Inferential Induction: A Novel Framework for Bayesian Reinforcement Learning

Bayesian reinforcement learning (BRL) offers a decision-theoretic solution\nfor reinforcement learning. While "model-based" BRL algorithms have focused\neither on maintaining a posterior distribution on models or value functions and\ncombining this with approximate dynamic programming or tree search, previous\nBayesian "model-free" value function distribution approaches implicitly make\nstrong assumptions or approximations. We describe a novel Bayesian framework,\nInferential Induction, for correctly inferring value function distributions\nfrom data, which leads to the development of a new class of BRL algorithms. We\ndesign an algorithm, Bayesian Backwards Induction, with this framework. We\nexperimentally demonstrate that the proposed algorithm is competitive with\nrespect to the state of the art.\n