See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220546066

S. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, Third

Article  in  Artificial Intelligence · April 2011

CITATIONS READS

9 35,801

2 authors, including:

Bart Selman
Cornell University
291 PUBLICATIONS   19,536 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Bart Selman on 08 September 2020.

The user has requested enhancement of the downloaded file.

Contents lists available at ScienceDirect

Artificial Intelligence
www.elsevier.com/locate/artint

Book review

S. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, Third Edition.

Ashish Sabharwal a , Bart Selman b,∗

a r t i c l e i n f o

Article history:
Received 12 January 2011
Accepted 12 January 2011
Available online 19 January 2011

Since the publication of its first edition in 1995, Artificial Intelligence: A Modern Approach has become a classic in our
field. Even researchers outside AI, working in other areas of computer science, are familiar with the text and have gained a
better appreciation of our field thanks to the efforts of its authors, Stuart Russell of UC Berkeley and Peter Norvig of Google
Inc. It has been adopted by over 1000 universities in over 100 countries, and has provided an excellent introduction to
AI to several hundreds of thousands of students worldwide. The book not only stands out in the way it provides a clear
and comprehensive introduction to almost all aspects of AI for a student entering our field, it also provides a tremendous
resource for experienced AI researchers interested in a good introduction to subfields of AI outside of their own area of
specialization. In fact, many researchers enjoy reading insightful descriptions of their own area, combined, of course, with
the tense moment of checking the author index to see whether their own work made it into the book. Fortunately, due in
part to the comprehensive nature of the text, almost all AI researchers who have been around for a few years can be proud
to see their own work cited. Writing such a high-quality and erudite overview of our field, while distilling key aspects of
literally thousands of research papers, is a daunting task that requires a unique talent; the Russell and Norvig author team
has clearly handled this challenge exceptionally well.
Given the impact of the first edition of the book, a challenge for the authors was to keep such a unique text up-to-date
in the face of rapid developments in AI over the past decade and a half. Fortunately, the authors have succeeded admirably
in this challenge by bringing out a second edition in 2003 and now a third edition in 2010. Each of these new editions
involves major rewrites and additions to the book to keep it fully current. The revisions also provide an insightful overview
of the evolution of AI in recent years.
The text covers essentially all major areas of AI, while providing ample and balanced coverage of each of the sub-
areas. For certain subfields, part of the text was provided by respective subject experts. In particular, Jitendra Malik and
David Forsyth contributed the chapter on computer vision, Sebastian Thrun wrote the chapter on robotics, and Vibhu Mittal
helped with the chapter on natural language. Nick Hay, Mehran Sahami, and Ernest Davis contributed to the engaging set
of exercises for students. Overall, this book brings together deep knowledge of various facets of AI, from the authors as well
as from many experts in various subfields. The topics covered in the book are woven together via the theme of a grand
challenge in AI — that of creating an intelligent agent, one that takes “the best possible [rational] action in [any] situation.”
Every aspect of AI is considered in the context of such an agent. For instance, the book discusses agents that solve problems
through search, planning, and reasoning, agents that are situated in the physical world, agents that learn from observations,
agents that interact with the world through vision and perception, and agents that manipulate the physical world through

* Corresponding author.
E-mail addresses: [email protected] (A. Sabharwal), [email protected] (B. Selman).

doi:10.1016/j.artint.2011.01.005
936 Book review / Artificial Intelligence 175 (2011) 935–937

electronic or robotic controls. This unifying theme goes a long way in tying together the vast variety of specialized subfields
of AI.
The third edition of the book is divided into seven main parts: foundations of AI and intelligent agents; problem-solving
through search; knowledge representation, reasoning, and planning; dealing with uncertainty; learning; perception and
action in the physical world through natural language, vision, and robotics; and philosophical issues surrounding AI. Each
part consists of typically four to six chapters, each with roughly half a dozen sections. All in all, the 1132 pages of this text
include 27 chapters, 177 sections, 416 figures, nearly 2000 citations, and a 38 page long index consisting of almost 6000
items. Literally an intellectual tour de force. Although the various parts of the book can be studied in isolation, a key factor
behind the success of this book is that it helps the reader realize the multitude of intricate connections between various
subareas of AI. When used in the classroom, students taking a one semester “Introduction to AI” course would benefit the
most from touching upon each of the seven main parts of the book but limiting the overall amount of material by focusing
on only a few chapters within each part.
In addition to clearly articulated technical material, the book also provides towards the end of every chapter a historical
perspective, researched in part by Douglas Edwards, and a brief literature overview. These perspectives and citations provide
a highly useful and engaging context for both students and AI researchers. The plentiful examples and figures are invalu-
able to the reader. Most topics are introduced from the ground up, covering derivations of nearly all foundational concepts
and results in AI. To mention just a few examples, the text explains the optimality of A*, the completeness of resolution,
the graphplan framework for planning, inference in Bayesian networks, particle filtering algorithms, back-propagation, PAC
learning, the EM algorithm, support vector machines, etc. It also discusses historical ideas that lie at the heart of computa-
tion and AI, such as the Turing test, completeness and incompleteness, and undecidability. One can also find in this book
amazingly clear and concise introductions to a wide range of core topics and concepts in computer science and even other
related areas of science. Such topics include computational complexity, logic, formal languages and grammars, probability
theory, control theory, game theory, decision theory, and notions from psychology and philosophy. The sheer breadth of the
topics the text covers illustrates the tremendous richness and diversity of our field, which is a large part of its appeal as a
field of study.
For the second edition, which appeared in 2003, Russell and Norvig expanded the text in a variety of areas including:
historical connections of AI to fields such as control theory, game theory, economics, and neuroscience; additional cover-
age of online algorithms, constraint satisfaction problems, and modern satisfiability (SAT) solvers; new planning approaches
including graph-based and SAT-based planning; probabilistic inference techniques such as for Bayesian inference, dynamic
Bayesian networks, hidden Markov models, and Kalman filters; and ethical implications of AI in human society. This consti-
tuted a substantial expansion of the first edition, providing good coverage of many significant contributions to the theory
and practice of AI during the late 1990s and early 2000s.
For the third edition of the text, Russell and Norvig again went back to the drawing board to incorporate the key develop-
ments in our field in the past seven to ten years. Nearly 20% of the citations in this edition are to works published after the
printing of the second edition. Probabilistic reasoning, statistical methods, and multi-agent approaches, in particular, have
grown yet further in importance and the new edition appropriately reflects this development. There is new emphasis on the
challenge of unifying probabilistic and logical reasoning and representations. Expanding upon the underlying theme of rep-
resentation by an intelligent agent, the authors now distinguish between atomic, factored, and structured representations,
highlighting trade-offs in terms of relevance, usefulness, and computational efficiency of working with these representations
in various contexts. This edition also provides a more comprehensive introduction to the area of machine learning, which
has witnessed tremendous growth over the past two decades, both in terms of theoretical developments and practical im-
pact. Partial observability has also received more attention in this edition in various contexts, such as in search, planning,
and game playing. Needless to say, sections such as ‘state-of-the-art game programs’ have been updated to reflect the latest
in the respective subfields. Finally, in keeping with the ever growing role vast amounts of information plays in our lives,
this edition includes expanded coverage of Web search engines, techniques for information extraction, and methods to work
with extremely large data sets. In terms of exposition, hundreds of new exercises have been added, along with substantial
new material on the book’s website, http://aima.cs.berkeley.edu, including sample source code, lecture slides, and suggested
syllabi.
As the field of AI continues to grow and expand at a rapid pace, there is a significant pressure to specialize within the
field. Most universities now offer a range of courses in various subfields of AI, such as machine learning, natural language
processing, graphical models, knowledge representation, computer vision, robotics, etc. There are also a number of good
textbooks written specifically for these subfields. Relying solely on these specialized courses and books, however, poses the
risk that new students entering the field of AI may no longer get a good overview of the richness and complexity of the
diverse research challenges in AI. Moreover, many themes of AI, such as trade-offs between expressiveness of representations
and their computational properties, for instance between propositional and first-order representations, cut across different
subfields of AI. We strongly believe that it is important for students to be aware of these general themes underlying AI
problems and to get a solid introduction to the broad range of topics in AI before embarking on a more specialized subfield.
In fact, it is likely that some of the future breakthroughs in AI will come from students taking concepts from different parts
of Russell and Norvig’s text and combining them in novel ways.
In summary, this true AI classic remains an exceptional resource for both new students and seasoned AI researchers
wishing to learn about a subfield of AI or to get a good understanding of the overarching scope of AI as an integral part of
 Book review / Artificial Intelligence 175 (2011) 935–937 937

computer science. Over the years, the text will continue to educate in a superb manner literally hundreds of thousands of
computer scientists about the achievements and the excitement of our field. The authors’ ability to distill the essence of AI
concepts and techniques as well as their rich inter-connections is simply unsurpassed. Moreover, the revisions — going from
the first edition in 1995 to the second edition in 2003 and now the third edition in 2010 — beautifully keep pace with our
fast moving and continuously expanding field.

View publication stats