Determination, Uniformity, and Relevance: Normative Criteria for Generalization and Reasoning by Analogy

Abstract

If an agent is to apply knowledge from its past experience to a presen episode, it must know what properties of the past situation can justifiably be projected onto the present on the basis of the known similarity between the situations. The problem of specifying when to generalize or reason by analogy, and when not to, therefore looms large for the designer of a learning system. One would like to be able to program into the system a set of criteria for rule formation from which the system can correctly generalize from data as they are received. Otherwise, all of the necessary rules the agent or system uses must be programmed in ahead of time, so that they are either explicitly represented in the knowledge base or derivable from it.