Learning control knowledge for AI planning domains

Sungwook Yoon, Purdue University

Abstract

We consider techniques for learning to plan in deterministic and stochastic Artificial Intelligence (AI) planning domains. Many AI planning domains have structure that can be exploited during planning. However, traditional automated techniques fail to capture and use such structure and as a result do not scale well as the size of the problems grows. Planning systems that exploit human-specified domain structure have shown impressive performance. In this thesis, we explore automated techniques that can identify domain-specific structure in AI planning domains. We apply machine-learning algorithms to solved problem instances to find control knowledge describing useful structure in the target planning domain. One key component in our approach is the use of concept-based knowledge representation for control knowledge. We develop learning algorithms suited for learning control knowledge. Our planning system using the learned control knowledge provides state-of-the-art performance in benchmark deterministic and stochastic planning domains. Finally, we eliminate the need for solved problem instances for learning by using a novel form of approximate policy iteration and a random-walk-based sampling algorithm. The result of this extension is a domain-independent system that automatically finds a domain-specific control knowledge by learning from the domain definition.

Degree

Ph.D.

Advisors

Fern, Purdue University.

Subject Area

Computer science|Electrical engineering|Artificial intelligence

COinS