Software¶
Actively maintained libraries. Archived and maintenance-only packages are listed under Unmaintained.
- Ax - Adaptive experimentation platform on top of BoTorch, with constraints, multi-objective, and mixed spaces.
- Vizier - Google's service-style black-box optimization, now open source (paper).
- BoTorch - Modular Monte Carlo BO library built on PyTorch and GPyTorch (paper).
- Trieste - TensorFlow/GPflow toolbox for batch, constrained, multi-fidelity, and multi-objective BO (paper).
- BoFire - Experimental design and BO for mixed spaces, used in chemical and pharmaceutical settings (paper).
- BayBE - Bayesian DoE with chemical encodings and transfer learning (paper).
- HEBO - Heteroscedastic evolutionary BO from Huawei Noah's Ark Lab (paper).
- SMAC3 - Sequential model-based algorithm configuration, using random forests or GPs (paper).
- OpenBox - Black-box optimization system with transfer, multi-fidelity, and distributed runs (paper).
- Dragonfly - Scalable BO for expensive functions, including parallel and high-dimensional variants (paper).
- BayesianOptimization - Small scikit-learn GP implementation of EI-style BO.
- PyBADS - GP-assisted mesh adaptive search for mildly expensive, nonsmooth, or noisy black-box fitting (paper).
- emukit - Multi-fidelity emulation, experimental design, and BO on top of GPy (paper).
- SMT - Surrogate modeling toolbox with kriging, mixed and hierarchical GPs, and EGO (paper).
- NUBO - Compact PyTorch BO package aimed at scientists writing their own loop (paper).
- Syne Tune - AWS HPO toolkit with BO, multi-fidelity, and transfer methods (paper).
- HyperMapper - BO for computer-systems autotuning, including constrained and multi-objective problems (paper).
- CAKE - LLM-driven evolution of GP kernels inside BO (paper).
- PlugBO - Modular agentic interface around a BO loop.
Related HPO tools¶
Not GP-BO, but often compared with it.
- Optuna - Define-by-run HPO. Default sampler is TPE, not a GP (paper).
- Hyperopt - TPE over complex search spaces (paper).