GPyOpt: available surrogate models and acquisitions

Written by Javier Gonzalez, University of Sheffield.

Reference Manual index

Last updated Friday, 11 March 2016.

1. Supported models

The surrogate models supported in GPyOpt are:

• Standard Gaussian Processes with standard MLE over the model hyperparameters: select model_type = GP in the GPyOpt wrapper.

• Standard Gaussian Processes with MCMC sampling over the model hyperparameters: select model_type = GP_MCMC in the GPyOpt wrapper.

• Sparse Gaussian processes: select model_type = sparseGP in the GPyOpt wrapper.

• Random Forrest: select model_type = RF. To illustrate GPyOpt modularity, we have also wrapped the random forrest method implemetented in Scikit-learn.

2. Supported acquisiitions

The supported acquisition functions in GPyOpt are:

• Expected Improvement: select acquisition_type = EI in the GPyOpt wrapper.

• Expected Improvement integrated over the model hyperparameters: select acquisition_type = EI_MCMC in the GPyOpt wrapper. Only works if model_type is set to GP_MCMC.

• Maximum Probability of Improvement: select acquisition_type = MPI in the GPyOpt wrapper.

• Maximum Probability of Improvement integrated over the model hyperparameters: select acquisition_type = MPI_MCMC in the GPyOpt wrapper. Only works if model_type is set to GP_MCMC.

• GP-Lower confidence bound: select acquisition_type = LCB in the GPyOpt wrapper.

• GP-Lower confidence bound integrated over the model hyperparameters: select acquisition_type = LCB_MCMC in the GPyOpt wrapper. Only works if model_type is set to GP_MCMC.