Analysis of the Impact of Prediction Accuracy on Search Performance in Surrogate-assisted Evolutionary Algorithms

When solving optimization problems with evolutionary algorithms (EAs), the optimization process can involve evaluating many solutions, which may lead to substantial computational time. To address this issue, surrogate-assisted evolutionary algorithms (SAEAs) have been proposed, utilizing solution evaluations predicted by a machine learning model (surrogate). However, previous studies have not analyzed the impact of prediction accuracy of a surrogate on the performance in SAEAs. This study aims at analyzing the impact of the surrogate model's prediction accuracy on search performance. For this purpose, the search performance of SAEAs is compared using pseudo-surrogate models whose accuracy can be adjusted. In the experiments, two types of SAEAs, the pre-selection strategy with classification-based surrogate (PS-CS) and the individualbased strategy with absolute fitness surrogate (IB-AFS), were addressed on solving the CEC2015 benchmark problems. EAs without surrogates were also compared as a baseline method. The results showed that, for PS-CS, the surrogate model with accuracy of 1.0 demonstrated the best search performance, and higher accuracy resulted in better search performance. At an accuracy of 50 %, the performance was equivalent to or lower than that of EAs without surrogates. On the other hand, for IB-AFS, using high accuracy surrogate models was still effective, but the trend of higher accuracy leading to better search performance was less observed than PS-CS.