基于深度学习的宽带MIMO AMP迭代检测译码

Broad band multiple-input multiple-output (MIMO) transmission is no doubt a key element of the future wireless communications. Approximate message passing (AMP) iterative detection and decoding that makes the approximation of the maximum-likelihood detection performance with less computational load for broad band MIMO detection has gained significant attentions. However, traditional AMP iterative detection and decoding converges relatively slow and is not robust for imperfect channel state information. We would like to employ the deep learning idea to improve it, and construct an AMP-inspired deep network to accelerate its convergence to further reduce its computational load and make it more robust. Two key issues will be investigated. 1, We will investigate the equivalent model of AMP algorithm for learning, and investigate AMP-inspired deep network accordingly. 2, We will investigate the method to implement the AMP-inspired deep network into iterative detection and decoding for broad band MIMO transmissions. This work may offer additional possibilities to construct deep networks and find the robust detection method with less computational load for broad band MIMO transmissions.

宽带MIMO无线传输是未来移动通信系统的关键技术。迫近消息传递（AMP，approximate message passing）迭代检测译码是一种传输误码率逼近极大似然检测的低复杂度宽带MIMO检测方法。但现有的AMP迭代检测译码收敛速度慢，复杂度有待进一步降低，且误码率性能易受信道状态信息误差的影响，鲁棒性不好。本项目拟用深度学习优化AMP算法以构建深度神经网络，获取收敛速度快且鲁棒性好的AMP软输入软输出检测以及相应的迭代检测译码方法。主要研究内容包括：1、研究AMP算法深度学习等效模型，进而构建AMP深度神经网络。2、将AMP深度神经网络应用于宽带MIMO迭代检测译码。通过本项目的研究，可以拓展构建深度神经网络的思路；获取复杂度更低、鲁棒性更好的宽带MIMO检测方法。

迫近消息传递（AMP，Approximate Message Passing）是一种分布式消息传递算法，广泛应用于压缩感知和MIMO无线信号检测等领域。本项目利用机器学习优化AMP算法，主要研究内容包括，将AMP检测算法中迭代传递消息的过程构建为多层消息传递网络，并对部分消息设置可供学习的参数进行训练，从而找到发散概率低同时均方误差小的检测方法。利用在机器学习过程中得到的大数据，分析AMP算法在消息传递过程中的数学机制，在理论层面证明保证AMP算法收敛的条件。本项目极大推进了关于AMP算法收敛性的研究，在理论层面证明了一个必然导致AMP发散的条件，并以此为基础提出一种基于机器学习的改进AMP检测算法，其发散概率低至1e-3，基本达到实用水平，顺利完成研究的主要目标。研究的扩展目标，包括迭代检测译码方法和收敛的充分条件等问题有待进一步研究。