pv-lab / benchmarking Goto Github PK

I recently found your paper and found it quite interesting. As I have been exploring the repository, I have been trying to reproduce the demonstrated performance of the Matern 52 ARD kernel with LCB_2 acquisition function on the crossed barrel dataset. Using the jupyter notebook Example use of framework with GP type surrogate models.ipynb and setting ARD_ = True, I am finding that the model struggles to optimize. After some refactoring, I've been able to get it working, but cannot match the performance shown in the paper.

Running five ensembles for 100 iterations each, I produce the following:

Compare this to the plot shown in the paper which suggests that most candidates should have been found by this point using Matern 52 ARD with LCB2.

An ensemble of five is a pretty small sample, but based on the comparison, this would suggest that all of these are several standard deviations below the expected mean for this kernel-acqusition function combination.

Has there been an update to GPy that might be affecting performance?

@HarryQL
Very interesting, thorough, and timely paper!

Adaptive design materials informatics benchmarking is very relevant to me (see e.g. mat_discover), and I'd like to use the methods described here, but it seems like I might spend somewhere between 10-30 hrs before I can decide if they're really what I want/need.

Would it be possible for you to make some tutorials on how to:

1. perform adaptive design on a different dataset than what's shown
2. assess benchmarks for a user-defined surrogate model
3. perform the above two simultaneously

Happy to assist with refactoring the Jupyter notebooks into a Python class or making it installable via pip and/or conda if that's of interest.

@ramseyissa

Sterling