Resources Allocation for Drones Tracking Utilizing Agent-Based Proximity Policy Optimization

This paper presents a reinforcement learning agent-based model that works by incorporating the MESA environment with the Stone Soup radar systems simulator. In particular, the Proximity Policy Optimization (PPO) reinforcement algorithm is used to discover a policy for sensor selection that results in optimum sensor resource allocation. In this work, one radar and one camera collaborate to generate tracks of a drone. The sequential measurements processing method is used to merge the inputs from the camera and the radar. An Extended Kalman Filter (EKF) is used to estimate the track. The learned system can apply sensor allocation online, works in real-time, and selects one radar and one camera at a time without having to reevaluate a cost function at every time step. Comparisons are done with the straight minimum entropy method and the random selection baseline method. The work demonstrates how machine learning techniques can capture resource allocation policy and help avoid the complexity of having to re-calculate cost function at every time step, especially when we have many radars and many cameras.