LOLLA: Deep Reinforcement Learning for Closed-Loop Link Adaptation Towards a GPU-Accelerated AI-RAN

Conventional Outer-Loop Link Adaptation (OLLA) in 5G NR relies on a simple, reactive staircase mechanism based on single-bit HARQ feedback. This approach struggles in high-mobility environments where channel conditions change rapidly. The authors investigate whether a deep reinforcement learning (DRL) agent, conditioned on rich physical and MAC layer telemetry, can outperform OLLA while maintaining standards-compliant Modulation and Coding Scheme (MCS) selection.

The authors propose LOLLA (Learned Outer-Loop Link Adaptation), which treats link adaptation as a Constrained Markov Decision Process (CMDP). Instead of replacing the entire MCS selection process, LOLLA uses a residual learning formulation: the agent outputs a continuous SINR offset that modulates the existing, standards-compliant SINR-to-MCS lookup table. This ensures the system retains a robust structural prior. The policy is trained using Proximal Policy Optimization (PPO) with a Lagrangian constraint to enforce specific Block Error Rate (BLER) targets automatically, eliminating the need for manual penalty tuning. The framework is implemented as a real-time dApp on a GPU-accelerated 5G NR stack, achieving end-to-end control latencies under 500 microseconds.

LOLLA demonstrates significant performance improvements over conventional OLLA across various 3GPP TDL channel models. Throughput gains range from 15% to 92%, with the most substantial improvements occurring at higher Doppler frequencies (up to 400 Hz). The learned policy successfully generalizes to unseen channel models and scales to support up to eight concurrent UEs under shared-resource scheduling. Furthermore, the Lagrangian formulation allows the system to strictly satisfy diverse reliability targets (1% to 15% BLER) without requiring retraining or manual calibration of penalty parameters.