Path Sampling - Error: Log Likelihood reaches Infinity - no calculation of marginal likelihood estimate possible

Sarah N.

unread,

Jun 23, 2016, 5:12:28 AM6/23/16

to beast-users

Dear Beast-Users,

I have a problem using PathSampling in BEAST2.

Everytime I am running a PathSampling Analysis that has more than 14 steps the LogLikelihood reaches infinity in the last steps. As a consequence it is not possible to calculate the marginal likelihood estimate etc.
It happens to me when I use the PathSampling via the appstore, but also when I generate the Xml.file myself.
Here for example I wanted to calculate 40 steps. It stopped at step 35. When I wanted to calculate 60 steps it stopped at 54.
I slightly increased the numbers of steps from 4 to 16 to see, what is possible. 14 steps worked fine but when I tried 16, it stopped at 15.

It looks like that

         45000    -73954.9235        25.2       -73562.0962      -392.8272 37m15s/Msamples
          46000    -74768.6350        25.7       -74365.7411      -402.8939 37m17s/Msamples
          47000    -73125.3805        26.3       -72724.1990      -401.1815 37m18s/Msamples
          48000    -83522.6890        27.8       -83128.9918      -393.6971 37m21s/Msamples
          49000   1246081.9561        45.0      1246488.0531      -406.0969 37m15s/Msamples
          50000       Infinity              N       Infinity      -413.1727 37m18s/Msamples
          51000       Infinity              N       Infinity      -413.1727 37m23s/Msamples
          52000       Infinity              N       Infinity      -413.1727 37m26s/Msamples
          53000       Infinity              N       Infinity      -413.1727 37m30s/Msamples
          54000       Infinity              N       Infinity      -413.1727 37m29s/Msamples

And at the end:

Loading /tmp/step/step35/likelihood.log, burnin 50%, skipping 500 log lines

|---------|---------|---------|---------|---------|---------|---------|---------|

*****************************marginalLs[35 ] = Infinity

******************************************************

Calculating statistics

|---------|---------|---------|---------|---------|---------|---------|---------|

********************************************************************************

Loading /tmp/step/step36/likelihood.logFile error: \tmp\step\step36\likelihood.log (Das System kann die angegebene Datei nicht finden)

BEAST has terminated with an error. Please select QUIT from the menu.

Does someone of you knows why this happens? And what could help to avoid it?

How many steps do I need to run to have reliable results? I guess 12 is not enough as the marginal L estimate is still changing.

And is there maybe a way to restart the analysis from where it stopped and then calculate the marginal L estimate?

Thank you for your answers and advices!

Bests,
Sarah

Marie

unread,

Dec 20, 2016, 4:18:11 AM12/20/16

to beast-users

Hi all,

I have the same problem, I try to get 100 steps, but it always stops at 66 for 1 of them and 69 for the other...and Marginal Likelihood = Infinity for the last step.

Does someone would have an idea of why this happens and what I could do to get my 100 steps ?

Thanks a lot for your help,

Best,

Marie

Remco Bouckaert

unread,

Dec 20, 2016, 12:22:23 PM12/20/16

to beast...@googlegroups.com

Hi Marie,

It is hard to say something specific, but large positive weighted posteriors indicate numeric instability, which is typically caused by parameters becoming very large or very small (near zero). Perhaps you can have a look at the trace file that is stored in the step-directory of the first step that misbehaves and see whether there is a parameter that behave like that at the point where the likelihood becomes very large, and revisit the priors on the parameter.

Cheers,

Remco

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at https://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.

Guy Baele

unread,

Dec 20, 2016, 4:25:38 PM12/20/16

to beast-users

You may want to double check if you have defined proper priors for all the parameters you're estimating.

Improper priors are known to lead to numerical instabilities more frequently.

Best regards,