CAREER: Accelerating Spatial Network Design: An Uncertainty-Driven Predict-and-Optimize Learning Framework

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。空间网络在自然界和人类社会中无处不在，例如交通网络、电网、食品供应网络和分子系统。空间网络的结构和配置决定了各自空间系统的重要属性。空间网络设计，即为预期结果设计空间网络结构和配置的问题，因此在许多领域都是迫切需要的。该项目将开发一个数据驱动的框架，可以实现快速和弹性的空间网络设计。该项目的独特之处在于，它将预测模型紧密集成到快速空间网络设计的优化算法中，同时考虑到系统固有的不确定性。该项目将有助于解决许多紧迫的社会挑战，例如优化交通网络以缓解拥堵，在人类流动网络上分发疫苗以遏制疾病传播，以及合成可产生环保材料的新分子。从技术上讲，该项目将开发一个“预测和优化”的学习框架，以实现快速和有弹性的空间网络设计。它将为此解决三个关键挑战。首先，它将通过建模复杂的时空依赖关系，同时捕捉系统固有的不确定性，为空间网络开发不确定性感知的深度预测模型。第二，将不确定性感知预测模型集成到优化生成算法中，有效搜索广阔的设计空间。第三，它将通过利用不确定性进行交互式数据收集和标签高效学习来解决空间网络设计中的数据稀缺问题。开发的工具将是开源的，并在各个领域传播空间网络设计问题。最后，该项目将培养下一代学生和劳动力，并促进数据科学教育的多样性。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。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Spatial networks are ubiquitous in nature and human society, examples include traffic networks, power grids, food supply networks, and molecular systems. The structures and configurations of spatial networks determine important properties of the respective spatial systems. Spatial network design, the problem of designing spatial network structures and configurations for desired outcomes, is thus in pressing need across many domains. This project will develop a data-driven framework that can achieve fast and resilient spatial network design. The uniqueness of the project is that it tightly integrates predictive models into optimization algorithms for fast spatial network design, while accounting for the inherent system uncertainty. The project will help address many pressing societal challenges, such as optimizing a traffic network to mitigate congestion, distributing vaccines over the human mobility network to contain disease spread, and synthesizing new molecules that lead to environment-friendly materials.Technically, this project will develop a "predict-and-optimize" learning framework to achieve fast and resilient spatial network design. It will address three key challenges to this end. First, it will develop uncertainty-aware deep predictive models for spatial networks by modeling complex spatiotemporal dependencies while capturing the inherent uncertainty of the system. Second, it will integrate uncertainty-aware predictive models into optimization and generation algorithms, to effectively search the vast design space. Third, it will address the data scarcity issue in spatial network design by leveraging uncertainty for interactive data collection and label-efficient learning. The developed tools will be open-sourced and disseminated for spatial network design problems in various domains. Finally, this project will train the next generation of students and workforce and also promote diversity in data science education.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.