Metamodels for Fast Multi-objective Optimization: Trading Off Global Exploration and Local Exploitation

摘要

Metamodels can speed up the optimization process. Previously evaluated designs can be used as a training set for building surrogate models. Subsequently an inexpensive virtual optimization can be performed. Candidate solutions found in this way need to be validated (evaluated by means of the real solver).rnThis process can be iterated in an automatic way: this is the reason of the fast optimization algorithms. At each iteration the newly evaluated designs enrich the training database, permitting more and more accurate metamodels to be build in an adaptive way.rnIn this paper a novel scheme for fast optimizers is introduced: the virtual optimization - representing an exploitation process - is accompanied by a virtual run of a suited space-filler algorithm - for exploration purposes - increasing the robustness of the fast optimizer.