EVOLUTIONARY ALGORITHMS WITH SURROGATE MODELING FOR COMPUTATIONALLY EXPENSIVE OPTIMIZATION PROBLEMS
Kyriakos C. Giannakoglou, Marios K. Karakasis, Ioannis C. Kampolis, Kyriakos C. Giannakoglou, Marios K. Karakasis, Ioannis C. Kampolis · 2006
Abstract. During the last decade, the development of efficient optimization tools that utilize Evolutionary Algorithms (EAs) as the core search tool gained particular attention and reached a certain level of maturity. These tools enabled the extensive use of EAs in large-scale industrial applications, in which the analysis (evaluation) tool is computationally expensive. A literature survey reveals that the majority of new, promising variants of EAs are conceptually based on the reduction of the otherwise excessive number of calls to the evaluation software. This reduction is possible through the use of various techniques such as: (a) the use of computationally cheap surrogate evaluation models or metamodels trained on samples collected suring or separately from the evolution (Metamodel-Assisted EAs, MAEAs), (b) the use of more than one evaluation tools, with different approximation errors and computing cost, according to a hierachical structure (Hierarchical EAs, HEAs) and (c) the use of Distributed EAs (DEAs), which subdivide the entire population into concurrently evolving, semi-isolated subsets, which regularly exchange promising individuals. These techniques and the most efficient combination of all of them in a single search method (Hierachical, Distributed Metamodel-Assisted EAs, HDMAEAs), are discussed in this paper. Due to space limitations, only three applications are presented; however, more