Multi-agent track confirmation utilising reinforcement learning and game theoretics

Abstract

This research investigates the problem of track conformation for a search and evade challenge within the context of the radar resource management domain. To analyse how agent collaboration affects the ability of multiple radar agents in confirming an evasive target, a high-fidelity radar simulation was designed. The simulated radar environment implements a realistic noise and clutter distribution and uses a generalised likelihood ratio test to make detections. The challenge is implemented as a limited information game with a highly evasive target attempting to reach one of four objective points before the track is confirmed by collaborative confirmation agents. Multiple Gaussian heuristic methods are compared with a reinforcement learning approach as the action selection agent. Three game theory strategies were also implemented: non-collaborative best response, collaborative best response, and leader-follower consensus. The results explore the effect of false alarms on confirmation performance and the impact of collaborative game theory applied to the agents versus isolated performance.