Contests
Standard Benchmark Functions
Classical benchmark functions (Sphere, Ackley, Rastrigin, Branin, etc.).
Real-World Sample Problems
Real-world-flavored black-box problems — chemical reaction yield, injection-molding defect rate, neural-net hyperparameter tuning.
Noisy Benchmark Functions
Noisy variants of classic benchmarks, mirroring the BBOB noisy suite (Noisy Sphere / Noisy Rastrigin / Noisy Ackley). Each call returns base_value + Gaussian observation noise, so evaluating the same x twice yields different y. Tests how robust your surrogate is to noise and whether your strategy re-evaluates near the optimum to denoise. A positive offset is added so values stay non-negative.
Human-Powered Aircraft (HPA) Design
A simplified human-powered-aircraft wing design contest inspired by the OptunaHub HPA benchmark (https://hub.optuna.org/benchmarks/hpa/). The real benchmark runs a scipy-backed aerodynamic + structural analysis; we ship analytic surrogates instead. Eight design variables (root / tip chord, span, dihedral, twist, spar diameter, CFRP layer count, additional payload) are normalised to [0, 1] and de-normalised inside the objective. Infeasible designs (CL_required > CL_max, sustained cruise power above human capacity) get added as soft-constraint penalties.
HPO Surrogate Benchmarks
Hyperparameter-optimization surrogates inspired by OptunaHub HPO suites (HPOBench / HPOLib). Each problem is an analytic surrogate that mimics the response surface of a real model on a real dataset — no GPU, no training, just the optimizer's strategy under test. Three flavours — XGBoost (continuous-heavy), Random Forest (mixed types), and SVM (log-scale-heavy).