In unknown emergency environments, agents use range sensors for localization and mapping, making optimal decisions among numerous uncertain candidate target locations. Exploration graphs can significantly reduce state space dimensionality. By employing deep reinforcement learning models with graph neural networks to predict actions, the agents can autonomously explore unknown environments with varying scales of landmarks. However, the opacity of strategy learning and prediction complicates understanding temporal and spatial correlations of state transitions and state-action relationships. To address this, we propose a multidimensional visualization framework to analyze high-dimensional state-action vector correlations. A post-hoc explanation tool based on feature attribution was designed to create importance subgraphs, describing causal effects among nodes and explaining the agent’s decision-making. The framework’s effectiveness was validated using correlation analysis, node uncertainty entropy reduction, and navigation performance metrics.
