Multimodality in Multi-Objective Optimization - More Boon than Bane?
Speaker: PD Dr.-Ing. Christian Grimme
Abstract: This talk addresses a new perspective on multimodality of multi-objective (MO) optimization landscapes. In Operations Research, we learn that local optima are traps for local opimizers. Interestingly, this fact is simply transfered to multi-objective optimization as certainty as well, although we do not have enough insights into multi-objective landscape characteristics. In our work, we use sophisticated visualization techniques, which rely on gradient field heatmaps and enable a perception of multi-objective landscapes for the first time. We show that local efficient sets in a multi-objective setting are not necessarily traps but can assist optimizers in finding global ecient sets. Finally, the simple MO local optimizer MOGSA is introduced, which exploits those observations by sliding down the multi-objective gradient hill and moving along the local efficient sets.
Short Bio: Christian is PostDoc Researcher at ERCIS and the Information Systems and Statistics chair at the University of Muenster. He received his PhD in computer science at TU Dortmund and completed his habilitation in information systems in December 2018, here at the WWU. Christian's research is focussed on data analytics, combinatorial multi-objective optimization, and scheduling theory, thus operations research. Additionally, he is principal investigator of the PropStop consortium for social media analytics and academic co-director of the ERCIS Competence Center on Social Media Analytics.