JMC@SAIL.STANFORD.EDU (John McCarthy) (08/03/87)
Like mathematics, philosophy and engineering, AI differs from the (other) sciences. Whether it fits someone's definition of a science or not, it has need of scientific methods including controlled experimentation. First of all, it seems to me that AI is properly part of computer science. It concerns procedures for achieving goals under certain conditions of information and possibility for action. We can even consider it analogous to linear programming. Indeed if achieving one's goals always consisted finding the values of a collection of real variables that would minimize a linear function of these variables subject to a collection of linear inequalities, then AI would coincide with linear programming. However, the relation between goals, available actions, the information initially available and that can later be acquired is sometimes more complex than in any of the branches of computer sciences the main character of whose scientific treatment consists of mathematical theorems. We don't have a mathematical formalization of the general problem faced in AI let alone general mathematical methods for their solution. Indeed what we know of human intelligence doesn't suggest that a conventional mathematical formalization of the problems intelligence is used to solve even exists. For this reason AI is to a substantial degree an experimental science. The fact that a general mathematical formalization of the problems intelligence solves is unlikely doesn't make mathematics useless in AI. Many aspects of intelligence are formalizable, and languages of mathematical logic are useful for expressing facts about the common sense world, and logical reasoning, especially as extended by non-monotonic reasoning is useful for drawing conclusions. In my view a large part of AI research should consist of the identification and study of intellectual mechanisms, e.g. pattern matching and learning. The problems whose computer solution exhibits these mechanisms should be chosen for reasons of scientific perspicuousness analogously to the fact that genetics uses fruit flies and bacteria. A. S. Kronrod once said that chess is the {\it Drosophila} of artificial intelligence. He might have been right, but the organizations that support research have taken the view that problems should be chosen for their practical importance. Sometimes it is as if the geneticists were required to do their work with elephants on the grounds that elephants are useful and fruit flies are not. Anyway chess has been left to the sportsmen, most of whom only write programs, not scientific papers and compete about who can get time on the largest computers or get someone to support the construction of specialized chess computers. Donald Norman's complaints about the way AI research is conducted have some validity, but the problem of isolating intellectual mechanisms and making experiments worth repeating is yet to be solved, so it isn't just a question of deciding to be virtuous. Finally, I'll remark that AI is not the same as cognitive psychology although the two studies are allied. AI concentrates more on the necessary relations between means and ends, while cognitive psychology concentrates on how humans and animals achieve their goals. Any success in either endeavor helps the other. Methodology in AI is worth studying, but acceptance of its results should be moderated by memory of the behaviorist catastrophe in psychology. Doctrines arising from methodological studies crippled the science for half a century. Indeed psychology was only rescued by ideas arising from the invention of the computer --- and at least partly ideas originating in AI.