Decision tree methods for finding reusable MDP homomorphisms

摘要: State abstraction is a useful tool for agents interacting with complex environments. Good state abstractions are compact, reuseable, and easy to learn from sample data. This paper combines extends two existing classes of methods achieve these criteria. The first class search MDP homomorphisms (Ravindran 2004), which produce models reward transition probabilities in an abstract space. second methods, like the UTree algorithm (McCallum 1995), compact value function quickly Models based on can easily be extended such that they usable across tasks similar functions. However, cannot this fashion. We present results showing new, combined fulfills all three criteria: resulting learned data, used