Monte Carlo Methods for Bayesian Reinforcement Learning and POMDP

Опубликовано 12 августа 2016, 2:13

Partially Observable Markov Decision Process is an elegant and general model for planning under uncertainty. Applications for POMDPs include control of autonomous vehicles, dialog systems, and systems for providing assistance to the elderly. Learning problems such as reinforcement learning, making recommendations and active learning can also be posed as POMDPs. Unfortunately, solving POMDPs is computationally intractable. When the state space is not too large, we give conditions under which solving POMDPs becomes computationally easier, and describe algorithms for solving such problems. We extend the algorithms to very large or infinite state spaces using Monte Carlo methods.