Giters

promised-ai / lace

A probabalistic ML tool for science

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Better samplers for prior parameters

BaxterEaves opened this issue · comments

commented

Prior parameter samples are rejected too much.

The following MH samplers for prior parameters:

  • Symmetric random walk, which rejects too much
  • Prior MH (the hyper prior is the proposal), which rejects too much but is really fast at rejecting least
  • Globally adaptive MCMC, which is finicky for unknown reasons

We also have:

  • Slice sampler, which uses a "stepping out" procedure to find the slice region, but which can step out forever (causing slowness and eventually a panic) under certain circumstances.
  • Importance MH, which is like prior MH, but uses a user-defined function

Better options:

  • HMC
  • PDMDP
  • Other adaptive MH schemes?