The uplink of 5G networks allows selecting the transmit waveform between cyclic prefix orthogonal frequency division multiplexing (CP-OFDM) and discrete Fourier transform spread OFDM (DFT-S-OFDM), which is appealing for cell-edge users using high-frequency bands, since it shows a smaller peak-to-average power ratio, and allows a higher transmit power. Nevertheless, DFT-S-OFDM exhibits a higher block error rate (BLER) which complicates an optimal waveform selection. In this paper, we propose an intelligent waveform-switching mechanism based on deep reinforcement learning (DRL). In this proposal, a learning agent aims at maximizing a function built using available throughput percentiles in real networks. Said percentiles are weighted so as to improve the cell-edge users' service without dramatically reducing the cell average. Aggregated measurements of signal-to-noise ratio (SNR) and timing advance (TA), available in real networks, are used in the procedure. Results show that our proposed scheme greatly outperforms both metrics compared to classical approaches.

翻译：5G网络上行链路允许在循环前缀正交频分复用（CP-OFDM）与离散傅里叶变换扩频正交频分复用（DFT-S-OFDM）之间选择发射波形。由于DFT-S-OFDM具有更低的峰均功率比，并能支持更高的发射功率，这对使用高频段的蜂窝边缘用户极具吸引力。然而，DFT-S-OFDM表现出更高的误块率，这增加了最优波形选择的复杂性。本文提出一种基于深度强化学习的智能波形切换机制。在该方案中，学习代理旨在最大化一个基于实际网络中可用吞吐量百分位数构建的函数。通过加权处理这些百分位数，可在不明显降低小区平均性能的前提下改善边缘用户的服务质量。该方法利用了实际网络中可获取的信噪比与定时提前量的聚合测量值。结果表明，与传统方法相比，我们提出的方案在两项指标上均显著优于现有方案。