Towards Global Reinforcement Learning

There has been substantial progress in the field of Reinforcement Learning (RL) in recent years. Many RL applications have achieved remarkable success by taking advantage of as much prior knowledge (PK) as possible. However, they often do so through problem-specific heuristics, due to the lack of a general framework capable of incorporating broad PK. We take the first steps towards a global RL framework one that incorporates PK over multiple learning components, to learn about all of them simultaneously. Specifically, we consider PK over both model dynamics and policy. We construct a prior belief that reflects both components of PK, use this belief to select actions and update the belief upon new evidence.