Hashing for Lightweight Episodic Recall

We demonstrate a supplemental episodic memory system that can help arbitrary Soar agents use reinforcement learning in environments with hidden state. Our system watches for learning bottlenecks and then specializes the agent's existing rules by conditioning on recent history. Because we avoid a full episodic retrieval, performance scales well regardless of the agent's lifespan. Our approach is inspired by well established methods for dealing with hidden state.