sandon@AI.WISC.EDU (Pete Sandon) (02/04/86)
From: sandon@ai.wisc.edu (Pete Sandon) This is not to defend the Dreyfus brothers, since I have yet to read their books. On the other hand, I think they make a good point, though with a bad example, in emphasizing learning as a process of refinement. The example related in Miles Murdocca's submission is that of learning to ride a bike through trial and error. The reason the example is a bad one, is that it fits into the category of skill refinement as AI researchers would use the term. This leads to the argument that Dreyfus and Dreyfus are missing the critical distinction between knowledge acquisition and skill refinement. My feeling is that too much is made of this distinction. Had the example been one of learning to distinguish fruits from vegetables, or one of learning the symptoms of a class of diseases well enough to diagnose them, this argument would not have arisen. Clearly these involve knowledge acquisition rather than skill refinement. And yet, it could be argued, and perhaps is argued by the Dreyfus's, that what the AI researchers consider to be knowledge acquisition should be just as much a refinement process guided by trial and error as learning to ride a bike. Whereas AI considers concept formation to occur as the acquisition of discrete chunks of knowledge, an alternative is to use the gradual acquisition of evidence to support one concept definition over another, in a manner similar to skill refinement. Of course, if this criticism of AI is correct, AI has already answered it. The use of connectionist models, and the corresponding learning mechanisms currently being studied, provide just the sort of cognitive models that support this refinement type of learning through trial and error. --Pete Sandon