State Occupancy Matching

State occupancy matching focuses on aligning the probability distributions of states visited by a learned agent with those of a target policy or observed data. Current research explores this concept across diverse applications, employing methods like Wasserstein distance minimization, $f$-divergence regularization, and occupancy-based model learning within frameworks such as Monte Carlo Tree Search and neural occupancy fields. This approach is proving valuable in various fields, including robotics (improving navigation and manipulation), reinforcement learning (preventing reward hacking and enhancing model-based learning), and building energy management (predicting room occupancy for optimized energy use).