Learning partially observable action models: efficient algorithms

摘要: We present tractable, exact algorithms for learning actions’ effects and preconditions in partially observable domains. Our maintain a propositional logical representation of the set possible action models after each observation execution. The perform any deterministic domain. This includes STRIPS actions with conditional effects. In contrast, previous rely on approximations to achieve tractability, do not supply approximation guarantees. take time space that are polynomial number domain features, can stays compact indefinitely. experimental results show we learn efficiently practically domains contain over 1000’s features (more than 2 states).