Broadband Anti-Jamming With Distributed Sensing and Deep Reinforcement Learning: Spectrum Compression and Reward Estimation

This article investigates the dynamic frequency selection problem in broadband anti-jamming communications through distributed sensing and deep reinforcement learning (DRL). In broadband anti-jamming scenarios, a single agent is often impractical for sensing the whole range of the band due to various restrictions on implementation. In this article, a novel distributed sensing architecture is proposed, where a number of distributed sensors are cooperated for sensing the whole band with each sensor only responsible for an individual sub-band. In this way, a fusion center should collect the sensing spectrum from each sub-band and the problem of spectrum compression is formulated under the framework of autoencoder. To cope with the problem of unavailable reward in distributed sensing, we propose a domain-adapted reward estimation method, with which the agent could work near perfectly under the framework of DRL. Simulation results show that the distributed sensing method could work effectively, and the compression ratio of the proposed autoencoder-based method could be 14 times higher than that of the direct quantization approach without any degradation on the final anti-jamming performance. In addition, the proposed DRL-based anti-jamming agent using reconstructed spectrum waterfall and estimated rewards achieves close-to-ideal performance in the environment with unknown jamming patterns.