search

stan-dev / gmo

Inference on marginal distributions using gradient-based optimization
BSD 3-Clause "New" or "Revised" License
13 stars 4 forks source link

parameter truncation during optimization #6

Open dustinvtran opened 7 years ago

dustinvtran commented 7 years ago

We do optimization in the unconstrained space. When this eventually gets mapped back to the constrained space, the values can be numerically unstable if the parameters in the unconstrained space are very far away. Right now, we arbitrarily truncate at 25 and -25.

Follow the convention in Stan's library.

dustinvtran commented 7 years ago

from bob carpenter:

The effective domain for the transforms from unconstrained to constrained depend on the constraint. For exp(), it's around +/- log(MAX_DOUBLE) (about 315 if I remember correctly) for example.