Distributed Synchronous and Asynchronous Stochastic Optimization Algorithms over Networks

2019年，物联网（IoT）设备的安装数量达到266.6亿，到2025年，这一数字将达到750亿。这些设备收集了大量的数据集，对这些数据集的分析可以显著改善我们的日常生活。然而，如何有效地处理这些数据集仍然是一个挑战。首先，由于数据量大和隐私问题，不可能将所有数据传输到一个位置。其次，这些设备形成的网络是复杂的。因此，现有的分布式方法不能直接应用于这种场景，必须开发基于这些设备之间通信的新算法来理解这些大规模数据集。在这个项目中，PI将解决现有去中心化共识算法的主要缺点，并极大地提高大规模去中心化算法的效率和可扩展性。为了实现这一目标，PI将系统地研究去中心化算法的理论认识和两大挑战，即大规模数据和大规模网络。有三个目标。第一个目标是为现有的和新的分散确定性算法提供更好的收敛速度。这一目标的成功将是为今后两个目标奠定基础的第一步。第二个目标是为大规模数据开发具有方差减少的分散随机算法。最后一个目标是异步分散算法。该项目将为处理大规模分布式数据集的新研究铺平道路，并在很大程度上推动分散优化在各种应用领域的研究边界。这项研究将影响分散算法在无线传感器网络、机器学习、物联网和医疗保健等领域的使用。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。

The number of installed internet of things (IoT) devices reached 26.66 billion in 2019, and this number will reach 75 billion by 2025. These devices collect massive volumes of datasets, and the analysis of these datasets can significantly improve our daily lives. However, how to process these datasets efficiently is still challenging. First, it is impossible to transfer all the data to a location because of the large volume and privacy concerns. Second, the networks formed by these devices are complex. Thus, existing distributed methods can not be directly applied to this scenario, and novel algorithms based on the communication between these devices have to be developed to make sense of these large-scale datasets.In this project, the PI will tackle the major drawbacks of existing decentralized consensus algorithms and greatly promote the efficiency and scalability of large-scale decentralized algorithms. To achieve this goal, the PI will systematically investigate the theoretical understanding of decentralized algorithms and two major challenges, i.e., large-scale data and large-scale networks. There are three objectives. The first objective is a better convergence rate for existing, and new, decentralized deterministic algorithms. The success of this objective will be the first step that will form the foundation of the next two objectives. The second objective is to develop decentralized stochastic algorithms with variance reduction for large-scale data. The last objective is asynchronous decentralized algorithms. This project will pave the way for new research endeavors to deal with large-scale distributed datasets and largely push the research boundaries of decentralized optimization in various application domains. This research will impact the use of decentralized algorithms in fields including wireless sensor networks, machine learning, the internet of things, and healthcare.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.