tpapp / DynamicHMC.jl

Implementation of robust dynamic Hamiltonian Monte Carlo methods (NUTS) in Julia.
Other
242 stars 21 forks source link

Simplify stepsize search. #181

Closed tpapp closed 1 year ago

tpapp commented 1 year ago
  1. use log acceptance ratios, to avoid numerical overflows.

  2. revert back to the original NUTS crossing algorithm. Sophisticated bracketing does not buy much, and occasionaly it can lead to step sizes which adapt "too well" to local information and recovering from them takes time.

tpapp commented 1 year ago

After testing extensively with some economic models, the original version works better, so closing. The log part will be cherry-picked for another PR.

Deciding to merge this, at least when the posterior is bad this fails more transparently than micro-optimizing the initial stepsize.