Enhancing Multi-Cell Dynamic TDD with Multi-Agent Deep Reinforcement Learning

Dynamic Time Division Duplex (D-TDD) is an important feature in 5G and future 6G networks. It allows flexible allocation of Uplink (UL) and Downlink (DL) slots. This helps to manage traffic demands dynamically. However, two key challenges exist. First, the system determines the best TDD pattern to match user traffic. Second, cross-link interference occurs when different cells use different TDD configurations. This interference degrades network performance. The 3GPP standard does not provide an optimal method for TDD configuration. It does not solve cross-link interference issues. To address these gaps, we proposed a Multi-Agent Deep Reinforcement Learning (MADRL) approach. This approach models the TDD problem as a linear programming problem. Introduced the Multi-Agent Deep Reinforcement Learning-based 5G RAN TDD Pattern (MADRP) framework. This method is decentralized. Each cell has an independent agent that learns the best TDD configuration. The system reduces control latency and signaling overhead. The MADRP model monitors the buffer states of uplink and downlink data. It exchanges messages with neighboring cells to minimize cross-link interference. Each agent uses reinforcement learning to determine the best TDD allocation. The model adapts to traffic variations and prevents buffer overflows. It highlights the limitations of MADRP. Performance is degraded in high-interference environments. Future work will focus on implementing MADRP in real-world 5G systems. This aimed to integrate the model with OpenAirInterface (OAI) to demonstrate real-time adaptability. This will provide insights into practical deployment challenges. This research introduces a novel DRL-based TDD adaptation approach. It efficiently manages UL and DL allocation while minimizing cross-link interference. The method enhances performance in multi-cell 5G environments. It provides a scalable and effective alternative to static TDD configurations.