Adaptive CQI and RI Estimation for 5G NR: A Shallow Reinforcement Learning Approach

In this paper, we study the problem of estimating rank indicator (RI) and channel quality indicator (CQI) under 5G new radio (NR) specification. The main challenge for estimating RI and CQI comes from the fact that their estimation result is affected by many factors, e.g., delay spread, Doppler spread, etc. We use a shallow reinforcement learning technique to estimate the expected spectral efficiency for different RIs and CQIs. In particular, Q-learning is applied to solve the problem with a neural network consisting of a single hidden layer. To update the neural network output when the channel behavior changes, we develop an online adaptation (OA) algorithm working on top of the trained neural network. As a result, the combined algorithm adaptively controls the reported RI and CQI values to satisfy a target block error rate (BLER). We show through simulation results that the proposed algorithm adapts to different channel conditions and maintains the BLER to be around a target value of 10%. The proposed algorithm also significantly outperforms a scheme of fixed CQI and fixed RI when channel varies slowly.

I. Introduction

Sending accurate channel state information (CSI) feedback from user equipment (UE) to next generation Node B (gNB) is essential to help better utilize the available resources. In order to maximize throughput at the UE, a gNB allocates the resources for future transmissions in accordance with CSI feedback. According to the 3rd generation partnership project (3GPP) 5G new radio (NR) specification [1], a CSI report consists of multiple indicators, e.g., channel quality indicator (CQI) and rank indicator (RI). CQI indicates the type of modulation and code rate for the physical downlink shared channel (PDSCH) transmission, while RI indicates the rank for that transmission. In this work, we study the estimation of both RI and CQI using a shallow reinforcement learning (RL) technique. Without loss of generality, this algorithm can be applied to estimate either RI or CQI only as well, when the other is already given or fixed to some value.