Collaborative Research: MLWiNS: Deep Neural Networks Meet Physical Layer Communications -- Learning with Knowledge of Structure

机器学习技术有望成为未来无线网络性能提升的主要推动力。在实践中，在通信系统中应用学习算法的困难来自方法论的范式转变：从传统的基于模型的分析解决方案到基于数据驱动的经验-结果的方法。虽然学习算法在使用计算资源方面提供了更好的灵活性和效率，但在性能、鲁棒性和安全性方面缺乏保证，在应用于工程设计时提出了一些问题。该项目的目标是将基于学习的数据驱动方法与传统的基于模型的分析方法相结合，设计新的学习算法，利用工程系统的结构知识，从而获得两种方法的优势。该项目侧重于无线网络的物理层，其中通常存在分析解决方案、明确定义的性能基准以及明显孤立的建模缺陷。他们的主要挑战是尽可能地利用分析结果，并将学习能力集中在传统方法失败的系统部分，包括非线性、非高斯和非平稳元素。本文确定了测试这种新的混合设计方法的三个问题领域，包括1)MIMO衰落干扰信道的符号检测，2)具有低分辨率模数转换器的大规模MIMO，以及3)非线性无线电信道中的MIMO- ofdm波形分析。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Machine learning techniques are expected to be the main driving force for performance gains in future wireless networks. In practice, the difficulty in applying learning algorithms in communication systems comes from a paradigm shift in the methodology: From conventional model-based analytical solutions to data-driven empirical-result-based approaches. While learning algorithms offer better flexibility and efficiency in using computational resources, the lack of guarantees in the performance, robustness and security raise issues when adopting into engineering designs. The goal of this project is to combine learning-based data-driven approaches with conventional model-based analysis, to design novel learning algorithms that can utilize the structural knowledge of engineering systems, thus reaping the advantages of both methods.This project focuses on the physical layer of wireless networks, where analytical solutions, clearly defined performance benchmarks, as well as clearly isolated modeling deficiencies are usually present. They key challenge is to utilize the analytical results whenever possible, and focus the learning power on those parts of the system where the conventional approach fails, including non-linear, non-Gaussian, and non-stationary elements. Three problem areas are identified to test such new hybrid design methods, including 1) symbol-detection for MIMO fading interference channels, 2) massive MIMO with low resolution analog-to-digital converters, and 3) MIMO-OFDM waveform analysis in non-linear radio channels.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.