From devika@cs.cornell.edu Wed Nov 30 17:20:44 EST 1994
Article: 25620 of comp.ai
Newsgroups: comp.ai,comp.ai.edu
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!news.kei.com!travelers.mail.cornell.edu!cornell!devika
From: devika@cs.cornell.edu (Devika Subramanian)
Subject: A review of Russell and Norvig's new AI text
Message-ID: <1994Nov30.151429.14332@cs.cornell.edu>
Organization: Cornell Univ. CS Dept, Ithaca NY 14853
Date: Wed, 30 Nov 1994 15:14:29 GMT
Lines: 93
Xref: glinda.oz.cs.cmu.edu comp.ai:25620 comp.ai.edu:2134

A brief review of

"Artificial Intelligence: A Modern Approach"
Stuart Russell and Peter Norvig
Prentice Hall, December 1994. ISBN 0-13-103805-2

by Devika Subramanian, Cornell University


While the enterprise of artificial intelligence has often been defined
around the dream of intelligent agents, Russell and Norvig's book is
the first attempt to present the technical accomplishments of AI to a
broad scientific audience in the context of embedded agents acting in
real-world environments.  The book is not merely an expositional
triumph; Russell and Norvig achieve a unique synthesis of concepts and
algorithms in AI that have evolved in very disparate sub-communities
of the field. The book draws on ideas from logic, decision theory,
control theory, Markov processes, economics, on-line algorithms,
complexity theory, probability and statistics and information theory,
to coherently present methods in AI in a jargon-free manner. This
makes the book an ideal introduction to newcomers to AI from computer
science as well as other branches of science and engineering. For
seasoned practitioners, it offers a new, thought-provoking way to
understand AI.

The book is organized into eight sections.  The first section begins
with a brief history of AI and introduces the basic vocabulary for
describing agents embedded in task environments.  The last section
(Section VIII) comprises a beautiful essay on the philosophical
foundations of AI and an engaging commentary on the current state and
future challenges facing AI.  The sections in between constitute the
technical meat of the book.  Section II highlights general
problem-solving methods for embedded agents and includes informed
search methods that take resource constraints into account.  The third
section emphasizes the role of knowledge in decision-making and
presents an array of methods for representing and reasoning with
logical or categorical knowledge.  Section IV presents planning as
reasoning about action choice; contemporary planning and replanning
methods are presented as specializations of the general methods of
logical reasoning introduced in the third section.  Section V
introduces probability and decision theory as tools for agents acting
under uncertainty. It explains how belief networks can be used to
represent uncertain knowledge and describes decision-making methods
based on them. The sixth section focuses on learning and adaptation in
intelligent agents. It presents a unified model of learning, a brief
introduction to computational learning theory, as well as specific
techniques such as decision-tree learning, neural networks, and a new
method for learning belief networks.  It also includes a tutorial
exposition of recent work in reinforcement learning, as well as the
knowledge-based inductive logic programming method.  Section VII
focuses on interactions of the agent with the external world: natural
language communication, perception and robotics.  Russell and Norvig
have recruited established experts (Jitendra Malik and John Canny) to
cover the specialized topics of perception and robotics, ensuring a
uniformly high quality to all of the technical material in the book.

The book is hefty: over 900 pages in all. However, almost 200 pages
are devoted to items sometimes missing from AI texts: a very thorough
index, a truly massive bibliography, "Historical Notes" sections that
are researched in depth and make fascinating reading, and a large
collection of excellent exercises.

This is perhaps not the place to go through all the book's chapters in
detail, but some deserve special mention.  The second chapter on
agents is brilliant; it puts the entire history of work in AI in
perspective and explains WHY people built the algorithms that were
built. This is the first question that most first-timers to AI have,
and this is answered up front. The chapters on reasoning about
uncertainty are by far the best tutorial exposition of material on
probability and belief networks: they make the original papers in the
area much more accessible.

Judged from all respects, this is a remarkably comprehensive and
incisive treatment of the field. The book is well-written and
well-organized and includes uniform and clear descriptions of all
major AI algorithms.  The authors have managed to describe key
concepts with technical depth and rigour without falling prey to
stodginess and Greek-symbolitis. AI is presented as a set of
inter-related design principles, rather than a grab bag of tricks.
The book brims with optimism and contagious excitement about the
frontiers of AI. I recommend it without reservation to anyone
interested in the computational study of intelligence, whether they be
undergraduate or graduate students or senior scientists in the field.


About the reviewer: Subramanian is an Assistant Professor at the
Computer Science Department at Cornell University. Her interests are
in AI, its theoretical foundations and practical applications in
design, scheduling and molecular biology.  She has been teaching AI at
the undergraduate and graduate levels for about five years.