Using a model other than GP

[eloiseum]

Hi everyone,

I wanted to try another model instead of GP (e.g. random forests), but I am a bit lost on how to implement it using BoTorch. Has anyone done it or any hints on how to make this change? I have read the documentation but found very little guidance on how to implement other models.

Many thanks in advance!

[Balandat]

Hi, how were you planning on addressing predictive uncertainty with your model (e.g. RFs)?

At the end of the day what you’ll need to make things work is a Model class that has a posterior() method that upon calling reruns a Posterior object. If you don’t have a closed form posterior then you can still use this setup with Monte Carlo acquisition functions so long as you can draw samples from your Posterior object. If you want to use gradient based optimization of the acquisition function then on top of that you’ll need to be able to backprop through those samples.