For background, @drezap is currently at Aalto in my group for 3.5 months. In addition there are approximately 5 other people using Stan for GP research in my group. We will make case studies of different types of flexible use of GPs during the following months and should have a few of them ready before StanCon in Helsinki.
Summary for the questions @drezap asked and @syclik’s comments :
- 1+2. It seems no-one objects this
gp_exp_quad_cov(x, l, sigma) + gp_exp_quad_cov(x, l, sigma, jitter)
so we go with that now.
3 . We follow @syclik’s suggestion for types.
We do nothing yet for distance metrics.
@drezap will change the pull request to follow R&W.
I’m committed to get better GP support in Stan. I think I have a quite good understanding of the capabilities and restrictions of current Stan and what we are doing the following months is based on that. I know the Stan roadmap includes some useful features coming in the future (e.g. functional types and Stan 3 sub-models), but it will be helpful to do sophisticated GP models already before these features are available. We might consider at some point functions in similar way as in ode, but we can already do a lot with what @drezap has now implemented.
I agree on this and I’m sorry if my writing is so confusing that it gave some other impression. You and I do disagree on what supporting existing users mean, but I’ll let the forthcoming case studies to speak. I’m looking forward for more research in principled priors for single-kernel, multi-kernel, hierarchical and chained GPs.