Collaborative research: Gaussian Process Frameworks for Modeling and Control of Stochastic Systems

不确定性下决策的定量模型继续吸引着自然科学和工程领域的大量研究。随着应用中越来越复杂的模型的出现，计算需求继续超过可行性，有效的数值方法的溢价仍然很高。研究人员将探索最新机器学习技术与控制范例之间的协同作用，这些范例在金融、能源存储和安全以及传染病的流行病学建模等多种应用中出现。开发的“智能”算法将为处理大规模/复杂设置的模拟提供必要的性能升级。该项目还将促进本科、研究生和博士后水平的数学科学跨学科培训。研究人员将研究用于非线性动态随机系统建模、分析和控制的统计学习技术。通过开发复杂随机模拟器的算法和统计模型，以及自主数据采集的主动学习策略，该项目将提高动态随机现象的数学分析能力和效率。该方法依赖于使用高保真的近似高斯过程替代物自适应分配计算资源，以便最大限度地提高建模目标的输入-输出关系的学习率或动态规划的输入-控制映射的学习率。通过将随机模拟与机器学习和非参数统计相结合，并与计算实现相结合，该项目将加强大规模模拟和优化设置中的知识发现。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。

Quantitative models for decision making under uncertainty continue to attract intense effort across natural sciences and engineering. With the advent of ever more sophisticated models in applications, computational demands continue to outpace what is feasible and the premium on efficient numerical approaches remains high. The investigators will explore synergies between the latest machine learning techniques and control paradigms, arising in applications as diverse as finance, energy storage and security, and the epidemiological modeling of infectious diseases. The developed "smart" algorithms will deliver performance upgrades essential for using simulations in tackling large-scale/complex settings. The project will also contribute to inter-disciplinary training in mathematical sciences across undergraduate, graduate and post-doctoral levels. The investigators will investigate statistical learning techniques for modeling, analysis and control of nonlinear dynamic stochastic systems. Through developing algorithms and statistical models for complex stochastic simulators, and active learning strategies for autonomous data acquisition, the project will achieve enhanced capabilities and efficiency in mathematical analysis of dynamic random phenomena. The approach hinges on the use of high fidelity approximate Gaussian Process surrogates to adaptively allocate computing resources in order to maximize the learning rate of the input-output relationship for modeling objectives or of the input-control map for dynamic programming. By connecting stochastic simulation with machine learning and non-parametric statistics, and integrating with the computational implementation, the project will enhance knowledge discovery in large-scale simulation and optimization settings.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.