基于元学习的MIMO接收机学习方法设计

Design of MIMO Receiver Learning Method Based on Meta Learning

摘要: 基于数据驱动的接收机通常以提高训练成本为代价提升性能。本文对基于深度神经网络(Deep Neural Network, DNN)的多输入多输出(Multiple Input Multiple Output, MIMO)接收机的学习方式进行设计，旨在降低基于DNN的接收机的训练成本并提升性能。解决包括周期性发送导频信号导致网络被频繁的训练，即需要大量的训练数据和训练时间以及网络收敛慢等问题。根据元学习的学习方式，对系统进行基于任务的构建。提出了包括元训练、元适应和测试三阶段的接收机学习算法。并在训练的过程中对学习过程进行了优化，包括梯度简化和参数继承，提升了网络的收敛速度和泛化能力。在实验设计中与传统的接收机在不同信道模型下的性能测试结果表明所提出的学习算法在收敛性和系统性能上的提升。

Abstract: Data-driven receivers often improve performance at the expense of training. This paper designs the learning method of Multiple Input Multiple Output (MIMO) receiver based on Deep Neural Network (DNN), aiming to reduce the training cost and improve the performance for the receiver. Solve problems such as periodic transmission of pilots that cause the network to be trained frequently, require a large amount of training data, training time, and slow network convergence. According to the learning idea of meta-learning, the system model is constructed task-based. A receiver learning algorithm including meta-training, meta-adaptation and testing is proposed. In the process of training, the learning process is optimized, including gradient simplification and parameter inheritance, which improves the convergence speed and generalization ability of the network. For the experimental design, the simulation results of the traditional receiver under different channel models show that the proposed learning algorithm for the receiver improves the convergence and system performance.