search

ebachelet / pyLIMA

GNU General Public License v3.0
31 stars 8 forks source link

MCMC convergence and mixing depends on eps parameter #81

Closed mpgh closed 1 week ago

mpgh commented 7 months ago

The epsilon parameter "eps" defning the cloud of start walkers as part of MCMC_fit.py line 58 can lead to poor mixing of chains. In many cases a smaller eps like 1e-2 would be more appropriate to keep chains in the vicinity of the best solution. The uniform distribution also contributes to runaway chains. I would suggest to either reduce it or better make it a user parameter.

ebachelet commented 7 months ago

Good catch Markus, I probably did this for some reasons but I will revert it.

ebachelet commented 1 week ago

I believe this is fixed now