Empirical Study of Surrogate Model Assisting JADE: Relation Between the Model Accuracy and the Optimization Efficiency
Konrad Krawczyk, Jarosław Arabas · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
This paper presents an empirical study of the combination of surrogate models with the JADE algorithm. It focuses on the connection between the accuracy of surrogate models and optimization efficiency. Our study uses surrogate models to approximate costly fitness functions that assist the optimization process. Our analysis primarily focuses on how the accuracy of surrogate models affects the effectiveness of the JADE algorithm in a search space when searching for optimal solutions. Experiments were performed using various surrogate models. The effectiveness of the different surrogate models in combination with JADE was tested on the CEC2013 benchmark. Our observations provide insights into how the dynamics of the error of surrogate models look and what their correlation with optimization efficiency is. Cases were also observed where models with higher error rates paradoxically contribute to better optimization results. Our findings reveal that higher surrogate model accuracy does not necessarily equate to more effective optimization. Instead, a model's ability to preserve the hierarchical order of solutions, which helps with effective mutation and selection, is also an important factor. The study underlines the importance of evaluating surrogate models not only on their approximation errors but also on their ability to maintain accurate rankings.