PROGRAM

Gabriele Pannocchia

Title : …
Session : ….
University of Pisa
https://people.unipi.it/gabriele_pannocchia/

Bio : ……………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………

Abstract : ……………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………….

Maciej Szymkat

Title : Numerical solution of dynamic optimization problems using systematic transformations of potentially optimal control structures
Session : …
ONT, Kraków,Poland
https://ont.com.pl

Bio : Maciej Szymkat graduated from AGH University of Science and Technology and Jagiellonian University. His Ph. D. thesis, written under supervision of Prof. Henryk Górecki, concerned stability criteria for time-delay systems. He is an author or editor of over 50 scientific publications. He wrote several books or book chapters in modeling, simulation and optimization of dynamical systems. He was involved in activities of Cracow Centre for Advanced Training in Information Engineering (CCATIE) resulting in publishing of 13 books series in 1995-2000. He co-organized a number of Polish and international conferences including the Computer Methods and Systems series (1995-2009, http://cms.agh.edu.pl), 23rd IFIP TC 7 Conference Krakow, Poland, 2007 (revised selected papers published in IFIP Advances in Information and Communication Technology, 109, ed. by A. Korytowski, K. Malanowski, W. Mitkowski and M. Szymkat, Springer, 2009), 16th French-German-Polish Conference on Optimization, Kraków, Poland, 2013 (extended plenary papers published in Advances in Mathematical Modeling, Optimization and Optimal Control, ed. by J-B. Hiriart-Uruty, A. Korytowski, H. Maurer and M. Szymkat, Springer, 2016). In recent years his research in colaboration with A. Korytowski concentrated on numerical and analytical methods of optimal control resulting in the development of the Monotone Structural Evolution (MSE) method.

Abstract : The paper begins with a brief review of various existing approaches to numerical solution of optimal control problems, starting from the symbolic pre-processing of problem description, through the definition of the preferred numerical setting, including the selection of collocation methods or ODE solvers, shooting formulations, and completed by an appropriate choice of the optimization algorithm. Special attention is paid to the representation of control discontinuities and the role of necessary optimality conditions of Pontryagin’s principle. The method of Monotone Structural Evolution (MSE, Szymkat, Korytowski, 2003, 2007, Korytowski, Szymkat, 2010, 2016, 2021), is a direct method for numerical solution of optimal control problems. In the MSE algorithm, controls are constructed by a concatenation of arcs of control procedures taken from a predefined finite set. The sequence of procedures composing a control is its structure. The length and contents of the control structure, the parameters of the procedures and the switching times are included in the sets of decision variables. The search for optimum consists of periods of gradient optimization, each period in a fixed decision space with constant dimension, separated by discrete changes of that space, called structure transformations. In each transformation, a new sequence of procedures composing the control is created in such a way that the control does not change as a function of time. In consequence, the cost functional monotonically decreases in the course of computations. Further, the practical use of the method is illustrated by a simple example exposing the rare property of the MSE that it can find the optimal control structure in a finite number of steps. In the following part, theoretical results on convergence of the MSE and its extensions for switched and state-constrained systems are discussed. The paper concludes with a more advanced example of application of the MSE method to optimal control problems for Hamiltonian systems where the spatiotemporal symmetries allow a reduction of the computational complexity.

Prof. Jerzy Baranowski

Beyond Gradients: Bayesian Approaches to Efficient Black-Box Optimization
Session : …
Laboratory of Computer Science in Control and Management of the Department of Automatic Control and Robotics, AGH University of Science and Technology in Krakow.
https://isz.agh.edu.pl
http://jerzybaranowski.pl

Bio : Prof. Jerzy Baranowski (M2012-SM2021) received his master’s degree, PhD, and DSc (Habilitation) in control engineering from the AGH University of Krakow in 2006, 2010 and 2017, respectively. In 2024, he was awarded the title of professor by the president of the Republic of Poland. He is currently a full professor of automatic control at the Department of Automatic Control and Robotics of AGH University in Krakow. He is the head of the Laboratory of Computer Science in the Control and Management department. He has authored over 200 publications, including books, journal papers and conference talks. He was awarded by the Polish Ministry of Science and Higher Education a stipend for the best young researchers and a medal of the National Education Committee. His research interests include the fields of process diagnostics, predictive maintenance, decision support, statistical inference, control theory and efficient methods of computation.

Abstract : Bayesian optimization provides a principled framework for the sequential optimization of expensive, black-box functions in the absence of analytic gradients. This talk reviews the underlying probabilistic inference mechanisms, with emphasis on Gaussian-process surrogate models and the construction of acquisition functions that mediate exploration and exploitation. Key implementation details—such as hyperparameter estimation, acquisition optimization, and stopping criteria—will be presented alongside extensions for noise handling, parallel evaluation, and multi-objective settings. Scalability challenges in high-dimensional spaces are addressed through sparse approximations, trust-region methods, and additive model structures. Representative applications in hyperparameter tuning, engineering design, and robotic control illustrate performance considerations and best practices. The session concludes with a survey of recent methodological advances and an outline of open problems in automated model selection and integration with deep learning.