Collaborative Research: CNS Core: Medium: Analytics and Online Optimization at Scale for Cellular Networks

蜂窝网络已经成为社会的关键基础设施之一，用户希望获得可靠的连接和性能。在幕后，运行这些网络需要更新数百个参数，时间范围从几个小时到几周不等，这对网络运营中心的工程师来说是极其昂贵和低效的。此外，当发生故障或执行效率低下时，检测和隔离根本原因也是一项具有挑战性的关键任务。这项提议的重点是有效地运行这些网络，并开发工具来检测异常，两者都使用机器学习技术。该方案的目标是开发基于在线学习、贝叶斯优化和深度学习的参数调整和异常检测算法。我们的算法和方法建立在与主要移动运营商的合作伙伴关系以及使用真实数据跟踪和试验台的基础上，具有现实世界的影响。利用无线理论、系统和机器学习领域的知识，开发样本高效的在线学习方法来优化单个蜂窝基站的多维调谐参数，然后应用转移学习来进一步支持多基站的分布式和协作性参数调谐。此外，还设计了用于蜂窝网络异常检测和诊断的深度压缩感知。这些推力是相辅相成的：异常检测有助于在参数调整期间提供安全检查，而从参数调整中获得的见解也有助于消除歧义和诊断异常。结合美国主要蜂窝网络、模拟和试验台实验的真实痕迹来验证设计。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Cellular networks have become one of the critical infrastructures for society, with users expecting reliable connectivity and performance. Behind the scenes, operating these networks require updating hundreds of parameters at time scales ranging from hours to weeks, which is extremely costly and inefficient for engineers at the network operations center. Further, when failures or inefficient performance occurs, detecting and isolating the root causes is again a challenging, but critical task. This proposal focuses on efficiently operating these networks and developing tools to detect anomalies, both using machine learning techniques. The goal of this proposal is to develop algorithms based on online learning, Bayesian optimization and deep learning for parameter tuning and anomaly detection. Building on partnerships with major cellular providers and the use of real data-traces and testbeds, our algorithms and approaches have real-world impact. The research outcomes are incorporated into the graduate and undergraduate curriculum.Using domain knowledge in wireless theory and systems and machine learning, this project develops sample-efficient online learning methods to optimize multi-dimensional tuning parameters in a single cellular base station, and then apply transfer learning to further support distributed and cooperative parameter tuning for multiple base stations. Moreover, it designs deep compressive sensing for anomaly detection and diagnosis in cellular networks. These thrusts are complementary to each other: anomaly detection helps to provide safety checking during parameter tuning while insights gained from parameter tuning also helps disambiguate and diagnose anomalies. A combination of real traces from a major US cellular network, simulation, and testbed experiments is used to validate the design.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.