Cost-sensitive reinforcement learning

We introduce cost-sensitive regression as a way to introduce information obtained by planning as background knowledge into a relational reinforcement learning algorithm. By offering a trade-off between using knowledge rich, but com- putationally expensive knowledge resulting from planning like approaches such as minimax search and computationally cheap, but possibly incorrect generalizations, the reinforcement learning agent can automatically learn when to apply planning and when to build a generalizing strategy. This approach would be useful for problem domains where a model is given but which are too large to solve by search. We discuss some difficulties that arise when trying to define costs that are semantically well founded for reinforcement learning problems and present a preliminary algorithm that illustrates the feasibility of the approach.