A Cognitive Multi-Carrier Radar for Communication Interference Avoidance via Deep Reinforcement Learning

Abstract

Spectrum sharing between the radar and communication systems has become increasingly prevalent in recent years, therefore reducing the communication interference is a critical issue for radar. Deep reinforcement learning (DRL) based frequency allocation is a popular approach to solving the problem, especially in the highly dynamic spectrum. However, most DRL based methods suffer from low training efficiency due to the limited channel state information (CSI). To address the challenge, we propose a cognitive multi-carrier radar (CMCR), which acquires more CSI in one transmission and thus can learn the spectrum evolution faster. The frequency allocation problem for the CMCR is formulated as a partially observable Markov decision process which is hard to solve due to the combinatorial action space. To this end, we use the Iteratively Selecting approach along with the Proximal Policy Optimization (ISPPO) to solve it. To further enhance the performance of the CMCR in a short-term task, we pre-train the policy with model agnostic meta learning (MAML). Simulation results show that the CMCR learns fast and achieves an excellent detection ability in a congested spectrum on the basis of the ISPPO method. Besides, we also illustrate the efficiency of the MAML pre-training.