Deep Reinforcement Learning Based Dynamic Sub-Band Full Duplex for 5G-Advanced and 6G

Sub-band full duplex (SBFD) is a novel duplexing scheme that enables concurrent uplink (UL) and downlink (DL) transmission on non-overlapping frequency resources within an unpaired carrier. Earlier studies indicated that to unleash the full potential gain of SBFD, the radio frame configuration must be adapted to traffic load condition and network characteristics. This paper proposes a centralized deep reinforcement learning framework, based on soft actor-critic for discrete action setting (SAC-Discrete), with a modified prioritized replay mechanism, to train a model that dynamically determines the best frame configuration based on cells' buffer statuses. This method aims to maximize the geometric mean of user throughput, ensuring fairness across all users in both link directions. Simulation results demonstrate that our proposed solution improves UL performance by maximizing the geometric mean of user throughput under varying offered loads.