ABSTRACT

Learning is one of the most difficult problems facing researchers in artificial intelligence. There appear to be two reasons for this. First, it is difficult to specify the problem in functional terms, such as the search for a process that will associate a set of possible inputs with a set of desired outputs. Since the “output” in this case is a modification of the system itself, and since the meaning of such a modification can only be determined by reference to a reasonably well-specified theory of how that system carries out its other tasks, the general lack of well-specified theories of how intelligent systems do things is a serious obstacle. Second, it seems quite possible that there will not be a theory of learning, but will instead be a number of distinct theories of learning. There is no a priori reason to doubt that there might be several different processes that could be employed to modify a complex mechanism in the direction of improved performance.