How to Learn Domain-Invariant Representations for Visual Reinforcement Learning: An Information-Theoretical Perspective | Synapse