Lecture
Joint Lectures on Evolutionary Algorithms - June 2025
- Date
- Wednesday 18 June 2025
- Time
- Location
- Online
To interact or not to interact in Bayesian multi-objective optimisation
After a brief introduction to Bayesian optimisation, this talk will start by discussing the potential benefits of interactively eliciting preference information in multi-objective optimisation. We demonstrate how to find knees in multi-objective optimisation, not requiring any involvement of the decision maker. Then we show that if we can ask the decision maker only once, it may be beneficial to let the decision maker choose from an approximated frontier instead of the final solution set. Finally, we propose a way to estimate the value of more frequent user interaction, thereby allowing Bayesian optimisation to decide when it should ask the decision maker for input rather than following a pre-set interaction scheme.