Distributionally Robust Control and Incentives with Safety and Risk Constraints

具有安全和风险约束的分布式鲁棒控制和激励

• 批准号: 1708906
• 负责人: Mihailo Jovanovic
• 金额: $ 31.89万
• 依托单位: University of Southern California
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2017
• 资助国家: 美国
• 起止时间: 2017-08-01 至 2022-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1708906&HistoricalAwards=false
• 关键词: Distributionally; Robust; Control; Incentives; Safety

Massive data collected from the Internet-of-Things and cyber-physical systems can have transformative impacts on our society, spanning from personalized medicine to urban infrastructure systems. However, several concerns related to robustness, safety, risk, and reliability have been raised centered on how to incorporate such large-scale data into solving critical decision-making problems, as the data and the estimated statistical models are often inaccurate. Thus, the proposed research will establish a control-theoretic foundation to resolve this issue by allowing distributional errors in the statistical models and by developing control strategies that are robust against the errors. The potential application domains of the proposed distributionally robust control tools include battery management systems, power grids, food supply chains, manufacturing systems and personalized medicine. With the successful implementations of the proposed control tools in such domains, we will be able to improve individual safety and quality of life, and the reliability of data-driven control systems, which would build high confidence in society. The research outcomes in this project will also be used for (i) the USC Chevron Frontiers in Energy Research Summer Camp which is one of our K-12 STEM outreach efforts; (ii) USC Women in Science and Engineering programs that provide hands-on research experiences to undergraduate and (iii) open house and workshops in Viterbi Center for Engineering Diversity to train and recruit educationally-disadvantaged underrepresented students. The overarching goal of our proposed research is to develop theoretical foundations and computational methods for distributionally robust control problems associated with safety-critical and/or non-cooperative systems that operate with limited information. The proposed tools can contribute to the following three fundamental areas: 1. Stochastic control theory: The proposed research aims to establish a game theoretical and algorithmic foundation of distributionally robust control methods for nonlinear stochastic systems when faced with ambiguous distributional information about uncertain variables. In particular, we will investigate a duality-based dynamic programming solution to alleviate the infinite-dimensionality issue in the control problem and combine it with (deep) neural network- and occupation measure based methods to systematically adjust computational complexity and solution accuracy.2. Safety and risk aware control theory: We will extend stochastic reachability analysis methods to cases with imperfect information about the probability distribution of disturbances. This distributionally robust reachability tool will be used to specify the worst-case probability that the system fails to stay in the safe range and the worst-case risk of system loss. Based on the safety and risk specifications, we will then propose a systematic approach to synthesize a safety preserving and risk-aware control law that is robust against disturbance distributional ambiguity. 3. Incentive contract theory: Incentive contracts under moral hazard can be used to coordinate noncooperative sub-systems controlled by local agents in which local control actions and uncertain variables cannot be monitored by a central coordinator. To broaden the applicability of the contracts to engineering and socio-technical problems, we will generalize the theory in two directions: (i) integrating engineering systems with nontrivial dynamics into contracts, and (ii) constructing incentive contracts in a distributionally robust fashion.

从物联网和网络物理系统收集的大量数据可以对我们的社会产生变革性影响，从个性化医疗到城市基础设施系统。然而，由于数据和估计的统计模型往往不准确，如何将如此大规模的数据纳入解决关键决策问题，引起了与鲁棒性、安全性、风险和可靠性相关的一些担忧。因此，所提出的研究将建立一个控制理论基础，通过允许统计模型中的分布误差并开发对误差具有鲁棒性的控制策略来解决这个问题。所提出的分布式鲁棒控制工具的潜在应用领域包括电池管理系统、电网、食品供应链、制造系统和个性化医疗。随着所提出的控制工具在这些领域的成功实施，我们将能够提高个人安全和生活质量以及数据驱动控制系统的可靠性，这将在社会中建立高度的信心。该项目的研究成果还将用于 (i) 南加州大学雪佛龙能源研究前沿夏令营，这是我们 K-12 STEM 外展工作之一； (ii) 南加州大学科学与工程领域的女性项目，为本科生提供实践研究经验；(iii) 在维特比工程多样性中心举办开放日和研讨会，以培训和招收教育弱势学生。我们提出的研究的总体目标是为与信息有限的安全关键和/或非合作系统相关的分布式鲁棒控制问题开发理论基础和计算方法。所提出的工具可以为以下三个基本领域做出贡献： 1.随机控制理论：所提出的研究旨在为非线性随机系统在面对不确定变量的模糊分布信息时建立分布鲁棒控制方法的博弈论和算法基础。特别是，我们将研究基于对偶性的动态规划解决方案，以缓解控制问题中的无限维问题，并将其与（深度）神经网络和基于占据测量的方法相结合，以系统地调整计算复杂性和解决方案精度。2．安全和风险意识控制理论：我们将把随机可达性分析方法扩展到干扰概率分布信息不完善的情况。这种分布式鲁棒可达性工具将用于指定系统无法保持在安全范围内的最坏情况概率以及系统丢失的最坏情况风险。基于安全和风险规范，我们将提出一种系统方法来综合安全保护和风险意识控制律，该控制律对干扰分布模糊性具有鲁棒性。 3.激励契约理论：道德风险下的激励契约可用于协调由局部代理控制的非合作子系统，其中局部控制行为和不确定变量无法由中央协调器监控。为了扩大合同在工程和社会技术问题上的适用性，我们将从两个方向推广该理论：（i）将具有非平凡动力学的工程系统集成到合同中，以及（ii）以分布鲁棒的方式构建激励合同。

项目成果

Distributed robust statistical learning: Byzantine mirror descent

分布式鲁棒统计学习：拜占庭镜像下降

• 发表时间: 2019-12
• 期刊: 2019 IEEE Conference on Decision and Control (CDC
• 作者: Ding, D.;Wei, X.;Jovanovic, M. R.
• 通讯作者: Jovanovic, M. R.

Independent policy gradient for large-scale Markov potential games: sharper rates, function approximation, and game-agnostic convergence

大规模马尔可夫势博弈的独立策略梯度：更锐利的速率、函数逼近和与博弈无关的收敛

• 发表时间: 2022-01
• 期刊: Proceedings of the 39th International Conference on Machine Learning
• 作者: D. Ding and C.

Distributionally robust stochastic control with conic confidence sets

具有圆锥置信集的分布鲁棒随机控制

• DOI: 10.1109/cdc.2017.8264292
• 发表时间: 2017-12
• 期刊: 2017 IEEE 56th Annual Conference on Decision and Control (CDC
• 作者: Yang; Insoon
• 通讯作者: Insoon

Performance of noisy higher-order accelerated gradient flow dynamics for strongly convex quadratic optimization problems

强凸二次优化问题的噪声高阶加速梯度流动力学性能

• 发表时间: 2023-07
• 期刊: 2023 American Control Conference
• 作者: Samantha Samuelson;Hesameddin Mohammadi;Mihailo R. Jovanovic
• 通讯作者: Mihailo R. Jovanovic

Policy gradient primal-dual mirror descent for constrained MDPs with large state spaces

具有大状态空间的受限 MDP 的策略梯度原始双镜像下降

• DOI: 10.1109/cdc51059.2022.9992419
• 发表时间: 2022-12-06
• 期刊: 2022 IEEE 61st Conference on Decision and Control (CDC)
• 作者: Dongsheng Ding;M. Jovanović
• 通讯作者: M. Jovanović