Active Inference for an Intelligent Agent in Autonomous Reconnaissance Missions

The researchers proposed an active reasoning path planning method for the autonomous control of agents, aiming to maintain a common understanding of the operational situation by reconnaissance of the geographical area. This method constructs an evidence map to reflect the current situational awareness, integrates sensor data collected over time, and updates it dynamically; it uses a generative model incorporating Dempster-Shafer theory and Gaussian sensor models to update the posterior probability distribution using Bayesian methods. The system calculates the variational free energy at all locations within the area by evaluating the divergence between the pignistic probability distribution of the evidence map and the posterior probability distribution of the target object, including the level of surprise caused by new observations. Based on this, the agent is guided in the simulation environment to move to incremental positions that minimize this value. This method effectively addresses the challenge of balancing exploration and utilization, enabling the agent to simultaneously search across a vast geographical area and track identified targets.