Integrating unsupervised learning and deep reinforcement learning for anti-jamming in backscatter-aided 6G communications
Abstract
This paper proposes a leakage-free unsupervised learning-aided D3QN framework for anti-jamming control in backscatter-aided 6G IoT communications. The proposed method first maps physical-layer observations, including SINR, RSSI, and estimated interference, into semantic states using unsupervised clustering, and then employs D3QN to jointly optimize transmission mode, transmit power, and reflection coefficient. To avoid information leakage, the scaler and clustering model are fitted only on training observations and kept fixed during evaluation. Simulation results over multiple independent runs show that the proposed framework converges within about 150-180 episodes and achieves approximately 3.2 bps/Hz average throughput and 0.65 bits/J energy efficiency. Under strong jamming, it improves throughput by about 30% over D3QN without clustering and more than 50% over DQN. Robustness tests under abrupt switching, unseen jammer patterns, noisy observations, and delayed feedback further confirm its effectiveness under challenging interference conditions.