jpiironen / gplite Goto Github PK

Hi, @jpiironen! I use gplite to take advantage of approximating GP by using LA.
I would model my data with a mean component that depends on another measured variable plus a latent variable GP. How can I combine the two things using gplite?
Thanks in advance

I don't see anywhere where the convergence results are stored. If we want to compute against this, we would have to handle the warning and pattern match the text for the appropriate message.

I'm thinking about using gplite for the Bayesian optimization done in the tune package. I'd like to pass in starting parameters since we run repeated GPs based in incremental updates to the data. My initial testing indicates that, for this application, taking a median or trimmed mean of the previous parameters is very helpful.

The issue is that it is difficult to know the parameter values for the covariance function (programmatically). I'd like to extract them per iteration but that might be difficult when there are compound covariance functions.

Can you make something like an extract_parameters function that could return a data frame of the parameters and their associated covariance function? An update method would also be nice.

For the example in the vignette, maybe something like this:

cf0 <- cf_const(magn = 5)
cf1 <- cf_lin(magn = 0.01)
cf2 <- cf_periodic(
  period = 12, 
  prior_period = prior_fixed(),
) * cf_sexp(
  lscale = 100,
  magn = 1, 
  prior_magn = prior_fixed()
)
cfs <- list(cf0, cf1, cf2)

data.frame(
  parameter = c("magn", "magn", "lscale", "magn"),
  value = c(5, 0.01, 100, 1),
  source = c("cf_const", "cf_lin", "cf_periodic", "cf_periodic"),
  index = c(1, 1, 2, 2)
)

or something similar to make them identifiable and updatable.