Solving POMDPs using selected past events
Abstract
A new algorithm for solving Partially Observed Markov Decision Processes build on theoretical results about exhaustive observable. It aims at building an extension of the state space so as to obtain a Markov Decision Process which can then be solved by classical methods. We present two versions of the algorithm, one using reinforcement learning when the evolution model is unknown, the other is quicker but requires the knowledge of this evolution model.