Deep Reinforcement Learning on Graphs for Autonomous Search-and-Track

This paper presents a novel reinforcement learning (RL) approach for the joint search-and-track problem. The proposed approach models relationships between objects in the environment (agents, tracks, and search regions) as a heterogeneous graph. The graph representation is dynamically scalable in scenarios with a time-varying number of objects, largely sensoragnostic, and can be efficiently processed in parallel using a graph neural network (GNN). We use the graph representation to train a model-based RL agent which plans future actions in the continuous domain without discretizing its control space. We verify the performance of the RL approach in a mobile surveillance scenario and show that the RL agent significantly outperforms a set of baseline heuristic policies.