search

lmoffatt / macro_dr

GNU General Public License v3.0
1 stars 1 forks source link

Discussion of next steps #111

Closed lmoffatt closed 2 months ago

lmoffatt commented 7 months ago

I am kind of stuck and I need to progess.

The problem I am having is two fold

  1. Scheme 4 does not seems to fit the data well, and it is worst for the 7 traces.
  2. Scheme 4 does not equilibrate in the allotted time.

An inspection in the emcee success rate and in the thermo jump success rate indicates that the temperatures are too separated and that the data is dispersed, a low success rate compared to the simulated data.

On the other hand, scheme 4 does not fit well the low frequencies components, so another possibility is to incorporate pink noise into the model, that is that there are two noise components one white noise and one pink noise of the same variance independent of the trace duration.

Finally, in order to properly compare thermo with cuevi, we should allow for the fractions to have less resolution in the ATP jump so the computational load is diminished. We can test that with the simulations.

So, one possiblity is to just run chains with more temperature and scouts, the other is to find a self regulating algorithm that increases the number of temperatures and scouts accordingly.

lmoffatt commented 7 months ago

So, the problems are

  1. Lousy fit. solution--> add pink noise. #114
  2. Slow execution. Solution--> simplify ATP profile. #112 #113

  3. slow chain mixing. Solutions

    • -> increase number of temperatures and scouts --> dynamic allocation of tempertures and scouts. #115

    --> implement Levenberg/Marquardt #116

lmoffatt commented 7 months ago

Situation as feb 27.

1) Inclusion of a "poissonian" variance term (i.e. proportional to the absolute current) improved the recovered values of the mean conductance and most of the paramters. The pink noise term does not seems to be that important. 2) The evidence did not reached equilibrium in these conditions either 3) although the emcee success rate is much better, it is still in the order of 1/20, a little low, that might make the evidence equilibration not reacheable in the 48 hours run. 4) cuevi and thermo algorithms perform comparable well.

lmoffatt commented 7 months ago

So, what are the next steps?

the main problem remains the time to reach equilibrium, it is too much. So, there are three things to do:

1) Implement the continuation.

2) Reformat to be able to use the 64 cores.

3) Implement an adaptive algorithm for emcee

4) implement Metropolis Levenberg Marquardt