CONTROLLING THE PROCESS OF LEARNING FROM AN EXAMPLE THROUGH ADAPTIVE GENERALIZATION OF EPISODIC MEMORY

Automizing acquisition of expert knowledge is one of the most important problems to be solved for knowledge-based systems such as expert systems or decision support systems. Machine learning is a subfield of artificial intelligence concerning the theory on how to learn concepts from examples. In this field, two types of learning methods have been proposed so far: explanation-based learning (EBL) as a theory of deductive learning and similarity-based learning (SBL) as a theory of inductive learning. This paper presents a methodology to acquire general knowledge from a single example (episode) by the use of a repertoire of prior cases based on a hybrid architecture of EBL and SBL. Being provided with a complex input stream of description on some episodic case as an example, the system selectively detects what aspects are coherent and which are unusual under multiple contexts. That is, at first, the system tries to understand or explain the input episode based on a knowledge structure called script, which is an expectation knowledge structure self-organized through the generalization of past cases stored in its memory. When the system detects something unexplainable in the input episode, it adaptively changes its strategy for organizing scripts from past cases in the memory, and evaluates the coherence of the input episode from a different viewpoint. The evaluation whether the input is significantly coherent in the context or not is performed by calculating its membership or typicality in the fuzzy category formed over a collection of prior cases. This procedure for typicality evaluation of an episode is formulated as a fuzzy reasoning process. Finally, the system outputs a general knowledge so that it can be applied to situations other than the input case.