Learning Weighted Rule Sets for Forward Search Planning

In many planning domains, it is possible to define and learn good rules for reactively selecting actions. This has lead to work on learning rule-based policies as a form of planning control knowledge. However, it is often the case that such learned policies are imperfect, leading to planning failure when they are used for greedy action selection. In this work, we seek to develop a more robust form of rule-based control knowledge, attempting to leverage the perceived utility of rules while allowing for imperfection. Specifically, we consider learning sets of weighted action-selection rules for a target planning domain, which are used to assign numeric scores to potential state transitions. These scores can then be used to guide forward search strategies for solving problems from the target domain. This approach allows for information from multiple rules to be combined to help maintain robustness to errors. Our learning approach is based on a combination of a heuristic rule learner and RankBoost, an efficient boostingstyle algorithm for learning ranking functions. We further show how to improve performance by incorporating FF’s heuristic and tuning the rule weights learned by RankBoost using a perceptron-style algorithm. Our initial empirical results show significant promise for this approach in a number of domains.