9:18 a.m.
(On behalf of Loris Roveda)
Dear all,
Friday May 6 at 10:30 in room B1.11 (Campus Est) we will have a seminar
"Surrogate-based methods for black-box and preference-based optimization: theory and
a controller tuning case study" from Davide Previtali (Università di Bergamo).
Please find the abstract and biography below.
See you there,
Loris
___________________________________________________
TITLE: Surrogate-based methods for black-box and preference-based optimization: theory and
a controller tuning case study
ABSTRACT: Many engineering and machine learning applications require the calibration of
some parameters. For example, in the context of control systems, there is often the need
to tune the controller parameters to achieve adequate closed-loop performances. Quite
often, the tuning is found by minimizing a suitable performance indicator, such as the
settling time of the closed-loop step response or the Integral Time Absolute Error (ITAE).
Usually, the mathematical relationship between the performance indicator and the
controller parameters is unknown; instead, the metric can only be computed from the
signals obtained by performing experiments on the system. Further complications arise if
no performance indicator is available and, instead, the quality of a tuning is assessed by
a decision-maker, who compares different calibrations for the controller parameters and
chooses the one which he/she prefers the most.
In this seminar, we address how to solve optimization problems where (i) no analytical
formulation of the objective function is available, but it can be measured by performing
experiments, or (ii) we need to find the optimal parameters’ calibration based only on the
preferences expressed by a decision-maker. The former case is referred to as black-box
optimization (BBO) while the latter is called preference-based optimization (PBO). We show
how BBO and PBO problems can be solved using surrogate-based optimization algorithms and
address their convergence to the global optimizer. Finally, we present a case study where
we apply black-box and preference-based optimization to tune the controller of a forming
press.
BIO: Davide Previtali was born in Bergamo, Italy in 1995. He received his M. Sc. degree
cum laude in Computer Science Engineering from Università degli Studi di Bergamo (Italy)
in 2019. Since October 2019, he is a a Ph.D student in Engineering and Applied Sciences at
the Control systems and Automation Laboratory of Università degli Studi di Bergamo (CAL
UNIBG). His research interests include black-box and preference-based optimization as well
as the application of adaptive control schemes in the industrial setting.