The code package implemented the surrogate models:
- GP
- DNGO (NN + Bayesian Linear Regression)
- MC Dropout
- Concrete Dropout
- BOHAMIAN (HMC-based BNN)
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
- python 3
- torch
- torchvision
- emcee
- gpy
- gpyopt