Achieving Consensus Among Autonomous Dynamic Agents using Control Laws that Maintain Performance as Network Size Increases

点击翻译按钮获取中文摘要

Recent advances in automation and robotics have created a pressing need for new "protocols," that is, for algorithms or control laws that allow teams of multiple autonomous agents to cooperate and accomplish complex tasks. Unfortunately, many of the best protocols for multi-agent coordination problems suffer from scalability issues, that is, while they perform well when the number of agents is small or moderate, their performance degrades sharply as the number of agents in the network grows. This project will develop new control laws for a range of multi-agent problems whose performance is maintained even as network size becomes very large. A number of tasks with broad practical importance will be considered, including optimal distribution of limited resources among agents, cooperative tracking and estimation, and adaptive positioning for optimal sensing. With these new protocols, large groups of autonomous agents(such as mobile robots or unpiloted aerial vehicles) will be able to quickly accomplish a number of useful and important tasks. These advances are needed to allow emerging technologies for autonomous vehicles and other networked autonomous systems to realize their potential economic and societal benefits.The main technical contribution will be to speed up a widely-used class of nearest neighbor interactions. It is common to optimize a global objective in multi-agent control by means of local updates that interleave the maximization local objectives with consensus terms that effectively couple these objectives. This project will develop techniques to speed up such consensus-like updates. By a judicious combination of weight-selection and extrapolation by each agent, the convergence time of consensus updates will be improved by one or several orders of magnitude. These speedups further imply quick convergence times for a number of multi-agent problems relying on consensus-like updates. The techniques applied mix recent advances from algebraic graph theory, optimization, switched dynamical systems, and the joint spectral radius.