Using the code on different benchmarks about constrained-bayesian-optimisation-for-automatic-chemical-design HOT 3 OPEN

Hi Alireza,

Apologies for the delay in responding, I've only just seen your message! I think it would just be a case of swapping out the samples from the Branin-Hoo function (under the folder containing the Branin-Hoo experiments) for samples from the higher dimensional benchmarks.

Best wishes,
Ryan

from constrained-bayesian-optimisation-for-automatic-chemical-design.

@Ryan-Rhys I didn't get it how to apply your implementation for a high dimensional Bayesian optimization. Branin-Hoo function is only dimension 2, but for ChemVAE, it is 196-dimensional latent space. Because they didn't upload their code on Gaussian process Bayesian optimization, I faced difficulties to replicate this part because of high dimension. Normally, GP with Bo deals with low dimensional input data, For GpyOpt library, the maximum dimension is 20. I need your help to figure out how to do this. Many many thanks.

from constrained-bayesian-optimisation-for-automatic-chemical-design.

@xuzhang5788 thanks for the comment!

In our code we use a sparse GP to cope with the high dimensionality of the latent space. If you're using vanilla BO (i.e. not constrained) then it should be relatively straightforward to adapt our code to your problem. Otherwise if you're keen on sticking with the GpyOpt, I believe they have an option to use sparse GPs as the surrogate model.

I'm not sure what model they use in the original Automatic Chemical Design paper. They don't mention sparse GPs in either the ArXiv or the ACS versions of the paper but I think when I inherited the code the model was a sparse GP.

from constrained-bayesian-optimisation-for-automatic-chemical-design.