Keynote Speakers

 

Prof. Edward Keedwell
University of Exeter, UK

Title: Hyper-heuristics and Interactive Metaheuristics for Real-World Optimisation Problems
Abstract: Optimisation methods, such as evolutionary algorithms, have been successfully applied to a wide range of problems across many industrial sectors. Although they typically work well in these scenarios, performance and generalisability can be improved through the use of learning and domain expertise. In this talk, I will explore the development of optimisation methods augmented by machine learning (hyper-heuristics) and human expertise (interactive optimisation) to improve generalisability and feasibility of the developed solutions. The methods will be illustrated by describing their application to a number of real-world problems, including those in the water and transportation sectors, with a focus on sustainability.

Bio: Ed is a Professor of Artificial Intelligence at the University of Exeter. He has worked at the University for over 20 years and leads a research group working on applied AI, attracting over £5m in research funding and yielding over 200 journal and conference publications. His research interests focus on the development of new machine learning and optimisation methods applied to problems in the biosciences and engineering. He has worked with partners in academia and industry to develop traditional metaheuristic, human-in-the-loop and hyper-heuristic optimisation methods with application to operational research, water systems and decarbonisation challenges. He is a member of the Association of Computing Machinery and a Fellow of the Higher Education Academy.

Prof. Michael Herrmann
University of Edinburgh, UK

Title: Perspectives of Physics-Informed Neural Networks: Competitiveness, Versatility, and Interpretability
Abstract: PINNs promise a fascinating merger between exact theory and real-world data. Their growing popularity has lead to a wealth of experiences and a number of impressive successes, and within less than a decade of research in this field, virtually every method in machine learning has been studied in combination with PINNs. Adaptive and exploratory variants are becoming practically interesting, and some theoretical advances have been made so that limitations in regard to scalability and compatibility of physical and learning dynamics are well understood. In addition to giving an overview of the field, we highlight several application cases, in particular beyond the physics domain, to demonstrate that PINNs can be show their benefits even for inexact knowledge and sparse data. Another focus will be set on the combination of PINNs with deep reinforcement learning where we can show that PINNs form an efficiently solvable multi-objective learning problem. Finally, we discuss interpretability of PINNs in comparison to the interpretability of neural networks in general.

Bio: Dr. J. Michael Herrmann graduated from the University of Leipzig, and took postdoctoral positions in Copenhagen, Tokyo, and Göttingen. He is now a faculty member of the University of Edinburgh’s School of Informatics, Institute of Perception, Action, and Behaviour (IPAB). He has co-authored more than 200 scientific papers, the majority of which are outputs from projects he led as the Principal Investigator (PI) in domains, including machine learning, data science, robotics, metaheuristic optimisation, computational finance, cognitive systems, and computational neuroscience. His work on self-organised criticality, data analysis and data interpretation as well as control of robotic behaviour is of particular importance.