Multi-Agent Deep Reinforcement Learning to Enable Dynamic TDD in a Multi-Cell Environment

Dynamic Time Division Duplex (D-TDD) is a promising solution to address newly emerging 5G and 6G services characterized by asymmetric and dynamic uplink (UL) and downlink (DL) traffic demands. However, there are two major issues: (i) determining the TDD scheme (i.e., the number of slots devoted to UL and DL) to meet the dynamic traffic demands of the Users Equipment (UE); (ii) cross-link interference between cells that use different TDD schemes. The 3GPP standard neither specifies algorithms or solutions to derive the TDD configuration nor solves the cross-link interference. To fill this gap, we model the dynamic TDD problem in 5G NR as a linear programming problem. Then, we design Multi-Agent Deep Reinforcement Learning based 5G RAN TDD Pattern (MADRP), a fully decentralized solution based on the Multi-Agent Deep Reinforcement Learning (MADRL) approach. Based on the simulation results, the algorithm effectively prevents buffer overflows, avoids cross-link interference, and adapts to changes in the traffic pattern, ensuring its versatility. We compared our solution with the optimal solution and different static TDD configurations. We found that MADRP outperforms the static TDD configurations. We finally discuss the algorithm's limitations in terms of the number of cells, traffic variance, and cross-link interference probability.