Multi Objective Regionalized Bayesian Optimization via Entropy Search
Keywords: Multi Objective Optimization, Bayesian Optimization
TL;DR: A novel approach using entropy selection to efficiently explore the Pareto front in multi-objective higher dimensional optimization problems
Abstract: Line search optimization methods fail with multiple objective functions whose gradients are
unavailable. The center of a crowded, trusted region is typically chosen as the point on the Pareto
front with the highest hypervolume contribution. The proposed approach uses an entropy selection
procedure to search the entire Pareto front, avoiding the computation of the Pareto front samples via
cheap multi-objective optimization. By reducing uncertainty in each region, the algorithm directs
its search towards areas with the highest potential for Pareto improvement. We tested the proposed
method on the DTLZ test suite and other real-world applications, such as the welded beam design
problem and the trajectory planning rover problem. The proposed approach yields results at par
with state-of-the-art methods for exploring the Pareto front
Submission Number: 24
Loading