Resource Allocation for Multi-Target Radar Tracking via Constrained Deep Reinforcement Learning

In this paper, multi-target tracking in a radar system is considered, and adaptive radar resource management is addressed. In particular, time management in tracking multiple maneuvering targets subject to budget constraints is studied with the goal to minimize the total tracking cost of all targets (or equivalently to maximize the tracking accuracies). The constrained optimization of the dwell time allocation to each target is addressed via deep Q-network (DQN) based reinforcement learning. In the proposed constrained deep reinforcement learning (CDRL) algorithm, both the parameters of the DQN and the dual variable are learned simultaneously. The proposed CDRL framework consists of two components, namely online CDRL and offline CDRL. Training a DQN in the deep reinforcement learning algorithm usually requires a large amount of data, which may not be available in a target tracking task due to the scarcity of measurements. We address this challenge by proposing an offline CDRL framework, in which the algorithm evolves in a virtual environment generated based on the current observations and prior knowledge of the environment. Simulation results show that both offline CDRL and online CDRL are critical for effective target tracking and resource utilization. Offline CDRL provides more training data to stabilize the learning process and the online component can sense the change in the environment and make the corresponding adaptation. Furthermore, a hybrid CDRL algorithm that combines offline CDRL and online CDRL is proposed to reduce the computational burden by performing offline CDRL only periodically to stabilize the training process of the online CDRL.