Venue, date, and registration

The workshop will be held on June 13th, 2023, at the European Control Conference 2023 in Bucharest, Romania. Venue and registration details here. The workshop is organized in the framework of the project ELO-X.

Workshop length

One full day

Speakers and Organizers

Prof. Lorenzo Fagiano 

Prof. Fredy Ruiz

Dr. Lorenzo Sabug, Jr.

Politecnico di Milano, Italy

Intended audience

The workshop is intended for both Ph.D. students and experienced researchers from both academia and industry. The first part will set the main fundamental concepts needed to fully get familiar with the SMGO approach, its strengths and limits.

Highlights

• A tailored, seamless 1-day experience to efficiently and effectively get acquainted with approaches to simultaneously learn and search for global solutions to general black-box problems from data;
• The morning session describes the main fundamental challenges and methods, and introduces a new algorithm to deal with the considered problem class – SMGO, Set Membership Global Optimization;
• The main afternoon session is hands-on and interactive: the participants will be challenged with solving constrained black-box problems with SMGO and other methods (e.g., Bayesian Optimization), thus having the possibility to compare their performance and merits;
• The last session points out current research directions and open challenges.

Overview

Black-box optimization problems arise in many practical scenarios, such as engineering design, management of complex processes, and control systems tuning. A common aspect of such problems is that the objective function is difficult to model mathematically in an accurate way and is rather evaluated using (possibly time-consuming) experiments or detailed simulations. Typically, these problems also feature several local minima, and they may be affected by uncertainty/disturbances, so that the value of the cost may change within certain intervals for the same value of the decision variables. These aspects call for global black-box optimization techniques, aiming to solve these problems by attempting to simultaneously learn the objective function and optimize it. Examples include swarm-based approaches [1], Mesh Adaptive Direct Search (MADS) [2-4], kriging-based methods [5-8] and Bayesian optimization (BO) [9,10]. Almost all such techniques try to balance exploitation and exploration, to achieve fast convergence and coverage of the search space, respectively. Moreover, many black-box problems also involve the satisfaction of one or more constraints, deriving for example from the fulfillment of norms and standards, the achievement of a minimum performance threshold, or state constraints. In the literature, most black-box optimization approaches assume that the constraint functions are known a priori and either analytically or numerically tractable, such as upper and lower bounds on each decision variables, convex sets, or nonlinear but known and differentiable constraint functions. However, there are problems where the constraint values are also black boxes, leading to uncertain and possibly disconnected feasible sets. Only a few approaches can deal with such problems.

In the described context, the objectives of this workshop are: