Data-efficient methods for model learning and control
Abstract: In this talk, I will present an overview of the research I have been working on during my Ph.D. project, focused on data-efficient methods in machine learning. The first part will be dedicated to learning models of dynamic systems. Models facilitate simulations, analysis of the system's behavior, decision making, and design of automatic control algorithms. Even inherently model-free control techniques such as reinforcement learning have been shown to benefit from the use of models. However, obtaining informative data for constructing dynamic models can be difficult, especially when the models are to be learned during task execution. To this end, symbolic regression proves to be a suitable method to automatically build such models. This technique, based on genetic programming, constructs from data parsimonious models in the form of analytic equations. It represents an alternative to the currently popular data-hungry deep learning methods, which typically produce black-box models. One of the challenges in continual model learning is posed by the large amount of data collected from the system. I will present a comparison of methods for selecting informative training samples and show that symbolic regression can be used to construct accurate models from very small informative data sets. Furthermore, I will explain how symbolic regression can be naturally extended to account for physical constraints and a partially known theoretical or empirical model of the system. However, data-efficient methods are needed not only to learn models of the dynamic systems but also to learn models of the environment. Therefore, the talk will be concluded by presenting a method from the field of computer vision for robotics. Ubiquitous changes of the scene structure and appearance are typical for real-world dynamic environments and make many conventional localization and navigation methods fail. I will present a method for change detection based on weighted local visual features that improves the localization accuracy by distinguishing between stable parts of the scene and potentially confusing changing regions.
Short bio: Erik Derner received the M.Sc. degree in artificial intelligence and computer vision from the Czech Technical University (CTU) in Prague, where he is currently pursuing a Ph.D. degree. Throughout his studies, he acquired a broad international experience during five stays at universities in four different countries, ranging from a few months to the full academic year. His research interests are focused on sample-efficient methods for model learning applied to dynamic systems such as mobile robots and their environments. The main areas of research comprise genetic programming, reinforcement learning, and computer vision. The central topic in his research is the use of symbolic regression to automatically construct nonlinear models of dynamic systems.
Presenter: Erik Derner
Date: 2021-12-13 12:00 (CET)
Location: Laboratorio de Grados Politecnica I, Carretera San Vicente del Raspeig s/n, San Vicente del Raspeig 03690, Alicante ES