Data-Driven Simultaneous Multibeam Power Allocation: When Multiple Targets Tracking Meets Deep Reinforcement Learning

In this article, we aim to study radar power allocation problem in multiple maneuvering targets tracking from a new perspective by deep learning, to develop a data-driven approach, which enables radar to learn a good (hopefully the best) power allocation policy from data. In order to enhance the utilization effectiveness of radar power resource, we present a cognitive design based on control framework incorporating an actor-critic variation of deep reinforcement learning called deep deterministic policy gradient. Specifically, at each tracking interval, the transmit power is allocated with control framework based on prior information, which is represented by the Fisher information matrix (FIM) of prior information of target's state computed using target's dynamic models. However, as targets are maneuvering, the dynamic models are always unavailable. Thanks to the sequential manner of target tracking, we consider to integrate a memory based on recurrent neutral network into the control framework for the estimation of FIM of prior information by characterizing the output as parameters for a high-dimensional Gaussian distribution of target's state instead of a deterministic state. Theoretical analysis has established that the derived covariance matrix is essentially the inverse of the FIM of prior information. Simulation results have shown the effectiveness of the proposed power allocation method.