now proceed to follow Rule 1. However, we will not continue moving

in a clockwise spiral at this point. Instead, the knight will now start to tra-
verse the board in a counterclockwise motion (Figures 94 and 95).

FIGURE 94 Visiting Squares 27–29. FIGURE 95 Visiting Squares 30 and 31.

We now have the opportunity to fill the final corner on the upper left of
the board. We will take advantage of it (Figure 96).

FIGURE 96 Upper Left Corner (Square 32).

Then we proceed (Figure 97).

FIGURE 97 Visiting Squares 33–36.

140 • Artificial Intelligence Problems and Their Solutions

Now that we have reached square 36, we will once again break the
rules. Move the knight to the squares shown in Figure 98.

FIGURE 98 Breaking the Rules—Second Time (Visiting Squares 37 and 38).

At this point, continue to follow the rules, with the knight once again
traversing the outer squares of the board clockwise, as shown in Figures 99
to 102.

FIGURE 99 Visiting Squares 39–41. FIGURE 100 Visiting Squares 42–44.

FIGURE 101 Visiting Squares 45–47. FIGURE 102 Visiting Squares 48–50.
 The Knight’s Tour Problem • 141

At this point, we have completed the knight’s traversal around the

board’s outer squares. You will notice that one of the squares has not been
visited. This is because it has been reserved so the knight can complete a
closed tour.
Now the knight must visit each of the remaining center squares. The
general idea now is to maintain subsequent moves in a clockwise direction
(Figures 103 to 105).

FIGURE 103 Visiting Center Squares 51–54. FIGURE 104 Visiting Center Squares 55–58.

FIGURE 105 Visiting Center Squares 59–61.

The rest is now straightforward (Figure 106).

FIGURE 106 Goal State (Visiting Squares 62–64).

142 • Artificial Intelligence Problems and Their Solutions

And thus we have completed our knight’s tour. Additionally, since we

are able to move the knight to its first position from its last position, it is a
closed tour.

8.4 HUMAN PROBLEM SOLVING

In the Spring of 2013, The Knight’s Tour Problem was presented to

several college students. They were not observed when they were given
the problem, so it is uncertain whether or not the students acquired their
solutions by fair means. The majority of the students took an approach simi-
lar to that of the “Simple Solution” (see Section 8.5.1). They began by mov-
ing the knight around the outer squares of the board. Then, when they hit
a point where they could not continue, they traversed some center squares
so they could reach an outer square and continue traveling to the outer
squares once again. Finally, after all the outside squares were visited, they
moved the knight to the center and visited the remaining squares.
One student used his own approach after discovering a solution after
some “trial and error.” As explained in his description, his main approach
was to “first eliminate the corner boxes [and] then the middle boxes.” How-
ever, he did not follow this heuristic specifically. First, he traversed the
outside squares, making any attempt to fill the corners. Then he entered
the center and filled random corners until every square had been visited.
Such an approach may work for the open tour (visit every square without
consideration of where we started and ended), but it is not likely to work for
the closed tour, since transition squares to extreme edge squares may well
have already been visited.

8.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

We found several solutions for The Knight’s Tour Problem. This is no

surprise, since there are so many methods to finding a tour and it is a rela-
tively popular problem. When analyzing the solutions (Table 13), we can
see that there is a mix between extensional and intensional solutions. This
is because solutions to The Knight’s Tour Problem can be expressed in two
possible ways: a solution can provide the general steps of a method that is
 The Knight’s Tour Problem • 143

Difficulty
5/10

Complex
O(n)

Prob.
Int or Intly/ Mode of
Name Rep. HW? Corr? Grn Sz Exec Compr Solv. Flex Opt? Tot
Ext? Extly Conv
Mthd
Y w/ Bcktrck
Algorithm Int. 8/10 Code Y Ideal 8/10 7/10 7/10 Txt Editor N 30/40
cycles w/ Heurist
Txt
Vis. Repr.
D&S Int. 6/10 Y Y Ideal 9/10 6/10 Subgoals 7/10 Editor; Y 28/40
w/ Descr.
Img Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Green Ext. 8/10 Y Y Ideal 5/10 7/10 8/10 Y 28/40
Img Subgoals Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Narayan Ext. 7/10 Y Y Ideal 8/10 7/10 7/10 Y 29/40
Img Subgoals Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Semi-Magic Ext. 7/10 Y Y Ideal 4/10 7/10 7/10 Y 25/40
Img Subgoals Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Simple Ext. 8/10 Y Y Ideal 8/10 9/10 8/10 Y 33/40
Img Subgoals Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Tessellation Ext. 8/10 Y Y Small 10/10 7/10 8/10 Y 33/40
Img Subgoals Mkr
Grid/Table; Vis. Repr.; Img/Tbl
Wiki Ext. 7/10 Y Y Ideal 8/10 7/10 8/10 Y 30/40
Img Subgoals Mkr

Key
Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?
Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?
Rep.: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is this solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 13 Ranking The Knight’s Tour Problem Solutions According to the Human Window.
144 • Artificial Intelligence Problems and Their Solutions

used to find a knight’s tour, classifying it as intensional, and a solution can

also be represented by showing how a method is used to find a tour step-
by-step, using actual diagrams of a chessboard and a knight, classifying it as
extensional.
From looking at Table 13, we can see that the majority of the solutions
are extensional. Additionally, they are ranked relatively high in extension-
ality. For an extensional Knight’s Tour solution, the more it appears as an
actual chessboard, the higher the extensionality. We can see that each of
them uses similar problem-solving methods and representations as well.
They also rank relatively identical in terms of comprehensibility. The
comprehensibility of the solutions depends on the amount of detail used for
their representation (e.g., chessboard, path, numbering). From the rank-
ings, it can be seen that no solution is highly comprehensible. However, it
is possible that, when combining specific attributes from each solution, the
MHWC representation can be produced.
Finally, we can see that the rankings for the executability of the solutions
are generally mixed. It is apparent that the executability of the solu-
tions depends on which method is used to find a knight’s tour. While there
are so many methods to use to find a tour, some can be carried out much
more easily than others. Therefore, one of the factors that may determine
if a solution is the MHWC is its executability, and thus finding the “best”
method for solving The Knight’s Tour Problem.

8.5.1 The Most Human Window–Compatible Solution

The solution we analyzed to be the MHWC solution is the “Simple

Solution.” This solution is the same as the one presented in Section 8.3.
This solution is very comprehensible. It is represented using a grid, which
makes the state space look like a chessboard. Also, like on a chessboard, the
colors of the squares alternate between dark and light, which makes each
square identifiable from others. Additionally, the entity that represents the
knight is a figure of the actual knight piece. This is a reminder that the solu-
tion must follow the same move restrictions as a knight. This is especially
helpful to who are those unfamiliar with chess. After each step, a line that rep-
resents the path the knight has taken so far is drawn, connecting the knight’s
previous positions to its current position. These lines remain for every step
of the solution. The final image in this solution summarizes the entire tour. It
 The Knight’s Tour Problem • 145

shows the entire path the knight took, as well as numbers on each square that
indicate the sequence of the moves the knight made in completing the tour.
Although the solution is comprehensible, it is possible that the repre-
sentation can be improved to increase its overall comprehensibility. For
example, another attribute that could be added are letters and numbers to
indicate the names of rows and columns, respectively, akin to those of an
actual chessboard. This can make the solution more traceable, and specific
sequences and squares can be traced, memorized, or just identified.
What makes this solution the MHWC is the method used. Not only is
the method highly executable, but it requires little to no memory to carry
out, which is ideal for most human beings. Although there are a few rules to
follow and some locations in which the heuristic must be broken, they are
relatively easy to remember.
You may notice from Table 13. that the “Tessellation Solution,” devel-
oped by Daniel E. Thomasson (see Section 8.2), has higher executability
than the “Simple Solution.” This is because the tessellation method not only
produces a closed tour, but it also guarantees that a tour will be produced
after moving the knight around the board so it traces the four polygons. No
matter where the knight starts on the board, if those four shapes are traced,
a tour will be created. Thus, the executability of the solution is very high.
Unfortunately, to carry out this method, a decent amount of memory is
required, as one would have to memorize each of the four shapes and which
squares correspond to those shapes. Therefore, the grain size of the solu-
tion is considered small. The “Simple Solution,” however, provides steps
that are easy to remember.

8.5.2 The Least Human Window–Compatible Solution

The solution assessed as LHWC is the “Semi-Magic Solution”

(Figure 107). The solution was also developed by Daniel E. Thomasson [6].
The method used to produce this knight’s tour is the square and diamonds
method (see Section 8.2), although it is implicitly used.
The solution shows the concept of the “Semi-Magic Tour,” which
is a knight’s tour that results in the formation of a semi-magic square.
An n × n matrix consisting of values from 1 to n2 is called a magic square if
the numbers in each row, column, and diagonal add up to a single number
146 • Artificial Intelligence Problems and Their Solutions

known as the magic constant [7]. If the values of each row and column add
up to this number, but the diagonals do not, it is considered a semi-magic
square.
Finding a knight’s tour can be thought of as producing an 8” × 8” square
whose values range from 1 to 82. Each square in the chessboard can be
assigned a value corresponding to the step number in which the knight
landed on that square. After the tour is completed, if the values of the
squares in each row, column, and diagonal add up to the same number,
the entire chessboard is considered a magic square. It has been proven,
however, that a magic square cannot be formed by a knight’s tour on an
8” × 8” board. However, it is possible to find a tour that results in a semi-
magic square [7]. The first original semi-magic square was discovered by
Leonhard Euler, where each row and column of the chessboard added up
to the magic constant of 260 [2].
The “Semi-Magic Solution” shows a method of how a semi-magic
square is formed with the same magic constant of 260. Using the squares
and diamonds method, two simultaneous knight’s tours are produced: one
that starts at square 1 and increases by step, and another decreasing by step.
Both tours always visit squares vertical to one other, of which the step num-
bers add up to 49 and 81. Although this solution produces both a completed
knight’s tour and a semi-magic tour and reveals the symmetry behind it, we
will analyze the representation of it in terms of the Human Window, as if it
were intended to be shown to any audience (i.e., those unfamiliar with The
Knight’s Tour Problem). The solution is also shortened to save space, as it
contains a total of 32 steps.
This solution may not be very executable for humans because there
are too many goals to remember. The simultaneous squares and dia-
monds method is repeated for all four quadrants, and the values of the
columns add up to 98. Afterward, the same process is continued, except the
steps must be numbered to add up to 81. To carry out this solution, a person
must remember exactly which shapes to produce at any given step and must
remember the order in which they are formed.
The solution may also not be very comprehensible either. This solution
does not have a knight to follow in the tour, nor is there a line indicating
the path. It simply shows the numbers of the squares the knight would land
on during a particular step. This may make the solution hard to follow and
could take a person away from the fact that this is a solution to The Knight’s
 The Knight’s Tour Problem • 147

Tour Problem. Finally, the squares are shaded differently to indicate the
particular shapes (i.e., diamonds and squares) being formed. The mixture
can make the solution hard to comprehend, even for the image that shows
the completed knight’s tour.

Initial State Square 1Q3 (1–4) Square 1Q4 (41–44)

Diamond 1Q3(45 and 46) Diamond 1Q4 (5–8)

Square 1Q1(9–12) Square 1Q2(33-36) Square 2Q1(17–20)

Diamond 1Q1(37–40) Diamond 1Q2(13–16) Diamond 2Q1(61–64)

Square 2Q2(21–24) Square 2Q3(53-56) Square 2Q4(49–52)

Diamond 2Q2(57–60) Diamond 2Q3(25–28) Diamond 2Q4(29–32)

Solution published with permission of Daniel E. Thomasson

http://www.borderschess.org/KTsemi.htm
FIGURE 107 The “Semi-Magic Solution” (Shortened).
148 • Artificial Intelligence Problems and Their Solutions

8.6 BEST MACHINE SOLUTION

Many algorithms and heuristics have been developed for finding a

knight’s tour. One algorithm that can be used is a divide-and-conquer algo-
rithm, which divides the chessboard into smaller pieces, finds sequences
for those pieces, and combines them (see Section 8.2) [8]. There are also
neural network solutions, in which the chessboard is implemented as a neu-
ral network and each square is represented as a neuron [9].
Given that there are about 4 × 1051 possible sequences on an 8” × 8”
board, it is best to guide a computer to find a knight’s tour through the use
of heuristics to avoid the need to analyze each of these sequences. One of
the most efficient and commonly used heuristics is one that is based on
Warnsdoff’s Rule, which was developed by H. C. Warnsdoff. The rule is
to analyze the squares the knight can visit from its current position, and
whichever square provides the knight with the fewest successor possibilities
is the square it should move to [2].
The flaw with Warnsdoff’s Rule, however, is that it does not always
succeed in finding a complete tour, as it usually results in dead ends. Using
this heuristic to find a knight’s tour may sometimes result in failures, espe-
cially when a tour is to be found on larger chessboards. As the value of n
for an n × n chessboard becomes large, the success rate of Warnsdoff’s
Rule declines [10]. One method to resolve this problem is to combine the
heuristic with a backtracking algorithm. This algorithm will follow the
Warnsdoff heuristic as usual, but when the knight reaches a dead end,
the algorithm will then backtrack to a previous step and attempt to find
a tour once again. This same type of algorithm is used in the “Algorithm
Solution” [11].
A second and more efficient method has been analyzed by Luis Paris.
This algorithm implements the Warnsdoff heuristic with the addition of a
second heuristic that, in the case of a tie between the squares a knight may
visit at a particular step, will “choose the lowest-valued square nearest to
any of the four corners of the board, independently of n.” According to an
analysis done by Paris, this heuristic is always successful in finding a tour,
no matter how large the chessboard is. This algorithm uses no backtracking
and requires O(n2) time [10].
 The Knight’s Tour Problem • 149

George Koltanowski, an International Chess Master with an eidetic

memory, was famous for his “Knight’s Tour” lectures. One of the coauthors
remembers the U.S. Open in Portland, Oregon, in 1987, where he gave a
demonstration in which he asked the audience to attribute serial numbers
on dollar bills, names of famous baseball players, and so on to the squares
on a chessboard. Koltanowski would then turn his back on the audience and
study the board for a few minutes. Then he would turn to the audience,
performing a perfect knight’s tour, while reciting all the information he had
been given for each square. Not surprisingly, Koltanowski also holds the
world’s official record for Blindfold Chess—56 boards—which took place
in San Francisco in 1960.

8.7 PLAYABLE PROGRAM

A playable knight’s tour game is available at http://www.kongregate.

com/games/evgenykarataev/knights-tour. In this game, you get to control
the knight and pick which square to visit relative to its current position. The
squares it can visit and those it has already visited are indicated. Addition-
ally, the chessboard and the knight are presented three dimensionally, giv-
ing the player the sense of using a real chessboard.
150 • Artificial Intelligence Problems and Their Solutions

8.8 REFERENCES

1. Jelliss, G. (2002). Early History of Knight’s Tours. Available at http://www.

mayhematics.com/t/1a.htm. Accessed on January 9, 2014.
2. Jelliss, G. (2002). Rediscovery of the Knight’s Problem. Available athttp://www.
mayhematics.com/t/1b.htm. Accessed on January 9, 2014.
3. Schwenk, A.J. (1991). “Which Rectangular Chessboards Have a Knight’s
Tour?” Mathematics 64: 325–332.
4. Frezco. (2009). Solving a Knight’s Tour Blindfolded. Available at http://www.
frezcogames.com/Knights_Tour.php. Accessed on January 9, 2014.
5. Thomasson, D. (2011). Knight Tour Tessellations. Available at http://www.
borderschess.org/KTtess.htm. Accessed on January 9, 2014.
6. Thomasson, D. (2011). Semi-Magic Square Knight Tours. Available at http://
www.borderschess.org/KTsemi.htm. Accessed on January 9, 2014.
7. Weisstein, E.W. (2003). There Are No Magic Knight’s Tours on the Chessboard.
Available at http://mathworld.wolfram.com/news/2003-08-06/magictours/.
Accessed on January 9, 2014.
8. Parberry, I. (1997). “An Efficient Algorithm for the Knight’s Tour
Problem.” Discrete Applied Mathematics 73: 251–260. http://larc.unt.edu/
ian/pubs/algoknight.pdf.
9. Takefuji, Y., and Lee, K.C. (1992). “Neural Network Computing for Knight’s
Tour Problems.” Neurocomputing 4 (5): 249–254.
10. Paris, L. (2013). Heuristic Strategies for the Knight Tour Problem. Available
at http://faculty.harrisburgu.net/~paris/papers/ppr_icai_knight_tour.pdf.
Accessed on January 9, 2014.
11. Micka, P. (2013). Knight’s Tour. Available at http://en.algoritmy.net/article/39907/
Knights-tour. Accessed on January 9, 2014.
 CHAPTER

9
MASTERMIND

A A A A A
B A B D B
E B C C E
F B C D E

FIGURE 108 Mastermind.

9.1 BACKGROUND

Mastermind (Figure 108) is a board game invented in 1970 by Morde-

cai Meirowitz, an Israeli postmaster and telecommunications expert. After
being turned down by several leading toy companies, the rights to Master-
mind were finally purchased by Invicta Plastics Ltd, a small English firm,
and released in 1972 [1]. According to G. Darby, “Over 50 million copies
later, the game is still marketed today” [2].
Mastermind is a type of “code-breaking” game. In the original Mas-
termind game, one player, known as the code-breaker, must guess a secret
152 • Artificial Intelligence Problems and Their Solutions

code chosen by another other player, known as the code-maker. This code
is a sequence of four pegs that can be any of six different colors, with same-
colored pegs allowed in the same sequence. The code-breaker must break
this code by determining the exact colors and order of these pegs. To do
so, the code-breaker must make a series of at most 10 guesses. After each
guess, the code-maker gives the code-breaker a score in the form of two
values: the number of pegs that are the right color and in the correct posi-
tion and the number of pegs that are the correct color but not in the correct
position. These numbers are represented by small black and white pegs,
respectively. If in 10 guesses the code-breaker cannot determine the cor-
rect sequence, the code-maker wins. [2]
Since its release in 1972, there have been a wide variety of alternate
versions of the original game. One version is called Mastermind44, which
includes six colors, five pegs to a sequence, and as many as three code-
makers [3]. Another version is Ultimate Mastermind, which includes eight
colors and five pegs to a sequence [4].

9.2 PROBLEM-SOLVING TECHNIQUES

Like The 15 Puzzle, Mastermind does not have a specified initial state.
However, unlike any of the problems we have analyzed in this book, this
problem does not have a specified goal state. Additionally, for the other
problems, we attempt to find methods of achieving the goal, while in Mas-
termind we try to find the goal itself. We could say that The Knight’s Tour
Problem is similar in that there is no specific set goal. However, we do know
that the knight must visit all 64 squares, so it is not the same.
In the original Mastermind game, there are a total of 6 × 6 × 6 × 6
= 1,296 possible combinations, and the code-breaker must determine the
one correct sequence of colors [5]! With only 10 guesses available, the key
is to reduce the number of possible combinations as much as possible with
each guess. This can be done by analyzing the score given at each guess and
using deduction to determine which colored pegs to use and which not to
use, and in which order they must be placed. Therefore, deduction is the
most valuable problem-solving technique to use in solving for the correct
sequence of pegs and winning Mastermind. This kind of logical deduction
 Mastermind • 153

is also performed for the solution to the n coins problems to determine how
to best split the number of coins for each weighing.

9.3 SOLUTION

Before beginning the solution, let’s go over some notations:

■ Instead of using colors, the pegs will be represented as letters from A
to F. This notation is better than using actual color names because the
different versions of the game may use different-colored pegs.
■ A sequence of pegs will be represented by placing four of these letters
in a line. For example, ABBC is a sequence of four pegs that contains
peg color A, peg color B, peg color B, and peg color C, in that precise
order.
■ The score will be represented as a set of numbers in parentheses:
(x, y). The variable x will represent the number of pegs that are
guessed correctly and in the correct location (i.e., small black pegs),
and the variable y will represent the number of pegs that are guessed
correctly but are in the wrong position (i.e., small white pegs). For
example, (2, 1) means that 2 pegs are correct colors and in the correct
location in the final sequence, and 1 other peg color is correct but is
not in the correct location in the final sequence. This sequence has a
total score of 3.
An example of a single guess may look like Figure 109.

• ABBC (2, 1)

FIGURE 109 Sample Guess.

Since there is no one specific goal to achieve, there is no general solu-

tion to winning Mastermind. By taking the following steps, however, the
code-breaker can solve the sequence in four to six guesses:

1. The first guess should always be the sequence AABB. This is not only the
simplest guess, but it is also very important because, even in the worst
154 • Artificial Intelligence Problems and Their Solutions

case, this guess greatly reduces the number of possibilities remaining (in
the worst case, it reduces the 1,296 possible guesses to 256). [5]
2. You must use the score of the first guess to make assumption(s) as to
which of the pegs from the previous guess (i.e., pegs A and B) should
be used for the next guess. For example, if the score is a total of 0, pegs
A and B should not be used again for this game. If the score is a total
of 1 or 2, use both an A and a B in the next guess to determine which A
or B is in the sequence. If the score totals 3, use two A pegs and one B
peg in the next guess. Finally, if the score happens to be 4, that means
all four pegs are in the sequence and must simply be arranged to the
correct order.
3. Repeat Step 2 for future guesses until all the peg colors have been
used (that is, if we do not wind up having a score of 4 before then). At
this point there should be enough information from all the guesses
already made to use deduction and solve for the sequence.
At first, the above steps may seem too broad and unclear. However, we
will be providing five examples so you can understand them better.

9.3.1 Example 1

Let’s start with an easy example. The goal sequence is CEDF. We start
with the required first guess and get Figure 110.

1 AABB (0, 0)

FIGURE 110 Ex. 1—Guess 1.

We know from this that neither A nor B is in the sequence. Therefore,

we will make a similar guess using peg colors C and D, reducing the num-
ber of choices even further (Figure 111).

1 CCDD (2, 0)

2 AABB (0, 0)

FIGURE 111 Ex. 1—Guess 2.

 Mastermind • 155

We now have a score of 2. Even better, we have two pegs that are in
their correct positions. However, because we don’t know which pegs are
correct, we need to make an assumption. We can pick any two, but it is best
to have both colors included in the next guess. We will keep the first C and
first D in the sequence, and, since there are two spots available, we will add
two E pegs (Figure 112).

Assumption

3 CEDE (3, 0) C_D_

2 CCDD (2, 0)

1 AABB (0, 0)

FIGURE 112 Ex. 1—Guess 3.

Our score has increased to 3, with all three pegs in the correct posi-
tion. From this, we can conclude three things: our assumptions about the
inclusion and location of the C and D pegs were correct; peg E is in the
sequence; and with one peg still incorrect and one color remaining, peg F
must also be in the sequence. This means that one of the E pegs must be
removed. We will make the assumption that the E in position 2 is correct
and the E in position 4 should be replaced with an F (Figure 113).

Assumption

4 CEDF (4, 0)* CEDF

3 CEDE (3, 0)

2 CCDD (2, 0)

1 AABB (0, 0)

FIGURE 113 Ex. 1—Guess 4.

156 • Artificial Intelligence Problems and Their Solutions

Thus, we have solved for the sequence. Of course, guess 4 was a lucky
guess. If we had assumed the E in position 4 was correct instead of position
2, we would have Figure 114.

Assumption

4 CFDE (2, 2) CEFE

FIGURE 114 Ex. 1—Alternate Guess 4.

However, from this information we could simply determine the solu-

tion, and we would only need one extra guess to solve it. When looking at
the score, we can see that all four pegs exist in the final sequence, with
two pegs in the correct position and two in the incorrect position. Since
we already know that pegs C and D are in the correct place, we can easily
conclude that pegs E and F must be switched.

9.3.2 Example 2

The goal sequence is EAFA. As always, we make our required first

guess (Figure 115).

1 AABB (1, 1)

FIGURE 115 Ex. 2—Guess 1.

We see that we have a score of 2, with one peg in the correct position
and one in the wrong position. Having a score of 2 for the first guess makes
it hard to determine which A or B is correct. Let’s make the assumption
that both colors are in the final sequence, and we will put both in the next
guess. Let’s assume the first A peg is the correct spot and that one of the B
pegs is out of place. Move the B to position 2 and add two of the next color
to the sequence (Figure 116).
 Mastermind • 157

Assumption

2 ABCC (0, 1) AB_ _

1 AABB (1, 1)

FIGURE 116 Ex. 2—Guess 2.

We now have our score reduced to 1, with one peg being out of place.
This gives us two conclusions: C is not in the final sequence, and either two
A pegs or two B pegs are in the final solution, but not both. We know this
because our score of 2 in the first guess has been reduced to 1. For guess
3, we will assume that two As are in the sequence. Now we will make some
more assumptions based around this assumption. Since our current score
indicates that we have one peg out of place, we know that A cannot be in
position 1. Additionally, from looking at our score in guess 1, one of the
A pegs must be in position 2, since there was a black peg. Therefore, let’s
assume that one A is in position 2. The second A can be placed in either
position 3 or 4, so let’s will assume it is in 3. Finally, we will now introduce
peg color D (Figure 117).

Assumption

3 DAAD (1, 1) _AA_

2 ABCC (0, 1)

1 AABB (1, 1)

FIGURE 117 Ex. 2—Guess 3.

We once again have a score of 2, with one peg in place and one out of
place. Unfortunately, with this guess, we cannot make any immediate con-
clusions. This guess can mean one of two things:
158 • Artificial Intelligence Problems and Their Solutions

1. Both As are in the sequence, and the Ds do not belong. We would just
need to move the A that is in position 3 to position 4.
2. The Ds are in the sequence, and our assumption about including the
As was wrong, leaving the two Bs as a part of the final sequence.
Let’s go with #1 and continue to assume that A is the correct color.
From looking at our score, we can assume that the A in position 2 is in place
because we have a black peg. We can also assume that A is in position 4, as
mentioned before. Finally, the other two positions will be replaced by an E
peg (Figure 118).

Assumption

4 EAEA (3, 0) _A_A

3 DAAD (1, 1)

2 ABCC (0, 1)

1 AABB (1, 1)

FIGURE 118 Ex. 2—Guess 4.

This guess was very helpful. We have now come to three conclusions:

1. D is not included in the sequence, since we have an increased score,

despite our removing them.
2. Thus, the two As are in the sequence and are in the correct place.
3. One E is in the sequence, and it is in the correct position.
With F being the only peg we have not used, we would normally con-
clude that F is the last missing peg. However, it is still possible that a third
A peg is a part of the sequence (we have not disproved it). So for our next
guess, let’s assume that the correct E peg was the one in position 1 and that
position 3 contains an F peg (Figure 119).
 Mastermind • 159

Assumption

5 EAFA (4, 0)* E*F*

4 EADA (3, 0)

3 DAAD (1, 1)

2 ABCC (0, 1)

1 AABB (1, 1)

FIGURE 119 Ex. 2—Guess 5.

Thus, we have found our sequence. If position 3 had another A instead

of an F, we would have only needed one extra guess to determine it.

9.3.3 Example 3

The goal sequence is EABC. As usual, the first guess is as shown in

Figure 120.

1 AABB (2, 0)

FIGURE 120 Ex. 3—Guess 1.

We begin with a score of 2, which means it may take a few guesses to

determine which A and/or B is in the sequence. For the next guess, we will
assume that both an A peg and a B peg are in the sequence. Additionally, since
we have two black pegs, let’s assume that the A peg is in position 1 and the B
peg is in position 3. The remaining two positions will be Cs (Figure 121).

Assumption

2 ACBC (2, 1) A_B _

1 AABB (2, 0)

FIGURE 121 Ex. 3—Guess 2.

160 • Artificial Intelligence Problems and Their Solutions

With this guess, we have increased our score to 3 with the addition of a
white peg. Due to this increase from the last guess, we can make an obvious
conclusion that at least one C peg is in the sequence. We can also use some
logic to determine less obvious conclusions.
Let’s say we keep our assumption that A and B are in the correct posi-
tion. By this, we would have to say that there is one C peg in the sequence
and it represents the white peg. However, this does not make sense. Think-
ing that switching the two Cs would get one in the correct position is unre-
alistic. In other words, switching the two Cs cannot suddenly produce a
black peg, because nothing will change. With this single logical inference,
we can come to four conclusions:
1. There is only one C in the sequence.
2. C represents the black peg and is in the correct position.
3. Based on #1 and #2 and the score for this guess, A and B are in the
sequence.
4. One A or B is represented by the black peg, and the other is repre-
sented by the white peg. That is, one is in the correct position, and the
other is in an incorrect position.
First, let’s assume that the C peg belongs in position 2. We know that
A and B exist in the sequence, and now we must determine where to place
the A and B pegs. By analyzing the score from guess 1, in which there are
two black pegs, we can assume that the A peg can only be in either position
1 or 2, and B can only be in either position 3 or 4. Based on conclusion #4,
one A or B will remain in place, while the other will be moved to its next
possible position. For our next guess, let’s assume that A is represented by
the black peg and is in place and that B is represented by the white peg and
should be moved to its next possible position 4. Finally, position 3 will be
given the D peg, as shown in Figure 122.

Assumption

3 ACDB (0, 3) AC_B

2 ACBC (2, 1)

1 AABB (2, 0)

FIGURE 122 Ex. 3—Guess 3

 Mastermind • 161

We continue to have a score of 3, but now every peg is out of position.

This is not necessarily bad news, because we can now make four conclu-
sions:

1. The C peg belongs in position 4.

2. The B peg belongs in position 3.
3. The A peg belongs in position 2.
4. Based on the fact that the score has not increased from the previous
guess, the D peg does not belong in the sequence.
Conclusions #1 to #3 were easily determined because the A, B, and C
pegs only had one of two possibilities, as we determined from our second
guess. Since we chose wrong for all three pegs, we now know the correct
positions for each. With this information, we can now construct our next
guess. Position 2 will have peg A, position 3 will have peg B, and position 4
will have peg C. The remaining position will have the next type of peg we
haven’t used yet, which is E (Figure 123).

Assumption

4 EABC (4, 0)* _***

3 ACDB (0, 3)

2 ACBC (2, 1)

1 AABB (2, 0)

FIGURE 123 Ex. 3—Guess 4.

And thus we have found the sequence.

9.3.4 Example 4

The goal sequence is FBFF. For our last example, we have a goal that
is a little unusual, since we are now dealing with three of the same kind of
peg.
162 • Artificial Intelligence Problems and Their Solutions

We start with our usual first guess (Figure 124).

1 AABB (0, 1)

FIGURE 124 Ex. 4—Guess 1.

We start a score of 1, with one peg out of place. We will assume that A
is in the sequence and is out of place. We will first assume it is in position 3.
The other three remaining positions will be filled with C pegs (Figure 125).

Assumption

2 CCAC (0, 0) _ _A_

1 AABB (0, 1)

FIGURE 125 Ex. 4—Guess 2.

With this guess we can make two conclusions: the A peg is not in the
sequence, so the B peg must be, and C is not in the sequence.
Remembering that guess 1 had a score of 1 with a white peg, the B peg
can only be in either position 1 or 2. We will first assume that peg B is in
position 1. The other positions will be given D pegs (Figure 126).

Assumption

3 BDDD (0, 1) B_ _ _

2 CCAC (0, 0)

1 AABB (0, 1)

FIGURE 126 Ex. 4—Guess 3.

With this guess, we have arrived at two conclusions: B is in position 2,

and D is not in the sequence.
 Mastermind • 163

For the next guess, we will place the B peg in position 2. The rest will
be given the E peg (Figure 127).

Assumption

4 EBEE (1, 0) _B_ _

3 BDDD (0, 1)

2 CCAC (0, 0)

1 AABB (0, 1)

FIGURE 127 Ex. 4—Guess 4.

We now have a black peg, which means our conclusion was correct and
the B peg was in fact located in position 2. Additionally, we can also make
two more conclusions: the E pegs are not in the sequence, and the remain-
ing positions are F pegs, since it is the last type of peg available (Figure 128).

Assumption

5 FBFF (4, 0)* ****

4 EBEE (1, 0)

3 BDDD (0, 1)

2 CCAC (0, 0)

1 AABB (0, 1)

FIGURE 128 Ex. 4—Guess 5.

We have found our sequence.

164 • Artificial Intelligence Problems and Their Solutions

9.4 HUMAN PROBLEM SOLVING

The Mastermind game was distributed digitally to students to analyze

how the average person approaches the game and methods used to find
the solution. The game was distributed to nine college students, and it did
not have the original rules. In this version, there were eight colors, and no
duplicate colors were allowed.
For the first guess, two students picked the first four colors that were
presented as choices, while the rest picked four random colors. However,
we cannot determine if any of the students would have picked the AABB
guess if they had played the original game, although they may have picked
the first four colors regardless.
Regardless of what the students picked as their first guess, all the stu-
dents did fairly well on the game. Seven of the students used the right
amount of deduction and logic to find the correct pegs and their order. One
student was able to figure out the correct pegs but had difficulty determin-
ing their correct order. Eventually, this student was able to find the correct
sequence in seven guesses. The other exception did a good job in deducing
the correct order of the pegs but did not do so well in figuring out the cor-
rect pegs. This student used the same pegs that could have been reasoned
out of the final sequence at an earlier guess. As a result, it took this student
nine guesses to figure out the correct sequence.
On average, we can see that the students had solved the sequence with
a fairly small number of guesses. One student solved the sequence in four
guesses, three students in five guesses, two students in six guesses, two stu-
dents in seven guesses, and one student in nine guesses. The mode is five
guesses, which is also the average number of guesses it would take to solve
the solution presented in Section 9.3.

9.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

Unfortunately, there was insufficient data for an analysis of the Human

Window for this problem. Also, there were not enough solutions found that
could be used to help in this analysis. What we did find were word descrip-
tions to general strategies on solving Mastermind but no concrete solutions.
 Mastermind • 165

This may be because, as mentioned before, this problem does not have a
specific goal, so only general tips and strategies can be used to help find
the goal. This, however, gives us an opportunity to use our Human Window
analysis from other problems in the book to determine what kind of solu-
tion can be considered the MHWC for Mastermind.
Since the Mastermind problem does not have one specific goal to
achieve, it is best to use a general, intensional solution, because this will
help find the end goal to a Mastermind game regardless of what it is. Like
the “Tile Solution,” combined with examples like those in Section 9.3, a
similar approach can be taken with a general solution to Mastermind. This
solution can contain one or more examples that demonstrate a step-by-
step method for solving for a particular goal sequence. Furthermore, there
should be enough examples so that every scenario that can possibly occur
can be explained. This will give the solution a high ranking of executability.
Additionally, to increase the comprehensibility of the solution, images of
the actual Mastermind game could be used at every step to give an idea of
what the game should look like.

9.6 BEST MACHINE SOLUTION

The best solution to the Mastermind game for a machine so far is known
as the Five-Guess Algorithm, which was developed and presented in 1977
in a paper titled “The Computer as Mastermind,” by Donald Knuth. Knuth
based his algorithm on a study he did on the original Mastermind game. It
is called the Five-Guess Algorithm because it guarantees that any solution
can be found in five or fewer guesses [6].
The algorithm is quite sophisticated and intricate. Here is a brief
explanation:

1. Create a set P of remaining possibilities. At the beginning, P is always

1,296.
2. Make the first guess of AABB (the first guess as used in Section 9.3).
Remove all possibilities from P that would give a lower score if they
were the answer.
166 • Artificial Intelligence Problems and Their Solutions

3. For each of the 1,296 possible guesses, calculate how many possibili-
ties from P would be eliminated for each possible score. The score of
the guess is the least of such values.
4. Play the guess with the highest score.
5. Repeat steps #2 to #4 until the sequence is found [6].

We repeat a historical note from Knuth [6]:

A game very similar to Mastermind, called “Bulls and Cows,” has been
popular in England for many years. The difference is that all the digits
of the code in Bulls and Cows must be distinct, but any digits 0 through
9 are allowed. This version of the game became a popular computer
demonstration after Frank H. King introduced a program for it in
August 1968 at Cambridge University.
F.H. King, The Game of M00, University of Cambridge, Computer
Laboratory, memorandum dated February, 1976.

9.7 PLAYABLE PROGRAM

A playable Mastermind game is available at http://www.grayman.de/

mastermind/index_en.php4. Luckily, it follows the rules of the original
game, so following the examples in Section 9.3 could be helpful. Addi-
tionally, the different pegs are not only colored individually, but they are
numbered additionally as well.

9.8 REFERENCES

1. Nelson, T. (2000). A Brief History of the Master Mind Board Game. Available
at http://www.tnelson.demon.co.uk/mastermind/history.html. Accessed on
January 9, 2014.
 Mastermind • 167

2. Darby, G. (2013). Mastermind Game. Available at http://www.delphiforfun.

org/Programs/Mastermind.htm. Accessed on January 9, 2014.
3. Mastermind44. (1972). Available at http://boardgamegeek.com/
boardgame/12467/mastermind44. Accessed on January 9, 2014.
4. Ultimate Mastermind. (1972). Available at http://boardgamegeek.com/
boardgame/3874/ultimate-mastermind. Accessed on January 9, 2014.
5. Nelson, T. (1999). Investigations into the Master Mind Board Game. Available
at http://www.tnelson.demon.co.uk/mastermind/. Accessed on January 9,
2014.
6. Knuth, D. (1976). “The Computer as a Master Mind.” Journal of Recreational
Mathematics 9: 1–6. http://www.dcc.fc.up.pt/~sssousa/RM09101.pdf.
 CHAPTER

10
THE MONTY HALL PROBLEM

FIGURE 129 The Monty Hall Problem.

10.1 BACKGROUND

The Monty Hall Problem (Figure 129) is one of the most well-known
and widely studied problems in the area of puzzles. It is a probability-based
puzzle where the chance of a favorable outcome is to be compared before
and after an event occurs, and then the better of the two situations must be
determined.
The problem is as follows: In a game show, the host asks a contestant to
pick one door out of three, given that one of the doors has a car behind it.
Each of the other two doors has a goat behind it. After the contestant picks
a door, the host opens one of the other two doors to show a goat behind it.
170 • Artificial Intelligence Problems and Their Solutions

The contestant is then asked if he or she wishes to switch to the unopened

door or stay with the original choice. Do you think the contestant has a bet-
ter choice of winning if he or she switches?
Initially, it seems switching makes no difference because after the host
opens the door with a goat, we still assume that the chances of the remain-
ing unopened doors having the car are ½, the same chance as the origi-
nal choice. Since the chances of winning seem equal with both remaining
doors, there seems to be no benefit in switching.
Surprisingly, the correct answer is the contestant has a better chance of
winning the car if he or she switches to the other door. We explain why this
is true in the next section.

10.2 PROBLEM-SOLVING TECHNIQUES

It is clear that either of the two choices gives only a chance or prob-
ability of winning the car and does not guarantee a win. Therefore, proba-
bility-based analysis is the key to solving this problem. Also, this problem
involves making decisions. Initially the contestant has to choose a door,
which is very likely a random choice because all three doors have an equal
probability (1/3) of having a car behind them. But after it is revealed that
one of the doors has a goat, the situation changes. The contestant now has
more information than when he or she picked one door out of three. Since
the contestant’s final decision of staying with or switching the chosen door
depends on the original choice and the host’s action, we create a decision
tree to analyze the problem. A tabular version of the decision tree is shown
in Table 14.
Also, since this problem has a small state space in terms of all possible
outcomes of the combination of the contestant’s original decision, the host’s
choice of door, and the contestant’s final decision, we can do an exhaustive
enumeration of the solution state space as shown in Table 14.
Table 14 shows all the possible scenarios for this puzzle. The numbers
1, 2, and 3 represent the three doors. We consider the three possibilities
of the car being behind each door one by one. In each of the three cases
stated above, we further consider three cases or possibilities for the original
choice.
 The Monty Hall Problem • 171

Door with Initial Correct/

Host’s Choice Final Choice Result
the Car Choice Incorrect
Stay Win
1 Correct 2 or 3
Switch Lose
Stay Lose
1 2 Incorrect 3
Switch Win
Stay Lose
3 Incorrect 2
Switch Win
Stay Lose
1 Incorrect 3
Switch Win
Stay Win
2 2 Correct 1 or 3
Switch Lose
Stay Lose
3 Incorrect 1
Switch Win
Stay Lose
1 Incorrect 2
Switch Win
Stay Lose
3 2 Incorrect 1
Switch Win
Stay Win
3 Correct 1 or 2
Switch Lose

TABLE 14 The Monty Hall Problem.

For each of the possible initial choices, the host will choose a door hav-
ing a goat, thus reducing the number of doors possibly having a car to just
two instead of three. This leaves the contestant with two choices: either to
stay with the initial choice or switch to the door not opened by the host.
The final choice made by the contestant and the ensuing result are
shown in the last two columns. It is evident that staying with the original
choice guarantees a win if the original choice was correct. And the prob-
ability of the original choice being correct is 1/3 because only one door out
172 • Artificial Intelligence Problems and Their Solutions

of three has a car behind it. On the contrary, the probability of the original
choice being incorrect is 2/3, so in this case switching leads to a better chance
of winning.
We can look at it another way. Say instead of 3 doors, there were 10
doors with a car behind 1 door and goats behind 9 doors. The contestant
chooses a door, and the host reveals 8 doors that have goats. In this case,
should the contestant stay with his or her original choice or switch?
It is clear that initially the probability of opening the correct door is 1/10,
which also means the probability of opening a door with a goat is 9/10. Since,
we are more likely to open a door with a goat initially, switching is more
likely to lead to a win—in this case, a ½ chance of winning.

10.3 SOLUTION

The solution as described in the previous section is that you have more
chances of winning if you switch. This is because you have double the
chances of having a wrong door with your initial choice. The host always
opens a door (that you did not choose) with a goat behind it (as described in
the problem). So after this event happens, we know the car is behind one of
the two other doors and thus there is a 50% chance of the car being behind
either of the two other doors. So staying with the original choice would
mean 1/3 = 33.3% chance of winning, and switching would mean 2/3 = 66%
chance of winning. This is clearly shown in Table 14.

10.4 HUMAN PROBLEM SOLVING

The Monty Hall Problem first appeared in Marilyn vos Savant’s column
in Parade Magazine. Vos Savant, who answers questions and solves puz-
zles submitted by readers, was once listed in the Guinness Book of World
Records as having the highest IQ in the world.
The problem became very popular because it is loosely based on the
popular game show Let’s Make a Deal and was named after the show’s host,
Monty Hall. In response to the question of whether it was better to stay
with the original choice or switch, Vos Savant stated that switching was the
 The Monty Hall Problem • 173

better strategy. Her answer triggered around 10,000 responses, including

from many PhDs, who insisted her answer was incorrect [1, 2]. This story
garnered so much attention that it was featured in the New York Times [2].

10.5 REFERENCES

1. Deaves, R.H. (2007). The Monty Hall Problem: Behind Closed Doors. Lulu.com
Available at http://www.amazon.com/The-Monty-Hall-Problem-Beyond/
dp/1847530788. Accessed on January 10, 2014).

2. Yanofsky, N. (2013). The Outer Limits of Reason: What Science, Mathematics

and Logic Cannot Tell Us. Boston, MA: The MIT Press.
 CHAPTER

11
RUBIK’S CUBE

FIGURE 130 Rubik’s Cube.

11.1 BACKGROUND

Rubik’s Cube (Figure 130) is one of the most popular puzzles in his-
tory. It is a three-dimensional combination puzzle that was invented in 1974
by Ernö Rubik, a Hungarian architect and professor [1]. The cube is made
up of 26 smaller cubes. When considering one of these smaller cubes as a
unit, we can say that the whole cube has dimensions of 3” × 3” × 3”, with 1
smaller cube missing from its center. Together, 9 smaller cubes that make
up one of the faces of the whole cube can be turned into a “layer” that has
the ability to rotate while staying in position. Rubik created the cube in an
176 • Artificial Intelligence Problems and Their Solutions

attempt to understand how blocks could move independently without [the

entire structure] falling apart [2].
When the cube is unscrambled, each of its six faces will have one color.
This is the cube’s goal state. To arrive at this state, each of the smaller cubes
must be moved into a certain position. However, this is not as straightfor-
ward as it sounds. A single small cube cannot be moved individually. Entire
layers of the cube must be rotated to position it correctly. Unfortunately,
this will inevitably reposition other smaller cubes, even if they are already
in their correct positions. Therefore, specific algorithms must be used to
correctly position smaller cubes without moving others out of place. After
scrambling the model, Rubik took a month of trial and error to solve the
puzzle himself [2].
In 1975, Rubik applied for a patent for his creation and began to look
for a partner to help him market it [3]. He entered into a contract with the
Hungarian toy manufacturer Politechnika to mass produce the cube. In
1977, the first version of the cube appeared in toy stores in Budapest as
the Büvös Kocka, also known as the Magic Cube [4]. Eventually, the cube
became popular throughout Hungary [3]. In 1980, the cube was taken to
an international market and sold under the patent name, Rubik’s Cube,
making it well known throughout the world. Today, the puzzle is currently
owned by Rubik’s Brand Ltd. and is still widely popular. Rubik’s Cube puz-
zle solvers, or “cubists,” still practice and research the puzzle to this day [2].
Some of them are so dedicated that they have been able to set records by
solving the cube in a matter of seconds.
Rubik’s original cube was 3” × 3” × 3”. Since then, several variations
have been introduced. Two of the most popular variations are Rubik’s
Revenge (also known as the Master Cube), which is a 4” × 4” × 4” cube,
and the Professor’s Cube, which is a 5” × 5” × 5” cube [5, 6].

11.2 PROBLEM-SOLVING TECHNIQUES

The most valuable problem-solving technique for Rubik’s Cube is solv-

ing for subgoals. There are many methods for solving Rubik’s Cube and more
still waiting to be discovered. However, regardless of what steps the methods
consist of and how different they are from one another, they will more than
likely require the use of subgoals in their solution. Specifically, a subgoal is
placing a particular set of smaller cubes into their correct positions.
 Rubik’s Cube • 177

As mentioned before, regardless of whether or not they are in their

correct positions, it is inevitable that the smaller cubes will be moved after
rotating the layers in an attempt to correctly position other small cubes. To
overcome this, specific “algorithms” must be used. In the context of Rubik’s
Cube, an algorithm refers to a sequence of layer rotations that moves par-
ticular small cubes to specific locations and does not ultimately displace
other small cubes that were already positioned correctly.
In general, a solution to Rubik’s Cube may proceed as follows:
1. Use algorithm(s) to maneuver a particular set of small cubes into their
correct positions.
2. Use other algorithm(s) to maneuver another particular set of small
cubes into their correct positions without moving the cubes that were
already arranged in the previous step(s). (It is all right if they are
moved during the algorithm as long as they are returned to their cor-
rect positions when the algorithm is completed.)
3. Repeat Step 2 until each of the smaller cubes is arranged in its correct
position and the cube is solved.
It is possible for more than one algorithm to be used at a particular
subgoal—in other words, one algorithm can be used to reorient the smaller
cubes, while another correctly positions them.

11.3 SOLUTION

While there are many ways to solve the standard 3” × 3” × 3” Rubik’s

Cube, this section describes one of the easiest and most popular methods.
This method uses seven subgoals to solve the cube. Before we begin, how-
ever, it is imperative that we provide some notations.

11.3.1 Cube Components

The Parts
As mentioned before, Rubik’s Cube is made up of 26 smaller cubes.
However, they are not identical, as there three different types of these
smaller cubes.
The Corners (in white) (Figure 131)
There are a total of 8 corner cubes, each consisting of three colors.
178 • Artificial Intelligence Problems and Their Solutions

FIGURE 131 Corner Pieces.

The Edges (Figure 132)

There are a total of 12 edge pieces, each consisting of two colors.

FIGURE 132 Edge Pieces.

The Center (Figure 133)

There are a total of 6 center pieces, each consisting of one color. The
center piece is a very important piece to take note of when solving the cube,
as it identifies the color of a particular face. This is because a center piece
will never change position relative to the other center pieces.
For this solution, we will name each smaller cube by the colors they
contain and their type. Additionally, we will refer to them as pieces instead
of smaller cubes. For example, an edge piece consists of the colors red and
blue. We will call this the Red-Blue edge piece.
 Rubik’s Cube • 179

FIGURE 133 Center Pieces.

The Faces (Figure 134)

Just like any cube, there are a total of six faces (or sides) to the Rubik’s
Cube. Each face is identified by a different color, and that color is deter-
mined by the color of its center small cube. The configuration of the colors
may differ from cube to cube, so for the solution here, we use a two-
dimensional view of the cube and its colors:
R = Red
W = White
Y = Yellow
B = Blue
G = Green
O = Orange

FIGURE 134 Cube Faces.

The Layers
By layers, we are referring to the sections of the cube that can be
rotated. A layer consists of 9 smaller cubes that can be rotated as much as
180 • Artificial Intelligence Problems and Their Solutions

360 degrees two-dimensionally. The cube has a total of six layers, and each
will be denoted by a single letter:
The Top (or Up) Layer = U (Figure 135)

FIGURE 135 Top Layer.

The Bottom (or Down) Layer = D (Figure 136)

FIGURE 136 Bottom Layer.

The Left Layer = L (Figure 137)

FIGURE 137 Left Layer.

 Rubik’s Cube • 181

The Right Layer = R (Figure 138)

FIGURE 138 Right Layer.

The Front Layer = F (Figure 139)

FIGURE 139 Front Layer.

The Back Layer = B (Figure 140)

FIGURE 140 Back Layer.

The layers are named relative to the faces of the cube that are considered
to be the front and top. For example, if we choose the White Face to be our
182 • Artificial Intelligence Problems and Their Solutions

Front Face and the Red Face to be our Top Face, then the layer considered
as the Left will be different from the layer considered the Left if say the
Blue Face was considered the Front Face.
These denotations will be used to show how a particular layer is to be
rotated in an algorithm. When a letter is shown, that means that the layer
it corresponds to is to be rotated 90 degrees clockwise. Furthermore, an
apostrophe (’) after the letter indicates that the same layer is to be rotated
90 degrees counterclockwise. A chain of these letters will make up an algo-
rithm, and they must be followed in the order given. Here is an example of
an algorithm:
U D F’
This algorithm indicates that the Top Layer should be moved clock-
wise, then the Bottom Layer, and finally the Front Layer should be moved
counterclockwise. You will see that as more small cubes are put into place,
the algorithms used to solve the remaining cubes become more complex.
To begin this solution, we choose the Red Side to be the Top Face. The
Front Face will be determined later.

11.3.2 Subgoal 1: The Top Cross

FIGURE 141 Subgoal 1.

The first subgoal (Figure 141) is to place the top 4 edge cubes into their
correct positions. As a result, the Top Side should look like a red cross.
 Rubik’s Cube • 183

Achieving this configuration is straightforward. The worst case occurs

when an edge piece happens to be on an incorrect side. For example, let’s
assume that the Red-Green edge is a part of the Blue Face. We want to
get it into its correct place on the Green Face. The first thing to do is
rotate the layer that contains the Blue Face until this edge is on the Bottom
Layer. Next, rotate the Bottom Layer until this edge piece is located on its
corresponding Green Face. Finally, rotate the layer with the Green Face
twice, and the Red-Green edge should be in its correct position. It could
have been the case that the Red-Blue edge could have been in its correct
position from the start, and our move displaced it. Fortunately, it is easy
to put the Red-Blue edge back into place. After rotating the layer with the
Blue Face enough to put the Red-Green edge to the Bottom Layer, rotate
the Bottom Layer once so the Red-Green edge is away from the layer with
the Blue Face, and then rotate that layer until the Red-Blue Side is back in
place. It is possible you may end up with Figure 142.

FIGURE 142 Inverted Edge.

Take note of the issue here. The Red-Green edge has its colors inverted.
To fix this, a particular algorithm must be used.
First, consider the Front Side as the Green Face, with the Top Side
remaining as the Red Face. Next, perform these moves:
F’ U L’ U’
This algorithm will reorient the Red-Green edge to its correct
configuration, while keeping the other three top edges in place.
184 • Artificial Intelligence Problems and Their Solutions

11.3.3 Subgoal 2: The Top Corners

FIGURE 143 Subgoal 2.

The sub-goal in Figure 143 will not only put the 4 top corner pieces
in the correct place, but it will also complete the Red Face and the entire Top
Layer!
Let’s say we want to begin by putting the Red-Blue-Green corner in its
correct place. Our first step is to bring the corner to the bottom if it is not
already there. First, locate one of the sides that the Red-Blue-Green corner
is on and call that the Front Side. In our case, let’s say that it is located in
between the Yellow and White Faces. We can either choose the Front Face
to be the Yellow Face or the White Face. Let’s choose the White Face to
be the Front Face. The Red-Green-Blue corner can now be located on
the upper left corner of the Front Face. To get it to the bottom without
displacing the Red-White edge, perform this algorithm:
F D’ F’
If we had chosen the Front Face to be the Yellow Face instead, the
Red-Green-Blue corner would be located on the upper left side, and we
would have to perform this algorithm instead:
F’ D F
Whichever algorithm was performed, our target corner should now be
on the Bottom Layer, and the edge we positioned in subgoal 1 should still
be on the Top Layer. Rotate the Bottom Layer until the corner piece is
located in between the Green and Blue Faces. At this point, we can have
one of the three possible states shown in Figures 144 to 146.
 Rubik’s Cube • 185

FIGURE 144 State 1. FIGURE 145 State 2. FIGURE 146 State 3.

State 1: Make the new Front Face the face that contains the correct
color (which in this case will be the Blue Face). Our target corner should
now be a part of the lower left corner of the Front Face. Then perform the
following algorithm:
L D’ L’
State 2: As with State 1, make the new Front Face the face that con-
tains the correct color (which in this case will be the Green Face). However,
the target corner should now be a part of the lower right corner of the Front
Face. Then perform the following algorithm:
R’ D R
State 3: In this orientation, we cannot properly place the Red-Green-
Blue corner in its correct position. We need to reorient the corner so it
matches State 1 or State 2. Let’s change it to State 1. First, pick the new
Front Face that will make the target corner be a part of its lower left corner
(in this case the Green Face). Then perform the following algorithm:
L D’D’ L’ D
After performing this, the corner piece will be back in between the
Green and Blue Faces and will be oriented as it appears in State 1. Now
simply perform the State 1 algorithm. Do these methods for all four cor-
ners, and you will have achieved subgoal 2.

11.3.4 Subgoal 3: The Middle Layer

Subgoal 3 (Figure 147) puts the four center edge cubes in their cor-
rect positions. By the end of this subgoal, the middle layer of Rubik’s Cube
186 • Artificial Intelligence Problems and Their Solutions

FIGURE 147 Subgoal 3.

will be completely solved. Before we begin, it is important to flip the cube

upside down. Therefore, the new Top Layer is the Orange Face, now mak-
ing the Red Face the Bottom Layer.
First, locate an edge on the Top Layer that does not contain the color
of the top face (in this case Orange). Let’s say we find the Blue-Green edge.
Also, let’s say that the vertical color of this edge happens to be Blue, while
the horizontal is Green. Rotate the Top Layer until the vertical side of the
edge piece matches its corresponding face. In this case, rotate the Top Layer
until the Blue-Green edge is located on the Blue Face, which will now be
considered the Front Face. The cube should now look like Figure 148.

FIGURE 148 Pre-Algorithm State.

Take note of the top color of this edge piece. From the Front Face,
look at the colors of the Left and Right Faces. Whichever color matches the
target edge’s top color determines which algorithm to use next. In this case,
 Rubik’s Cube • 187

the Green on the top of the Blue-Green edge matches the Green Face,
which is located on the Right Layer. Therefore, use the following algorithm:
U R U’ R’ U’ F’ U F
After performing this algorithm, the Blue-Green edge will be placed in
its appropriate position in the middle row.
Alternatively, if it were the case that the Green Face was on the Left
Layer, we would use this algorithm:
U’ L’ U L U F U’ F’
One possible issue that can occur is that an edge piece will be placed in
its correct position but oriented incorrectly, as in Figure 149.

FIGURE 149 Incorrect Orientation.

To fix this issue, it must first be moved back into the Top Layer. This is
done simply by using any one of the two algorithms above to place another
edge piece in that spot. Specifically, this edge should have Orange on it,
since it will be replaced anyway and will not misplace any other important
edges. After it is removed from the middle row, appropriately use the above
algorithms to put it back into its correct orientation. Do this for the four
edges to successfully solve the middle layer.

11.3.5 Subgoal 4: The New Top Cross

We have successfully completed two-thirds of the entire cube. How-

ever, solving this Top Layer requires four subgoals.
This next subgoal is to make a cross on the now Top Face (Figure 150).
It does not matter which edge pieces are used as long as the top of the
edges are Orange and an Orange cross is made.
188 • Artificial Intelligence Problems and Their Solutions

FIGURE 150 Subgoal 4.

After the previous subgoal, the Top Face will be arranged in one of the
four ways shown in Figures 151 to 154.

FIGURE 151 State 1. FIGURE 152 State 2. FIGURE 153 State 3. FIGURE 154 State 4.

In any of these four states, do not consider the color of the corners
whether they are Orange or not. Also, this should still be considered the
Top Face and not the Front Face. In carry out the following algorithms,
you must pick a Front Face so the Top Face appears as one of these figures
when you look at the cube from above.
State 1: The cross is already formed, and you can move on to subgoal 6.
State 2: The only part of the cross is the center. Use the following
algorithm:
F U R U’ R’ F’
State 3: Only three pieces of the cross are in their correct places and
form something like a right angle or a backward “L.” Use the same algorithm
as State 2 for this configuration.
 Rubik’s Cube • 189

State 4: Again, only three pieces of the cross are in their correct places,
but this time they form a straight line. For this configuration, use the
following algorithm:
F R U R’ U’ F’
Additionally, you may find that the Top Face looks like Figure 155.

FIGURE 155 Alternate State.

This is essentially like a combination of States 3 and 4. For this configu-

ration, use one of the two above algorithms and it will reconfigure into one
of the previous states. When this happens, use the appropriate algorithm
that corresponds to that state. This completes this subgoal.

11.3.6 Subgoal 5: The New Top’s Corners

FIGURE 156 Subgoal 5.

Subgoal 5 (Figure 156) is to make the corners of the Top Face

Orange, which will complete the Orange Face. Again, it does not matter
if the corners are in the correct place or not as long as the tops of them
are Orange.
After the previous subgoal, the Top (Orange) Face will have one of the
three configurations shown in Figures 157 to 159.
190 • Artificial Intelligence Problems and Their Solutions

Again, this is not the Front Face but the Top Face. However, make sure
that you pick a Front Face so any of these configurations match what the
Top Face looks like while looking at the cube from above.

FIGURE 157 State 1. FIGURE 159 State 3.

FIGURE 158 State 2.

For each state, the following algorithm will be used:

R U R’ U R U U R’
State 1: Only the cross is present. It is important that you position the
cube so the lower left corner piece has the Orange color on the outside left
of it, just as the diagram depicts. In other words, while looking at the Left
Face, you should see Orange on its upper right corner.
State 2: Any of the two corners on the Top Face has Orange. It does not
matter how those two corners are oriented. What matters is that the lower
left corner piece has Orange on the outside bottom of it, just as the diagram
depicts. In other words, while looking at the Front Face, the upper left
corner should be Orange.
State 3: Along with the cross, there is one single Orange corner. This is
the state we are looking for. Orient the cube so this corner piece is on the
lower left corner of the Top Face. Then use the above algorithm.
After continuously using the algorithm on these states, the entire Top
Face will eventually be all the same color, and the subgoal will be achieved.

11.3.7 Subgoal 6: Correct Corner Positioning

Subgoal 6 (Figure 160) is to position the four Top Layer corner pieces
into their correct orientations. At this point, there will be exactly two cor-
ners in correct orientation. From an overhead view of the cube, the Top
Layer will be arranged in one of the two ways shown in Figures 161 and 162.
 Rubik’s Cube • 191

FIGURE 160 Subgoal 6.

The “C” indicates the corner pieces in their Correct orientation. They
are indicated as such because this can apply to any two random corners in
the Top Layer. To see which corners are in correct orientation, the entire
cube must be analyzed. Fortunately, doing this is an easy task. First, analyze
the four corner pieces as they appear. If you do not see that they are in their
correct spots, then rotate the Top Layer 90 degrees and look again. If you
still cannot determine them, rotate the Top Layer 90 degrees once again.
Eventually, it will become obvious which corners are in correct orientation.
However, it is important that once you discover the correct corners, you
must rotate the Top Layer so they are placed in their correct locations rela-
tive to the entire cube.

FIGURE 161 State 1 (Adjacent). FIGURE 162 State 2 (Diagonal).

For this subgoal, the following algorithm must be used in either case:
R’ F R’ B B R F’ R’ B B R R U’
State 1: It is possible that the two correctly oriented corner pieces
will be directly next to each other. First, double check to see if they are in
their correct positions relative to the cube. If not, you must rotate the Top
Face until they are. This is easy to determine, since they are both adjacent
and share two of the same colors. Then pick a Front Face so when you are
192 • Artificial Intelligence Problems and Their Solutions

looking at the Top Face from above, it will appear as in Figure 161 (i.e., the
upper corner pieces). When looking at the Front Face, these pieces should
be all the way in the back of the cube at the top.
State 2: It is also possible that the correctly oriented cubes are directly
across from each other on the Top Layer. At this point we need to bring
the cube into State 1. If the correctly oriented corner pieces appear as they
do in Figure 162 simply use the algorithm above. It does not matter how
the Top Layer is positioned or which face is chosen to be the Front Face as
long as the Top Face we have been using remains as such (in this case the
Orange Face). At this point, your cube will be ready for the final subgoal.

11.3.8 Subgoal 7: The Top Edges

FIGURE 163 Subgoal 7 (Goal State).

The final subgoal, subgoal 7, is to correctly position the four remain-

ing Top Layer edge pieces (Figure 163). There are two algorithms that can
solve the cube. At this point, the cube can be in one of two states:
State 1: One edge cube is in the correct position. It may be the case
that one of the Top Layer edges is already in its correct position. The first
thing to do is examine the cube and analyze the entire Top Layer to see
which edge is in position. This is easy to determine because the edge that is
in correct position will have completed one of the Side Faces. Then, orient
the cube so the completed face is a part of what we will now consider the
Back Layer.
The next step is to determine which algorithm to use. To do so, we must
analyze the incorrect edge pieces and how they are positioned in relation
 Rubik’s Cube • 193

to each other. First, look at two of the incorrect edges, preferably the one
on the Front Layer and one on any Layer adjacent to it (i.e., Left Layer or
Right Layer). At this point you will notice that one of the edges will have to
either be shifted to the left or right to be correctly placed.
Let’s say, for example, we have found the Blue Face to be a part of the
Front Layer. Assume we have the configuration in Figure 164.
We see that the Orange-Blue edge piece needs to be moved to the
Front Layer from the Right Layer, meaning we need to shift it left. There-
fore, use the following algorithm:

FIGURE 164 Orange-Blue Edge to the Right.

F F U L R’ F F L’ R U F F
This will shift the three target edge pieces in a clockwise direction rela-
tive to the Top Layer, which is why we can call this the “clockwise” algo-
rithm
It may also be the case that the cube has the configuration in Figure
165.

FIGURE 165 Orange-Green Edge to the Right.

194 • Artificial Intelligence Problems and Their Solutions

Now, we see that the Orange-Green edge needs to be moved from the
Front Layer to the Right Layer and shift it right. Therefore, we use what we
call the “counterclockwise” algorithm:
F F U’ L R’ F F L’ R U’ F F
This will shift the three target edge pieces counterclockwise. Using
either of these algorithms once should place these edges in their correct
positions
State 2: None of the Top Layer edges are correct. It may be the case
that all four Top Layer edges are not in their incorrect positions. If so, then
use any one of the two algorithms above. It will not put all four edges in
their correct positions, but it will put one edge in its correct position, bring-
ing the cube into State 1. At this point, Rubik’s Cube will be completely
solved.

11.4 HUMAN PROBLEM SOLVING

Unfortunately, we have not been able to obtain a collection of test sub-

jects to apply our Human Problem Solving analysis to Rubik’s Cube. How-
ever, if you know anyone who is determined to solve a puzzle, that puzzle
is probably Rubik’s Cube. It may just be out of curiosity, but if you put a
Rubik’s Cube in a room with some people, someone (probably even a few
people) is going to pick it up and try to solve it. Even if you are with a friend
and we are having a conversation, he or she will pick it up and begin to solve
it, either consciously or unconsciously.
Personally, we have never seen anyone solve the cube. If we observe
someone trying, he or she always makes the fatal mistake of trying to solve
one face at a time. First, without an attempt to use any algorithms, they
struggle to solve one face. As soon as they solve it, they try to solve a second
one, only to find that their first face becomes unsolved. Then they simply
give up. If they had attempted to correctly position the small cubes instead
of simply the faces, maybe they would have been more successful.
 Rubik’s Cube • 195

Difficulty

7/10

Complex.

O( n2/log )

Prob.
Int or Intly/ Mode of
Name Rep. HW? Corr? Grn Sz Exec Compr Solv. Flex Opt? Tot
Ext? Extly Conv
Mthd
Img/Txt
Adam Ext. 3/10 Txt w/ Pics Y Y Ideal 8/10 4/10 Subgoals 5/10 N 20/40
Editor
Img/Txt
Animation Ext. 9/10 Txt w/ Pics Y Y Ideal 9/10 8/10 Subgoals 7/10 N 33/40
Editor
Img/Txt
C&P Ext. 4/10 Txt w/ Pics Y Y Ideal 8/10 6/10 Subgoals 6/10 N 24/40
Editor; Hand
Cartoon Img/Txt
Ext. 5/10 Txt w/ Pics Y Y Ideal 8/10 7/10 Subgoals 7/10 N 27/40
Pics Editor; Hand
Img/Txt
Enemy Ext. 8/10 Txt w/ Pics Y Y Ideal 7/10 9/10 Subgoals 7/10 N 31/40
Editor
Magic Img/Txt
Ext. 7/10 Txt w/ Pics Y Y Ideal 7/10 8/10 Subgoals 7/10 N 29/40
Cube Editor; Hand
Img/Txt
Short Ext. 6/10 Txt w/ Pics Y Y Ideal 2/10 3/10 Subgoals 6/10 N 17/40
Editor; Hand
Simple Img/Txt
Ext. 7/10 Txt w/ Pics Y Y Ideal 8/10 9/10 Subgoals 7/10 N 31/40
Cube Editor
Img/Txt
Square Ext. 7/10 Txt w/ Pics Y Y Ideal 6/10 4/10 Subgoals 7/10 N 24/40
Editor

Key

Int or Ext?: Intensional or Extensional—Is the solution intensional or extensional?

Intly/Extly: Intensionality/Extensionality—How intensional or extensional is the solution?
Rep.: Representation—How is the solution represented?
HW?: Human Window—Does the solution exist in the Human Window?
Corr?: Correctness—Is the solution correct?
Grn Sz: Grain Size—How much computation (large) or memory (small) does one need to solve the solution?
Exec: Executability—How executable is this solution?
Compr: Comprehensibility—How comprehensible is this solution?
Prob. Solv. Mthd: Problem-Solving Method—What method is used in the solution to solve the problem?
Flex: Flexibility—How flexible is this solution (i.e., can this solution be represented in other ways)?
Mode of Conv: Mode of Conveyance—In what ways can this solution be reproduced/replicated?
Opt?: Optimal—Is the solution optimal?

TABLE 15 Ranking Rubik’s Cube Solutions According to the Human Window.

196 • Artificial Intelligence Problems and Their Solutions

11.5 HUMAN WINDOW ANALYSIS OF SOLUTIONS

Since Rubik’s Cube is such a popular puzzle with so many methods to

solve it, it only makes sense that such a large number of solutions exist for it.
However, regardless of the fact that such a large number of solutions exist,
their representations are substantially identical.
When analyzing the solutions (Table 13), it can be seen that there are
three key features for each that make up each of their representations:
Cube Depiction: Whether it is to show a goal arrangement, a sub-
sequent step, or an arrangement to avoid, it is common for a solution to
present images of the actual cube itself. Displaying images of the cube will
show a clear depiction of what the cube should or should not look like at a
particular step. This makes a solution highly comprehensible, especially to
those who are using a physical cube, as they can simply match their cube to
the image. It is possible for the cube to be described using words, but that
would make things unnecessarily complex.
Text Descriptions: While pictures are beneficial, it is also important
that a solution contains some description as well. This will help to empha-
size and explain important features of a solution, such as its notations, the
sides of the cube to focus on at a particular step, and what to look for or
avoid at a step. A textual description provides a sort of guidance through the
solution. A solution containing only images without descriptions will make
it less organized and harder to follow and therefore may be less compre-
hensible.
Movement Notation: The feature that seems to be the most dissimilar
between solutions is the representation of rotating layers of the cube when
describing algorithms. The most commonly used representation is actual
images of the cube, with indications as to which layer needs to be rotated and in
which direction. Generally, this seems to be the best representation. Another
representation is a small square with lines that represent layers and arrows
on those lines that represent the direction of rotation (i.e., , ), as with
the “Adam Solution” and the “C&P Solution.” Finally, another representation
uses letters that represent the layer to rotate and apostrophes to indicate that
they will be moved counterclockwise, as with the “Cartoon Pics Solution,” the
“Short Solution,” and the solution presented in this chapter (Section 11.3).
 Rubik’s Cube • 197

These features are the main determining factors of how comprehen-

sible a solution is. As you can see from the table, the rankings are generally
mixed. Some solutions had more diagrams and no text, others had too much
text and too few diagrams, some were just right, and so on. For a solution to
be ranked high in comprehensibility, it should have just the right amount of
text and diagrams. The movement notation, however, can vary, depending
on how the solution is conveyed.
The most commonly used algorithms and subgoals are those used in
the solution presented in this chapter. Some solutions, however, provide
similar subgoals but with different algorithms. Others give completely dif-
ferent algorithms and subgoals altogether. It is these criteria that determine
a solution’s executability ranking. Longer and more complex algorithms that
are used may generally be harder to carry out correctly and may require
more memory.
Finally, none of the analyzed solutions are optimal. Since the solutions
are found by achieving particular subgoals, the number of moves needed
to achieve these subgoals may very well be more than the shortest number
of possible moves. More advanced solutions are most likely to be optimal,
since they show more precise algorithms (see Section 11.6).

11.5.1 The Most Human Window–Compatible Solution

The solution assessed as MHWC is the “Animation Solution,” which

was developed by Ryan Heise. This solution contains animations, so it is
difficult to reproduce it on paper. However, it can be found in reference [7].
This solution uses the same subgoals, steps, and algorithms as the solution
depicted in this section. It depicts the subgoals so they are easy to under-
stand, while the descriptions are short and clearly explain how to achieve
these subgoals, making this solution highly comprehensible. The solution
also includes every possible case at particular subgoals and how to correctly
resolve them.
The most notable feature of this solution, and one of the main reasons
this solution is considered the MHWC, is the fact that both the Cube
Depiction and Movement Notation are combined together. The solution
depicts three-dimensional images of the cube to show final subgoal arrange-
ments and what the cube would look like at particular cases, increasing the
solution’s comprehensibility and extensionality. The most noteworthy fea-
198 • Artificial Intelligence Problems and Their Solutions

ture of these images is that they are also animated, so when they are clicked,
the layers rotate to show how the specific algorithms are carried out on
command. This increases the solution’s comprehensibility and executability
even further.
To improve this solution even more, the steps of each algorithm should
be depicted along with the animated cubes. This will give the solver a choice
between watching the animation and following the steps of the algorithms
as they are commonly followed.

11.5.2 The Least Human Window–Compatible Solution

Step 1: The Cross

Step 2: Bottom Layer Corners

Each algorithm moves a corner from the top layer to the bottom layer
right below.
 Rubik’s Cube • 199

Step 3: Middle Layer

• First, position the edge from the top layer to the middle layer in
the front.
• Second, position the edge from the top layer to the middle layer in
the back.

Step 4: Edges
If Orange pieces form a line on top, apply the first algorithm. Then apply
the second.

FIGURE 166 The “Short Solution.”

200 • Artificial Intelligence Problems and Their Solutions

Step 5: Corners
View the top layer.
Algorithm: R U R’ U R U2 R’

If ONE corner is Orange, move it to the bottom left. Then apply the
algorithm.

or

If TWO corners are Orange, rotate top layer until a Yellow sticker is
on the bottom left, facing front. Then apply the algorithm.

If ZERO corners are Orange, rotate top layer until an Orange sticker
is on the bottom left, facing left. Then apply the algorithm.
 Rubik’s Cube • 201

Step 6: Permutations
Roatate the top layer until the same color is in the back. If you don’t
have this, apply the algorithm to get them.

The solution assessed as LHWC is the “Short Solution” (Figure 166).

(A representation similar to this can be found in reference [8].) Although it
was intended to be a reference to those who already understand the solu-
tion, we will view this solution in terms of the Human Window as if it were
intended for those who do not know how to solve the puzzle.
This solution also uses the common subgoals like those presented in
Section 11.3. However, unlike the “Animation Solution” and the solution
presented in Section 11.3, this solution is excessively shortened. The sub-
goals are briefly described and have little detail in their description, which
may possibly decrease both the comprehensibility and executability. It also
shows the possible layer arrangements for a subgoal and the algorithm used
to resolve it, but it fails to show every possible case, which may also reduce
executability.
Finally, it is important to take note that the first subgoal does not depict
any description or algorithms. It merely shows the final arrangement and
the arrangement to avoid. If a person is new to solving Rubik’s Cube, he or
she will be unaware of which algorithms to use.

11.6 BEST MACHINE SOLUTION

Rubik’s Cube has approximately 43 quintillion possible arrangements,

so from the start we could disregard any kind of brute force algorithm [9].
Despite this incredibly large number, we have already shown that the cube
202 • Artificial Intelligence Problems and Their Solutions

can be solved in a reasonable amount of steps by solving for subgoals. The

question now is how efficiently can a machine solve the cube?
In 1981, mathematician Morwen Thistlethwaite developed a com-
plex algorithm that could solve the cube with 52 or fewer moves. Since
a fair amount of memory is required, this algorithm is more efficient for
a machine. Unlike the algorithms we have seen in this chapter, Thistle-
thwaite’s algorithm “works on all the pieces at the same time, restricting
them to fewer and fewer possibilities until there is only one possible posi-
tion left for each piece” [10]. There are four phases:
G0 = {L, R, F, B, U, D}
G1 = {L, R, F, B, U2, D2}
G2 = {L, R, F2, B2, U2, D2}
G3 = {L2, R2, F2, B2, U2, D2}
G4 = {1}
These groups represent the subgoals that are to be achieved. They are
depicted as sets, which represent every possible arrangement that can be
achieved, given certain restrictions on layer movements. The letters con-
tained in the brackets represent these restrictions. Each letter represents a
layer of the cube, and those with an exponent next to them are restricted to
180-degree turns only.
To achieve a particular subgoal, the cube must be arranged in one of
the positions contained in the set of positions of the next subgoal. For exam-
ple, say the cube is in one of the arrangements of set G1. This means that to
achieve subgoal G2, the cube needs to be repositioned into a position con-
tained in subgoal G2, which essentially means a position in which the Front,
Back, Up, or Down layers no longer need to be rotated 90 degrees. How-
ever, to achieve such a position, we must follow the rules of G1, and when-
ever the Up or Down face is to be moved, they are strictly to be turned 180
degrees and not 90 degrees. To determine whether or not the next subgoal
is met, the current position of the cube is matched with a table [10].
Despite the improvements made to Thistlethwaite’s algorithm, its
biggest flaw was that the number of phases meant the table had to be
quite large. One person who addressed this problem and improved the
algorithm was Herbert Kociemba. First, he reduced the algorithm to just
two phases:
 Rubik’s Cube • 203

G0 = {U, D, R, L, F, B}
G1 = {U, D, R2, L2, F2, B2}
G2 = {1}
The consequence of reducing the number of phases is that it greatly
increases the number of moves available and the size of the table per phase.
To fix this, he decided “to make pruning tables and use IDA* to solve each
phase,” greatly reducing the state space [11]. To improve the algorithm fur-
ther, Kociemba added a feature that continues the IDA* search even after
the solution is found for a particular phase. This allows the algorithm to find
the solution with the shortest number of steps possible [11].
Although Kociemba’s algorithm is efficient, it is not optimal. In 1995,
Michael Reid analyzed this algorithm and discovered that it takes at most 29
moves to solve the cube with the algorithm [12]. However, in 2010, Tomas
Rokicki, Herbert Kociemba, Morley Davidson, and John Dethridge proved
that Rubik’s Cube can be solved in at most 20 moves, no matter what the
position is. Although there is no specific algorithm that can solve the cube in
20 moves, this was proven by doing extensive research to solve all 43 quin-
tillion possible arrangements. By using simple mathematics programming
and by partitioning these arrangements into particular sets, it took approxi-
mately 35 CPU years to solve every position of Rubik’s Cube, discovering
that it only takes at most 20 moves to solve the cube from any arrangement.
This number is now referred to as “God’s Number,” since it is the opti-
mal number of moves, indicating that “God’s Algorithm” would result in 20
moves [12]. The term “God’s Algorithm” refers to an algorithm that solves
Rubik’s Cube, or any other combinatory puzzle, in the shortest number of
steps. Some algorithms have come close, but none have received the title.
Regardless of which algorithm is used to solve Rubik’s Cube, the complex-
ity for any algorithm in the worst case was found to be O( n2/log(n) ), accord-
ing to a research paper on algorithms for Rubik’s Cube titled “Algorithms for
Solving Rubik’s Cubes” [13].

11.7 PLAYABLE PROGRAM

A playable Rubik’s Cube game is available at http://activeden.net/item/

cube-assembler/full_screen_preview/3131480. The game depicts a three-
dimensional cube with layers that can be rotated with a simple stroke of
204 • Artificial Intelligence Problems and Their Solutions

the mouse. Also, the cube itself can be rotated at will by shifting one of two
scrollbars, allowing the player to get a full view of the cube. Additionally,
the player can choose between solving the easy 2” × 2” × 2” cube up to the
more difficult 5” × 5” × 5” cube. Finally, the player has the option to change
the colors of the cube faces to any color he or she desires.

11.8 REFERENCES

1. The History of Rubik’s. (2013). Available at http://www.rubiks.com/world/

history.php. Accessed on January 9, 2014.
2. MSKIFF. (2013). The Rubik’s Cube: A Brief History. Available at http://
azscitechfest.org/blogs/rubik’s-cube-brief-history. Accessed on January 9,
2014.
3. Goudey, Christophe. (2013). The Rubik’s Cube History. Available at http://
cubeland.free.fr/infos/infos.htm. Accessed on January 9, 2014.
4. Rosenberg, J. (2013). History of the Rubik’s Cube. Available at http://
history1900s.about.com/od/1980s/a/rubikscube.htm. Accessed on January 9,
2014.
5. The Rubiks 4 x 4 x 4—“Rubiks Revenge.” (2013). Available at http://www.
rubiks.com/products/3d_puzzles/rubiks_cube-4x4.php. Accessed on January 9,
2014.
6. The Rubiks 5 x 5 x 5—”Professor’s Cube.” (2013). Available at http://www.
rubiks.com/products/3d_puzzles/rubiks_cube-5x5.php. Accessed on January
9, 2014.
7. Heise, Ryan. (2013). Beginner’s Rubik’s Cube Solution. Available at http://
www.ryanheise.com/cube/beginner.html. Accessed on January 9, 2014.
8. BADMEPHISTO. (2013). Rubik’s Cube Solution: Printable Cheat Sheet.
Available at http://badmephisto.com/begsoln/. Accessed on January 9, 2014.
9. Dillow, C. (2010). “God’s Number Revealed: 20 Moves Proven Enough to
Solve Any Rubik’s Cube Position.” Popular Science. Available at http://www.
popsci.com/science/article/2010-08/gods-number-revealed-20-moves-will-
solve-any-rubiks-cube-position. Accessed on January 9, 2014.
10. Jaap. (2013). Thistlethwaite’s 52-Move Algorithm. Available at http://www.
jaapsch.net/puzzles/thistle.htm. Accessed on January 9, 2014.
 Rubik’s Cube • 205

11. Jaap. (2013). Computer Puzzling—Kociemba’s Algorithm. Available at http://

www.jaapsch.net/puzzles/compcube.htm. Accessed on January 9, 2014.
12. Rokicki, T., Kociemba, H., Davidson, M., and Dethridge, J. (2013). God’s
Number Is 20. Available at http://www.cube20.org/. Accessed on January 9,
2014.
13. Demaine, E.D., Demaine, M.L., Eisenstat, S., Lubiw, A., and Winslow,
A. (2011). Algorithms for Solving Rubik’s Cubes. Cambridge, MA: MIT
Computer Science and Artificial Intelligence Laboratory.
 CHAPTER

12
THE PRISONER’S DILEMMA

FIGURE 167 The Prisoner’s Dilemma.

12.1 THE TRADITIONAL PROBLEM*

In the movie The Postman Always Rings Twice,** the two lead charac-
ters fall in love and decide to “eliminate” the husband of the female char-
acter. The police have insufficient evidence to obtain convictions. However,
after the murder, the lovers are apprehended by the police and held for

* This section is based on Lucci and Kopec (2013). Artificial Intelligence in the 21st Century.
Mercury Learning, Dulles, Virginia, pp. 130-132.
** The 1981 version is a reproduction of an earlier movie based on the novel by James M.
Cain (1934) that stars Jack Nicholson and Jessica Lange.
208 • Artificial Intelligence Problems and Their Solutions

questioning in separate interrogation rooms, where each is offered the same

proposition: “Tell the truth and incriminate your partner, and we’ll go easy
on you.” Acquiescing to this request is called “defecting.” Both perpetrators
know that their accomplice is being offered this same deal. What should
each prisoner do? This predicament is intriguing in that neither person is
privy to their counterpart’s thoughts (they are in isolated cells). In an ideal
world, both would remain loyal to each other (as per their original agreement
in perpetrating the crime) and without the existence of further evidence and
supporting testimony against them would probably be convicted of a lesser
crime. However, if either accomplice defects, then the other would certainly
be better off by defecting as well, rather than serving a murder sentence.
This so-called Prisoner’s Dilemma (Figure 167) was first formulated
in game-theoretic terms by Merrill Floyd and Melvin Dresher at the RAND
Corporation in 1950 [1]. At the time, most countries were involved in a Cold
War with two nuclear super-powers: the Soviet Union and the United States.
Should these two countries cooperate with each other and work toward
mutual disarmament (aware at all times that the other side might renege),
or should each continue to create new and more lethal armaments? This is
the dilemma that confronted our planet for four decades, from the end of
World War II until the eventual end of the Cold War in 1989 with the fall of
the Berlin Wall. The Prisoner’s Dilemma aptly models that era of mistrust.
This dilemma can also be modeled by the payoff matrix in Figure 168.

Prisoner B

Cooperate (remain silent) Defect (betray partner)

(remain silent)

A: 1 Year A: 10 Year
Cooperate
Prisoner A

B: 1 Year B: 0 Year
(betray partner)

A: 0 Years A: 5 Years
Defect

B: 10 Years B: 5 Years

FIGURE 168 Payoff Matrix for The Prisoner’s Dilemma.

 The Prisoner’s Dilemma • 209

A payoff matrix specifies the return to each player for every combination of
actions by the two game participants.
Assume that the two players (that is, the prisoners) in this game are
rational and want to minimize their jail sentences. Each prisoner has two
choices (as did the two characters in The Postman Always Rings Twice):
cooperate with their partner in crime and remain silent or defect by con-
fessing to the police in return for a lesser sentence.
You might notice that this game differs in an important aspect from the
games discussed earlier in this book. To determine a course of action in
the other games and puzzles discussed, you need to know your opponent’s
course of action. For example, if you are moving second in a game of tic-tac-
toe, you need to know where the other player has placed the initial X. This
is not the case in The Prisoner’s Dilemma. Suppose you are the A player
and you choose to defect. However, the B player decides to remain loyal
and chooses the cooperation strategy. In this case, your decision results in
no prison time as opposed to the one-year term had both of you chosen to
cooperate. If your partner chooses to defect, your outcome is still superior
if you choose to defect. In game-theoretic terms, defecting is a dominant
strategy. Because you assume that your opponent in this game is rational,
he or she will arrive at the same strategy.
The strategy {Betray, Betray} shared by the two participants is referred
to as a Nash Equilibrium. A change in strategy by either player results in
a lesser return to them (i.e., more jail time).
As shown in Figure 168, if each player acts more on faith than rational-
ity (faith that their partners would remain loyal), then the total payoff would
exceed the total of 10 prison years accorded by the Nash Equilibrium of
{Defect, Defect}. This strategy of {Cooperate, Cooperate} yields the best
possible outcome in terms of total payoff to the two players. This optimal
strategy is referred to as Pareto Optimal. It should be noted that The
Prisoner’s Dilemma is not a zero-sum game. Why not? In such games, Nash
Equilibrium does not necessarily correspond to a Pareto Optimal game [2].

12.2 THE ITERATED PRISONER’S DILEMMA

If you play The Prisoner’s Dilemma only once, defecting is a domi-

nating strategy for either player. In another version of this game, you play
210 • Artificial Intelligence Problems and Their Solutions

repeatedly—n times—where there is some memory of previous actions.

When having more than one turn, each player is not as quick to defect,
knowing that revenge from the opponent is forthcoming. One strategy
might be to start with a single cooperation to give your opponent a chance
to act compassionately. If your opponent chooses to defect anyway, you can
counter by continually defecting. If your opponent eventually chooses to
cooperate, you can return to a more magnanimous policy. It is interesting
that with The Prisoner’s Dilemma and its iterations or variations, we can
ascribe words to numbers representing behaviors. Behaviors could include,
for example, betrayal, defection, cooperation, trust, revenge, magnanimity,
compassion, and so on. These concepts are fully discussed in Robert Bram’s
excellent treatise “Theory of Moves” [3]. The exercises at the end of this
chapter discuss other two-person games that are similar to The Prisoner’s
Dilemma.

12.3 APPLICATIONS IN DIVERSE AREAS

The Prisoner’s Dilemma allows us to address and depict problem situ-

ations in the social sciences—for example, economics, political science,
psychology, philosophy—in a numerical way such that they may be better
understood and solutions can be effectively represented in a convenient way.
In addition, through the “Iterated Prisoner’s Dilemma,” generations of the
problems under consideration can be simulated. The difference between The
Prisoner’s Dilemma and the other problems that we have discussed to this
point is that it is concerned with considerations of “What if?” In other words,
actions and responses to them are considered and evaluated in as realistic a
way as the “matrix system” allows. Chapter 13 presents the problem “Ten
Pirates and Their Gold.” This problem is similar to The Prisoner’s Dilemma,
since to determine what course of action should be taken, you must consider
a number of possible “moves” and scenarios several steps ahead.
The following are well-known problems that have been posed in Lucci
and Kopec’s textbook (page 132). Consider whether they are instances of
essentially the same problem as The Prisoner’s Dilemma. If you conclude
that a problem is equivalent to The Prisoner’s Dilemma, then design an
appropriate payoff matrix. Comment on the existence of a Nash equilib-
rium and Pareto Optimal in each case.
 The Prisoner’s Dilemma • 211

UNIX and GPL Cooperate Defect

Cooperate (5, 5) (10, 0)

Defect (0, 10) (1, 1)

TABLE 16 Example 1.

Example 1: Linux is a version of Unix that has been developed

under the GPL. Under this agreement, you are given free software
and may study the source code, which is modifiable. You can defect
by keeping these improvements to yourself, or you can cooperate by
distributing the improved version of the code. In fact, cooperation is
forced by making it illegal to distribute only the source code without
your improvements.
Solution: In this problem, higher numbers mean more success.
If you cooperate/cooperate by distributing the source code with your
improvements, there is mutual benefit. This is essentially a variation of The
Prisoner’s Dilemma. If you defect either by making no improvements or by
distributing the software illegally, these actions are represented by cooper-
ate/defect (10, 0) or defect/cooperate (0, 10), respectively. Finally, if you
distribute the software AND make no improvements to it, you are perform-
ing defect/defect, which means you are a thief and a liar.

Cigarette Advertising Advertise No Advertising

Advertise (1, 1) (10, 1)
No Advertising (1, 10) (1, 1)
TABLE 17 Example 2.

The Nash Equilibrium is represented by (0, 10) and (10, 0) as in the

original Prisoner’s Dilemma.
Example 2: Cigarette companies were at one time allowed to adver-
tise in the United States. If only one company decided to advertise, an
increase in sales invariably followed. However, if two companies launched
advertising campaigns, their ads would essentially cancel each other out,
and no increase in revenue resulted.
212 • Artificial Intelligence Problems and Their Solutions

Solution:
This is akin to the original Prisoner’s Dilemma, but with a twist. If both
companies advertise, the Nash Equilibrium is if either company advertises.
If both companies advertise or neither company advertises, then the result
is the status quo (1, 1).
Example 3: In New Zealand, newspaper boxes are left unlocked. One
can easily steal a paper (defect). Naturally, if everyone did this, no papers
would remain [2]. Clearly this is an example of the Iterated Prisoner’s
Dilemma, because even though at first it may seem that the “stolen”
newspapers have little effect, over time there would be no newspapers
left to sell.
Example 4: (Tragedy of the Commons): A village has n farmers,
and grassland is limited. Each of these farmers may decide to keep a sheep.
Each farmer obtains some utility from these sheep in the form of wool and
milk. However, the common grassland (the Commons) will suffer some-
what from the sheep grazing there.
Solution: The solution is described below. It was first described by
Hardin in The Tragedy of the Commons [5].
When multiple individuals act independently and they apply their
own self-interest, they will deplete a shared resource. This will occur even
though it is in no one’s long-term interest. Originally formulated in terms
of sheep and common grassland, it can readily be applied to many con-
temporary conundrums. Global climate is one such arena. It is cheaper to
manufacture products if little or no attention is paid to pollution and emis-
sion control, but it is certainly detrimental to the earth’s long-term health.
A nice example for this scenario is available at www.skepticalscience.com

12.4 REFERENCES

1. Merrill, F., and Dresher, M. (1961). Games of Strategy: Theory and

Applications. Upper Saddle River, NJ: Prentice Hall.
2. Poundstone, W. (1992). Prisoner’s Dilemma. New York: Doubleday.
3. Brams, S. (1994). Theory of Moves. Cambridge, UK: Cambridge University
Press.
 The Prisoner’s Dilemma • 213

4. Lucci, S., and Kopec, D. (2013). Artificial Intelligence in the 21st Century.
Dulles, Virginia: Mercury Learning Inc.
5. Hardin, G. (1968). “The Tragedy of the Commons.” Science (162) 3859:
1243–1248.
 CHAPTER

13
MISCELLANEOUS PROBLEMS

T
his chapter contains five smaller problems from different areas, such
as probability, logic, and mathematics. We believe they are sufficiently
intriguing to be of interest and value to any dedicated problem solver.

FIGURE 169 Cards Laid Out on a Table.

13.1 CARDS/COINS IN THE DARK

Cards/Coins in the Dark is a math puzzle and has two variants: one with
a deck of cards and one with a set of 100 coins (Figure 169).

13.1.1 Cards in the Dark

The problem is as follows: You are in a dark room and are given a
deck of 52 cards. You are told that 13 of the 52 cards are faceup and are
distributed throughout the deck. You are asked to divide the deck into two
216 • Artificial Intelligence Problems and Their Solutions

piles consisting of the same number of faceup cards. You cannot distin-
guish between the faceup and facedown cards by feeling them, and also,
since the room is dark, you cannot see the cards. How can you accomplish
this task?
Solution
There are 13 faceup cards. If we randomly divide the whole deck of
cards into two piles, the 13 faceup cards will be divided between the two
piles. Say if the first pile has n faceup cards, then the second pile will have
13 – n faceup cards. So if the first pile had 13 – n faceup cards instead of n
faceup cards, the two piles will have an equal number of faceup cards.
Thus, if the first pile has a total of 13 cards, then it would have 13 – n
facedown cards, and if we flip all the cards in the first pile, we would have
13 – n faceup cards. So the solution is to divide the cards into two piles, one
with 13 cards and one with 39 cards and then flip all the cards in the first
pile. Figures 170 to 172 show this solution.

FIGURE 170 The Deck of Cards Laid Out on a Table.

 Miscellaneous Problems • 217

Number of faceup cards in Pile 1 = 3.

Pile 1

Pile 2

Number of faceup cards in Pile 2 = 13 – 3 = 10.

FIGURE 171 The Two Piles of Cards.

Number of faceup cards in Pile 1 = 10.

Pile 1

Pile 2

Number of faceup cards in Pile 2 = 10.

FIGURE 172 After Flipping All the Cards in Pile 1.

218 • Artificial Intelligence Problems and Their Solutions

13.1.2 Coins Version of the Problem

The coins version of this problem has 100 coins on a table in a dark
room. Of the 100 coins, 10 coins have heads-up, and the goal is to divide the
coins into two piles with an equal number of coins heads-up.
The solution is similar to that of the cards version of the problem. We
choose any 10 coins to form the first pile and flip all of them over. The
remaining 90 coins form the second pile.
If the first pile originally had n coins with heads-up, the second pile will
have 10 – n coins heads-up. After all the coins in the first pile are flipped
over, the first pile will also have 10 – n coins.

13.1.3 References

1. http://www.calpoly.edu/~mcarlton/riddles.html
(July 10, 2013)
2. http://wordplay.blogs.nytimes.com/2013/05/20/
dark-2/ (July 10, 2013)

FIGURE 173 Pirate and Gold

13.2 TEN PIRATES AND THEIR GOLD

The Ten Pirates and Their Gold Problem has been posed for over 20
years (Figure 173), but it gained attention when some serious mathema-
ticians (Ian Stewart) and computer scientists (Stephen M. Omohumdro)
tackled it with some variations in the late 1990s. The problem is as follows:
Ten pirates find a buried treasure of 100 pieces of gold (Figure 174). The
challenge is to divide the gold in some desirable way according to some con-
straints (rules). The first constraint is that Pirate 1 is the lead pirate; Pirate
2 is the second in charge; Pirate 3 is the third most powerful; and so on.
The pirates have a scheme for dividing the money. They agree that the first
pirate (P1) will make a proposal for how the money is to be divided. If 50%
or more of the pirates agree with the P1 system, it will be put into effect. If
not, then P1 will be killed, and the next most powerful pirate becomes the
lead pirate. Now, again with one fewer pirate, the process repeats. The new
lead pirate, P2, now suggests a new process for divvying up the gold. It will
 Miscellaneous Problems • 219

be voted on, with a 50% vote needed for the leader’s suggestion to pass; less
than 50% results in the death of the lead pirate.
All the pirates are very greedy and savvy, so they will vote against a
proposal if it means they will get more gold if the proposal fails, and thus a
lead pirate is killed. They will never vote for a proposal that will give them
less gold or no gold.

When you first hear this problem and consider possible solutions, you
might think the following: Let’s say the lead pirate (P1) wants to be pretty
fair and takes 19 pieces of gold, offering 81 pieces of gold to the remaining
9 pirates (9 each). The problem is that he needs 4 pirates to agree with him.
And if the 4 pirates see a way to get more gold, they will not vote to go with
him.
The lead pirate taking 19 and the rest taking 11 seems pretty fair, but
the 4 voting pirates will see a better way for them. If P1 is eliminated (no
vote in favor), then a nice scenario might seem to be P2 gets 12 and the
others get 8 × 11 (= 88). This seems very reasonable and equitable, but
pirates are not known for being equitable. So the pirates might eliminate
P2, seeing a scenario like P3 = 16, and the others get 7 × 12 (= 84). Still,
there will be pirates who see a way for them to get more gold for themselves
who would exercise their veto vote to possibly achieve this.

FIGURE 174 Ten Pirates and their Gold

220 • Artificial Intelligence Problems and Their Solutions

Using this approach, we can see the following logic could lead to some
real results: Imagine if there were only two pirates, P1 and P2 (Figure 175).
P1 would want to take all the gold (100 pieces), leaving P2 with no gold and
no say about it (the 50% rule).

P1 P2

100 Pieces 0 Pieces

FIGURE 175 Two Pirates

Now let us consider the situation with three pirates: P1, P2, and P3
(Figure 176). Pirates are not stupid. They would realize that if they vote
against you, P2 gets all the gold and P3 gets no gold after you are gone.
Knowing this, P3 will accept your offer of 1 piece of gold, with no gold to
P2, P2 will not be happy, but your proposal will pass. With four pirates, the
same proposal (P3 gets 1 coin) will pass, with P3 giving you the one vote you
need to pass (50%) (Figure 177).

P1 P2 P3

99 Pieces 0 Pieces 1 Piece

FIGURE 176 Three Pirates

 Miscellaneous Problems • 221

P1 P2 P3 P4

99 Pieces 0 Pieces 1 Piece 0 Pieces

FIGURE 177 Four Pirates

The same logic applies to up to 10 pirates, whereby you would give one
piece of gold to P3, P5, P7, and P9 and keep the other 99 pieces for your-
self. Those four pirates would support your proposal, since otherwise they
would receive no gold if your proposal fails.
About the time of Stewart’s article in 1999, Steven Omohundro took
the problem to another level by asking what happens if there are not only
10 pirates but as many as 200 or 500? He concluded that with 200 pirates,
P200 will continue by offering nothing to odd-numbered pirates from P1 to
P199 and one gold coin to even-numbered pirates P2 through P198.
However, the situation gets complicated if we go beyond 200 pirates,
as P201 and P202 will find themselves lacking sufficient “bribing power”
to get support. Matters get even more complicated as we consider 203,
204, . . . , 207 pirates. Omohundro’s conclusion, which Ian Stewart shares
in the article below, is that as we approach 500 pirates, a pattern that con-
tinues indefinitely emerges.
Pirates who can make winning proposals (always to give themselves
nothing and to bribe 100 fellow pirates) are separated from one another by
ever-longer sequences of pirates who will be thrown overboard no matter
what proposal they make—and whose vote is therefore ensured for any
fiercer pirate’s proposal. The pirates who avoid this fate are P201, P202,
P204, P208, P216, P232, P264, P328, P456, and so on—pirates whose num-
ber equals 200 plus a power of 2.

REFERENCE

Stewart, I. (1999). “A Puzzle for Pirates.” Scientific American, 280 (5): May
1999, 98–99. Available at http://omohundro.files.wordpress.com/2009/03/
stewart99_a_puzzle_for_pirates.pdf /. Accessed on January 10, 2014.
222 • Artificial Intelligence Problems and Their Solutions

FIGURE 178 Halmos’s Handshake Problem

13.3 HALMOS’S HANDSHAKE PROBLEM

Paul Halmos was a famous Hungarian mathematician. He wrote a

charming biography of the great Jon von Neumann, who is generally cred-
ited with developing the architecture of the serial computer still used
today. He also developed the following logic problem, called Halmos’s
Handshake Problem (Figure 178).
As is common, academics will occasionally attend dinner parties.
Halmos and his wife attended such a party with four other couples. Dur-
ing the cocktail hour, some of those present shook hands but in an unsys-
tematic way, with no attempt to shake all the guests’ hand. Of course, no
one shook his or her own hand, no one shook hands with his or her spouse,
and no one shook hands with the same person more than once. During
dinner, Halmos asked each of the nine other people present (including
his own wife) how many hands that person had shaken. Under the given
conditions, the possible answers ranged from 0 to 8 hands shaken. Halmos
noticed that each person gave a different answer. One person claimed not
to have shaken anyone else’s hand, one person had shaken exactly one
other person’s hand, one person had shaken exactly two hands, and so on
up to one person who claimed to have shaken hands with all the others
present, except his or her spouse—so 8 handshakes in total. So, in sum-
mary, of the 10 people present, people gave answers from 0 to 8 hands
shaken (Figure 179).
 Miscellaneous Problems • 223

FIGURE 179 Ten People and Handshakes

224 • Artificial Intelligence Problems and Their Solutions

Now, here is the question: How many hands did Halmos’s wife shake?

FIGURE 180 How Many Shook His Wife’s Hand?

Let’s see what we already know (Figure 180). We’ll start by considering
the person who claims to have shaken 8 hands and call this person A. Whose
hands did A shake? Necessarily, A shook hands with everyone else present,
apart from his or her spouse (whom we call A(w)). So everyone besides A’s
spouse {A(w)} shook at least one hand—namely, A’s hand (Figure 181).

FIGURE 181 Person A—8 Shakes

 Miscellaneous Problems • 225

So who is the person who shook 0 hands? Since everyone except A’s
spouse shook at least one hand, the person who shook 0 hands must be A’s
spouse, A(w) (Figure 182).
Now consider the person who
shook 7 hands, whom we’ll call B.
B didn’t shake hands with himself or
herself or with his or her spouse {B(w)}
0 Handshakes or with A(w) (because A(w) shook no
hands). That leaves only 7 other people,
FIGURE 182 Person A’s Wife—0 Shakes so B shook hands with everyone else
except B and his or her own spouse {B(w)}
(Figure 183). And each of these people also shook hands with A, mean-
ing that all of these people shook at least two hands: A’s and B’s. So the
only person left who could have shaken just 1 hand is B’s spouse, B(w)
(Figure 184). So just as the person who shook 8 hands is married to the
person who shook 0 hands, the person who shook 7 hands is married to
the person who shook only 1 hand. We can now see a pattern evolving in
the solution.

FIGURE 183 Person B—7 Shakes

Similar analysis can be applied to

the person who shook 6 hands, C, who
could not have shaken hands with him-
self or herself or his or her spouse, A(w)
(who shook 0 hands), or B(w) (who
shook only A’s hand) (Figure 185). So
C shook hands with everyone else,
and since C also shook hands with A
and B, the only person left who could
have shaken just 2 hands is C’s spouse,
FIGURE 184 Person B’s Wife—1 Shake {C(w)} (Figure 186).
226 • Artificial Intelligence Problems and Their Solutions

FIGURE 185 Person C—6 Shakes

FIGURE 186 Person C’s Wife—2 Shakes

Applying the same logic one more time shows that the person who
shook 5 hands, D, is married to the person who shook 3 hands, {D(w)}
(Figures 187 and 188).

FIGURE 187 Person D—5 Shakes

FIGURE 188 Person D’s Wife—3 Shakes

 Miscellaneous Problems • 227

What does this leave us? We know that all the possible people who gave
the answers 0, 1, 2, 3, 5, 6, 7, and 8 are all married to other people who
answered Halmos’s question. The only person who is not married to one
of the people who answered the question is the one who answered 4—and
that must be Halmos’s wife {H(w)}. So Halmos’s wife (H(w)) shook 4 hands,
and it follows that Halmos (H) did, too (Figures 189 and 190).

FIGURE 189 Halmos—4 Shakes

FIGURE 190 Halmos’s Wife—4 Shakes

This problem illustrates that by using logical inferences and

given the small number of permutations for the problem, it is pos-
sible to exhaustively enumerate what the number of handshakes was for
each person who attended the party. Also, given the small number of
possible permutations, it is possible to use a simple notational system and/
or graphs to represent solutions.

REFERENCE

Halmos, P. (1973). “The Legend of John von Neumann.” The American

Mathematical Monthly 80 (4). Available at http:// www.jstor.org/stable/2319080
May 5. Accessed on January 10, 2014.
228 • Artificial Intelligence Problems and Their Solutions

FIGURE 191 Airline Seats

13.4 RANDOM AIRLINE SEATS PROBLEM

The Random Airline Seats Problem is a type of probability and logic

problem whose results are surprising and unexpected (Figure 191). Here
is one phrasing of the problem: One hundred people line up in the waiting
area to board a plane that has 100 seats available. The first person in this
line has lost her boarding pass. Without knowing what her assigned seat is,
she randomly chooses a seat. As a result, for each person entering the plane,
there are only two possibilities: either his assigned seat is available or it’s
not. If his seat is not available, he chooses an unoccupied seat randomly.
Based on this information, when the 100th person finally enters the plane,
what is the probability that he will find his assigned seat unoccupied [1]?
This problem requires a fair amount of logic to determine the solution.
Even with the solution revealed, it is still not quite obvious as to how it is
obtained. However, those who understand the laws of probability may have
a better understanding of how to approach this problem and why the solu-
tion is correct.

Solution
There is a 50% chance that the 100th passenger will get to sit in his or
her assigned seat. With such a large number of people entering the plane,
and the 100th passenger entering last, this answer seems implausible!
However, with an analytical approach and some logic, the solution can be
easily understood.
A common way to solve this riddle is to mathematically determine
the chance that each person sits in his or her assigned seat, but the math
gets complicated quickly. However, we can easily come to our answer by
analyzing the possibilities that can occur and the effects caused by those
possibilities.
 Miscellaneous Problems • 229

For this solution, we’ll call a passenger Pn, where n represents that per-
son’s place in line. For example, the person who is fifth in line is P5.
We begin by making two observations:

1. If any of the first 99 people sit in P100’s assigned seat, P100 WILL NOT
get to sit in his seat. This observation should be obvious, since somebody
else sitting in her seat means he cannot sit in it.
2. If any of the first 99 people sit in P1’s seat, P100 WILL get to sit in his
seat. This observation is not as obvious. To make things clear, let’s say, for
example, that P1 sits in P2’s seat. This means that P2 cannot sit in his as-
signed seat, so he must sit in another. Potentially, P2 will sit in any other
seat he chooses. Let’s say P2 sits in P3’s seat. Then P3 will be in the same
predicament that P2 was in and, without an assigned seat, will randomly
sit anywhere he wants. Let’s then say that P3 sits in Pn’s seat. All of the
passengers between P3 and Pn will get to sit in their assigned seats, but
Pn will find his seat occupied and therefore will randomly sit in any other
passenger’s unoccupied seat. This will continue until a passenger—say,
P50—happens to sit in P1’s seat. Once this happens, the remaining pas-
sengers will not find their seats occupied and will now be able to freely
sit in their assigned seats. P51 will sit in his assigned seat, P52 will sit in
her assigned seat, and so forth. Therefore, P100 will get to sit in his or
her own seat.
So to summarize these two observations, if P100’s seat becomes occu-
pied, P100 will not get her seat. If P1’s seat becomes occupied, P100 will
get her seat. Therefore, only these two factors determine whether or not
P100 will get his or her seat. As for the rest of the seats, we can see from our
observations that they make absolutely no difference in determining P100’s
last seat. When the other passengers choose seats that do not belong to P1
or P100, they are just prolonging the final determination.
It is important, however, to understand that the probability of a passen-
ger randomly choosing a specific seat is exactly the same for every seat. For
example, when P1 enters the plane, he has 100 seats to choose from and
therefore has a 1/100 chance that he will sit in his own seat and 1/100 chance
that he will sit in P100’s seat. If he happened to sit in P2’s seat, then P2 will
have 99 seats to choose from. Therefore, there will be a 1/99 chance that he
will sit in P1’s seat and a 1/99 chance that he will sit in P100’s seat. Gener-
ally, if there are n remaining seats available on the plane, the passenger that
is next in line will have a 1/n chance he will sit in P1’s seat and a 1/n chance
230 • Artificial Intelligence Problems and Their Solutions

he will sit in P100’s seat. Therefore, the chance of any passenger sitting in
either P1’s seat or P100’s seat will always be equally likely.
To summarize this solution:

1. P100 will get her assigned seat if any other passenger sits in P1’s seat.
2. P100 will not get her assigned seat if any other passenger sits in P100’s
seat.
3. 1 and 2 are the only two choices that determine P100’s seating, and they
are always both equally likely to happen.
Therefore, P100 will have a 50% chance that she will be seated in her
assigned seat if P1 randomly sits anywhere.

REFERENCE

Galvin, D. (2013). “Probability Puzzler 3—The TSA Wouldn’t Like This.”

Available at http://www3.nd.edu/~dgalvin1/Probpuz/probpuz3.html. Accessed
on January 9, 2014.

FIGURE 192 The Birthday Problem

13.5 THE BIRTHDAY PROBLEM

The Birthday Problem (or The Birthday Paradox) (Figure 192) is

another probability problem whose solution contains rather unexpected
results. The original problem is as follows: Suppose you have a random
 Miscellaneous Problems • 231

group of people in a room. How many people do you need in this group so
there is more than a 50% chance that two of these people have the same
birthday (month and day)? Unlike The Random Airline Seats Problem, this
problem requires a better understanding of the laws of probability.

Solution
The answer is 23. In a group of 23 random people, there is a 50.76%
chance that two or more of them have the same birthday! Although this
does not seem to make sense at first, it should become clearer when we
present some probability theory.
The question is asking about the probability with regard to two random
people in a group of people. A common misconception is that the problem
asks for the odds that two specific people will have the same birthday. A
direct computation to solve the problem is relatively complex. However,
to simplify the problem, we will first calculate the probability that no two
people have the same birthday [1].
Let’s assume that February 29 does not exist and that there are only
365 days in every year, meaning there are a total of 365 possible birth dates.
Say we have only 1 person, whom we will call Person A. Because Person A
is only one individual, her birthday could be any one of 365 days, and there
is no chance that her birthday will be the same as anyone else’s. Therefore,
there is a 365/365 = 1 = 100% probability that Person A will not have the
same birthday as someone else.
Let’s now add a second person to our group, named Person B. Since
Person A already exists in the group, there is one fewer possible birthday
that Person B could have whereby B does not match any other birthdays in
the group, which is basically Person A’s birthday. Therefore, the probability
that Person B does not have the same birthday as A is 364/365. Now we
need to find the probability that the group does not have matching birth-
days. It is important to take note that a person having a certain birthday
does not depend on another person having that same birthday. Therefore:

P(None) = (365/365) × (364/365) = .9973 = 99.73%

This means there is a 99.73% probability that no matching birthdays
exist in the group. If we continue to add a Person C, that person will have a
363/365 chance that she does not have the same birthday as the other two
people and therefore:
232 • Artificial Intelligence Problems and Their Solutions

P(None) = (365/365) × (364/365) × (363/365) = .9918 = 99.18%

From these three cases, we can see a pattern emerging. It can be con-
cluded that in a group with n persons the probability that no people will
share a birthday is:
P(None) = [(365 × 364 × ……… × (365 – (n – 1)) )] / (365n)
Simplifying this, we get:

P(None) = [365! / (365 – n)!] / (365n)

Now that we have developed a general formula, it is simple to calculate
the probability that two or more people WILL have the same birthday in
the group. All that needs to be done is to subtract 1 from the result of the
above formula. Therefore:

P(~None) = P(2 or more) = 1 – [365! / (365 – n)!] / (365n)

Using our previous example where Person C is added to the group, the
probability that two or more people will have the same birthday is:

P(2 or more) = 1 – .9918 = .0082 = .82%

This means there is less than a 1% chance that two or more people will
have the same birthday. This seems very small, but by using our formula
and substituting 23 for n:

P(2 or more) = 1 – [365! / (365 – 23)!] / (36523) = 1 – 0.4927

= .5073 = 50.73%
Thus, we can conclude that in a group of 23 people there is approxi-
mately a 50% chance that at least two people will have the same birthday.

REFERENCE

Strogatz, S. (2012). “It’s My Birthday Too, Yeah.” New York Times, October 1.
Available at http://opinionator.blogs.nytimes.com/2012/10/01/its-my-birthday-
too-yeah/?_r=0. Accessed on January 9, 2014.
 CHAPTER

14
CONCLUSION: TOWARD A THEORY
FOR PROBLEM SOLVING

14.1 HUMAN WINDOW STUDY

Thus, the study of comparison of Human Window solutions reveals some

important properties that good solutions possess, with the key being extension-
ality and a good choice of representation. The more visual or graphical a rep-
resentation is, the more effective it will be as an aid to most problem solvers.
Animations or playable games comprise some of the most extensional represen-
tations that people can enjoy. These two factors also affect the comprehensibil-
ity of the solution—another important aspect of the solutions ranked highest in
our study of the Human Window. (See Appendix C on the CD.) A theory about
the importance of the Human Window may well serve as a bridge between
the information age (with its overabundance of data) and the knowledge age,
whereby knowledge can be used for effective learning and decision making.
These results are consistent with a paradigm for science education
and problem solving known as the “constructivist epistemology.” The basic
premise of this school of learning is that knowing is doing. If you can do
something, then you know it. The best teachers are those who know how
to do something and can most effectively explain to others what they know
with words, graphics, and methods.

14.2 LESSONS LEARNED

In this section, we discuss some of the lessons learned from studying

and analyzing the most of the problems in this book as a possible starting
234 • Artificial Intelligence Problems and Their Solutions

point for developing a theory for problem solving. Table 18 summarizes

what we have learned about problem solving by examining the 10 major
problems in this book.

Knowledge Problem-Solving
No. Problem Problem Type
Representation Strategies
The Search tree,
Logic, constraint Solving subgoals,
Missionaries state transition
1 satisfaction backtracking,
and Cannibals diagram, problem
problem symmetry, search
Problem state space matrix
The 12 Coins Logic, Problem reduction,
2 Search tree, table
Problem mathematical recursion
Mathematical, Generate and test,
Knowledge
constraint deduction, arithmetic,
3 Cryptarithms table, constraint
satisfaction backtracking, forward
hypergraph
problem checking, heuristics
The Red Don- Solving subgoals,
4 Pictorial Sliding block puzzle
key Puzzle bidirectional search
Solving subproblems;
numerous methods,
Graphical, row
5 The 15 Puzzle Sliding block puzzle including disjoint
by row
patterns, database
heuristics
Graph—
The Knight’s Graphical; Tessel- Patterns and
6 Hamiltonian Cycle
Tour Problem lations heuristics
Problem
Characters or Logic, Deduction, Five-
7 Mastermind
Graphic combinatorics Guess Algorithm

The Monty Hall Pictorial, logic,

8 Pictorial Logic, probability
Problem probability

Subgoals, steps,
9 Rubik’s Cube Pictorial 3 blocks; patterns
algorithms
Logic, constraint
The Prisoner’s Deduction, iterative,
10 Table satisfaction
Dilemma strategical thinking
problem

TABLE 18 Problem-Solving Strategies.

 Conclusion: Toward A Theory For Problem Solving • 235

14.2.1 The Missionaries and Cannibals Problem

This problem is simpler when compared to the other three problems

and can be solved by brute force or by trial and error. However, the most
important lesson is that the choice of representation greatly affects the
problem-solving process. Therefore, it is advised that before attempting to
solve a problem, students take time to get familiar with the problem, under-
stand its nature, and choose a representation of the problem that most
closely resembles the learner’s or problem solver’s internal repre-
sentation and understanding of the problem space. For example, this
problem always makes us create a mental image (graphic) of the scene in
a real-world setting; hence, a pictorial representation makes sense. The
more realistic the pictorial representation, the more suitable it seems in the
problem-solving process.
An interesting aspect of this problem is that the goal state is very clear.
It is the same as the initial state but on the other side of the riverbank. This
indicates a sort of symmetry in terms of time and space. Thus, solving a
problem backward from the goal state is a useful strategy when the
problem possesses the property of symmetry.
Usually, a problem where an initial state is to be transformed to a goal
state, identifying and formulating subgoals can be helpful. However, this
can be also be dangerous because people can tend to move only toward
the subgoal without considering the possibility of moving to a perceived
bad state—that is, one that moves away from the goal and might be the
only way to move toward the goal. Sometimes you have to take a step back
to move forward. This is evident from the state transition diagram that
demonstrated that moving away from a goal temporarily is sometimes
necessary. However, the ability to recognize when it is necessary to move
away (and hence step “backward” from a solution path) may be particular to
expert problem solvers or those who are especially familiar (involved) with
a problem domain. Finally, one of the preliminary testing subject’s strate-
gies of using trial and error to test various moves to improve learn-
ing and planning for the next move is also an important tool.

14.2.2 The 12 Coins Problem

This problem illustrates the importance of employing problem

reduction when feasible. The problem itself is very difficult if one tries to
236 • Artificial Intelligence Problems and Their Solutions

solve it right away without looking into its smaller versions. As discussed in
Chapter 4, we started from the most elementary form of this problem to
understand the constraints of the problem and the best way of dividing the
coins. The pitfalls of dividing the coins into two equal groups were evident
from the four coins version of the problem. Solving the reduced versions
of the problem helped us get familiar with the nuances of the prob-
lem and differentiate between good and bad weighings.
Furthermore, one of the important techniques used was to maximize
the amount of knowledge gained from earlier weighings. Due to the
constraint of trying to minimize the number of weighings, sometimes we
had to find the odd coin and its relative weight in a single weighing.
Recursion was another important techniques used as we had solved
the smaller and elementary versions of The 12 Coins Problem. Recursion
helped reduce the problem continually until it was converted to its most
elementary form and solved directly.
Furthermore, the use of recursion facilitates the conversion of exten-
sional solutions for subproblems (e.g., 3, 4, 5, 6 coins) to a compact inten-
sional formula that can be generalized for all numbers of coins [17].

14.2.3 Cryptarithms

Due to the nature of arithmetic operations, there are some clues inher-
ent in cryptarithms that can be spotted easily, such as the occurrence of 0, 1,
and 9. These clues serve as a starting point for solving the problem. Graphs
representing the constraints of the problem help in uncovering the
relationships and dependencies between the variables of the prob-
lem. Problems like cryptarithms call for the use of a variety of techniques in
various stages of the problem-solving process. This is also true for real-life
problems where one technique may not offer a complete solution.
Initially, clues and carryovers can help make some progress. After
that, the problem solvers veer toward narrowing down the domains of the
variables, eliminating some values that will ultimately lead to dead ends.
This saves time and avoids unnecessary calculations. When the domain of
some variables is low enough to allow an exhaustive search, the generate-
and-test technique can be used to facilitate some assumptions, proceed,
and backtrack if further search is impossible. Such values are eliminated,
and the remaining values are tested.
 Conclusion: Toward A Theory For Problem Solving • 237

In addition, columns in the problem were considered one by one as a

result of the strategy of solving subproblems. Furthermore, a structure
like the Knowledge Table can be very useful in keeping track of solved vari-
ables that can further help in deducing the remaining variables. Experience
has shown that many student problem solvers hamper their progress by not
being sufficiently organized about what they have ascertained at a given
stage in the problem-solving process.

14.2.4 The Red Donkey Puzzle

This puzzle involves the transformation of the initial state to a goal

state and thus can be formulated as a series of subgoals. The subgoals also
serve as milestones that can be backtracked to in case one gets lost in the
puzzle. As previously mentioned regarding The Missionaries and Cannibals
Problem, there are times when we have to move away from the goal tempo-
rarily to be able to move forward.
If the exact goal state of the solution is known, then a backward search
or a bidirectional search could prove a useful technique. This is akin to
chess endgames known as Queen and Pawn Endings. Many chess players
unnecessarily tend to avert such endings because they wrongly assume a
large branching factor and great complexity, when in fact the number of
possibilities is often actually limited, leading to a long, narrow tree.
Additionally, as evident from all four problems, having a playable
(either physical or virtual) model of the problem may greatly reduce the
learning time, keeping the problem solver’s interest alive and helping to
test various theories and techniques. Furthermore, the problem solver is
relieved of cumbersome complex calculations and memory-intensive tasks
and can concentrate on decision making.
In the context of problem solving, the strategies discussed above are
the basis of the human decision-making process, but at the same time,
humans are restricted by limitations on their computation speed and
memory capacities (see Lucci and Kopec, Table 6.3). These limitations can
be overcome with the use of computers to aid or complement the human
problem-solving process. Furthermore, computers can be used to test
different decisions, hypotheses, or ideas quickly to validate their fitness
for solving the problem in question, thus eliminating invalid moves and
reducing search time.
238 • Artificial Intelligence Problems and Their Solutions

14.2.5 The 15 Puzzle

The 15 Puzzle has been around for nearly 150 years and has probably
been subjected to more applications of AI search techniques than any other
problem in this book. Heuristics such as “the number of tiles out of place”
or “The Manhattan Distance,” coupled with algorithms such as depth-first
search, breadth-first search, A* algorithm, and, most recently the disjoint
pattern database method, have routinely used The 15 Puzzle as a test bed.
We have learned from the MHWC solution that systematic approaches
developing the solution row by row are most welcome and easy to compre-
hend.

14.2.6 The Knight’s Tour Problem

The Knight’s Tour is also a very old problem that has been the subject of
study for thousands of years. The remarkable fact is that solutions must be
identified from 4 × 1051 possible sequences. A number of techniques have
been used to identify solutions, including subproblems, exhaustive enumer-
ation, heuristics, patterns, and others. There is also some reminder of the
90 – 10 rule in AI. That is, a few heuristics take care of most of the problem
space (90%), and then a number of new heuristics (or rules) must be devel-
oped to handle 10% of the problem space. In this case, the heuristic that
always suggests choosing edge squares is effective most of the time—until
it isn’t. The problem solver must learn those special cases.

14.2.7 Mastermind

Mastermind is one of the more familiar logic problems subjected to

algorithms and computer methods for many years. It is a very logical and
intuitive problem that people (even children) can easily comprehend.
Understanding the problem is much easier than deciding on the logical
methods (or choices) that can be used to make the fewest guesses.
Donald Knuth, perhaps the greatest computer scientist of all time in
the United States, decided to address the problem in the mid-1970s. He
produced the “Five-Guess Algorithm,” which seems to be accepted as the
final word on the subject. We believe that one of our contributions here is
not only an analysis of the various possible solutions but also a study of these
solutions from a Human Window perspective, as well as the study of human
 Conclusion: Toward A Theory For Problem Solving • 239

subjects’ problem-solving activity for Mastermind. (See Appendix A on the

accompanying CD.)

14.2.8 The Monty Hall Problem

James Surowiecki’s book Wisdom of the Crowds discusses how when a

heterogeneous crowd (audience) is asked about its opinion, that opinion is
more intelligent (informed, correct) than the most intelligent (by IQ) person
in the room. In this problem, as described by Yanofsky in his recent book
The Outer Limits of Reason, the opinion of some 10,000 Mensa members
differs from that of Marilyn vos Savant—and they were wrong! Sometimes
the ability to abstract and represent (draw) images of a solution bears more
significance than anything else.

14.2.9 Rubik’s Cube

This too has been a familiar problem for many years. It seems
like Rubik’s Cube, akin to cell phones, texting, apps, and so forth, is a
problem more suitable for young people than adults. We believe that the
presentation and discussion of solutions to “The Cube” in Chapter 11 can
provide insights for those heretofore stupefied by it. Of course, no learn-
ing can come without some attention and work. It is a classic example of
illustrating the importance of being able to tackle problems by identifying
and solving subgoals.

14.2.10 The Prisoner’s Dilemma

The ability to apply problems, ideas, and solution methods from

Game Theory to practical and real social problems is of great interest
to all. The Prisoner’s Dilemma is one such problem from Game Theory
that can be applied in many ways to sociopolitical problems, economic
situations, and so on.

14.2.11 Miscellaneous Problems

The five miscellaneous problems in Chapter 13 demonstrate that

problem solving is in many cases a matter of perspective and experience.
Problems need not be long or intricate to be interesting and instructive.
240 • Artificial Intelligence Problems and Their Solutions

The Ten Pirates and Their Gold, for example, sets the stage for thinking
about and representing solutions to The Prisoner’s Dilemma.

14.3 RETROSPECTIVE, CONCLUSIONS, AND FUTURE

WORK

Donald Michie and Danny Kopec first proposed the concept of the
Human Window.* It states that the solutions to difficult, AI-type problems
must meet certain criteria to fall within the bounds of the Human Window.
As originally presented in Shweta’s thesis, they need to be correct, compre-
hensible, executable, and of a moderate grain size. It was to Dr. Kopec’s
great joy that some 30 years later, Chris Pileggi performed the study of
Human Window solutions on The Missionaries and Cannibals Problem,
cryptarithmetics, The 12 Coins Problem, The Knights Tour Problem,
The 15 Puzzle, and The Red Donkey Puzzle) for his senior thesis project.
Pileggi’s work was extremely impressive and provided the first application
of Michie’s ideas to real, mainstream problems and their solutions.
Shweta Shetty assembled and studied the solutions to the problems in
this book. As we can see from the Human Window analysis of these solu-
tions, humans are comfortable with solutions that are graphical, not too
short or too long, easy to comprehend and execute, testable, and applicable
to every problem state.
Furthermore, we now understand that even though we believe we have
identified the best solutions to a well-known set of AI problems, this is not
enough. Future work needs to focus on how Human Window solutions can
best be presented to human subjects. Clearly, with the current and likely
future generations, even graphical approaches with rules (on paper) will not
suffice. Notions of active participation in learning (patterns) and execution
(performance) are necessary. Video (possibly interactive) or animated
presentation of our solutions would seem like the best next step in pursuing
this research.

* Michie liked to credit Michael Clarke for the notion, but it was Michie who indeed wrote
more about it and developed the concept via Dr. Kopec’s Ph.D. thesis.
 Conclusion: Toward A Theory For Problem Solving • 241

Further research is needed to continue this work by delving deeper

into the subject of studying human problem-solving processes by using
computers as a tool for experimentation. In addition, the thought processes
of human subjects should be recorded and analyzed for all such problems,
as Luger did for cryptarithm problems. A comparative study of the human
learning process, with analysis of how well solutions are actually under-
stood (both before and after solutions are found) seems to be a necessary
and important step toward gaining a better understanding of the problem-
solving process and how it can be successful.

14.4 SUPPLEMENTAL REFERENCES

1. Amarel, S. (1968). On Representations of Problems of Reasoning about

Actions. Donald Michie, editor, Machine Intelligence 3, pages 131–171.
American Elsevier Publishers, New York, NY.
2. Anderson J.R., Reder L. M., and Simon, H. (2000). “Applications and
Misapplications of Cognitive Psychology to Mathematics Education,” Texas
Education Review (1) 2: 29–49.
3. Anzai, Y., and Simon, H.A. (1979). The Theory of Learning by Doing.
4. Bryan, J. (2012). The 12 Coin Problem. http://www.jerrybryan.net/
math/12coins.html. Accessed on April 22, 2013.
5. Bogomolny, A. (2013). Odd Coin Problems: the 120 Marble Problem—Five
Weighings from Interactive Mathematics Miscellany and Puzzles. http://www.
cut-the-knot.org/blue/OddCoinProblems.shtml. Accessed on April 22, 2013.
6. Burton, L., and Burton, M. (1980). “Problems and Puzzles.” For the Learning
of Mathematics (1) 2: 20–23. Published by: FLM Publishing Association.
Stable http://www.jstor.org/stable/40247710. Accessed on December 23, 2012.
7. Cerveny R.P., Garrity, E.J., and Sanders, G.L. (1990). “A Problem-Solving
Perspective on Systems Development.” Journal of Management Information
Systems (6) 4: 103–122. Published by: M.E. Sharpe, Inc. Stable http://www.
jstor.org/stable/40398777. Accessed on December 23, 2012. For the Learning
of Mathematics (1) 2: 35–42. Published by: FLM Publishing Association. Stable
http://www.jstor.org/stable/40247714. Accessed on December 23, 2012.
242 • Artificial Intelligence Problems and Their Solutions

8. Coleman, P.T. (2011). The Five Percent: Finding Solutions to Seemingly

Impossible Conflicts. New York: Public Affairs, Perseus Books.
9. Elitmus, G. (2012). Guide to Solve Cryptic Multiplication Question.
Available at http://elitmusguide.blogspot.in/2012/07/guide-to-solve-cryptic-
multiplication.html. Accessed on January 10, 2014.
10. Eysenck, M.W. (2002). Simply Psychology, 2nd ed. New York, NY,
Psychology Press Ltd. / Taylor Francis
11. Gunzelmann, G., and Anderson, J.R. (2002). Problem Solving:
Increased Planning with Practice .Cognitive Systems Research 4(1): 57-76..
12. Logicville. (2012). http://www.logicville.com/cryptarithm.htm. Accessed on
April 22, 2013.
13. Lucci, S., and Kopec, D. (2013). Artificial Intelligence in the 21st
century. Dulles, Virginia: Mercury Learning, Inc.
14. Luger, G.F. (2008). Artificial Intelligence: Structures and Strategies
for Complex Problem Solving, 6th ed. Reading, MA: Addison-Wesley.
15. Manvel, B. (1977). “Counterfeit Coin Problems.” Mathematics Magazine 50:
90–92.
16. Minsky, M. (1960). Steps Toward Artificial Intelligence. In Computers and
Thought, (Eds. E. Feigenbaum and J. Feldman), McGraw-Hill, New York,
pp. 406-450.
17. Newell, A., and Simon, H.A. (1972). Human Problem Solving. Englewood
Cliffs, NJ: Prentice-Hall.
18. O’Connor, J.J., and Robertson, E.F. (2012). Propositiones ad acuendos
iuvenes by Alcuin. http://jnsilva.ludicum.org/HMR12_13/AlcuinStA.pdf.
Accessed on December 4, 2013.
19. Polk, T.A., and Newell, A. (1995). “Deduction as Verbal Reasoning.”
American Psychological Association (102) 3: 533–536.
20. Polya, G. (1957). How to Solve It, 2nd ed. Princeton, NJ: Princeton University Press.
21. Ruggiero, V.R. (1998). The Art of Thinking. Reading, MA: Addison-Wesley.
Burton, L., and Burton, M. (1980). “Problems and Puzzles.” For the Learning
of Mathematics (1) 2: 20–23. Published by: FLM Publishing Association.
Stable http://www.jstor.org/stable/40247710. Accessed on December 23, 2012.
 Conclusion: Toward A Theory For Problem Solving • 243

22. Soares, J. (2003). Will Cryptarithmetic Survive Innovation? http://

cryptarithms.awardspace.us/survival.htm. Accessed on April 22, 2013.
23. Unterrainer, J.M., and Owen, A.M. (2006). “Planning and Problem Solv-
ing: From Neuropsychology to Functional Neuroimaging.” Journal of
Physiology—Paris 99: 308–317.
24. Brandeis. (2013). Red Donkey. http://www.cs.brandeis.edu/~storer/Jim-
Puzzles/ZPAGES/zzzRedDonkey.html. Accessed on April 22, 2013.
25. Bsswebsite. (2013). Red Donkey Puzzle. http://www.bsswebsite.me.uk/
Puzzlewebsite/Reddonkeypuzzle/Reddonkeypuzzle.htm.
26. Sweller, J. (1988). “Cognitive Load During Problem Solving: Effects on
Learning.” Cognitive Science 12: 257–285.
27. Kose, E. (2012). Comparing AI Search Algorithms and Their Efficiency When
Applied to Path Finding Problems. Ph.D. Thesis, New York, NY: CUNY,
Graduate Center.
28. Bryan, J. (2012). The 12 Coin Problem. http://www.jerrybryan.net/
math/12coins.html. Accessed on April 22, 2013.

14.5 COMPOSITE PROBLEM PLAYABILITY SITES

The Missionaries and Cannibals Problem – http://www.smartestgames.com/

game/missionaries-and-cannibals/
The 12 Coins Problem – http://www.mathplayground.com/coinweighing.html
Cryptarithmetic – http://www.letsgetwordy.com/#am_app
The Red Donkey Puzzle – http://www.bsswebsite.me.uk/Puzzlewebsite/Reddon-
keypuzzle/Reddonkeypuzzle.htm
The 15-Puzzle – http://migo.sixbit.org/puzzles/fifteen/
The Knight’s Tour Problem – http://www.kongregate.com/games/­
evgenykarataev/knights-tour
Mastermind – http://www.grayman.de/mastermind/index_en.php4
Rubik’s Cube – http://activeden.net/item/cube-assembler/full_screen_
preview/3131480
ON THE COMPANION DISC
Appendix A - Solutions for Missionaries and Cannibals Problem referred
to in Chapter 3.
Appendix B - An Algorithm for The Twelve Coins Problem
Appendix C - A Study Comparing Human Solutions
Appendix D – Programmed Solutions to Problems
■ Appendix D.1 - A Program for the “Missionaries and Cannibals Prob-
lem” Solution in Picat
■ Appendix D.2 - Programs for Cryptarithmetic

1. Appendix D.2.1 - A Program for a General Alphametic

(cryptarithmetic) Solver in Picat
2. Appendix D.2.2: A Program for the Cryptarithm SEND+MORE =
MONEY written in Prolog
3. Appendix D.2.3: Another Program for the Cryptarithm
SEND+MORE = MONEY written in Prolog
■ Appendix D.3 - A Program for The 15 Puzzle written in Picat
■ Appendix D.4 - A Program for The Knight’s Tour written in Pascal
■ Appendix D.5 - Programs for the Red Donkey Puzzle
246 • Artificial Intelligence Problems and Their Solutions

1. Appendix D.5.1: A Program for the Red Donkey Puzzle written in

Java
2. Appendix D.5.2: A Program for the Red Donkey Puzzle written in
Picat
■ Appendix D.6 - A Program for Rubik’s Cube written in Python
■ Appendix D.7 – Extra: Program for the Towers of Hanoi Problem writ-
ten in Picat
Appendix D presents several programs for some of the problems pre-
sented in the book. They are written in a variety of languages:
Picat — to compile and run a program written in Picat, go to the main
Website http://www.picat-lang.org/ to install Picat and its compiler. A user’s
guide can be found on this Website with instructions on how to compile
Java — to compile and run a Java program, you must download the
Java Runtime Environment. To download this and other necessary software
for running Java programs, visit http://www.oracle.com/technetwork/java/
javase/downloads/index.html
Prolog — to compile and run a program written in Prolog, visit http://
www.probp.com/download.html for the download and a user’s guide.
Python — to compile and run a program written in Python, you must
download the Python interpreter. The download can be found at http://
www.python.org/getit/ with an option to choose from a number of operating
systems.
 I NDEX
A Contradiction, 13
Abduction, 12 Correctness of solution, 22, 25
Admissible heuristic, 129 Crypt-arithmetic, 71
Algorithm solution, 148 Cryptarithms
Alphabetic App, 90 background, 71–72
American Agriculturist magazine, 71 best machine solution, 89–90
American Mathematical Monthly, 51 human problem solving, 84
Analogous problem (analogy), 16 Human Window analysis of solutions,
Anderson, J.R., 4 84–89
Anzai, Y., 4 overview, 2
playable program, 90–91
B problem-solving techniques, 72–73
Backtracking algorithm, 18, 84, 148 solution, 73–84
Best-first heuristic, 129 Theory for Problem Solving, 236–237
Bidirectional search, 38, 98–99 Cube depiction, 196
The Birthday Problem, 230–232
Boundary conditions, 17–18 D
Bram, Robert, 210 Dad’s Puzzler, 93
The Brothers and Sisters Problem, 31 Darby, G., 151
Davidson, Morley, 203
C Decomposition (equivalence classes), 18
Cards in the Dark, 215–217 Deduction, 12
Cerveny R.P., 4 de Moivre, Abraham, 135
Chapman, Noyes, 109 Dethridge, John, 203
Coins version of the problem, 218 Disjoint pattern database heuristic, 129
Comprehensibility of solution, 22, 25–26 Divide and conquer, 14
Constraint hypergraph, 83 Divide-and-conquer algorithm, 148
Constraint rules, 47, 48 Dominant strategy, 209
Constraint satisfaction problems, 72 Donald solution, 88
248 • Index

Dresher, Melvin, 208 How to Solve It (Polya, George), 4

D&S Solution, 135 Human problem solving
The 12 Coins Problem, 62–63
E cryptarithms, 84
Euler, Leonhard, 134, 146 The Knight’s Tour Problem, 142
Executability of solution, 22, 25 Mastermind, 164
Exhaustive enumeration, 13, 134 The Missionaries and Cannibals Problem,
Extensional representations, 10, 23 38–41
The Monty Hall Problem, 172–173
F The Red Donkey Puzzle, 99–100
The 15 Puzzle Rubik’s Cube, 194–195
background, 109–110 Human Problem Solving (Newell and
best machine solution, 128–129 Simon), 4
corner technique, 116–118 Human Window analysis of solutions, 22–24
first row, solving the, 111–116 comprehensibility of solution, 22, 25–26
fourth row, solving the, 122–124 correctness of solution, 22, 25
Human Window analysis of solutions, Cryptarithms, 84–89
124–128 executability of solution, 22, 25
overview, 2 The 15 Puzzle, 124–128
playable program, 130 grain size of solution, 22, 25
problem-solving techniques, 110–111 The Knight’s Tour Problem, 142–147
second row, solving the, 118–119 Mastermind, 164–165
solution, 111–124 The Missionaries and Cannibals Problem,
Theory for Problem Solving, 238 42–48
third row, solving the, 119–122 The Red Donkey Puzzle, 100–106
Floyd, Merrill, 208 Rubik’s Cube, 196–201
Forward checking, 84 Theory for Problem Solving, 233
The 12 Coins Problem, 63–68
G Hunter, J.A.H., 71
Gardner, Martin, 94
Garrity, E.J., 4
I
Generic table solution, 65
Incubation, 18
God’s algorithm, 203
Induction, 11
Grain size of solution, 22, 25
Intensional representations, 10, 23
Guinness Book of World Records, 172
Intuition, 17
Gunzelmann, G., 4
Invicta Plastics Ltd, 151
H Iterative-deepening-A* search, 129
Halmos’s handshake problem, 222–227
Hamiltonian Cycle Problem, 134 J
Heuristics, 18–20 Jealous Husbands problem, 31, 32
Hierarchy, imposing a structure, 16 Johnson, Woosly, 110
Hordern, L. Edward, 93 Joseph, Andrew, 135
 Index • 249

K Mastermind
Kavyalankara (Sanskrit work), 134 background, 151–152
Kinsey, Ernest, 110 best machine solution, 165–166
The Knight’s Tour Problem human problem solving, 164
background, 133–134 Human Window analysis of solutions,
best machine solution, 148–149 164–165
human problem solving, 142 overview, 3
Human Window analysis of solutions, playable program, 166
142–147 problem-solving techniques, 152–153
overview, 2–3 solution, 153–163
playable program, 149 Theory for Problem Solving, 238–239
problem-solving techniques, Meirowitz, Mordecai, 151
134–135 MHWC. See Most Human
solution, 135–142 Window-Compatible
Theory for Problem Solving, 238 Michie, Donald, 22
Knowledge representation, 10 Minimum remaining values (MRV)
Knuth, Donald, 238 heuristic, 83
Kociemba, Herbert, 202, 203 The Missionaries and Cannibals Problem
Kopec, D., 23, 210 analysis of solution, 41–42
background, 31–32
L best machine solution, 48
LCV. See Least constraining value choosing an appropriate representation,
Least constraining value (LCV), 83 32–34
Least Human Window-Compatible human problem solving, 38–41
(LHWC), 23 Human Window analysis of solutions,
Cryptarithms, 87–89 42–48
The 15 Puzzle, 126–128 overview, 2
The Knight’s Tour Problem, 145–147 playable program, 49
The Missionaries and Cannibals Problem, solution, 34–38
45–48 Theory for Problem Solving, 235
The Red Donkey Puzzle, 104–106 The Monty Hall Problem
Rubik’s Cube, 198–201 background, 169–170
The 12 Coins Problem, 66–68 human problem solving, 172–173
LHWC. See Least Human overview, 3
Window-Compatible problem-solving techniques, 170–172
Lloyd, Samuel, 110 solution, 172
Logically enumerated solution, 66–68 Theory for Problem Solving, 239
Most Human Window-Compatible
M (MHWC), 23
Magic constant, 146 Cryptarithms, 85–87
Manhattan distance heuristic, 129 The 15 Puzzle, 126
Master cube, 176 The Knight’s Tour Problem, 144–145
250 • Index

Most Human Window-Compatible Problem playability sites, 243

(MHWC) (Continued) Problem reduction technique, 15, 52
The Missionaries and Cannibals Problem, Problem solving
43–45 classification, 27–28
The Red Donkey Puzzle, 100–103 heuristics, 18–21
Rubik’s Cube, 197–198 Human Window, 22–24
The 12 Coins Problem, 64–66 Human Window criteria and ranking of
Movement notation, Rubik’s Cube, 196–197 solutions, 24–27
MRV heuristic. See Minimum remaining Polya’s five steps, 7–11
values heuristic techniques, 11–18
Problem-solving techniques
N abduction, 12
Nash equilibrium, 209 analogous problem (analogy), 16
Newell, A., 4 bright idea, 17
consider boundary conditions, 17–18
O contradiction, 13
Omohumdro, Stephen M., 218, 221 decomposition (equivalence classes), 18
Optimal solutions, 126, 127 deduction, 12
The Outer Limits of Reason, 239 divide and conquer, 14
exhaustive enumeration, 13
P generate and test, 15–16
Parade Magazine, 172 hierarchy, imposing a structure, 16
Parallel genetic algorithm, 90 incubation, 18
Pareto optimal, 209 induction, 11
Payoff matrix, 208 intuition, 17
Pearl, Judea, 18 problem reduction, 15
Pileggi, Chris, 240 recursive, 13–14
Planning, 10 related problem, solving a, 15
Polya, G., 7 subgoal, solving, 14–15
Polya’s five steps, for problem solving subproblem, solving, 14
develop an algorithm, 8 using prior knowledge, 15
implement the algorithm, 9 working backward (retrograde analysis), 18
revise the algorithm, 9–11 Problem state space matrix, 34, 36
test the algorithm, 9 Propositiones ad Acuendos Juvenes, 31
understand the problem, 7–8
The Postman Always Rings Twice, 207, 209 R
The Prisoner’s Dilemma RAND Corporation, 208
diverse areas, applications in, 210–212 Random Airline Seats Problem, 228–230
iterated prisoner’s dilemma, 209–210 Ranking of solutions
overview, 3 complexity, 27
Theory for Problem Solving, 239 Cryptarithmetic solutions, 86
traditional problem, 207–209 difficulty level, 27
 Index • 251

The 15 Puzzle Solutions, 125 subgoal 3, middle layer, 185–187

flexibility of solutions, 26 subgoal 4, new top cross, 187–189
human window, constraints of, 25–26 subgoal 5, new top’s corners, 189–190
intensionality/extensionality, 24 subgoal 6, correct corner positioning,
intensional vs. extensional, 24 190–192
The Knight’s Tour Problem, 143 subgoal 7, top edges, 192–194
knowledge representation, choice of, 24–25 Theory for Problem Solving, 239
The Missionaries and Cannibals Problem
Solutions, 44 S
mode of conveyance, 26 Sanders, G.L., 4
possible optimality, 26–27 Schell, E. D., 51
problem-solving method, 26 Schwenk, Allen, 134
The Red Donkey Puzzle Solutions, 101 Scientific American, 94
Rubik’s Cube, 195 Semi-magic solution, 145
The 12 Coins Problem Solutions, 64 Shetty, Shweta, 240
The Red Donkey Puzzle Simon, H.A., 4
background, 93–94 Sliding Piece Puzzles
best machine solution, 106–107 (Hordern, L. Edward), 93
human problem solving, 99–100 State transition diagram, 34, 36
Human Window analysis of solutions, Stewart, Ian, 218, 221
100–106 Story, William, 110
overview, 2 Surowiecki, James, 239
playable program, 107 Symmetric solution, 46
solution, 94–98 Symmetry under time reversal, 37
Theory for Problem Solving, 237
Retrograde analysis, 18 T
Rokicki, Tomas, 203 Ten Pirates and Their Gold Problem,
Rubik, Ernö, 175 218–221
Rubik’s Brand Ltd., 176 Tessellation, 135
Rubik’s Cube Text descriptions, 196
background, 175–176 Theory for Problem Solving
best machine solution, 201–203 Cryptarithms, 236–237
cube components, 177–182 The 15 Puzzle, 238
human problem solving, 194–195 The Knight’s Tour Problem, 238
Human Window analysis of solutions, Mastermind, 238–239
196–201 The Missionaries and Cannibals
overview, 3 Problem, 235
playable program, 203–204 The Monty Hall Problem, 239
problem-solving techniques, 176–177 The Prisoner’s Dilemma, 239
solution, 177–194 The Red Donkey Puzzle, 237
subgoal 1, top cross, 182–183 Rubik’s Cube, 239
subgoal 2, top corners, 184–185 The 12 Coins Problem, 235–236
252 • Index

Theory of Moves, 210 U

Thomasson, Daniel E., 135, 145 Understand the problem, 7–8, 10
The Tragedy of the Commons (Hardin), 212 Uniform cost search, 129
Tree diagram, 35
Turagapadabandha (Sanskrit work), 134 V
The 12 Coins Problem Vatriquant, Simon, 71
background, 51–52 von Neumann, Jon, 222
best machine solution, 69 Von Warnsdorff, H. C., 134
human problem solving, 62–63 vos Savant, Marilyn, 172, 239
Human Window analysis of solutions,
63–68 W
overview, 2 Warnsdoff, H. C., 148
playable program, 69 Warnsdoff’s rule, 148
solution, 56–62 Wert, Jack, 59
solving a smaller problem, 52–56 Wisdom of the Crowds
Theory for Problem Solving, 235–236 (Surowiecki, James), 239