search

NICTA / stateline

Distributed Markov Chain Monte Carlo
Other
28 stars 11 forks source link

Adaption mechanism rework #84

Closed AlistaiReid closed 8 years ago

AlistaiReid commented 8 years ago

Al is prototyping a new parameter-free adaption. The adaption learns a linear regressor mapping sigma to accept rate that incrementally trains on the noisy accepts and rejects wrt squared error. The model has an analytic inverse so it can then be used to predict which sigma will achieve a target accept rate.

darrnshn commented 8 years ago

This sounds exciting!!

AlistaiReid commented 8 years ago

Yeah, I've got a prototype working and it looks promising!

Hows the internship?

darrnshn commented 8 years ago

Yeah it's going well! This is pretty cool...and it has the added bonus of clearing up the configuration file. Could this be extended to take into account the location of the chains as well, so you get different sigmas for different regions?