Abstract
Many open problems involve the search for a mapping that is used by an algorithm solvR ing an MDP. Useful mappings are often from p the state set to some other set. Examples ino clude representation discovery (a mapping to b a feature space) and skill discovery (a mapf ping to skill termination probabilities). Difs ferent mappings result in algorithms achievb ing varying expected returns. In this paper w we present a novel approach to the search for e any mapping used by any algorithm attempte ing to solve an MDP, for that which results m in maximum expected return.