Cyber Valley Evening Colloquium on Autonomous Systems

Autonomous multi-agent systems are important for various applications towards a sustainable future. However, without an adequate theoretical foundation, multi-agent learning applications are not only hindered by missing interpretability, their safe and beneficial application for a sustainable future cannot be ensured. In this talk, I will give an overview of my research on the link between evolutionary game theory and reinforcement learning, i.e., replicator reinforcement learning (RRL) dynamics. RRL is a semi-formal method to study idealised multi-agent learning behaviour for improved interpretability. Focusing on the example of social dilemmas, I will show how the agents’ caring for the future parameter alone can turn a tragedy of the commons into a comedy, given a sufficiently severe environmental threat. Regarding the safe and beneficial application of autonomous multi-agent systems, I will show conditions under which an optimisation paradigm is neither sustainable nor safe and reinforcement learning is neither stable nor predictable. I will end with a brief outlook on future work.

Wolfram Barfuss is a postdoctoral research scientist working on collective learning for a sustainable future. He recently moved to the Tuebingen AI Center (Uni Tuebingen) and holds guest research positions at the Potsdam Institute for Climate Impact Research and Princeton University. Previously, he was a research fellow at the School of Mathematics (Uni Leeds) and the Max Planck Institute for Mathematics in the Sciences, Leipzig. He obtained his PhD from the Humboldt University of Berlin and the Potsdam Institute for Climate Impact Research in 2019.

Link to talk details

Transportation systems are undertaking rapid transformation, thanks to the advances of disruptive technologies such as connected and automated vehicles. My research aims to leverage these advances to address the challenges in transportation systems. First, I will present my work on improving multimodal traffic operations using the information provided by connected vehicles. In particular, I will develop a multi-scale and multimodal perimeter control approach for large-scale traffic networks in a connected vehicle environment. Second, I will exploit the benefits of automated vehicles and devise strategies to innovate traffic operations in mobility-on-demand systems in a scenario with mixed autonomy.

Dr. Kaidi Yang is currently a postdoctoral scholar with the Autonomous Systems Laboratory in the Department of Aeronautics and Astronautics at Stanford University. He obtained a Ph.D. degree from ETH Zurich in 2019, with a concentration on intelligent transportation systems. Before this, he received a BEng. in Automation, a BSc. in Pure and Applied Mathematics, and an MSc. in Control Science and Engineering from Tsinghua University. His main research interest lies in the design and operations of future mobility systems utilizing disruptive paradigms such as connected and automated vehicles, shared mobility, etc. He has published 11 papers in academic journals and made presentations at multiple international conferences, including a lectern presentation at International Symposium on Transportation and Traffic Theory (ISTTT). He is the recipient of two projects from the Swiss National Science Foundation (SNSF) Postdoc Mobility Fellowship. He has won the Best Paper Award for a paper published in 2019 on the journal Omega, the 2019 Chinese Government Award for Outstanding Self Finance Students Abroad, and the Best Student Paper (2nd place) in IEEE Intelligent Transportation Systems Conference. He serves as a Review Editor for Frontiers in Future Transportation.

Link to talk details

The human brain is one of the most complex autonomous systems on earth, with billions of neurons operating in parallel, giving rise to mental faculties such as reasoning, problem solving and the use of symbolic language. Despite the enormous capacity that the brain holds for parallel processing, humans are remarkably limited in the number of tasks they can execute simultaneously. Limitations in our ability to multitask are not only apparent in daily life. They are also universal assumptions of most general theories of human cognition. Yet, a rationale for why the human brain is subject to these constraints remains elusive. In this talk, I will draw on insights from neuroscience, psychology and machine learning to suggest that limitations in the brain’s ability to multitask result from a fundamental computational dilemma in neural architectures. Using a combination of graph-theoretic analysis, neural network simulation and behavioral experimentation, I will demonstrate that neural systems can face a tradeoff between learning efficacy, that is promoted through the shared use of neural representations across tasks, and multitasking capability, that is achieved through the separation of neural representations between tasks. This work suggests that it can be optimal for a neural system to sacrifice multitasking capability, to learn single tasks more quickly by sharing rather than separating representations between tasks. I will conclude by demonstrating that this tradeoff can explain a variety of behavioral and neural phenomena related to human multitasking and that it can inform the design of autonomous artificial agents tasked to navigate this tradeoff.

Sebastian Musslick is a Fellow of the Cognitive Science Program at Princeton University where he is about to receive his PhD degree in Quantitative and Computational Neuroscience. His research program involves the study of fundamental computational dilemmas in neural systems, to explain limitations of human and artificial cognition. Sebastian received his diploma in Psychology at the Technische Universität Dresden in 2014, as well as a Master’s degree in Neuroscience at Princeton University. During his diploma studies, he joined the University of Colorado in Boulder as a short-term research scholar from 2012 to 2013 where he developed biologically inspired models of human task switching performance. Sebastian amassed a strong record of publishing and collaborating with partners from academia and industry across disciplines, including psychology, neuroscience, mathematics, physics and computer science. In an effort to facilitate interdisciplinary exchange about the study of human and artificial cognition, he co-organized several international workshops and conferences, including the annual Conference on the Mathematical Theory of Deep Neural Networks and the annual Workshop on Mental Effort.

Link to talk details

Autonomous robots operating in complex and contact-rich environments are becoming essential for industry. These robots would benefit from a soft tactile skin that detects the location and strength of contacts over the entire robot's body surface. Implementing such tactile skin is challenging as the skin should seamlessly cover large and curved surfaces to monitor unexpected physical contacts. Human tactile skin efficiently resolves this challenge using an overlapped receptive field structure and cognitive processing. My research mimics this efficient feature by designing a piezoresistive structure with sparsely distributed electrodes and indirectly inferring contact locations and magnitudes from the electrodes. This approach considerably simplifies wholebody tactile sensor design and fabrication while achieving its contact sensing performance comparable to that of human skin. This talk will introduce key design components of a fabric-based tactile sensor and evaluation results, showing its potential toward efficient whole-body tactile skin applications.

Dr. Hyosang Lee is currently a research scientist at the Max Planck Institute for Intelligent Systems, Stuttgart, Germany. He received a B.S. in Mechanical Engineering from Korea University in 2010. His passion for robotics led him to obtain an M.S. in Robotics from Korea Advanced Institute of Science and Technologies (KAIST), Daejeon, South Korea, in 2012, and a Ph. D. in Mechanical Engineering from KAIST in 2017 where he developed soft tactile skin using piezoresistive nanocomposites. Moved by his interest in haptics, he joined Max Planck Institute for Intelligent Systems as a postdoctoral research fellow in 2017. Since then, he works on the development of a feasible whole-body tactile skin and its possible applications.

Link to talk details