Criteria for choosing SGLD params

[aelmokadem]

Hi,

I am playing with the code to run Bayesian SGLD inference on a UDE example similar to the one in file BayesiaNUDE_PSGLD_LV.jl. I just have two questions:

• The SGLD predictions seem to have a much tighter uncertainty using the default params in the LV example than the NUTS sampler (also seen in the paper)! Is this expected?
• It seems that the argument "a" to the function "train_neuralLV" has a big impact on the uncertainty band for the prediction! What are the criteria for picking this SGLD param (as well as b and gamma)?

Thank you.