Final Report Summary - CDMAN (Control of Spatially Distributed Complex Multi-Agent Networks)
One of the most remarkable trends of the modern era is the rapid growth in connectivity of society and technology. Accompanying this is an equally fast growth in complexity of the systems we rely upon, resulting in a whole new set of challenges that engineers must overcome in the analysis of large dynamic networks and the design of systems to work or interact with them. This helps explain the recent boom in research on networked systems and control theory in applications ranging from distributed sensing and robotics to epidemic control and human decision making. Control design for networked multi-agent systems traditionally involves solving optimization problems, sometimes centralized and other times distributed, but very often with a shared global objective. However, when the agents have different or even competing objectives, as is often the case, each agent must take into account the actions of its competitors and single-objective optimization methods fail. Through the ERC starting grant project, CDMAN, some major progresses have been made for introducing control actions into a large population of autonomous agents interacting with each other organized in networks. The idea of influencing leaders and thus influencing potentially the whole population has proven not just theoretically feasible, but practically effective in teams of autonomous robots. For theoretical development, the research team has looked into how network topologies and individual agents' dynamics affect the overall collective behavior of the complex multi-agent networks. It is found that when the network topologies belong to some special classes, e.g. trees and regular graphs, one can construct explicit, closed-form bounds on the number of "controlled" leader agents in order to achieve the complete control of the network-level global dynamical behavior. The team has looked into agent dynamics that are of an evolutionary nature, and popular in theoretical biology and sociology study. The team has provided some surprisingly insightful global convergence analysis that indicates that it is possible to predict over long period of time, where the collective dynamics of a large population of interacting agents will evolve to and in which way. Such findings have greatly improved the known results and opened up the new ground for designing controlled leaders to affect all the other agents as followers when the agents evolve according to their interactions over time with their peers. For control scientists and engineers, these results facilitate the study of timely and challenging issues related to social, economic, and biological sciences from a control theoretic perspective. For experimental development, new design of teams of autonomous robots has been implemented. The testbed's trajectory tracking and localization system is completely redeveloped through real time data association. The resulted distributed cooperative control based on local sensing has become robust against the noisy surrounding of the testbed.
http://www.rug.nl/staff/m.cao
http://www.rug.nl/staff/m.cao