Closed gtng92 closed 7 years ago
I agree, this feels like there's more of an issue with sampling particles than with an inability to inform the prior distribution with data --- for example, in the first case the prior distribution has the correct mean (shape/rate=0.1) but the SMC trace is locked into a divergent value.
We need to dig into the small case with n=20 particles where the SMC analysis gets locked into a fixed value after about 10 iterations. What exactly is happening to particles to cause this degeneracy? Are they failing to be resampled? Perturbed? Examine the printouts at each iteration.
After discussion with @gtng92, it sounds like the prior distributions being used may be too informative (peaked) at the incorrect values. @gtng92 will re-evaluate with flatter priors to see if this issue is reproduced.
From previous discussion, I mentioned that I ran a test with more relaxed priors, with 500 particles, and the result showed a beta
parameter that was not converging. I checked again and I actually ran that with 20 particles (oops).
When I changed to n.particle = 500
, approximately 2.5 hours into the run, Rstudio aborts and the following message pops up:
Problem in rstudio
Sorry, the program "rsutdio" closed unexpectedly
Your computer does not have enough free memory to automatically analyze the problem and send a report to the developers.
Currently running same seeded run on command line in the background, and it's been a little over 3 hours without any aborting / errors. I will update if the run ends up finishing.
The run finished, using the following parameters in commit 4d37ca4 There are separated issues that are more related to #42 so I will continue posting there. I think this is resolved (my resulting plots no longer locks in on a fixed value), so I will be closing this and merging branch issue92-5 into master.
In issue #42 parameters beta and gamma were shown to be estimable parameters, not confounded with N or mu. Depending on the priors I would set, the estimated value of beta would alter.
For example, as I estimate
beta
:seed=50, shape=1, rate=10
(plot correction:nparticle=50
) Changing shape and rate:seed=50, shape=0.3, rate=5
(plot correction:nparticle=50
) And if I increase tonparticle=500
Could be related to issue #92 part 2, when resampled particles aren't being further perturbed.