turbo is a modular Bayesian optimisation framework which focuses on gathering and storing the intermediate optimisation steps to give insight into the decision making process. turbo is capable of producing a wide variety of plots and supports many variations of the basic Bayesian optimisation algorithm.
- Acquisition Functions
- Probability of Improvement (PI)
- Expected Improvement (EI)
- Upper/Lower Confidence Bound (UCB/LCB)
- Pre-Phase 'naive' selectors
- Random
- Latin Hypercube Sampling (LHS)
- Manual
- Surrogate Models
- Scikit-Learn Gaussian Process
- GPy Gaussian Process
- Latent Space
- Fixed warping (e.g. log-transformed or linear map to
[0,1]etc)
- Fixed warping (e.g. log-transformed or linear map to
- Fallback
- Scheduled random samples ("Harmless" Bayesian Optimisation)
- de-duplication
- Misc
- able to use the same storage and plotting functionality with random search or any of the available 'naive' samplers
all dependencies can be installed with pip, see requirements.txt
- Black Box Optimisation Benchmarking Procedure: http://coco.lri.fr/COCOdoc/bbo_experiment.html
- python implementations of many benchmarking functions https://github.com/andyfaff/ampgo/blob/master/%20ampgo%20--username%20andrea.gavana%40gmail.com/go_benchmark.py
- https://github.com/fmfn/BayesianOptimization
- https://github.com/thuijskens/bayesian-optimization
- https://github.com/befelix/SafeOpt
- https://github.com/scikit-optimize/scikit-optimize
- https://hyperopt.github.io/hyperopt/
- https://github.com/automl/RoBO
- https://github.com/resibots/limbo