ladkin@KESTREL.ARPA (Peter Ladkin) (03/08/86)
After reading Stuart Russell's commentary on Dreyfus's talk, I could hardly believe I'd heard the same talk that he had. A summary: Dreyfus is arguing that the rule-based expert system paradigm cannot, in some cases, codify the behaviour of an expert. They may be able to reproduce the behaviour of a proficient practitioner (in his taxonomy) who is not an expert (e.g. chess programs). He allows that there are some domains where a rule-based system may fare better than a human (and mentioned the backgammon program, but was corrected by members of the audience who said it wasn't nearly as good as he had been led to believe). The concept of expert behaviour as internalised rules goes back to Plato, and he can trace the influence of this idea through Descartes and Kant, even to Husserl. He believes it is fundamentally mistaken, and provided few arguments in the talk (some of them may be found in *What Computers Can't Do*). He presented a proposal for a taxonomy of skilled behaviour, which is consistent with the phenomenology of the domain, and which he believes is a testable conjecture for explaining skilled behaviour. This he credits to his brother Stuart. He illustrated some of the ideas from the domain of driving a car (it was originally a study of pilot skills for the Air Force). He discussed at some length his experiments with Julio Kaplan, a former Junior World Champion at chess. He regards the conclusions they would wish to draw as *an anecdote* [his words] because of the difficulty of obtaining suitable subjects to perform controlled experimants. Most highly expert chess players (grand masters?) are so concerned with the game that their concentration is hard to break. Kaplan is an exception, and they are able to get him to concentrate on counting beeps while playing. Others, he said, tend to ignore the test in favor of the game. Dreyfus thinks the current connectionist work is exciting, and may have possibilities that the rule-based *Traditional AI* [his words] work does not have. [End of summary]. I address some of Russell's points, omitting the loaded terminology in which they are expressed, and some of Russell's less professional speculations. I use his numbering. 1) The discussion was free of dissent because there was little to disagree with. He's not submitting a cognitive model for AI as a whole, he's addressing expert systems, and claiming (as he has done for many years) that not all expert behaviour admits of rule-based mimicry. 2) I have been unable to find a reference to Dreyfus believing *human experts solve problems by accessing a store of cached, generalised solutions*, probably because that is not a reasonable representation of his views. It is certainly not consistent with the views in *What...*. 3) His view that humans use *intuitive matching processes based on total similarity* is argued in *What...* with evidence from the domain of gestalt psychology. It's surprising that Russell thought he couldn't be more specific, as he had been 7 years ago. I suspect inexact communication. 4) Russell says, referring to the above, that *this mechanism doesn't work*. This is a misapprehension. Dreyfus is referring to a phenomenon, observed by some researchers. I presume Russell is denying the existence of this phenomenon, without argument. Dreyfus does make the claim that whatever mechanism may be underlying the phenomenon cannot be implemented in a rule-based system. (Is this the same as *a system which uses symbolic descriptions*? After all, I am such a system, witness the present posting.) A quick re-reading of *What....* has convinced me that many contributors to this debate have not read it carefully for its arguments. I recommend reading it if you haven't done so. Incidentally, it is truly embarrassing to see some of the quotations from pre-1979 AI workers. Surely, no-one could have said those things.....but then, that's why he wrote the book, and our current attitudes have been molded in part by the resulting debate. Peter Ladkin