Implementation and derivation of "Variational Bayesian inference for a nonlinear forward model." [Chappell et al. 2008] for arbitrary, user-defined model errors.
We currently have Gamma.Noninformative() --> shape=0, scale = 1/3.
Several alternatives are discussed here that point out that a
Gamma(alpha=eps, beta=eps) (or Gamma(shape=eps, scale=1/eps) with eps-->0
is a more reasonable choice for a vague prior, but with its own difficulties at the limit eps=0. Plus the mean is always 1.
I am tempted to just remove the method such that every user has to make up his own mind...
We currently have
Gamma.Noninformative() --> shape=0, scale = 1/3
.Several alternatives are discussed here that point out that a
Gamma(alpha=eps, beta=eps) (or Gamma(shape=eps, scale=1/eps)
witheps-->0
is a more reasonable choice for a vague prior, but with its own difficulties at the limiteps=0
. Plus the mean is always 1.I am tempted to just remove the method such that every user has to make up his own mind...
What do you think?