Career: Learning for Strategic Interactions in Societal-Scale Cyber-Physical Systems

在社会规模的网络物理系统（SCPS）中，机器学习算法正日益成为利益相关者之间的接口——从拼车平台上的司机和乘客匹配，到电动汽车（EV）充电站的能源实时调度。这些系统中不同的利益相关者有不同的目标，这一事实导致了战略互动，这可能导致整个SCPS的效率低下和负外部性。NSF CAREER项目旨在发展对SCPS中出现的战略相互作用的基本理解，它们对社会福利的影响，以及它们如何影响算法决策。目标是改变工程师为SCPS设计算法的方式。目前，学习算法是孤立地训练和开发的，不确定性和战略互动（如果有的话）被视为对抗性或最坏情况。相比之下，拟议的研究旨在开发能够在决策时考虑经济相互作用、人类行为和不确定性的算法。通过本项目开发的理论和算法将在两个物理测试台上进行验证：1。一个电动汽车充电测试平台，司机经常错误地报告他们对更快充电的偏好；加州理工学院社会科学实验实验室将在这里进行控制实验，以了解人们对算法的反应。该提案还包括一项综合教育和外展计划，其中包括向K-12学生和关于SCPS学习复杂性的新本科和研究生课程进行外展。该项目的主要目标包括开发一种统一的设计方法，用于在SCPS中存在战略行为的情况下学习，以及系统地研究个人用户和政策制定者可以使用的控制行动和控制权力，以实现社会目标。SCPS中的战略操纵是通过（错误的）数据报告或算法决策来实现的，这一事实将这些问题与博弈论和经济学中经典研究的问题区分开来。此外，与计算机科学和经济学中研究战略互动的现有工作相比，该项目旨在从动态角度看待SCPS，它利用动态系统理论和随机过程的工具和思想来补充机器学习、博弈论和行为经济学的思想。这一观点将允许对重复互动如何影响SCPS中的战略决策以及哪些设计决策影响博弈论设置中的学习提供新的见解。这为以前被忽视的控制旋钮在SCPS中实现社会目标的新见解和分析打开了大门。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。

In societal-scale cyber-physical systems (SCPS), machine learning algorithms are increasingly becoming the interface between stakeholders---from matching drivers and riders on ride-sharing platforms to the real-time scheduling of energy resources in electric vehicle (EV) charging stations. The fact that the different stakeholders in these systems have different objectives gives rise to strategic interactions which can result in inefficiencies and negative externalities across the SCPS. This NSF CAREER project seeks to develop a foundational understanding of the strategic interactions that arise in SCPS, the impacts they have on social welfare, and how they affect algorithmic decision-making. The goal is to shift how engineers design algorithms for SCPS. Currently, learning algorithms are trained and developed in isolation, and uncertainty and strategic interactions are treated---if at all--- as adversarial or worst-case. In contrast, the proposed research aims to develop algorithms that can consider economic interactions, human behavior, and uncertainty when making decisions. The theory and algorithms developed through this project will be validated on two physical testbeds: 1. an EV charging testbed where drivers routinely mis-report preferences for faster charging, and 2. the Caltech Social Science Experimental Laboratory where controlled experiments will be conducted to understand how people respond to algorithms. The proposal also includes an integrated education and outreach plan, which includes outreach to K-12 students and new undergraduate and graduate courses on the complexities of learning in SCPS.Key goals of this project include developing a unified design methodology for learning in the presence of strategic behaviors in SCPS and the systematic study of the control actions and control authority that individual users and policymakers can wield to achieve societal goals. The fact that strategic manipulations in SCPS are played out through the (mis)-reporting of data or through algorithmic decision-making distinguishes these problems from those classically studied in game theory and economics. Furthermore, in contrast with existing work in computer science and economics that study strategic interactions, this project aims to take a dynamic view of SCPS, which leverages tools and ideas from dynamical systems theory and stochastic processes to complement ideas in machine learning, game theory, and behavioral economics. This perspective will allow for new insights into how repeated interactions affect strategic decision-making in SCPS and which design decisions impact learning in game theoretic settings. This opens the door to new insights and the analysis of previously overlooked control knobs for achieving societal goals in SCPS.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.