Inference and Learning for Wireless Communications

该研究计划的目的是帮助理解基于贝叶斯理论的推理和学习算法的设计中的基本原则、限制和权衡，并强调它们在通信网络中的信息传输中的作用。 贝叶斯形式主义在于将所有相互作用的实体建模为随机变量。它被成功地应用于科学和工程的许多领域，近几十年来见证了贝叶斯推理的进步。特别是，使用图形描述变量之间关系的基于图形的模型获得了显著的恶名，导致了使用消息传递原理开发贝叶斯算法的良好结构的方法。使贝叶斯方法更接近实践的另一个步骤是变分贝叶斯技术，其显式地使用消息到预定义函数的因式分解，以简化在消息传递期间所需的边际化。 虽然贝叶斯框架是众所周知的，但必须在实践中使用近似方法，并且必须在了解应用环境的情况下设计和分析这些方法；在这里，这些方法是指无线网络中的信息传输。我们还将利用在机器学习或模式识别的一般领域取得的进展。然而，为了消除无线通信环境高度多变的特点对数据量的要求，我们将使用产生式模型方法。 该程序将重点解决两个主要问题：1)近似推理算法的分析和校正；2)从二元观测中学习。作为一种特殊场景的推理问题的研究，我们将以海量MIMO接收机为背景。学习问题将在a)确定点对点传输中的误差源；以及b)多用户无线网络中的干扰建模的情况下进行研究。我们将同时考虑模型选择策略的问题，这与严重依赖先前模型的问题2)有关。 这项研究计划将有助于知识创造，并将建立技术专业知识，并有助于培训学生在贝叶斯推理和学习与应用到无线通信领域。贝叶斯技术在许多应用科学领域的应用日益广泛，无线网络的持续发展也证明了这一点，这样的专业知识变得越来越重要，将使参与该项目的学生/实习生受益。

The objective of this research program is to contribute to the understanding of the fundamental principles, limitations, and trade-offs in the design of the Bayesian-theory based algorithms for inference and learning with emphasis on their role in transmission of information in communication networks. The Bayesian formalism consists in modelling of all interacting entities as random variables. It was successfully applied in many areas of science and engineering, and the recent decades witnessed a progress in Bayesian reasoning. In particular, the graph-based models which describe the relationship between the variables using graphs obtained significant notoriety leading to a well structured approach to the development of the Bayesian algorithms using the principle of message passing. Yet another step bringing the Bayesian methods closer to practice are the variational Bayesian techniques which explicitly use factorization of the messages into predefined functions to simplify the marginalization required during the message passing. While the Bayesian framework is well known, the approximate approaches must be used in practice and these must be devised and analyzed knowing the application context; here, the transmission of information in wireless networks. We will also leverage the progress made in the general area of machine learning or pattern recognition. However, to remove the requirement for large amount of data, which is impractical for highly variable environment characteristic of wireless communications, we will use to the generative-model approach. The program will focus on two main problems: 1) the analysis and correction of approximate inference algorithms, and 2) the learning from the binary observations. As a particular scenarios to study the problems of inference, we will use the context of massive MIMO receiver. The questions of learning will be study in the context of a) identification of the source of errors in point-to-point transmission; and b) modelling of interference in multiuser wireless networks. We will simultaneously consider the issue of model selection strategies, related to problems 2) which rely heavily on the prior models. This research program will contribute the knowledge creation, as well as, will build technical expertise and contribute to the training of students in the area of Bayesian inference and learning with application to wireless communication. As witnesses by the growing presence of Bayesian techniques in many areas of applied science, as well as persistent growth of wireless networks, such an expertise is increasingly important and will benefit the students/interns involved in the project.