July 6, 2012, 2pm, A.05.001
by Michael Emmerich, Leiden University
Indicator-based multiobjective optimization seeks to find a set of Pareto optimal solutions that has a high performance in a pre-defined performance indicator, for instance the hypervolume measure. A common choice is the hypervolume indicator. The usage of a unary indicator based methods a to generalize paradigms from single objective optimization to set-oriented multiobjective optimization.
In optimization with computational expensive black-box function it is common to use surrogate models to identify promising solutions for further evaluation. These surrogate models are typically learned from past evaluations. For instance Gaussian Processes are surrogate models that are based on Bayesian statistics and not only make it possible to approximate function values but also to assess the uncertainty of fuction value approximations. This talk will give an overview on existing work and recent trends on how Gaussian processes can be combined with indicator based multiobjective optimization in order to design highly efficient algorithms for Pareto front approximation.