Levels of Pattern Description in Learning

Abstract

A learning system in a complex, real-world domain will require a significant amount of knowledge to be used in order to (1) deal with large numbers of features, most of which are irrelevant, and (2) find similarities between the concepts that are inferred from the observed data. Use of knowledge-free, syntactic approaches to generalization _in complex environments will result in a combinatorial explosion in the number of possible generalizations. Moreover, the important semantic features are not "in" the data; rather they must be hypothesized using prior knowledge. The learning system described in this paper uses a multi-level knowledge-directed approach in order to cope with these problems. This paradigm is explored in the action-oriented game of baseball. The system attempts to interpret observed activity in terms of general knowledge provided about competitive games. This approach to learning can be viewed as a type of recognition, where the level of initial knowledge is general and where the specific observations mold a particular structure from the general knowledge. The system is organized into multiple levels of pattern descriptions, processing, and knowledge, reflecting the logical structure of the problem. In moving through those levels of description, the system filters out irrelevant features, hypothesizes additional semantic features (goals and relationships) and forms a hierarchy of generalized classes that extract the similarities in the descriptions. Examples of learning by a working computer program are presented.