search

RJT1990 / pyflux

Open source time series library for Python
BSD 3-Clause "New" or "Revised" License
2.11k stars 240 forks source link

BBVI for unstable likelihoods (sensitive parameters) #44

Closed RJT1990 closed 7 years ago

RJT1990 commented 7 years ago

Some models - such as GAS models - have likelihoods that are very sensitive to poor parameterization. This results in nan likelihoods, and thereby nan gradients for BBVI. Possible solutions include gradient clipping for GAS models, and a warm-up stage for BBVI that is specific to some model types.

RJT1990 commented 7 years ago

This issue is mostly resolved - in the sense that we now give the user the chance to tweak the parameters for BBVI. But more documentation needed on BBVI tuning perhaps.