CIF:Small:Collaborative Research:Distributed Fog Computing for Non-Convex Big-Data Analytics

在我们这个数据泛滥的时代，无处不在的传感器不断收集的大量信息，通过分布式计算架构进行交流和处理。为了解决新兴的大数据计算问题，该项目启动了一项雄心勃勃的多学科研究工作，旨在通过大规模分布式数据集数据分析的通用算法框架，推进最先进的网络/分布式大数据处理。所提出的算法框架能够在广泛的计算架构上对各种异构数据集进行完全分布式和并行的大数据分析。发展的研究方向也有利于远远超出大数据分析的领域，如信号处理、机器学习、下一代无线通信、智慧城市和智能电网网络。研究成果通过档案出版物、课程、本科生研究机会、教程和会议报告进行分发。所开发的方案依赖于一种新的凸化/分解技术，该技术适用于具有不可分离目标函数的丰富的非凸、非结构化和随机优化任务。算法是为数据分布在大量多核计算节点上的设置而设计的，这些节点位于具有（可能）时变甚至随机链接的任意拓扑网络中。这类新算法通过(i)完全控制处理器/网络节点之间计算/信号的并行度和分布，以及（ii）通过提供大量的凸近似、正则化条款、步长规则和通信协议，解决了当前（非并行和非分布）凸化技术的缺点。该框架是为时变甚至随机网络拓扑设计的，它还展示了分布式计算的另一个理想属性：对（随机）网络故障的弹性。

In our data-deluge era, massive chunks of information, perpetually collected by pervasive sensors, are communicated and processed by distributed computational architectures. To address emergent big-data computational issues, this project embarks on an ambitious multidisciplinary research effort that aims at advancing the state-of-the-art in-network/distributed big-data processing via a general algorithmic framework for data analytics over massively distributed data sets. The proposed algorithmic framework enables fully distributed and parallel big-data analytics, for a variety of heterogeneous data sets over a wide range of computational architectures. The developed research directions are beneficial also to domains far beyond big-data analytics, such as signal processing, machine learning, next-generation wireless communications, smart-city and smart-grid networks. Research results are distributed through archival publications, courses, undergraduate research opportunities, tutorials and conference presentations.The developed scheme relies on a novel convexification/decomposition technique which accommodates a rich class of non-convex, unstructured and stochastic optimization tasks with non-separable objective functions. Algorithms are designed for settings where data are distributed across a large number of multi-core computational nodes, within a network of arbitrary topology with (possibly) time-varying and even random links. This new class of algorithms addresses shortcomings of current (non-parallel and non-distributed) convexification techniques via (i) full control of the degree of parallelism and distribution of the computation/signaling among processors/network nodes, and (ii) by offering a plethora of convex approximants, regularization terms, step-size rules, and communication protocols. Designed for time-varying or even random network topologies, the advocated framework demonstrates also another desirable attribute for distributed computations: resiliency to (random) network failures.