Regression ability of gaussian process about edbo HOT 4 OPEN

@shenggong1996 any update on this? Did you try adjusting the hyperparameters or using a random forest? Would be great to see any results (e.g. before and after parity plots) and/or an example of the "jumps" you were describing if you have them available.

Hi Sterling,

I played around with hyperparameters and random forest, and no significant improvement appeared. My solution is using a neural network instead, but since it is not straightforward to use NN to estimate uncertainty, I use NN in the exploitations and GP in explorations. Our data will be presented in a preprint soon, with the possible title "Bayesian Optimization Assisted Laser Reduction of poly(acrylonitrile) for Electrochemical Applications".

from edbo.

Hi,

Sorry for the delayed response. One possibility is that the default hyperparameters for the GP model are not good for your application. For example, when there isn't much data the length scale prior will tend give longer length scales and a function distribution which may not fit your discontinuous data well. I would suggest that you play with the GP hyperparameters or build your optimizer using edbo.bro.BO_express which attempts to select priors appropriate for the dimensionality of the data. In addition, you could use a random forest as the surrogate model for edbo (edbo.models.RF_Model).

from edbo.

@shenggong1996 any update on this? Did you try adjusting the hyperparameters or using a random forest? Would be great to see any results (e.g. before and after parity plots) and/or an example of the "jumps" you were describing if you have them available.

from edbo.

Hi Sheng, thanks for the info! That seems like an interesting approach. When the preprint is posted, I'm interested in checking it out.

from edbo.