基于AI的信道估计的泛化性能提升方法

Approaches for Improving Generalization Capability of AI Based Channel Estimation

人工智能(Artificial Intelligence,AI)技术在未来无线通信中将发挥重要作用,其中信道估计是一个典型的AI与无线通信的结合点。基于AI的信道估计技术可以显著提高估计性能,尤其是针对低信噪比和非线性信道的估计问题。然而,基于AI的方案具有泛化能力不足的通病,尤其是在信道估计这种变化频繁、标签难获得的场景。针对泛化问题,提出了结合迁移学习、联合训练和模型无关的元学习(Model-Agnostic Meta-Learning,MAML)的基于AI的信道估计方案,并以信道场景变换为例验证了上述三种方案的泛化以及迁移性能。结果表明,相比于不做任何处理,三种方案均可以提高信道估计的泛化性能,且随着微调次数的增加,性能增益也会变大。其中,基于MAML的方案以最少的微调次数实现了最高的信道估计精度,是一种非常有潜力的训练方案。

Artificial intelligence(AI)technology will play an important role in future wireless communications,in which channel estimation is a typical combination of AI and wireless communication technologies.AI-based channel estimation techniques can significantly improve estimation accuracy,especially for cases with low signal to noise ratio(SNR)and nonlinear impacts.However,AI based schemes have the common problem of insufficient generalization capability,especially in the scenario where channel estimation is frequently changed and labeled data is difficult to obtain.To improve generalization performance,an AI-based channel estimation scheme combining transfer learning,joint training and model-agnostic meta-learning(MAML)is proposed.The generalization and migration performance of above three schemes are verified by differentiating the channel model of training and finetuning.Results show that,compared with the case of no finetuning,above three schemes can improve the generalization performance of channel estimation,and the performance gain will be larger with the increase of fine-tuning times.Among these three schemes,the MAML-based scheme achieves the highest channel estimation accuracy with the fewest fine-tuning times,which is a very promising training scheme.