Browse Books

Reviewer: H. Van Dyke Parunak

Practitioners of artificial intelligence (AI) sometimes have difficulty explaining to outsiders just what it is. Since the term was adopted at Dartmouth in 1956, its definition has varied widely depending on whom one asks and when one asks the question. Robotics and vision, originally at the heart of the enterprise, gradually developed into separate fields. Neural networks certainly are artificial and promise intelligence, but mainstream AI in its early years took a stance against them, resulting in a schism that is only recently healing. In contrast with the close interdependency between theory and practice in many disciplines, AI's theorists and experimentalists sometimes do not agree that they are part of the same enterprise. This introductory AI text bridges these chasms with a single coherent theme. For Russell and Norvig, AI is the design of systems that act (as opposed to just thinking), and do so in a rational (as opposed to a limited human) way. That is, AI is the design of agents, systems that perceive and act. The idea of agents unlocks the doors separating many of the factions in AI's past. If an agent is to perceive and act, vision and robotics must receive at least summary attention. Russell is on the Stanford faculty, and the orientation to logical formalisms is evident, but agents require their logic to be coupled to their actions, giving ample scope for discussion of classic algorithms. Even neural nets are welcomed back into the fold as an important mechanism for enabling agents to learn. The bulk of the book consists of eight parts, of which the first is an orientation and the eighth is a review and prospectus. Part 1, “Artificial Intelligence,” summarizes the history and foundations of AI, reaching back to psychology, mathematics, and Greek philosophy (chapter 1). The philosophical relevance of the enterprise is a theme throughout the book, and comes into prominence again in Part 8, “Conclusions,” where the authors engage the major critics of AI in a courteous and articulate discussion (chapter 26) and examine the field's present state and future directions (chapter 2 7). Part 1 also introduces two taxonomies to support the agent theme. The first taxonomy is of agents, ranging from simple “tropistic” agents that react directly to their stimuli, to agents that model their environments, maintain goals, and develop goals based on utilities. The second taxonomy is of the environments in which agents live. The intervening six parts show how these successive layers of sophistication can be implemented. At every step, the agent theme remains clearly in view, showing the reader why the new information is relevant, how it overcomes the weaknesses of systems developed with the techniques presented up to that point, and what opportunities it leaves open for further development. Although comparatively complex functions can be provided with communities of tropistic agents, the authors leave them behind in the introduction. Part 2, “Problem-solving,” introduces the concept of search, which assumes that agents have goals, in chapter 3. Chapter 4 reviews the major algorithms, including simulated annealing, which often is overlooked in AI, and chapter 5 shows how they can be applied in game playing. Part 3, “Knowledge and Reasoning,” shows how an agent can build and maintain a model of itself and of the world in which it operates. This objective provides the motivation for a brief description of propositional logic in chapter 6; a much fuller one of first-order predicate calculus in chapter 7; the application of first-order logic to building a knowledge base in chapter 8; a discussion of inference procedures in chapter 9; and their application in various reasoning systems in chapter 10. Part 4, “Acting Logically,” enhances the agent concept a step further to support planning. Its three chapters describe planning both theoretically and practically, and give special attention to integrating planning with action. Part 5, "Uncertain Knowledge and Reasoning,” introduces probability theory and Bayes's rule in chapter 14 and the application of these concepts to reasoning systems in chapter 15. This framework permits agents to be given utilities (presented in chapter 16) that support more complex decisions (covered in chapter 17) than those accessible to agents guided only by fixed goals. Part 6, “Learning,” describes how agents can improve their behavior on the basis of experience. After a general review of this vast topic in chapter 18, the book offers a detailed discussion of neural and belief networks in chapter 19. Chapter 20 is an overview of reinforcement learning, including genetic algorithms, and chapter 21 discusses the value of prior knowledge in guiding learning. Part 7, “Communicating, Perceiving, and Acting,” interfaces agents with their environments. The topic of communicating provides the context for a discussion of formal models of language in chapter 22 and their application in practical natural language applications in chapter 23. The section concludes with chapter 24, on perception, and chapter 25, on robotics. This book is encyclopedic, making it a valuable desk reference for the researcher as well as a flexible introductory text that will support many approaches to the subject. Nonetheless, the agent theme raises important issues that are not discussed. The behavior of a collection of agents is far more than the sum of the parts, exhibiting complex and even formally chaotic dynamics, enabling groups of tropistic agents (to which this book gives only cursory attention) to act robustly with apparent collective purpose, and showing self-organization in apparent disregard for the second law of thermodynamics. Following traditional AI research on agents, the authors focus on the architecture and design of individual agents, but give little or no attention to system-wide issues such as dynamics, emergent properties, and self-organization. These topics are of concern to researchers in artificial life, and a future edition of this book will do readers a service if it can build the kind of bridge to these topics that the current edition does to neural networks and evolutionary programming. The book is attractively produced, with keywords highlighted in a wide margin, over 1100 references through 1995, and a full index. Each chapter has bibliographical and historical notes that point to further information, and a variety of exercises (more than 300), including some that require interaction with the literature and some with significant programming. The authors support the book with an Internet site that gives access to a repository of working code, lists of errata, and clarifications and supplementary material.

Reviewer: Pavol Navrat

Many books on artificial intelligence have been published, and the competition among them is quite strong. This seems to have been an additional source of motivation for the authors, besides the inherent challenge of a discipline that attempts to understand and build intelligent entities—an exciting effort that has been occupying people's minds for centuries, but became a scientific field of study called “artificial intelligence” only 40 years ago, with the advent of computers. The authors have not only written an excellent textbook, but have distinguished themselves by their approach to presenting all of the major themes of AI. The common theme that unifies their presentation is intelligent agent design. First, they describe the desired abilities of an intelligent agent. Then, they present ways to achieve them. The book is organized into a preface and eight parts. Part 1 introduces artificial intelligence and the authors' approach to presenting it. Part 2 covers problem solving: goals, searching, informed search methods, and game playing. In Part 3, knowledge and reasoning are treated: first-order logic, building a knowledge base, inference in first-order logic, and logical reasoning systems. Part 4, “Acting Logically,” presents planning. Part 5 introduces uncertain knowledge and reasoning: uncertainty, probabilistic reasoning systems, and making simple and complex decisions. Part 6 is devoted to learning: learning from observation, learning in neural and belief networks, reinforcement learning, and knowledge in learning. Part 7 deals with communication, perception, and action by an agent in an environment. It covers agents that communicate, practical natural language processing, perception, and robotics. The conclusions drawn in Part 8 address philosophical foundations and the future of artificial intelligence. Two appendices present short summaries of complexity analysis, languages, and algorithms. An extensive bibliography (45 pages) and index (28 pages) are also included. One of the distinguishing features of the book, besides its unified presentation of a single theme, is its comprehensive coverage of the field. It is both broad and deep. The level of detail provided, however, is not always sufficient for the subject treated. For example, fuzzy set theory is dealt with in a single page, where it obviously cannot be properly introduced. In natural language processing, only understanding (but not generation) is dealt with. In problem solving, searching decompositions is not covered. On the other hand, some issues in decision networks and natural language parsing are treated in more detail than a one-semester course would typically allow. The breadth relates not only to the topics covered, but to the period of time discussed. The authors begin with the birth of Plato in 428 BC, and devote the next 20 pages to the development of ideas related to AI in the history of science. The authors should be credited for their efforts not only to present the field in a historical perspective, but to perfect the presentation by including detailed chapter summaries and bibliographical and historical notes. The writing is clear, concentrating more on ideas than on definitions or algorithms. Most of the notions are introduced informally and not through formal definitions. The book definitely reads well, at least for someone who already knows the field. I am positive that creative students will also enjoy it. The book is primarily intended for use in an undergraduate course. This does not mean that it can be covered in one semester, however. On the contrary, two semesters are probably needed . The book has many exercises, some of which can be assigned as term projects.