An Efficient DRL-Based Link Adaptation for Cellular Networks with Low Overhead

Link Adaptation (LA) that dynamically adjusts transmission parameters to accommodate time-varying channels is a critical technology in the Long-Term Evolution/New Radio system. Previous deep reinforcement learning (DRL)-based LA techniques directly select the Modulation and Coding Schemes (MCS) for each transmission. However, the frequent inference results in a high computational resource cost, leading to significant decision delays and potentially compromising the overall performance. To address these challenges, we present a new algorithm named TD3-OLLA, building upon separating the frequent selection of MCS from the time-consuming DRL model inference process. In particular, we introduce a two-level control framework that makes the real-time MCS selection using the traditional Outer Loop Link Adaptation (OLLA) algorithm and employs the DRL algorithm to tune OLLA's parameters. Our algorithm adopts the advanced Twin Delayed Deep Deterministic Policy Gradient (TD3) model and methods like classified experience replay to enhance the performance against rapidly changing link conditions. The simulation results demonstrate that TD3-OLLA achieves higher throughput than state-of-the-art LA techniques with an ultra-low overhead—the computation time is reduced by 80% compared to other DRL-based algorithms. Furthermore, it exhibits a high tolerance for model decision delays, making it well-suited for practical communication systems with strict latency requirements.