Open belledon opened 4 years ago
Maybe the proposal width is too tiny for MH?
In any case, PF with 300 or 1000 particles might be just fine as the IO model.
Would love to see how low- and high-sensory noise compares in IO.
I emailed Niko and will let you know
Maybe the proposal width is too tiny for MH?
I thought so too but the width is the same as for the PF
I have posted the set of traces for 300 particles with either 0.1 or 0.8 observation noise in https://yale.app.box.com/folder/108897606057
There you can view the complete set of inference traces and visualizations. I've included a couple below as a demonstration for the trials I've been using for debugging
Very interesting and very promising! Thanks!
We would like to also have a 10 or 4 particle particle filter with 0.1 noise and obtain summary statistics in terms of fit to behavior.
@iyildirim @eivinasbutkus
we are one or two days away from having BO model fits. Below are a few highlights
inference chain analysis and visualization is working
particle filter procedure is verified (thank god)
MH is implemented but the results hint towards a potential bug. Also limited to 6k samples due to a recursion limit (could play with in the future).
While I think it is possible to rescue MH (maybe using HMC) the results from the high particle count PF are very promising. I think it is reasonable to go forward with the PF as both the RPM and IO with IO cranking up towards 300+ particles.
Here is a PF of 100 particles
I'm running two batches of the IO PF with "high" and "low" obs noise to see if there are any quantitative differences between t1 and t2 (the first two vertical lines)
Here is what I have for MH for a congruent trial with 3000 samples (showing/ plotting the top 100 samples).. seems to be stuck on burn in.. or perhaps the prior is too loose.
Tomorrow I'll process the IO results so we can get some model judgements and @eivinasbutkus and I can get BO running.
@iyildirim let me know if you need anything else before you email Niko