Relational State-Space Feature Learning and Its Applications in Planning

Jia-Hong Wu, Robert Givan

We consider how to learn useful relational features in linear approximated value function representations for solving probabilistic planning problems. We first discuss a current feature-discovering planner that we presented at the International Conference on Automated Planning and Scheduling (ICAPS) in 2007. We then propose how the feature learning framework can be further enhanced to improve problem solving ability.

Subjects: 12. Machine Learning and Discovery; 11. Knowledge Representation

Submitted: Sep 11, 2007