What is the best way to choose inducing points in a multiple-output GP model where in the input data is multi-dimensional?
I have 370 observations with 847 dimensions (each observation is a spectrum) with all values in the range 0..1. I have 10 response variables I wish to predict per observation.
I have modelled the data with independent GPs, and have shown that my data is able to be modelled, and so now want to create a multi-output model with separate independent multi-output kernels and shared independent inducing variables.
From the GPFlow example, I think I need an inducing point array of shape (M, 847), where M is the number of inducing points.
What is the best way in which I can choose values for the 847-length dimension required? The example is simply M values in the range min(X)..max(X), which works when X has one dimension.