burn-in points in bayes.dat?

[Xin Wang]

Hi all,

Could I please know whether the data points recorded in the “bayes.dat” file are all after the burn-in phase? Below please find a plot showing chi2 values versus Nsample from the “bayes.dat” file given by one of my modeling attempts. Do you think the funky multi-furcation feature stems from incomplete burn-in phase? Thanks you guys so much!

Best regards,

Xin Wang
--------------------------------------
Department of Physics and Astronomy
University of California, Los Angeles
CA 90095-1547, USA

[Johan Richard]

Hi,

Yes, the bayes.dat only covers the sampling phase, i.e. does not include any of the burnin. We have seen the effect you mention in the past, and it happens

when the full convergence is not achieved during the burnin phase. The bifurcation seems to come from the different MCMC chains.

To prevent this effect there are several possibilities:

- reduce the burnin rate in the "inverse" command, to explore more the parameter space in the burnin phase.

- adjust your priors if they are too large and include possible local minima in chi2.

- combination of the above.

Hope this helps,

Johan

--
--
_____________________________________________________________________
LENSTOOL: http://projets.lam.fr/projects/lenstool/wiki
To post to this group, send email to lens...@googlegroups.com
To unsubscribe from this group, send email to lenstool-unsubscribe@googlegroups.com
For more options, visit this group at http://groups.google.com/group/lenstool?
---
You received this message because you are subscribed to the Google Groups "lenstool" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lenstool+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Xin Wang]

Hi Johan,

Thanks for your reply! I’ll definitely try combining the two possibilities. Just to double check, for the bayesian optimization method, i.e.,

inverse 3 float1 float2

(by default float1=0.5, float2=1000?)

So your suggestion no.1 is to reduce float1 from the default value to lower value, in order to explore more the parameter space in the burnin phase?

Best regards,

Xin Wang
--------------------------------------
Department of Physics and Astronomy
University of California, Los Angeles
CA 90095-1547, USA

On Mar 22, 2017, at 01:11, Johan Richard <nova...@gmail.com> wrote:

Hi,

Yes, the bayes.dat only covers the sampling phase, i.e. does not include any of the burnin. We have seen the effect you mention in the past, and it happens

when the full convergence is not achieved during the burnin phase. The bifurcation seems to come from the different MCMC chains.

To prevent this effect there are several possibilities:

- reduce the burnin rate in the "inverse" command, to explore more the parameter space in the burnin phase.

- adjust your priors if they are too large and include possible local minima in chi2.

- combination of the above.

Hope this helps,

Johan

On Wed, Mar 22, 2017 at 7:40 AM, Xin Wang <albert...@gmail.com> wrote:

Hi all,

Could I please know whether the data points recorded in the “bayes.dat” file are all after the burn-in phase? Below please find a plot showing chi2 values versus Nsample from the “bayes.dat” file given by one of my modeling attempts. Do you think the funky multi-furcation feature stems from incomplete burn-in phase? Thanks you guys so much!

<Screen Shot 2017-03-21 at 23.34.43.png>

Best regards,

Xin Wang
--------------------------------------
Department of Physics and Astronomy
University of California, Los Angeles
CA 90095-1547, USA

--
--
_____________________________________________________________________
LENSTOOL: http://projets.lam.fr/projects/lenstool/wiki
To post to this group, send email to lens...@googlegroups.com
To unsubscribe from this group, send email to lenstool-unsubscribe@googlegroups.com
For more options, visit this group at http://groups.google.com/group/lenstool?
---
You received this message because you are subscribed to the Google Groups "lenstool" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lenstool+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
--
_____________________________________________________________________
LENSTOOL: http://projets.lam.fr/projects/lenstool/wiki
To post to this group, send email to lens...@googlegroups.com

To unsubscribe from this group, send email to lenstool-u...@googlegroups.com

For more options, visit this group at http://groups.google.com/group/lenstool?
---
You received this message because you are subscribed to the Google Groups "lenstool" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lenstool+u...@googlegroups.com.

[Johan Richard]

Hi,

Yes, that's the suggestion. Note that adjusting the 'float2' value (increasing it) will only increase the number of MCMC

samples in the *sampling* (bayes.dat) phase, not the burn-in.

Johan

[Xin Wang]

Got it Johan! Thanks so much!

Best regards,

Xin Wang
--------------------------------------
Department of Physics and Astronomy
University of California, Los Angeles
CA 90095-1547, USA

To unsubscribe from this group, send email to lenstool-u...@googlegroups.com

For more options, visit this group at http://groups.google.com/group/lenstool?
---
You received this message because you are subscribed to the Google Groups "lenstool" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lenstool+u...@googlegroups.com.