Belief-Rewards to Learn Radar Scheduling Policies

Beam-agile radars have the potential to learn scheduling policies requiring only a holistic belief constructed from past observations. Rewarding improvements to belief directly enables an agent to learn policies to track known targets and search for unknown targets, without explicitly rewarding this behavior. In this paper, we model a radar controller as an information-gathering agent in a Belief Markov Decision Process. Using a Reinforcement Learning solver to develop a Monte Carlo Tree Search policy, actions are chosen considering only belief and a model estimating the improvement in belief-reality distance. Importantly, our approach is agnostic to the number of targets and adapts to dynamic environments. We present preliminary results as proof of our underlying concept.