Low-cost genetic optimization based on inexact pre-evaluations and the sensitivity analysis of design parameters

Giannakoglou, KC; Giotis, AP; Karakasis, MK

dc.contributor.author	Giannakoglou, KC	en
dc.contributor.author	Giotis, AP	en
dc.contributor.author	Karakasis, MK	en
dc.date.accessioned	2014-03-01T01:16:41Z
dc.date.available	2014-03-01T01:16:41Z
dc.date.issued	2001	en
dc.identifier.issn	1068-2767	en
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/14170
dc.subject	Artificial neural networks	en
dc.subject	Genetic algorithms	en
dc.subject	Optimization	en
dc.subject.classification	Engineering, Multidisciplinary	en
dc.subject.classification	Mathematics, Applied	en
dc.title	Low-cost genetic optimization based on inexact pre-evaluations and the sensitivity analysis of design parameters	en
heal.type	journalArticle	en
heal.identifier.primary	10.1080/174159701088027771	en
heal.identifier.secondary	http://dx.doi.org/10.1080/174159701088027771	en
heal.language	English	en
heal.publicationDate	2001	en
heal.abstract	A new method, which is capable of reducing the number of evaluations required by conventional Genetic Algorithms in order to reach the optimum solution, is described. For this purpose, an inexact, low-cost, pre-evaluation phase is introduced. In the proposed method, the inexact pre-evaluation aims at distinguishing the most promising among the population members in each generation, which will only be re-evaluated through the exact, and thus costly, tool. To effect this low-cost screening, artificial neural networks are employed which, in the case of inverse design or optimization of aerodynamic shapes, serve as Computational Fluid Dynamics substitutes. Here, Radial Basis Function networks are used, after being trained on a number of input-output pairs; inputs stand for shapes, i.e., candidate solutions to the problem, whereas outputs are their payoff values. Additionally, this paper introduces the concept of locally trained networks and the autocatalytic use of new sensitivity measures, the so-called importance factors. During the networks' training, the role of importance factors is to distinguish and give priority to the most important among the design parameters. Using several test problems, it will be demonstrated that the genetic optimization enhanced through the proposed inexact pre-evaluation, which is based on neural networks trained using dynamically computed importance factors, reduces the number of exact evaluations by approximately one order of magnitude, compared to its conventional counterpart.	en
heal.publisher	GORDON BREACH PUBLISHING, TAYLOR & FRANCIS GROUP	en
heal.journalName	Inverse Problems in Engineering	en
dc.identifier.doi	10.1080/174159701088027771	en
dc.identifier.isi	ISI:000171024800004	en
dc.identifier.volume	9	en
dc.identifier.issue	4	en
dc.identifier.spage	389	en
dc.identifier.epage	412	en