Decentralized Control of Networked Systems

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

This project is focused on decentralized control of networked systems, including both problems of control of communication networks, and also control of spatially dis- tributed systems with specifications on their mutual interconnection. A key common feature to the problems under consideration is the mandatory decen- tralization in the control algorithms. Despite the well-known difficulties in imposing such information constraints to control systems design, we are motivated by recent successes in special problems, which indicate that perhaps the time has come for a more fruitful interac- tion between control and networking. In particular, in some of our recent work on Internet congestion control, we have been able to devise decentralized algorithms that have prov- able properties over arbitrary network topologies, overcoming some limitations of currently deployed TCP protocols. This work, which we plan to continue in this project, is already leading to prototype implementations which are being tested in high-speed networks. One of our main goals in this project is to expand this area of research to a number of other network resource allocation problems that involve decentralized feedback, such as multicast flow control and routing control in the Internet, and an even broader set of questions for wireless networks: power control, bandwidth allocation and their interaction with flow control. A second thrust of this project concerns a different kind of relationship between control and networks: namely, the decentralized control of a network of spatially distributed, mutu- ally interacting dynamical systems. We also build here on our previous work on distributed arrays, which identified a procedure that renders tractable decentralized control synthesis, at the expense of some conservatism. Our plan here is to extend these methods to control- ling networks of systems in more general topologies, such as vehicle formation problems, and also to combine this approach with model predictive control (MPC) running at each unit. The broader impacts of this project are of two kinds: one, judging from our recent successes in Internet congestion control and the ongoing experimental collaborations, there is a distinct possibility that this kind of alternative algorithms could be deployed, or at least influence the development of next-generation protocols; the impact of this would be substantial. Secondly, the interdisciplinary quality of the work described here can have great educational impact, not only for the students involved but also through more unified curricular offerings in control and networks. Indeed, the significant common ground between these and other related disciplines (optimization, game theory) makes this a very fertile area, and also one that reinvigorates the presence of Controls in Electrical Engineering departments such as our own. 1