Closed kaare-mikkelsen closed 7 years ago
All hyperparameters which must be positive are constrained to be positive and thus when t-distribution with location parameter 0 is used for those hyperparameters, it effectively sets a half-t-distribution prior. We'll try to clarify this in the next release.
so, to be clear, the initial guess does not interact with the prior?
correct
thanks
Looking at the regression demos, there is a clear trend that the means of all priors are kept at the default (0), while the initial guesses are different (see, for instance, demo_regression_additive1.m). Is this supposed to indicate that the actual prior used in gp_optim is prior+initial guess? Otherwise it seems somewhat odd to use a student's t distribution centered around 0 as a prior for a length scale, which necessarily must be positive.
if I have deduced the behavior correctly, I think it should be described a bit more prominently. Possibly appended to the help-texts for the priors functions.