mbway/turbo

Bayesian optimisation for global black box function optimisation

turbo

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.

Algorithm Features

• 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)
• 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

Dependencies

all dependencies can be installed with pip, see requirements.txt

Links

• 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

Some other Bayesian optimisation libraries

• 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