Domain-Adapted Reward Estimation in DRL-based Anti-Jamming Communications

Deep reinforcement learning for dynamic frequency selection is widely used in the field of anti-jamming communications, and its performance relies heavily on the reward. In most existing works, the reward is calculated via the signal-to-interference-plus-noise ratio (SINR), which, however, is not available in practical scenarios. To address this challenge, a novel domain adaption approach for reward estimation is proposed, where the source domain is automatically labelled with simulations under known jamming patterns and the target domain is assumed to be totally unlabeled with unknown jamming patterns. In particular, an asymmetric tri-training method for reward estimation is proposed for transferring from the totally controllable source domain to the unknown target domain. Simulation results show that proposed method achieves close-to-ideal performance in the environment with unknown jamming patterns.