Quantifying Individual and Joint Module Impact in Modular Optimization Frameworks
Authors
A. Nikolikj, A. Kostovska, D. Vermetten, C. Doerr, T. Eftimov
Publication
IEEE Congress on Evolutionary Computation IEEE CEC 2024
Yokohama, Japan, 1-5 July, 2024
Abstract
This study explores the impact of modules in modular optimization frameworks for single-objective black-box optimization with numerical decision variables. We use the functional ANOVA (f-ANOVA) framework to assess the impact of modules and their interactions on the performance of two algorithms, the modular Covariance Matrix Adaptation (modCMA) and the modular Differential Evolution (modDE). We analyze performance data from 324 modCMA and 576 modDE variants on the BBOB benchmark collection in two problem dimensions. Noteworthy findings include the identification of modules that strongly influence the algorithm performance, such as the weights_option and mirrorred modules in modCMA for low problem dimension, and the base_sampler for high problem dimension. For modDE, the \textit{lpsr} module is the most impactful regardless of the problem dimension and budget. In the comparison between modCMA and modDE, both algorithms show varying shifts in the significance of individual modules and their interactions with changes in dimensionality and budget. Notably, modDE undergoes a more pronounced shift from individual to interaction effects, while modCMA follows the opposite pattern.
BIBTEX copied to Clipboard