Training Length Adaptation for Reinforcement Learning-Based Detection in Time-Varying Massive MIMO Systems With One-Bit ADCs

This study elucidates a reinforcement learning (RL)-based detection with one-bit analog-to-digital converters (ADCs) in time-varying massive multiple-input and multiple-output channels. In one-bit ADCs, conventional channel estimation exhibits poor performance owing to nonlinear quantization. The RL-based detection alleviates this degradation by learning the true likelihood probability during data transmission. However, in time-varying channels, the learned likelihood probability is inconsistent with the true likelihood probability due to temporal channel variations. This inconsistency can cause severe performance degradation. To effectively exploit the learned likelihood probability, we propose a training length adaptation method that determines an appropriate training length based on the channel conditions. To achieve this, we consider an optimization problem that minimizes the training length while guaranteeing the performance of RL-based detection. The solution of the optimization problem is obtained by an explicit form based on simple approximations, and it reveals that the optimal training length depends on the change in the likelihood probability. Simulation results demonstrate that the proposed method efficiently reduces the training length when a rapid change in likelihood probability is produced in fading channels. Moreover, this reduction contributes to improving the spectral efficiency by an avoiding undesirable learning process. Consequently, the spectral efficiency of the proposed method can be significantly increased compared to that of conventional RL-based detection. For instance, the proposed method achieves 1.76 times higher spectral efficiency than the conventional method at 30 km/h.