[GENERAL SUPPORT]: How to deep dive into Ax reasoning method ?

[arnaudlahouati]

Question

Hello !
I am using Ax for a MOBO problem in a material science problem, aiming to predict new materials compositions to try out in the lab. I am working on 240 samples DB with 5 features, trying to maximise 2 targets. The optimisation/prediction works super well without specifying any models or aq functions. However I can't manage to find how to read the specs Ax is using, and to obtain the precise Surrogate Model with hyperparameters etc.
Would someone know what function to use ? Tried to save the JSON file but I only got information about the data pretreatment !

Please provide any relevant code snippet if applicable.

Code of Conduct

• I agree to follow this Ax's Code of Conduct

[mpolson64]

Hi @arnaudlahouati -- I'm glad to hear Ax is working well for your mat sci use case. Admittedly, Ax does not do a fantastic job of surfacing this information. Assuming you haven't passed in anything particularly special to Client.configure_generation_strategy Ax is using Botorch's SingleTaskGP with default settings and the qLogNoisyExpectedHypervolumeImprovement acquisition function. You can extract all this info from the Client and even look at the Botorch objects directly by calling client._generation_strategy._curr._fitted_adapter.generator._acquisition.__dict__

Thanks for reaching out about this; its good signal to us that it would be helpful to our users to provide an easier way to access this information. Let me know if this was helpful or if you have any follow up questions.