rubinxin / BNN_BO

Implement BO with BNNs and GPs

Bayesian Optimisation with Bayesian Neural Network

The code package implemented the surrogate models:

• GP
• DNGO (NN + Bayesian Linear Regression)
• MC Dropout
• Concrete Dropout
• BOHAMIAN (HMC-based BNN)

Running a BO experiment

Run Bayesian optimisation experiments: python bo_general_exps.py followed by the following flags:

• -f Objective function: default='egg-2d'
• -m Surrogate model: 'GP'(default), 'MCDROP',MCCONC, 'DNGO' or 'BOHAM' or 'LCBNN'
• -acq Acquisition function: 'LCB'(default) or 'EI'
• -bm Batch option: 'CL'(default) or 'KB'
• -b BO Batch size: default = 1
• -nitr Max BO iterations: default = 40
• -s Number of random initialisation: default = 20
• -uo Utility function type for LCBNN: 'se_yclip' or 'se_y'

E.g. python bo_general_exps.py -f='egg-2d' -m='GP' -acq='LCB' -bm='CL' -b=1 -nitr=60 -s=10

Requirement

• python 3
• torch
• torchvision
• emcee
• gpy
• gpyopt