Special Seminar in Computing + Mathematical Sciences

Friday February 5, 2016 3:00 PM

Informational bottlenecks in control systems

Speaker: Gireeja Ranade, Microsoft Research, Redmond
Location: Annenberg 213

The trends set by the Internet of Things (IoT) are pushing us to understand the interactions between communication, computation and control in systems. This talk will discuss simple models that can provide insights on the "informational" limits on control of high-performance systems with uncertain system parameters that are difficult to track.

The talk will develop a novel notion of "control capacity," as the fundamental limit on the ability of a noisy actuator to control a system. Control capacity builds on the traditional notion of communication capacity in information theory. The models considered involve multiplicative noise on the control signals, and the notion of control capacity can be thought of as partially generalizing the traditional uncertainty threshold principle from control. 

 
The formulation provides an information-theoretic perspective that allows us to understand (and in some examples explicitly compute) the value of side information in the context of control systems. Some of the calculations are inspired by the well-studied dropped control problem.
 
If time permits, I will also discuss the stabilizability limits on systems with multiplicative observation noise. The converse bounds here use a non-standard approach to get around the analytical challenges posed by multiplicative noise. 

 

Contact: Sheila Shull at 626.395.4560 sheila@caltech.edu