A GNBG-Inspired Multi-Component Benchmark Suite with Controlled Difficulty Axes for the IOH BenchDesign Challenge
Abstract
We submit a deterministic, scalable suite of 25 continuous benchmark problems for the Benchmark Design Challenge. The suite is designed to maximize performance diversity among five reference solvers (CMA-ES, Differential Evolution, Particle Swarm Optimization, BFGS with restarts, and COBYLA) under a fixed evaluation budget. Each problem is defined as the minimum over multiple transformed quadratic components (basins). Diversity is created by systematically varying interaction structure, conditioning spectra and patterns, GNBG-style oscillatory warping and asymmetry, and basin overlap/sharpness to induce deceptive landscapes.
Authors: Rohit Salgotra, Kalyanmoy Deb, Amir H. Gandomi
Published in: Genetic and Evolutionary Computation Conference Companion (GECCO Companion) (2026)