Strategy Learning for Reasoning Agents

We present a method for knowledge-based agents to learn strategies. Using techniques of inductive logic programming, strategies are learned in two steps: A given example set is first generalized into an overly general theory, which then gets refined. We show how a learning agent can exploit background knowledge of its actions and environment in order to restrict the hypothesis space, which enables the learning of complex logic program clauses. This is a first step toward the long term goal of adaptive, reasoning agents capable of changing their behavior when appropriate.