I have a problem where I am trying to model the output of some complicated process using surrogate models (I’ve asked several questions in this regard). My surrogate model is a gaussian process fit to data generated by my expensive model. The GP is parameterized such that the parameters of my expensive model are the inputs and measurements of interest are the outputs.
The measurements of my model are known with a high degree of certainty compared to the uncertainties on my parameters so that if my model variation in output due to my parameters is order 1, the variation in my measured test data is order 1e-3 - 1e-4. I tried defining my likelihood like this:
unknown_sigma = pm.HalfNormal('unk_sigma',sd=some_value) sigma = pm.Deterministic('sigma',unknown_sigma + observed_sigma) y = pm.Normal('y',mu=surrogate_out,sd = sigma, observed=data)
but I have a lot of problems with divergence. Does anyone have any ideas on how to reparameterize this model so that I could avoid this problem?