Link Adaptation in 5G Networks : Reinforcement Learning Framework based Approach

Abstract: Link Adaptation is a core feature introduced in gNodeB (gNB) for Adaptive Modulation and Coding (AMC) scheme in new generation cellular networks. The main purpose of this is to correct the estimated Signal-to-Interference-plus-Noise ratio (SINR) at gNB and select the appropriate Modulation and Coding Scheme (MCS) so the User Equipment (UE) can decode the data successfully. Link adaptation is necessary for mobile communications because of the diverse wireless conditions of the channel due to mobility of users, interference, fading and shadowing effects, the estimated SINR will always be different from the actual value. The traditional link adaptation schemes like Outer Loop Link Adaptation (OLLA) improve the channel estimation by correcting the estimated SINR with some correction factor dependent on the Block Error Rate (BLER) target. But this scheme has a low convergence i.e., it takes several Transmission Time Intervals (TTIs) to adjust to the channel variations. Reinforcement Learning (RL) based framework is proposed to deal with this problem. Deep Deterministic Policy Gradient (DDPG) algorithm is selected as an agent and trained with several states of the channel variations to adapt to the changes. The trained model seems to show an increase in throughput for cell edge users of about 6-18% when compared to other baseline models. The mid-cell user throughput is increased up to 1-3%. This RL model trained is constrained with average BLER minimization and throughput maximization which makes the model perform well in different radio conditions.