Per the discussion at the meeting today I wrote down the required function implementations for adding a Gaussian process in the Stan Math Library, https://github.com/stan-dev/stan/wiki/Adding-a-Gaussian-Process-Covariance-Function . I tried to keep the same naming prefix convention that the cov_exp_…

Okay, I get it. Thank you so much! Best, Lu Zhang

@betanalpha can you clarify the naming convention with an example? Are you envisioning deprecating cov_exp_quad and encouraging exp_quad_cov? It just wasn’t clear enough in the doc for me to answer that question.

Not encouraging, requiring. The suffix notation is consistent with the probability densities which takes priority over the single GP covariance function implementation in the language.

Cool. We’ll need an issue to deprecate cov_exp_quad() and rename it exp_quad_cov(). And make everything else follow that convention.

Can we prepend the functions with something like kernel, or Gram? exp_quad seems too vague to me.

I talked to Rob about this the other day – he was referring to the fact that there’s nothing in the naming convention that specifies that these are related to Gaussian processes. We considered also requiring a _gp_ tag, as in exp_quad_gp_cov, exp_quad_gp_chol, and exp_quad_gp_lpdf, but then we have…

I was thinking a bit more about @avehtari ’s request for a summation function over GP covariance functions. Again there are various technical issues and it’s not clear how often people are running into memory issues that might motivate an explicit function, but if we do want something then I think t…

Mike is right, I was worried about the fact that there wasn’t anything explicit about the functions exp_quad_cov, exp_quad_chol and exp_quad_lpdf being related to GPs. As we expand the kernel options (Matern 3/2, 5/2, etc.) it’ll be hard for users to find these functions in the doc appendix easily.…

If we add an additional identifier I’d prefer that it go with the suffixes so we don’t have an awkward sandwich of naming requirements. Do people think that ```…_gp_lpdf```` is too long or okay for a suffix?

Adding Gaussian Process Covariance Functions

Developers

Bob_Carpenter February 9, 2017, 7:07pm 12

Are you worried about alphabetization? What do you recommend as the whole name for the function, gram_exp_quad, exp_quad_gram, or was it exp_quad you objected to?

Topic		Replies	Views
Gaussian process roadmap Developers gaussian-process	48	4808	April 12, 2019
Gaussian Process Predictive Posterior Function Developers	7	1790	July 26, 2018
Could you please help to find out why my GP model got slower Modeling gaussian-process	12	1363	April 7, 2021
Case study: NNGP based models in Stan Modeling	34	3740	January 30, 2018
Gaussian process kernals Modeling gaussian-process	9	947	June 21, 2020

Adding Gaussian Process Covariance Functions

Related topics