ME Seminar: Fresh Perspectives to Classical Problems in Identificatio...
  • FENS
  • ME Seminar: Fresh Perspectives to Classical Problems in Identificatio...

You are here


Speaker: Suat Gümüşsoy

Title: Fresh Perspectives to Classical Problems in Identification & Control and Connections with Reinforcement Learning

Date/Time: March 12, 2018  /  12:40-13:30

Place: FENS G035


Abstract: This talk presents new solution approaches on three different domains: identification, control, and a practical application on multicore architectures. First, we introduce a new black and grey box transfer function estimation algorithm using frequency response measurements. This algorithm is recently integrated to tfest function in the MATLAB 2016b System Identification Toolbox with state-of-the-art speed-up and accuracy. We briefly mention its applications to many problems in identification and control.


Second, we bridge the gap between optimal control and time delay systems. Stability of delay systems is an active research area, yet there are few results on classical optimal control such as linear quadratic regulator (LQR). This makes delay systems difficult to control in practical applications. By identifying the structure of optimal LQR control law, we provide the existence and uniqueness conditions and analytical solution of underlying Lyapunov functional equation. This allows to solve the corresponding Riccati equation for delay systems iteratively and opens the door to optimal control field.

Third, we illustrate application of classical and modern control techniques to manage power and temperature dynamics in multicore architectures. We present complete stability analysis of this power/temperature dynamics and conditions under which the power-temperature trajectory converges to a stable fixed point. By computing this point at run-time with small error, we ensure thermally safe operation and guard against thermal threats.

Finally, we discuss how to apply my research on system identification and control to reinforcement learning. We also talk about applications of reinforcement learning to multicore architectures and wearable technologies. We conclude the talk with my current and near future research plans.


Contact: Bekir Bediz