Collaborative Research: A Distributed Approximate Dynamic Programming Approach for Robust Adaptive Control of Multiscale Dynamical Systems

本项目将开发一种用于管理和控制复杂系统的新型通用多智能体系统，在面对非线性和复杂性的情况下，智能体以完全合作的方式一起工作，以最大化全局性能。作为证明新方法价值的试验台，他们将模拟以下挑战：(1)地震反应，这是一类发生速度极快、交货时间短、地理范围有限的灾害；(2)干旱饥荒救济，这类灾害的预测和提前期较长，地理范围较大。研究结果将被广泛传播，并将用于教育和推广项目，包括杜克大学本科生研究经验（REU）网站和nsf资助的无线智能传感器网络IGERT，以及暑期学校和国际合作伙伴关系。这里的关键挑战是开发一种新版本的自适应近似动态规划（ADP），它是完全分布式的，以解决多尺度动态系统的情况。这项工作建立在首席PI最近在分布式最优控制（DOC）方面的工作基础上，包括开发用于系统部分降维的最优限制算子，以及利用偏微分方程（PDE）和随机微分方程（SDE）领域的方法。

This project will develop a new type of general multiagent system for management and control of complex systems, in which the agents work together in a fully cooperative way to maximize global performance over time, in the face of nonlinearity and complexity. As a testbed to prove the value of the new approach, they will simulate the challenges of: (1) earthquake response, representing a class of disasters with very rapid occurrence, short lead times and restricted geographic extent; and (2) drought-induced famine relief, representing the class of disasters with longer forecast and lead times, and larger geographic extent. The results will be widely disseminated and will feed into programs for education and outreach, including a Research Experience for Undergraduates (REU) site at Duke and the NSF-funded IGERT on Wireless Intelligent Sensor Networks, feeding into summer schools and international partnerships.The key challenge here is to develop a new version of adaptive, approximate dynamic programming (ADP) which is fully distributed,to address the case of multiscale dynamical systems. The work builds on recent work of the lead PI on Distributed Optimal Control (DOC),and includes development of optimal restriction operators for dimensionality reduction in parts of the system, and exploitation of methods from the field of partial differential equations (PDE) and stochastic differential equations (SDE).