Hierarchical Attention and Policy Regularization for Context-Aware Agent Decision-Making

Abstract

This study addresses the challenges of contextual fragmentation, temporal instability, and limited policy generalization in agent perception, reasoning, and decision-making under complex environments. A multi-level context-aware agent behavior algorithm is proposed, which integrates three key mechanisms-hierarchical feature encoding, cross-layer attention aggregation, and policy regularization optimization-within a unified framework to achieve deep semantic coupling from environmental perception to behavior generation. Specifically, the hierarchical encoder extracts environmental features at multiple scales to jointly represent local dynamics and global semantics. The cross-layer attention mechanism selectively aggregates contextual information across layers, allowing the model to adaptively capture key semantic dependencies through dynamic information flow. Furthermore, a policy optimization module based on state updating and distributional regularization ensures stable and consistent strategy outputs in non-stationary environments. Experimental evaluations demonstrate that the proposed method achieves significant improvements across multiple metrics, outperforming existing approaches in contextual generalization, multi-scale consistency, and policy robustness. These results confirm that multi-level contextual modeling effectively enhances the agent's environmental understanding and behavioral coordination, enabling stable and reliable decision-making in complex dynamic scenarios. This research provides a theoretical foundation and technical pathway for multi-level semantic modeling and context-integrated decision-making, laying the groundwork for robust behavior learning in high-dimensional and non-stationary environments.