With the proliferation of deep learning techniques for wireless communication, several works have adopted learning-based approaches to solve the channel estimation problem. While these methods are usually promoted for their computational efficiency at inference time, their use is restricted to specific stationary training settings in terms of communication system parameters, e.g., signal-to-noise ratio (SNR) and coherence time. Therefore, the performance of these learning-based solutions will degrade when the models are tested on different settings than the ones used for training. This motivates our work in which we investigate continual supervised learning (CL) to mitigate the shortcomings of the current approaches. In particular, we design a set of channel estimation tasks wherein we vary different parameters of the channel model. We focus on Gauss-Markov Rayleigh fading channel estimation to assess the impact of non-stationarity on performance in terms of the mean square error (MSE) criterion. We study a selection of state-of-the-art CL methods and we showcase empirically the importance of catastrophic forgetting in continuously evolving channel settings. Our results demonstrate that the CL algorithms can improve the interference performance in two channel estimation tasks governed by changes in the SNR level and coherence time.
翻译:随着无线通信深层学习技术的激增,一些工作采用了基于学习的方法来解决频道估算问题。虽然这些方法通常在推论时间的计算效率方面得到推广,但其使用仅限于通信系统参数方面的具体固定培训环境,例如信号对噪音比率和一致性时间。因此,这些基于学习的解决方案的性能将在模型在与培训不同的环境中测试时降低。这促使我们开展工作,调查持续监督的学习(CL)以缓解当前方法的缺陷。特别是,我们设计了一套频道估算任务,其中我们改变了频道模型的不同参数。我们侧重于对高斯-马尔科夫·雷利格频道的淡化估计,以评估非常态对平均平方差(MSE)标准方面业绩的影响。我们研究了最先进的CL方法的选择,并用经验展示了灾难性的遗忘在不断演变的频道环境中的重要性。我们的成果表明,CL算法可以改进受SNR水平和时间一致性变化制约的两种频道估算任务中的干扰性表现。