Paper Info
Title
Predictability on performance of surrogate-assisted evolutionary algorithm according to problem dimension.
Abstract
As the demand for computationally expensive optimization has increased, so has the interest in surrogate-assisted evolutionary algorithms (SAEAs). However, if a fitness landscape is approximated using only a surrogate model, thereby replacing a fitness function, it is possible for a solution to evolve toward a false optimum based on the surrogate model. Therefore, many conventional studies have been carried out in which the real fitness function and surrogate model are dealt with simultaneously. Nevertheless, such an approach leaves much to be desired because studies should be performed for real fitness function evaluation and surrogate model-aware search mechanisms. In this study, we discovered that the approximation error of the surrogate model at low dimensions has a significant relationship with the performance of SAEAs at high dimensions for three binary encoding problems and three real encoding problems. Therefore, if the approximate error is sufficiently small in the low dimension, then high GA performance can be obtained even when the real fitness function is not used, because a high-quality surrogate model can be obtained in the high dimension.
Year
DOI
Venue
2019
10.1145/3319619.3326775
GECCO
Keywords
Field
DocType
Surrogate-assisted evolutionary algorithm, genetic algorithm, machine learning
Predictability,Evolutionary algorithm,Computer science,Artificial intelligence,Machine learning
Conference
ISBN
Citations 
PageRank 
978-1-4503-6748-6
0
0.34
References 
Authors
0
2
Name
Order
Citations
PageRank
Dong-Pil Yu101.69
Yong-Hyuk Kim235540.27