Dennis Bernstein, Jan 2014
Dennis Berstein and James Forbes from U. Michigan will be visiting Caltech on 29 January 2014. If you would like to meet with Dennis and Jim during their visit, please sign up below.
- 12:30 - Lunch and meeting with Richard
- 2:00 - Seminar, 121 Annenberg
- 3:00 - CDS tea
- 3:30 - Doug MacMartin
- 4:00 - Yilin Mo, 310 Annenberg
- 4:30 - Matanya Horowitz, 335 Annenberg
- 5:00 - Eric Wolff, 331 Annenberg
- 5:30 - Done for the day
How Much Modeling Information Is Really Needed for Feedback Control?
University of Michigan
Modeling for control is often expensive and time-consuming—not to mention futile, especially when a plant changes unpredictably. Our research is therefore aimed at the following fundamental question: What is the minimal modeling information (order, parameters, nonlinearities, noise spectra, etc.) that must be known—and how *well* must it be known—so that a controller can reliably meet performance specifications? The approach we are developing is based on retrospective cost adaptive control (RCAC), which uses retrospective optimization for online learning. RCAC is easy to implement, and requires extremely limited modeling information. In this talk I will explain the rationale for RCAC, its applicability to various types of plants (stable/unstable, minimum-phase/NMP, SISO/MIMO, linear/nonlinear), the modeling information it can operate with and (especially) without, and the status of its theoretical foundation. For flight control, we will apply RCAC to the extreme case of totally unknown control-surface faults, such as a stuck rudder or severe rate saturation. Additional examples are taken from missile control, noise and vibration control, and spacecraft attitude control with nonlinear actuation such as CMGs.