A property-aware empirical analysis of differential evolution variants using the generalized numerical benchmark generator

Benchmarking studies in continuous optimization often report aggregate rankings, yet provide limited insight into why particular algorithms succeed or fail under specific landscape conditions. This paper presents a property-aware empirical assessment of ten competitive Differential Evolution (DE) variants using the Generalized Numerical Benchmark Generator (GNBG), a configurable framework that enables controlled manipulation of problem characteristics. Ten state-of-the art variants of DE are tested across three distinct datasets. The first dataset, GNBG-45, featuring 45 problem instances, each designed to isolate a single landscape property such as basin linearity, conditioning, variable interactions, symmetry or deceptiveness. The remaining datasets, GNBG-I and GNBG-II, are well established in the literature and include 24 functions each, covering unimodal, multimodal single component and multimodal multiple component problems. Together, these data sets provide a layered and comprehensive benchmarking environment. Experimental results show that ALSHADE exhibits the strongest overall robustness and consistency across diverse landscape categories and benchmark suites, while LSHADE-SPACMA exhibits the strongest scalability and robustness in multiple dimensional settings. Other variants, including IMODE, EA4eig, APSM-jSO, and TSALSHADE, demonstrate competitive behavior on selected problem classes but lower overall consistency. Overall, the study shows the ability of GNBG to distinguish algorithmic behavior in a variety of problem structures, supporting its use as a detailed and property-aware benchmarking suite.