Bayesian approaches provide a principled solution to the exploration-exploitation trade-off in Reinforcement Learning. Typical approaches, however, either assume a fully observable environment or scale poorly. This work introduces the Factored Bayes-Adaptive POMDP model, a framework that is able to exploit the underlying structure while learning the dynamics in partially observable systems. We also present a belief tracking method to approximate the joint posterior over state and model variables, and an adaptation of the Monte-Carlo Tree Search solution method, which together are capable of solving the underlying problem near-optimally. Our method is able to learn efficiently given a known factorization or also learn the factorization and the model parameters at the same time. We demonstrate that this approach is able to outperform current methods and tackle problems that were previously infeasible.

We describe an approximate dynamic programming algorithm for partially observable Markov decision processes represented in factored form. Two complementary forms of approximation are used to simplify a piecewise linear and convex value function, where each linear facet of the function is represented compactly by an algebraic decision diagram. ln one form of approximation, the degree of state abstraction is increased by aggregating states with similar values. In the second form of approximation, the value function is simplified by removing linear facets that contribute marginally to value. We derive an error bound that applies to both forms of approximation. Experimental results show that this approach improves the performance of dynamic programming and extends the range of problems it can solve.