RUSSELL@SUMEX-AIM.ARPA (Stuart Russell) (03/02/86)
To address some of the actual content of Dreyfus' recent talk at Stanford, delivered to an audience consisting mostly of AI researchers: 1) The discussion after the talk was remarkably free of strong dissent, for the simple reason that Dreyfus is now making a sloppy attempt at a cognitive model for AI, rather than making any substantive criticism of AI. Had his talk been submitted to AAAI as a paper, it would probably have been rejected as containing no new ideas and weak empirical backing. 2) The backbone of his argument is that human *experts* solve problems by accessing a store of cached, generalized solutions rather than by extensive reasoning. He admits that before becoming expert, humans operate just like AI reasoning systems, otherwise they couldn't solve any problems and thus couldn't cache solutions. He also admits that even experts use reasoning to solve problems insufficiently similar to those they have seen before. He doesn't say how solutions are to be abstracted before caching, and doesn't seem to be aware of much of the work on chunking, rule compilation, explanation-based generalization and macro-operator formation which has been going on for several years. Thus he seems to be proposing a performance mechanism that was proposed long ago in AI, acts as if he (or his brother) invented it and assumes, therefore, that AI can't have made any progress yet towards understanding it. 3) He proposes that humans access previous situations and their solutions by an "intuitive, holistic matching process" based on "total similarity" rather than on "breaking down situations into features and matching on relevant features". When I asked him what he meant by this, he said he couldn't be any more specific and didn't know any more than he'd said. (He taped our conversation, so he can no doubt correct the wording.) In the talk, he mentioned Roger Shepard's work on similarity (stimulus generalization) as support for this view, but when I asked him how the work supported his ideas, it became clear that he knew very little about it. Shepard's results can be explained equally well if situations are described in terms of features, but more importantly they only apply when the subject has no idea of which parts of the situation are relevant to the solution, which is hardly the case when an expert is solving problems. In fact, the fallacy of analogical reasoning by total similarity (which is the only mechanism he is proposing to support his expert phase of skilled performance) has long been recognized in philosophy, and also more recently in AI. Moreover, the concept of similarity without any goal context (i.e. without any purpose for which the similarity will be used) seems to be incoherent. Perhaps this is why he doesn't attempt to define what it means. 4) His final point is that such a mechanism cannot be implemented in a system which uses symbolic descriptions. Quite apart from the fact that the mechanism doesn't work, and cannot produce any kind of useful performance, there is no reason to believe this, nor does he give one. In short, to use the terminology of review forms, he is now doing AI but the work doesn't contain any novel ideas or techniques, does not report on substantial research, does not properly cite related work and does not contribute substantially to knowledge in the field. If it weren't for the bee in his bonnet about proving AI (except the part he's now doing) to be fruitless and dishonest, he might be able to make a useful contribution, especially given his training in philosophy. Stuart Russell Stanford Knowledge Systems Lab -------