Sequential memory improves sample and memory efficiency in episodic control

Abstract

Deep reinforcement learning algorithms are known for their sample inefficiency, requiring extensive episodes to reach optimal performance. Episodic reinforcement learning algorithms aim to overcome this issue by using extended memory systems to leverage past experiences. However, these memory augmentations are often used as mere buffers, from which isolated events are resampled for offline learning (for example, replay). In this Article, we introduce Sequential Episodic Control (SEC), a hippocampal-inspired model that stores entire event sequences in their temporal order and employs a sequential bias in their retrieval to guide actions. We evaluate SEC across various benchmarks from the Animal-AI testbed, demonstrating its superior performance and sample efficiency compared to several state-of-the-art models, including Model-Free Episodic Control, Deep Q-Network and Episodic Reinforcement Learning with Associative Memory. Our experiments show that SEC achieves higher rewards and faster policy convergence in tasks requiring memory and decision-making. Additionally, we investigate the effects of memory constraints and forgetting mechanisms, revealing that prioritized forgetting enhances both performance and policy stability. Further, ablation studies demonstrate the critical role of the sequential memory component in SEC. Finally, we discuss how fast, sequential hippocampal-like episodic memory systems could support both habit formation and deliberation in artificial and biological systems. Previous studies have explored the integration of episodic memory into reinforcement learning and control. Inspired by hippocampal memory, Freire et al. develop a model that improves learning speed and stability by storing experiences as sequences, demonstrating resilience and efficiency under memory constraints.