Towards a Unified A roach to Come

Rule induction (either directly or by means of decision trees) and case-based learning (forms of which are also known as instance-based, memory-based and nearest-neighbor learning) arguably constitute the two leading symbolic approaches to concept and classification learning. Rule-based methods discard the individual training examples, and remember only abstractions formed from them. At performance time, rules are applied by logical match (i.e., only rules whose preconditions are satisfied by an example are applied to it). Case-based methods explicitly memorize some or all of the examples; they avoid forming abstractions, and instead invest more effort at performance time in finding the most similar cases to the target one. There has been much debate over which of these two approaches is preferable. While each one can be extended to fit the results originally presented as evidence for the other, it typically does so at the cost of a more complex, less parsimonious model. In classification applications, each approach has been observed to outperform the other in some, but not all, domains. In recent years, multistrategy learning has become a major focus of research within machine learning. Its main insight is that a combination of learning paradigms is often preferable to any single one. However, a multistrategy learning system typically operates by calling the individual approaches as subprocedures from a control module of variable sophistication, and again this is not completely satisfactory from the point of view of parsimony. In my thesis work I argue that rule induction and case-based learning have much more in common than a superficial examination reveals, and can be unified into a single, simple and coherent model of symbolic learning. The proposed unification rests on two key observations. One is that a case can be regarded as a maximally specific rule (i.e., a rule whose preconditions are satisfied by exactly one case). Therefore, no syntactic distinction need be made between the two. The second observation is that rules can be matched approximately, as cases are in a case-based classifier ( i.e., a rule can match an example if it is the closest one to it according to some similarity-computing procedure, even if the example does not logically satisfy