An Investigation of Inherent Structural Bias in Established Benchmark Sets

Stať ve sborníku v databázi WoS či Scopus

en

In this paper, we investigate the inherent structural bias of BBOB and CEC benchmark sets. First, we theoretically derive the advantage of central evaluation against random search on the Sum of Squares-like functions. We then provide an experimental analysis of the effectiveness of the central evaluation on functions from BBOB, CEC2014, and CEC2017. We utilize four variants of the Differential Evolution (DE) algorithm, in which by changing the boundary handling technique, we introduce different types of structural bias in the method and show the link between the performance of the biased DE variants and the inherent biases in the benchmark sets. We find that the high use of Sum of Squares-like functions and the BBOB-specific location of the optima give a substantial advantage to the center-biased DE variants, especially in high dimension on low function evaluation settings. On the selected CEC functions, this effect was much less pronounced.

Benchmarking; Bound Handling Techniques; Differential Evolution; Structural Bias

2025-08-11

Association for Computing Machinery, Inc

9798400714641

Gecco 2025 Companion Proceedings of the 2025 Genetic and Evolutionary Computation Conference Companion

123–126

4