Efficient Global Optimization with Adaptive Target Setting

摘要

An adaptive target algorithm was proposed to allow EGO to work well with probability of improvement instead of expected improvement. EGO-AT was shown to be more efficient in global optimization compared to a constant target for Sasena, Hartmann 3, and Hartmann 6 functions. A comparison of EGO-AT with other surrogate-based optimization techniques showed that they produce superior or comparable results. Adding multiple points per cycle with the EGO-AT method is very easy and is shown to lead to a much faster convergence with more confidence in the final result, where it was comparable to EGO using ideal target based on knowledge of global optimum.