Unified Structure-Aware Representation Learning for Agent Perception and Decision Making

Abstract

This paper proposes a structure-aware agent representation learning algorithm to address the limitations of agents in structural understanding and semantic modeling within complex dynamic environments. The algorithm centers on structural dependency modeling and constructs a topology-aware latent representation space that unifies state features, relational structures, and policy generation. Methodologically, the model first employs a structural graph encoding mechanism to extract temporal dependencies and spatial correlations from the environment, capturing multi-level contextual information. It then aligns local features with global structures through a semantic representation learning module, forming interpretable structural-semantic mappings in the feature space. During optimization, structural consistency constraints and representation regularization losses are introduced to ensure stable training and robust expression under noise interference, sparse feedback, and structural drift. Experimental results demonstrate that the proposed method achieves superior performance under multidimensional sensor noise, graph perturbations, and time-varying conditions, significantly improving semantic consistency, path planning, and policy convergence speed. Compared with existing methods, the proposed structure-aware framework attains higher consistency in structural information extraction and semantic fusion accuracy, showing strong generalization and interpretability, and providing a unified structured modeling paradigm for agent perception and decision-making in complex scenarios.