search

PecanProject / pecan

The Predictive Ecosystem Analyzer (PEcAn) is an integrated ecological bioinformatics toolbox.
www.pecanproject.org
Other
204 stars 237 forks source link

PEcAnRTM: Algorithm ideas #1446

Closed ashiklom closed 7 years ago

ashiklom commented 7 years ago

Description

Context

These changes may improve convergence of Metropolis Hastings sample in PEcAn RTM, which is becoming more important as the sampler is put to more complex challenges (e.g. EDR inversion).

Attn: @serbinsh

ashiklom commented 7 years ago

One potential algorithm/approach: https://en.wikipedia.org/wiki/Parallel_tempering

ashiklom commented 7 years ago

A book containing valuable information on additional approaches: https://books.google.com/books?id=TRXrMWY_i2IC&lpg=PA89&ots=7h_rqnKtnx&dq=improve%20mcmc%20chain%20mixing&lr&pg=PA104#v=onepage&q=improve%20mcmc%20chain%20mixing&f=false

mdietze commented 7 years ago
  1. Could you explain better exactly what the actual issue is?
  2. Rather than monkeying around with alternative ways to adapt the adaptive M-H, why not try alternative algorithms that converge faster and are already implemented (e.g. DE and DREAM in BayesianTools, which is a package already being used in PEcAn)
  3. I thought you were looking at trying a SMC approach?
ashiklom commented 7 years ago

  1. Sorry, was just making a quick note before I forgot. Basically, @serbinsh and I were monitoring some EDR inversions and were seeing extremely slow/bad convergence -- chains were stuck in relatively narrow local minima for a very long time. (Shawn -- maybe you could link to one of the figures you had up?) .

  2. Good idea. Every time I've tried to do PROSPECT inversion using any of the functions from that package, I've been unable to do the sampling, probably because of operator error. Now that the package is more mature (?), I would definitely be interested in trying again. I just wanted to throw up these ideas while they were fresh on my mind.

  3. SMC is good for PROSPECT inversion, where the number of parameters is low, but I thought it might become less tractable for something like EDR, where the number of parameters is 7-12 x the number of unique PFTs (so for three cohorts, 21 - 36). But maybe not -- there may be a clever way of doing it that of which I'm not aware. You mentioned there were clever ways of leveraging Reimann integration (?) and similar interpolation techniques to reduce the number of samples required while still covering the likelihood space. Maybe we can discuss more in person tomorrow?

ashiklom commented 7 years ago

Closing because BayesianTools was implemented in #1569. Will re-open if this becomes a problem again.