On-line learning of macro planning operators using probabilistic estimations of cause-effects

Abstract

In this work we propose an on-line learning method for learning action rules for planning. The system uses a probabilistic approach of a constructive induction method that combines a beam search with an example-based search over candidate rules to find those that more concisely describe the world dynamics. The approach permits a rapid integration of the knowledge acquired from experience. Exploration of the world dynamics is guided by the planner, and – if the planner fails because of incomplete knowledge – by a teacher through action instructions.