CASE-BASED LEARNING OF STRATEGIC KNOWLEDGE BEATRIZ LOPEZ ENRIC PLAZA Institut d'Investigaci6 en Intel.ligbncia Artificial CEAB-CSIC, Camf de Santa B~bara, 17300 Blanes, Catalunya, Spain bea@ceab.es, plaza@ceab.es Research partially supported by CICYT 801/90 Massive Memory Project and Esprit I12148 Valid Project. Abstract In this paper we describe BOLERO,a case-based reasoner that learns strategic knowledge (plans) to improve the problem-solving capabilities of an expert system. As a planner, BOLERO is a reactive planner that when gathering new observations can immediately generate a new plan to cope with the new situations. As a learner, BOLERO is capable of learning strategies from observation of the problem-solving performed by a teacher. From this experience, BOLERO plans strategies that solve new problems. BOLERO learns from success and failure during its problem-solving process. An evaluation to measure the efficiency of BOLERO is performed by comparing the system's results against both the correct solution of a case and the solutions provided by different domain experts. BOLERO has been proved useful in acquiring strategic knowledge and in refining existing strategies, in a real-life expert system for pneumonia diagnosis. Keywords: Strategy learning, case-based learning, planing, control knowledge. 1. Introduction When approaching the application of machine learning techniques to the knowledge acquisition process of expert systems it is important to know what are the main problems today for building expert systems. In a recent study developed at Carnegie Group on the building of expert systems the surprising result was that the major effort (50%) of knowledge engineers was dealing with the uncontrolable interaction among rules and assuring that the proper sequence of goals and rule chainings is achieved during problem solving (Carbonell 1990). In this paper we show how this issue can be solved through the automatic acquisition of strategies that control the behavior of a rule-base. Strategies are acquired by case- based learning techniques by observing the problem-solving behavior of an expert and by learning from the system's own ease-based problem-solving. In this paper we describe BOLERO,a system that learns strategies to improve solving capabilities in expert systems. Problem solving strategies avoid to explore the whole search space for a problem,