New RL Framework Improves Learning with Brain-Inspired Feature Fusion

Neuroscientific studies suggest that the brain processes complex behaviors by utilizing structured, low-dimensional representations and dynamically integrating multiple information sources through adaptive gating. Drawing inspiration from these biological mechanisms, a new reinforcement learning (RL) framework has been developed. This framework aims to disentangle features specific to environmental dynamics from those relevant to rewards. It employs locally linear embeddings (LLEs) to capture the inherent local linear structure often found in environments, mirroring the smoothness observed in neural activity. Simultaneously, reward-specific features are derived using standard RL objectives. A crucial component of this approach is an attention mechanism, analogous to cortical gating, which adaptively fuses these complementary representations on a per-state basis. Experimental evaluations on benchmark tasks demonstrate that this neuroscientifically-grounded method significantly enhances learning efficiency and overall performance compared to conventional RL techniques, underscoring the benefits of explicitly modeling local state structures and adaptive feature selection.