Collaborative Research: CPS: Small: Co-Design of Prediction and Control Across Data Boundaries: Efficiency, Privacy, and Markets

如今，蜂窝网络和电网运营商测量了大量数据，这些数据可以为整个城市的交通和拥堵模式提供丰富的见解。与独立的外部实体（如出租车车队运营商）共享这些实时社会趋势，可以增强城市规模的资源分配和控制任务，例如电动出租车路线和电池存储优化。但是，富时间序列的所有者和外部控制机构必须跨越数据边界进行通信，这限制了他们可以共享的数据的范围和数量。该项目将开发新的算法和系统，以一种跨组织边界仅共享最小的任务相关数据的方式，共同压缩、匿名化和定价丰富的时间序列数据。通过强调通信效率，开发的算法将激励未来智慧城市的数据共享和协作。这项工作的关键动机是，今天的时间序列数据的表示是独立于最终控制任务而设计的，这通常会导致不必要的时间特征被发送，私有特征被揭示，最显著的趋势被低估。因此，该项目将开发一种统一的方法来共同设计丰富的时间序列数据的简洁，私有表示以及最终控制任务。在这里，协同设计意味着预测表示是在控制目标的更广泛背景下学习的，同时考虑到带宽限制、隐私、经济成本和数据处理的激励。这些算法将计算控制器对预测误差的敏感性，这些预测误差可能来自数据压缩、预测不确定性以及现代隐私工具注入的人工噪声。至关重要的是，控制器的灵敏度将反过来传递给网络运营商，以指导其优化和学习（例如，共同设计）一个简洁的、与任务相关的预测表示，该预测表示掩盖了私有属性，并自然地根据它们对控制的重要性对时间特征进行定价。例如，该研究将使运营商能够灵活地使用相同的底层蜂窝需求数据来强调出租车路线的高峰时段可变性，同时无缝地向移动视频流媒体公司提供细粒度的吞吐量预测，而不会泄露私人用户的移动性。最后，该项目的案例研究将被整合到德克萨斯大学奥斯汀分校和康奈尔大学的基于学习的控制课程中。更广泛的影响还包括外展和包容努力，以吸引来自历史上在STEM领域代表性不足的群体的学生。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。

Today, operators of cellular networks and electricity grids measure large volumes of data, which can provide rich insights into city-wide mobility and congestion patterns. Sharing such real-time societal trends with independent, external entities, such as a taxi fleet operator, can enhance city-scale resource allocation and control tasks, such as electric taxi routing and battery storage optimization. However, the owner of a rich time series and an external control authority must communicate across a data boundary, which limits the scope and volume of data they can share. This project will develop novel algorithms and systems to jointly compress, anonymize, and price rich time series data in a way that only shares minimal, task-relevant data across organizational boundaries. By emphasizing communication efficiency, the developed algorithms will incentivize data sharing and collaboration in future smart cities.The key motivation of this work is that today's representations of time series data are designed independently of an ultimate control task, which often causes unnecessary temporal features to be sent, private features to be revealed, and the most salient trends to be under-valued. Accordingly, this project will develop a unified approach to co-design succinct, private representations of rich time series data along with an ultimate control task. Here, co-design means that the forecast representation is learned within the broader context of a control objective while accounting for bandwidth constraints, privacy, and economic costs and incentives for data processing. The algorithms will compute a controller's sensitivity to prediction errors, which can arise from data compression, forecast uncertainty, as well as artificial noise injected by modern privacy tools. Crucially, the controller's sensitivity will in turn be relayed to a network operator to guide its optimization and learning (e.g., co-design) of a concise, task-relevant forecast representation that masks private attributes and naturally prices temporal features by their importance to control. The research will, for example, enable operators to flexibly use the same underlying cell demand data to emphasize peak-hour variability for taxi routing, while seamlessly delivering fine-grained throughput forecasts to a mobile video streaming company without revealing private user mobility. Finally, the case studies in this project will be integrated into courses on learning-based control at UT Austin and Cornell. Broader impacts also include outreach and inclusion efforts to engage students from groups that have historically been under-represented in STEM fields.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.