Seminar Details
2006-11-14 (14.00 pm) : "Control performance criteria in data-based control design"
At Euler building
Organized by Mathematical Engineering
Speaker :
Bazanella, Alexandre
Abstract :
Control performance criteria are a key element in control systems theory.
Not only they are fundamental from a conceptual point of view but this concept also
leads to a large variety of design methods, which are formulated as optimization problems.
For model-based design methods, the relevant properties of such criteria (convexity, uniqueness of minima, etc)
have been largely explored - although not always explicitly. On the other hand, it has been found that a limiting
factor in the application of data-based control design methods (such as IFT, but not restricted to it) is the convergence
to local minima of the criterion. Mainly - but not exclusively - motivated by that, we have been studying the "shape" of
different control performance criteria: model following, reference tracking, minimum variance, cheap control, etc, as
well as their combinations. We have determined conditions under which the criteria are "well-behaved". These conditions
involve mainly the data set, the initial condition for the optimization and the performance criterion itself. The results indicate
how to design the experiments and/or initialize the optimization so that it will succeed. They also suggest the use of cautious
control in the way to achieve the desired performance.