Uncertainty-Aware Agent Learning for Robust Decision-Making in Dynamic Environments

Abstract

This paper addresses the challenges of uncertainty modeling and robustness control in the perception, reasoning, and decision-making processes of agents operating in complex and dynamic environments. A unified uncertainty-aware agent framework is proposed, which integrates probabilistic state representation, uncertainty estimation, and risk optimization within a unified modeling system. By jointly modeling multi-source noise, feature missing, and observation occlusion, the framework enables robust perception and adaptive decision-making in dynamic environments. Methodologically, the model is built upon probabilistic inference to perform distributed modeling of latent states and incorporates information bottleneck and mutual information constraints to achieve a balanced trade-off between feature compression and semantic selection. At the policy level, the algorithm performs risk-based policy optimization through confidence distribution constraints, allowing the agent to adaptively balance exploration and robustness in dynamic environments. During training, structural regularization and temporal consistency constraints are combined to ensure continuity and interpretability of features and policies over time. Experimental results demonstrate that the proposed algorithm exhibits superior robustness and generalization under various uncertainty conditions, maintaining high average returns and success rates while significantly reducing calibration error and tail risk in noisy, incomplete, or occluded settings. This study provides a systematic modeling approach and practical pathway for uncertainty-driven agent decision-making, offering important theoretical and application value for autonomous systems, reinforcement learning optimization, and intelligent control in complex dynamic environments.