This document discusses information theoretic aspects of reinforcement learning from a PAC learning perspective. It introduces PAC (Probably Approximately Correct) learning and how the error bounds decrease with more data or a narrower hypothesis space. It also discusses limitations of PAC bounds when applied to neural networks. Model-based reinforcement learning algorithms like MBIE-EB aim to achieve PAC guarantees by incorporating exploration bonuses to restrict the changing rate of the policy distribution. Contrastive multi-skill reinforcement learning methods learn multiple skills to avoid distributional shift problems.