[Question] Constraining inducing points

[fleskovar]

I am trying to train a variational GP that has multidimensional inputs. I would like to constraint the learned inducing points during training within the convex hull of the training data to avoid "extrapolating" outside of the observed data. What would be the right way to implement this?

My initial thought was to penalize the objective function with a term that is 0 if the points are within the convex hull or a very large number if they are outside, but maybe there is a better way to do this.

Thanks!

[mochar]

This might be helpful: https://github.com/cornellius-gp/gpytorch/issues/1598

[fleskovar]

Thank you for the information! Quick question about using the raw inducing points: inside the inducing_points property I should implement a function what will map the raw values to the actual values, right?. Is it a problem if the transform that I use is not bijective?

[gpleiss]

inside the inducing_points property I should implement a function what will map the raw values to the actual values, right?

Correct!

Is it a problem if the transform that I use is not bijective?

I don't think this should be an issue.