Control Architecture Discovery and Multi-Objective Reinforcement Learning

Dr. Suat Gumussoy
Siemens Technology
Princeton, New Jersey

Abstract: This talk overviews our research agenda on learning & control, modeling and autonomous systems. While connections and synergies among these topics are highlighted, two important directions are focused: 1) distributed control architecture discovery for fixed-topology autonomous systems and 2) a novel algorithm on multi-objective reinforcement learning (MORL) for generative engineering design. Due to growing size of autonomous systems, it is not feasible to connect all systems with each other. On the other hand, the existing control and communication architecture is either determined by locality or expert knowledge. We present our search algorithm starting (optionally) from an initial architecture and performing smart exploration to find better architectures improving the desired performance objectives. We validate our approach on benchmark examples. Generative engineering design has been popular in industrial design problems aiming to find the “surprising design” for industrial autonomous system. The goal is to choose the optimal design among many design configurations by factoring both cyber and physical aspects. We present our MORL approach towards this direction where we compute a single neural network policy approximating the Pareto front solution of the design space. This allows the designer to traverse through different design points and adapt the optimal one for changing conditions. Finally, we conclude our talk by discussing possible extensions of these two topics and connections of other relevant topics to autonomous systems and generative engineering design.

Biography: Dr. Gumussoy is a senior key expert on data-driven control at Autonomous Systems & Control group at Siemens Technology in Princeton, NJ. His general research interests are learning, control, autonomous systems with particular focus on reinforcement learning, data-driven modeling and control, and development of commercial engineering design tools. Dr. Gumussoy received his B.S. degrees in Electrical & Electronics Engineering and Mathematics from Middle East Technical University at Turkey in 1999 and M.S., Ph.D. degrees in Electrical and Computer Engineering from The Ohio State University at USA in 2001 and 2004. He worked as a system engineer in defense industry (2005-2007) and he was a postdoctoral associate in Computer Science Department at KU Leuven (2008-2011). He was a principal control system engineer in Controls & Identification Team at MathWorks where his contributions ranges from state-of-the-art numerical algorithms to comprehensive analysis & design tools in Control System, Robust Control, System Identification and Reinforcement Learning Toolboxes. He is a Senior Member of IEEE since 2020 and a member of IFAC Technical Committee on Linear Control Systems since 2022. He served as an Associate Editor in IEEE Transactions on Control Systems Technology (2018-2022) and IEEE Conference Editorial Board (2018-2022).