Browse Books

Reviewer: Samuel J. Zaslavsky

A unified theory of cognition is defined as a unique set of mechanisms that explains the following elements of cognitive behavior: problem solving, decision making, and routine action memory, learning, and skill language motivation and emotion imagining, dreaming, and daydreaming The order of items in this list reflects Newells view of their importance in cognitive behavior. He views the list as a set of prioritized objectives for the building of a unified theory. This list is as important to a unified theory of cognition as a system of axioms is to a formal theory. Just as there is a place for different sets of axioms, so there are other, justified variations on this list. For example, I would give memory the highest priority, followed by learning and then decision making. As long as the elements in the list can be explained in different ways, more than one unified theory of cognition can be built from the current state of knowledge. These theories would differ in their level of detail and in how far down the list they would go. This variety is why Newell uses the plural in the title of this book. The book, based on the 1987 William James lectures delivered by the author at Harvard University, presents an example of a unified cognitive theory. It is based on the problem-solving system Soar, developed by the author and his colleagues. The book contains eight chapters, a bibliography, and name and subject indexes. Chapter 1 is an introduction that discusses unified cognitive theories and ways to build them. Chapter 2 presents the basic concepts of cognitive systems, including behaving system, knowledge system, compu tational system, intelligence and intelligent behavior, search in the problem space, and knowledge search. These are the major concepts needed to describe cognitive systems in general. Chapter 3 discusses the basics of human cognitive architecture. Newell assumes that this architecture is constructed out of neural technology and is capable of intelligent behavior. These assumptions lead to what he calls the real-time constraint on human cognition; the limitations are neural speed and the response time of the cognitive symbol system. Chapter 4 describes the symbolic processing system Soar, which Newell uses as the basis for a unified theory of cognition in the rest of the book. He pays special attention to Soars problem-space structure, production system, decision cycle, and ability to learn using a chunking mechanism. A paragraph discusses a mapping between the Soar architecture and that of the human brain. The rest of the book deals with Newells unified theory of cognition. It contains an overview of different areas of cognitive behavior and attempts to explain them in view of the organization of Soar. Chapter 5 focuses on immediate behavior—tasks that take a second or two from presentation to completion—and on discrete perceptual-motor behavior. Chapter 6 covers a theory of human memory (short- and long-term, declarative and procedural, episodic and semantic), learning, and skill acquisition. Chapter 7 gives an overview of tasks that take minutes or longer. These tasks include problem solving, logical reasoning, and sentence verification. Chapter 8 discusses which elements of human cognitive behavior could actually be modeled within Soar and which can only be extrapolated from it. The latter include concept learning, emotion, and motivation. A list of recommendations for the development of unified theories of cognition concludes the book. This volume can be considered a how-to book. It defines steps that can be followed while trying to find a unique explanation of cognitive behavior and gives an example of how to use those steps. While this is the main value of the book, the theory it presents is interesting in its own right. The wide framework of the book, which discusses cognitive behavior from different points of view and shows how theoretical assumptions and concepts lead to a practical implementation, makes it a good survey of the current state of that part of AI that deals with problem solving and expert systems. Its completeness makes it a good introduction to the subject. This book is up-to-date, well written, and interesting. It has no flaws, but I can still feel the loneliness of the long-distance runner within it. I recommend it highly.

Reviewer: Bradley Scott Stewart

Newell attempts to convince the reader that cognitive scientists are now prepared to pursue what he refers to as unified theories of cognition (UTCs). Newell defines a UTC as an integrated system of mechanisms that produce the full range of human cognitive behavior. Recognizing the current limitations of the state of the art in cognitive science, however, he allows any cognitive theory that makes substantial progress toward this goal to be included as a UTC. Newell's notion of a mechanism is equivalent to what most people would call a computer system, including both hardware and software. Thus, Newell is attempting to create a computer system that can faithfully reproduce the full range of human behaviors. The strongest message in the book is directed toward cognitive scientists themselves, who appear to Newell to be reluctant to pursue such UTCs. The book synthesizes significant amounts of prior research in a readable and coherent whole. It would be useful in the classroom as a cognitive science text, or as an introduction to cognitive science for computer scientists or those studying artificial intelligence. It does not include any problems or other supporting materials for instructors. Outside the classroom, anyone interested in cognitive science would find the book informative and refreshing; Newell's writing style in much of the book is more akin to that of popular fiction writers than that of scholarly journal authors. The book is divided into eight chapters. In the first chapter, Newell presents his ideas on the nature of theories, using some specific examples from the psychology literature. He then describes what he means by UTCs and makes the argument that psychology is ready to pursue UTCs. He makes the case for readiness by using specific examples of systems that he calls “harbingers” of UTCs. A major point introduced in this chapter and carried throughout the book concerns the existence of massive amounts of experimental psychological data. Newell believes that this empirical work has reached a critical mass that will support the development of UTCs that synthesize most prior work in psychology. Chapter 2 contains Newell's account of the foundations of cognitive science. This chapter is the most interesting and best written in the book. It is organized as a sequence of extended definitions of fundamental elements in the cognitive science picture. Starting with his notion of behaving systems, Newell proceeds to discuss knowledge systems, representation, machines and computation, symbols, architectures, intelligence, search and problem spaces, and preparation and deliberation. Considering that an entire book is required to adequately discuss any of these topics individually, Newell does a remarkable job of presenting this material concisely, readably, and yet not totally superficially. In chapter 3, Newell uses time constraints on human cognition to draw some interesting conclusions about the human cognitive architecture. These observations are summarized in a chart entitled “Time Scale of Human Action” that divides human activities into 12 basic categories according to the time frame in which they occur, ranging from 100 microseconds to months. These categories are further grouped into what Newell refers to as “bands.” The fastest human activities are included in the biological band, while the most time-consuming activities fall into the social band. In between, the cognitive band covers actions less lengthy than those in the rational band. In some sense, this chart provides the foundation for the rest of the book. In chapter 4, Newell presents the basic architecture of SOAR, his candidate UTC. Chapters 5, 6, and 7 then explain how SOAR behaves in immediate behavior; memory, learning, and skill; and intentionally rational behavior, respectively. These three chapters take common cognitive science topics and organize them by the way in which they fit with Newell's bands. The final chapter then cleans up some loose ends by discussing the boundary issues—SOAR and its relation to the fastest activities in the biological band and the most extended activities in the social band. Although Newell is careful to point out that he does not consider SOAR to be the only possible UTC, he does present a strong case that makes it appear exemplary. None of the alternative candidate UTCs that Newell mentions comes close to the breadth and depth of SOAR. The book contains some controversial points. For example, Newell states that, “There is never much agreement on foundations—they are always less secure intellectually than what they support” (p. 42). This seems to me more like a comment on premises, axioms, or assumptions than on foundations. His chapter 2 contains enough commonly agreed upon foundational material to provide something of a counterexample to this statement. The controversial points in the book add substantially to the presentation by making the reader think and evaluate to a greater extent than is usually required in less adventuresome scholarly prose. The loose writing style of the book makes it easy and entertaining reading, but is not without its drawbacks. One term that I found continually disturbing was that Newell kept saying that he “listened” to his architecture. Although he does attempt to provide some discussion of this notion in Section 5.2, it still requires a leap of faith to accept the notion of passively receiving information from an intellectual construct. This way of discussing SOAR seems to deemphasize the specific interactions between the architecture and its creators to a greater extent than is warranted. While a lot of useful information can certainly be derived from a good architecture, the derivation process is generally less trivial than Newell would have us believe. A final comment concerns the SOAR architecture as a complete theory of cognition. A fundamental problem that is given insufficient attention by Newell concerns the creation of problem spaces. In SOAR, all activities are conducted in the context of problem spaces. These are representations of situations in which SOAR is attempting to achieve some goal or subgoal. Problem spaces are created dynamically as SOAR proceeds toward attaining its single top-level goal. The basic problem that needs more attention is how to generate initial problem spaces starting from scratch. What “scratch” is in the context of SOAR is difficult to define, but Newell generally assumes that substantial entry of problem structure into the SOAR computer system has occurred before any activity takes place. Initial problem representation is often the biggest challenge in problem solving, and it is not clear from Newell's discussion how much of a contribution his SOAR architecture currently makes in this area.