Probabilistic Inference-based Reinforcement Learning

Abstract

We introduce probabilistic inference-based reinforcement learning (PIReL), an approach to solve decision making problems by treating them as probabilistic inference tasks. Unlike classical reinforcement learning, which requires explicit reward functions, in PIReL they are implied by probabilistic assumptions of the model. This would enable a fundamental way to design the reward function by model selection as well as bring the potential to apply existing probabilistic modeling techniques to reinforcement learning problems.

Publication
Benelearn 2017 conference
Date
Links
PDF