GNBG-III: A Property-Aware Benchmark Suite for Diagnosing Continuous Black-Box Optimizers

The choice of benchmarks critically shape conclusions in evolutionary computation and continuous black-box optimization. However, many widely used suites are either (i) collections of fixed test functions with limited control over problem properties, or (ii) generators whose flexibility can complicate standardized comparisons and reproducibility. Building on the Generalized Numerical Benchmark Generator (GNBG), we introduce GNBG-III, a property-aware benchmark suite of 24 box-constrained continuous problems designed to diagnose algorithmic strengths and weaknesses under controlled landscape characteristics and robustness instances. The suite systematically varies separability, conditioning, deceptiveness, interaction structure, and hybrid composition, and includes robustness tests via explicit noise and dynamic landscape shifts. We provide a standardized evaluation protocol featuring multi-target success, runtime-to-target metrics (including ERT), and convergence checkpoints, alongside a Python/MATLAB reference harness and reproducible seeds. To demonstrate the suite's diagnostic utility, we outline baseline results for five canonical optimizers, DE, PSO, GA, ES, CMA-ES, and ACO, intended as reference points rather than state-of-the-art claims.