SYMBOLIC COMPUTATION

Artificial Intelligence

Managing Editor: D. W. Loveland

Editors: S. Amarel A. Biermann L. Bole
A. Bundy H. Gallaire P. Hayes
A. Joshi D. Lenat A. Mackworth
E. Sandewall 1. Siekmann W. Wahlster
Springer Series
SYMBOLIC COMPUTATION -Artificial Intelligence

N.l Nilsson: Principles of Artificial Intelligence. XV, 476 pages, 139

figs., 1982.
lH. Siekmann, G. Wrightson (Eds): Automation of Reasoning 1. Classi-
cal Papers on Computational Logic 1957-1966. XXII, 525 pages, 1983.
lH. Siekmann, G. Wrightson (Eds): Automation of Reasoning 2. Classi-
cal Papers on Computational Logic 1967-1970. XXII, 638 pages, 1983.
L. Bole (Ed.): The Design of Interpreters, Compilers, and Editors for
Augmented Transition Networks. XI, 214 pages, 72 figs., 1983.
R.S. Michalski, lG. Carbonell, T.M. Mitchell (Eds.): Machine Learn-
ing. An Artificial Intelligence Approach. XI, 572 pages, 1984.
L. Bole (Ed.): Natural Language Communication with Pictorial Infor-
mation Systems. VII, 327 pages, 67 figs., 1984.
lW. ,Lloyd: Foundations of Logic Programming. X, 124 pages, 1984.
A. Bundy (Ed.): Catalogue of Artificial Intelligence Tools. XXV, 150
pages, 1984. Second, revised edition, IV, 168 pages, 1986.
M.M. Botvinnik: Computers in Chess. Solving Inexact Search
Problems. With contributions by A.1. Reznitsky, B.M. Stilman, M.A.
Tsfasman, A.D. Yudin. Translated from the Russian by A.A. Brown.
XIV, 158 pages, 48 figs., 1984.
C. Blume, W. Jakob: Programming Languages for Industrial Robots.
XIII, 376 pages, 145 figs., 1986.
L. Bole (Ed.): Natural Language Parsing Systems. XVIII, 367 pages,
155 figs., 1987.
L. Bole (Ed.): Computational Models of Learning. IX, 208 pages, 34
figs., 1987.
N. Cercone, G. McCalla (Eds.): The Knowledge Frontier. Essays in the
Representation of Knowledge. 552 pages, 93 figs., 1987.
G.Rayna: REDUCE. Software for Algebraic Computation. IX, 330
pages, 1987.
D. McDonald, L. Bole (Eds.): Natural Language Generation Systems.
XI, 400 pages, 84 figs., 1988.
L. Kanal, V. Kumar (Eds.): Search in Artificial Intelligence. X, 488
pages, 1988.
Laveen Kanal Vipin Kumar
Editors

Search in
Artificial Intelligence

With 67 Illustrations

Springer-Verlag
New York Berlin Heidelberg
London Paris Tokyo
Laveen Kanal
Department of Computer Science
University of Maryland
College Park, MD 20742
USA

Vipin Kumar
Computer Science Department
University of Texas at Austin
Austin, TX 78712-1188
USA

Library of Congress Cataloging-in-Publication Data

Search in artificial intelligence.
(Symbolic computation. Artificial intelligence)
Bibliography: p.
Includes index.
1. Artificial intelligence - Data processing.
2. Algorithms. 3. Information storage and retrieval
systems. I. Kanal, Laveen N. II. Kumar Vipin.
III. Series.
Q336.S4 1988 006.3 88-12247

© 1988 by Springer-Verlag New York Inc.

Softcover reprint of the hardcover 1st edition 1988

All rights reserved. This work may not be translated or copied in whole or in part without the written per-
missionofthe publisher (Springer-Verlag, 175 Fifth Avenue, New York, New York 10010, USA), except
for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc. in this publication, even if the
former are not especially identified, is not to be taken as a sign that such names, as understood by the
Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone.

Camera-ready text prepared by the editors.

9 8 7 6 5 432 1

ISBN -13:978-1-4613-8790-9 e- ISBN -13:978-1-4613-8788-6

DOl: 10.1007/978-1-4613-8788-6
 Preface

Search is an important component of problem solving in artificial intelligence (AI)

and, more generally, in computer science, engineering and operations research.
Combinatorial optimization, decision analysis, game playing, learning, planning,
pattern recognition, robotics and theorem proving are some of the areas in which
search algbrithms playa key role.
Less than a decade ago the conventional wisdom in artificial intelligence was that
the best search algorithms had already been invented and the likelihood of finding
new results in this area was very small. Since then many new insights and results
have been obtained. For example, new algorithms for state space, AND/OR graph,
and game tree search were discovered. Articles on new theoretical developments
and experimental results on backtracking, heuristic search and constraint propaga-
tion were published. The relationships among various search and combinatorial
algorithms in AI, Operations Research, and other fields were clarified. This volume
brings together some of this recent work in a manner designed to be accessible to
students and professionals interested in these new insights and developments.
The first three papers are concerned with developing general formulations of
search and investigating relationships among various search techniques of AI and
Operations Research. The paper by Kumar and Kanal presents a model for discrete
optimization problems and shows relationships among dynamic programming,
branch-and-bound (B&B) and heuristic search. The paper by Helman presents a
different model for discrete optimization and shows relationships between dynamic
programming and B&B. This model applies to a smaller class of problems than the
first paper, but allows an analysis of the complexity within a restricted class of com-
putations. The third paper, by Kumar, Nau and Kanal, presents a general B&B
procedure for AND/OR graph and game tree search, and shows that many important
existing AI search procedures can be treated as special cases of this general B&B
procedure.
The next three papers are concerned with the performance of the well known A*
algorithm and its variations. Although the worst-case complexity of A* is probably
exponential, the average case complexity depends upon the underlying nature of the
search tree and the heuristic function. The paper by Bagchi and Sen shows that for
certain characteristics of the search tree, the complexity of A* is exponential unless
 vi

the heuristic function is very accurate. It is interesting to note a related result in

Douglas Smith's 1979 dissertation. He showed that, for a different kind of search
tree, the average search complexity for a best-first B&B procedure (which is essen-
tially A*) can be polynomial (see JACM 1984).
Ever since its original development by Hart, Nilsson and Raphael, the optimality
of A* over other heuristic search algorithms has been a controversial topic. The
paper by Dechter and Pearl examines conditions under which A* is no better than
other algorithms, and also under which A* is optimal. A recent paper by Mero
presented a variation of A* in which the heuristic lower bounds are revised (to
become more accurate) dynamically. Intuitively, it seems that the new "improved"
bounds would only make the algorithm better, a claim also made in Mero's paper.
The paper by Mahanti and Ray shows that this claim is incorrect; i.e., revising
heuristic bounds can seriously hurt the performance in some cases. They present a
new algorithm that revises heuristic bounds dynamically and does not have this
drawback.
The paper by Korf deals with the problem of finding a path between two different
states in the presence of different sources of knowledge. He shows that if knowledge
about the domain is available only in terms of heuristic lower bounds, then an
algorithm that performs cost bounded depth-first search repeatedly over a search
space is indeed asymptotically optimal over all other admissible search algorithms.
If more knowledge is available (in the form of subgoals, macro-operators or abstrac-
tions), then more powerful search methods can be used.
Means-end analysis is an important problem solving technique, which is embod-
ied in the General Problem Solver (GPS) of Newell, Shaw and Simon. Banerji and
Ernst examine a number of developments related to GPS. In particular they set forth
the strengths and weaknesses of work by Goldstein and by Korf. The paper con-
cludes with a negative view of heuristic search techniques based on numerical
methods, and states the authors' belief that symbol manipulation or problem
descriptions represents a more promising direction. While this belief is reasonable
and is not uncommon in AI, it remains to be backed up by success on more than toy
problems. On the other hand, heuristic search techniques have been used success-
fully in many of the areas listed at the beginning of the preface.
The next four papers deal with algorithms for solving the constraint satisfaction
problem (CSP). Nadel presents a unified view of a large number of constraint satis-
faction algorithms that were developed by many researchers at different times. He
shows that all these algorithms can be viewed as having different degrees of tree
search and consistency checking components. Freuder examines the conditions
under which a CSP can be solved by doing a bounded amount of search. Dechter and
Pearl show that the idea of solving a relaxed problem to create heuristic guidance
can be effectively used in solving constraint satisfaction problems. Traditional for-
mulations of the CSP deal only with binary constraints. Montanari and Rossi present
an abstract formulation of the CSP which allows the constraints to be n-ary. They
show that many properties known to be true of binary constraint networks are also
true of the n-ary constraint networks. .
The final paper, by Chi and Nau, presents an investigation of competing back-up
rules for game tree search. Most game playing programs generate the game tree only
up to a limited depth, and use the minimax back-up rule to propagate the heuristic
values of the leaf nodes. Since these heuristic values can be in error, the selected
move is not guaranteed to be the best. But one would still expect the overall perfor-
 vii

mance (in terms of selecting the best move) to improve as the search depth is
increased. Surprisingly, investigations by Nau and by others have shown that the
discriminating power of the minimax back-up rule can decrease, rather than
increase, with increasing depth of the search tree. This has led to investigations of
alternatives to the minimax rule, such as the product rule. Chi and Nau show that
there are real games for which the product rule does better than the minimax back-
up rule, and investigate a possible way to predict the relative performance of the
product rule against minimax.
The articles in this book should provide the reader a broad view of recent develop-
ments on search in AI and some of the relationships among different search
algorithms. Even as this preface is being written, highly interesting work on parallel
implementations of search algorithms and on their experimental performance on
parallel computer architectures is appearing. The potential of parallel processing for
AI search algorithms is so great that from now on it is likely to be the focus of work
on search in AI. The new insights and developments provided by this volume should
also be helpful to researchers interested in parallel search algorithms.
The editors thank all those who submitted papers for possible inclusion in this
book. Even though some of the selected papers have been published elsewhere, we
hope that having the papers revised for this book has resulted in greater readability
and cohesiveness. We are grateful to the authors for their cooperation in revising
their papers. We also thank the staff of Springer-Verlag, for their cooperation and
help in producing the book.

Laveen Kanal and Vipin Kumar

 Contents

Preface v

Chapter 1. The CDP: A Unifying Formulation for Heuristic

Search, Dynamic Programming, and Branch-and-Bound 1
V. Kumar and L. Kanal

Chapter 2. An Algebra for Search Problems and Their Solutions 28

P. Helman

Chapter 3. A General Branch-and-Bound Formulation for

AND/OR Graph and Game Tree Search 91
V. Kumar, D.S. Nau and L. Kanal

Chapter 4. Average-Case Analysis of Heuristic Search

in Tree-Like Networks 131
A. Bagchi and A.K. Sen

Chapter 5. The Optimality of A* 166

R. Dechter and 1. Pearl

Chapter 6. Network Search Algorithms with Modifiable Heuristics 200

A. Mahanti and K. Ray

Chapter 7. Optimal Path Finding Algorithms 223

R. Korf

Chapter 8. Developments with GPS 268

R.B. Banerji and G.w. Ernst

Chapter 9. Tree Search and Arc Consistency in Constraint

Satisfaction Algorithms 287
B. Nadel
 x

Chapter 10. Backtrack-Free and Backtrack-Bounded Search 343

E. Freuder

Chapter 11. Network-Based Heuristics for

Constraint-Satisfaction Problems 370
R. Dechter and J. Pearl

Chapter 12. Fundamental Properties of Networks of Constraints:

A New Formulation 426
U. Montanari and F. Rossi

Chapter 13. Comparison of the MINIMAX and PRODUCT

Back-Up Rules in a Variety of Games 450
P. Chi and D.S. Nau

Index 473