Learning Shaping Rewards in Model-based Reinforcement Learning

This paper shows how, in the absence of knowledge to deflne potential manually, the potential function can be learned online in parallel with the actual reinforcement learning process, using the prototypical model-based R-max algorithm.