A new guide to artificial intelligence

In this unslim volume I have endeavored to cover the large and growing field of AI in some detail (although gruesome details are clearly marked as such by parenthetical curly brackets and can be omitted by the busy reader who does not wish to be burdened by them but is nevertheless comforted by the knowledge that they are there). At the other end of the spectrum we have the genuine AI texts: Winston was the force to be reckoned with here (and Rich's text to a lesser extent), but Charniak and McDermott now looks like it might carry all before it (although Tanimoto's new text may snatch a significant share of the spoils). All the gruesome detail is there. Last year's birds may not be in this year's nests, but the same cannot be casually asserted about old AI programs (many are no more than abandoned Ph.D. thesis projects) and current expositions of the field. The moribund AI programs of yesteryear are all too frequently still used as the examples of AI today. The known inadequacies and the general problems are largely ignored or just mentioned in passing. AI is, or should be, much more than currently fashionable programs, tools, and paradigms. We do need these detailed compendiums of the current wisdom, and Charniak and McDermott, and Tanimoto, supply this need. We need a broader perspective, especially in AI, a field that has always suffered from a myopic satisfaction and delight in the antics of a few performing programs. Elaine Rich's book is the only one that comes close.

There are books that examine and discuss global perspectives on AI Boden's and Hofstadter's, for example. But these books make no attempt to cover the diversity of theories and programs that is AI. The book you are reading attempts to provide a global perspective but in conjunction.

It attempts to cover all of the recognized AI work in sufficient detail to allow a critique from general concerns to be anchored whenever possible in the structure of specific AI programs. So I view this book as one that covers AI in some detail but needs to be supplemented by examinations of current projects published in the voluminous literature. It is in some sense a companion to the major AI texts, providing a broader perspective on the wealth of details that such texts contain. Taking vision work as an example: Tanimoto provides an exhaustive survey of AI vision work, but in 75 pages we get all of the details on the early processing operators but virtually nothing on perception as a process that intimately involves contextual information - a rich and varied fund of information whose application must be central to intelligent seeing. Horn expands the coverage of vision to 500 pages but with no significant change of emphasis (except more pure mathematics, in exchange for Tanimoto's LISP code). But there are virtually no implementations involving context, a crucial source of information for intelligent vision systems. Both Tanimoto and Horn practically ignore it (as do Charniak and McDermott) - despite the fact that Horn's book has a C3PO-ish robot playing the piano on the cover. What it cannot be illustrating is a humanlike skill acquired and refined by constant practice, for Horn makes absolutely no mention of these developmental and cognitive aspects of vision and motor control. Industrial robots are often impressively lizard-swift, but only when they are working blindly in a very precisely configured environment.

If you want Ramer's recursive algorithm for approximating a polygon from a set of vertices, then look in Tanimoto's book; if it's the details of a relaxation algorithm for finding intrinsic images that you're after, then Charniak and McDermott's book has them; but if you want to know something about what will be necessary in an intelligent vision system the problems of, and approaches to solutions for, applying and integrating top-down and bottom-up information - then you have almost no other choice within the set of broad AI texts but to look in this book.

To my way of thinking, the other panoramic AI texts focus too closely on particular programs and mechanisms that can be cheerfully viewed as the first steps towards truly intelligent systems. They examine, explore, and exhibit the wealth of work wherein the ultimate goals of AI have not been (and cannot be) achieved. They dissect the body of work in great detail, which is necessary but not sufficient on its own to advance the frontiers of current AI towards something more like real AI. It's like attempting to understand the neuroanatomical basis for intelligence by microscopically examining a hand. We must explore the larger picture with all its flaws, holes, and general lack of specificity, as well as dissect the current projects and explore inadequate formalisms. I have attempted to pursue both of these goals within a broad-based view of AI. I do labor the details when I'm dealing with work within the Connectionist Paradigm (CP), because they are not collectively available elsewhere. If the CP gets knocked on the head shortly by some latter-day duo like Min sky and Papert (or, if this pair bestir themselves again and can do a demolition job in the spirit of Perceptions), then this book may well go down in history as different but largely misguided. symbolic paradigm that has dominated (not to say stagnated) most AI work for many years. The importance of the CP derives from the healthy competition that it has engendered for the symbolic paradigm. It offers another viewpoint on many AI issues, and the new perspective serves to generate questions about the significance of "fundamental" AI problems as well as about the validity of basic assumptions. "The road to AI is shrewn with potholes; connectionism may be a less travelled route, but it is not any better maintained- (Pollack, 1987a, p. 108). But it does give us a new way to approach, and thus evaluate, fundamental problems and outcomes of current AI work.

Thus search strategies, which are a fundamental problem if all AI is viewed in terms of symbolic search spaces, become inconsequential if the generic representation is a subsymbolic network. Similarly, the important AI subfields of knowledge representation and of machine learning undergo a metamorphosis in the switch from the symbolic to the connectionist paradigm as the overarching conception of AI. Within the expert systems work, important little subproblems like conflict resolution just disappear in the CP, and so on.

As for the validity of basic assumptions: Knowledge bases (as I've said) become transmuted in form, and thus the basis of validity for logical inference as the mechanism to apply has vanished (the role of logical inference may be challengeable within the symbolic paradigm, but in the CP it has hardly a role to challenge). It sometimes worries me that my view of AI could perhaps be described as jaundiced. For, if I have replaced the rose tint of the customary AI perspective with a definitely yellowish one, then it will serve, on the whole, to redress the balance. At times I display little more than a torpid enthusiasm for many much-vaunted AI techniques and demonstration programs. I prefer to harp on the problems. Again, I think that this attitude may add a measure of balance when set against the current climate of overexuberance which surrounds the term AI.

Finally, the style of this book may seem too frivolous for a serious textbook. Nonsense has a useful role to play in explanations of the scientific. This transmutation of status is a threat to progress in any science; it is a particularly serious one in a field as young and rapidly developing as is AI. I hope to provoke disagreement. Work in AI is only just beginning to uncover the serious problems, so disagreement and debate are both in order. It is not clear to me that a lighthearted approach to the so-called facts and accomplishments of AI is necessarily any more misinformative, than dry, hard statements of the "truth." This book will be no exception - you can't buck theorems.

The writing of this book was split across the Atlantic: half in Las Cruces, New Mexico, and half in Exeter, Devon. I thus thank the guys (of both sexes) in the Department of Computer Science and in the Computing Research Laboratory, both part of New Mexico State University, and my new colleagues in the Department of Computer Science at the University of Exeter. In particular, Dan Fass, Dave Farwell, Paul Jeremaes, Antony Galton, and Colin Beardon gave me detailed, written reports on many of my mistakes. Ellis Horwood deserve special mention for their permission granted to me for the reuse of chunks of my earlier book.