无已知信道下基于残差辅助的通信系统端到端学习
摘要
利用强大的深度学习技术,通信系统的端到端(E2E)学习能够优于经典通信系统。遗憾的是,在无已知信道的情况下,该通信系统无法由深度学习训练。为解决此问题,近期提出了一种基于生成对抗网络(GAN)的训练方案以模拟真实信道。然而,GAN 的梯度消失与过拟合问题会导致通信系统 E2E 学习的严重性能退化。为缓解这两个问题,本文提出一种基于残差辅助 GAN(RA-GAN)的训练方案。具体而言,受残差学习思想启发,我们提出一种残差生成器,通过实现更鲁棒的梯度反向传播来缓解梯度消失问题。此外,为应对过拟合问题,我们通过添加正则化项重构训练损失函数,以限制 RA-GAN 的表示能力。仿真结果表明,训练后的残差生成器比传统生成器具有更好的生成性能,且所提基于 RA-GAN 的训练方案在理论信道模型和基于射线追踪的信道数据集上均以可忽略的计算复杂度增加实现了近最优块错误率(BLER)性能。
引用
@article{arxiv.2102.10786,
title = {Residual-Aided End-to-End Learning of Communication System without Known Channel},
author = {Hao Jiang and Shuangkaisheng Bi and Linglong Dai and Hao Wang and Jiankun Zhang},
journal= {arXiv preprint arXiv:2102.10786},
year = {2022}
}
备注
This paper has been accepted by IEEE TCCN. To solve the gradient vanishing and overfitting problems of the GAN-based training scheme in E2E system, this paper proposes a residual aided GAN (RA-GAN) based training scheme. Simulation codes will be provided to reproduce the results presented in this paper after publication: http://oa.ee.tsinghua.edu.cn/dailinglong/publications/publications.html