EARS: Utilizing Diverse Spectrum Bands in Cellular Networks - A Unified Information Learning and Decision Making Approach

在高数据速率移动服务需求飙升的推动下，以及监管和技术进步的推动下，蜂窝服务提供商正在增加或正在使用各种补充频段来增强他们自己的许可频谱，包括未经许可的频段、轻度许可的频段、辅助频段和高频频段。本项目研究如何有效地利用这些频段，特别是如何了解不同频段上的服务可用性和质量，以及如何在预算限制下有效地使用这些频段。为了应对这一挑战，研究人员提出了一个结合背景和预算的联合信息学习和决策框架。这个数学框架连接了两个重要但大多是独立研究的研究领域--信息学习和最优决策。这项研究为理解、设计和分析联合信息学习和决策算法提供了重要的见解。此外，由于这类问题的普遍性和重要性，所提出的方法可以应用于其他领域，如无线网络控制、众包、在线广告分配等。虽然是一种直观的方法，但由于预算限制引入了上下文之间和跨时间的耦合，因此设计联合学习和决策算法并分析其性能是具有挑战性的；信息学习和决策是紧密耦合的联合进化过程。该项目由三个主要推动力组成：1)总体框架：研究人员不仅争取对联合学习和决策的基本理解，而且争取具有实用简单性和理论性能保证的算法。这样的算法能够在蜂窝网络中有效地利用补充频谱；2)稀疏性和结构：诸如时间、位置和应用类型的上下文信息在选择适当的补充频带时提供有用的信息。然而，学习曲线(成本)随着情境-动作对的数量增加而增加。研究人员开发的算法可以通过利用领域知识、系统结构和压缩技术来加快学习速度；3)实际问题：当考虑实际无线网络中的动态频谱接入时，出现了实际问题。特别是，研究人员研究了系统统计数据随时间变化的非平稳系统，以及可以获得某些先验信息的热启动系统。

Driven by the skyrocketing demand for high data-rate mobile services and enabled by regulatory and technology advances, cellular service providers are augmenting or in the processing of augmenting their own licensed spectrum with a variety of supplemental bands, including unlicensed bands, lightly-licensed bands, secondary bands, and high frequency bands. This project studies how to effectively utilize such bands, in particular, how to learn the service availability and quality on different spectrum bands, and how to effectively use these bands under budget constraints. To address this challenge, the researchers propose a joint information learning and decision making framework with context and under budget. This mathematical framework connects two important yet mostly independently studied research areas, information learning and optimal decision making. The investigation provides important insights in understanding, designing, and analyzing joint information learning and decision making algorithms. In addition, because the generality and importance of such problems, the proposed approaches can be applied in other areas, such as wireless network control, crowd-sourcing, and online-ad allocation.While an intuitive approach, it is challenging to design the joint learning and decision algorithms and to analyze their performance because budget constraints introduce coupling among contexts and across time; and the information learning and decision making are closely coupled and jointly evolving processes. This project consists of three main thrusts: 1) General framework: The researchers strive for not only the fundamental understanding of joint learning and decision, but also algorithms with practical simplicity and theoretical performance guarantees. Such algorithms enable efficient supplemental spectrum utilization in cellular networks; 2) Sparsity and Structure: Context information, such as time, location, and application type, provides useful information in selecting the appropriate supplemental bands. However, the learning curve (cost) increases as the number of context-action pairs increases. The researchers develop algorithms that can accelerate the learning by exploiting the domain knowledge, system structure, and compressing techniques; and 3) Practical Issues: When considering the dynamic spectrum accessing in real wireless networks, practical issues arise. In particular, the researchers study non-stationary systems where the system statistics vary over time, and warm-start systems where certain prior information is available.