Levels of pattern description in learning

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