coalescentLikelihood -INF when using UPGMA starting tree

[GoogleCodeExporter]

Hi Walter,
I have been trying to get BEAST v1.5.1 to work and am having problems.  I
have tested it using three .xml files that have all worked fine in v1.4.8.
 These three files differ only in that one uses a birth-death model, one
uses a constant population size coalescent model and the third uses an
expansion growth coalescent model.  The birth-death one runs in v1.5.1
fine, but the other two cause BEAST to crash.  The output is suggesting the
coalescent likelihood is -Infinity, which I guess is causing the crash.  In
one of the runs where it is reported among the error messages the
treelikelihood is a sensible number, so it is not that my starting tree is
too unlikely.  I have attached the output files to this email so you can
check the various error messages.
Cheers,
James

Original issue reported on code.google.com by dong.w.xie@gmail.com on 21 Sep 2009 at 3:57

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

Hi, James,

There maybe some changes in coalescentLikelihood calculator between 1.4.8 and 
1.5.1.
We need at least partial data with XML for debug.

Cheers,
Walter

Original comment by dong.w.xie@gmail.com on 24 Sep 2009 at 3:53

[GoogleCodeExporter]

CoalescentLikelihood is -Infinity

Original comment by dong.w.xie@gmail.com on 29 Sep 2009 at 4:06

[GoogleCodeExporter]

The initial posterior is zero: CompoundLikelihood(compoundModel;=(OneOnX(gtr1.ac
)=0.0, OneOnX(gtr1.ag)=0.0, OneOnX(gtr1.at)=0.0, OneOnX(gtr1.cg)=0.0, OneOnX(gtr
1.gt)=0.0, OneOnX(gtr2.ac)=0.0, OneOnX(gtr2.ag)=0.0, OneOnX(gtr2.at)=0.0, OneOnX
(gtr2.cg)=0.0, OneOnX(gtr2.gt)=0.0, OneOnX(gtr3.ac)=0.0, OneOnX(gtr3.ag)=0.0, On
eOnX(gtr3.at)=0.0, OneOnX(gtr3.cg)=0.0, OneOnX(gtr3.gt)=0.0, OneOnX(constant.pop
Size)=-1.9315, CoalescentLikelihood(coalescentLikelihood;=-Inf), CompoundLikelih
ood(compoundModel;=(TreeLikelihood(treeLikelihood;=-3904.0276, TreeLikelihood(tr
eeLikelihood;=-1123.2908, TreeLikelihood(treeLikelihood;=-536.886)
For more information go to <http://beast.bio.ed.ac.uk/>.
Exception in thread "main" java.lang.RuntimeException: ErrorLog: Maximum number
of errors reached. Terminating BEAST
        at dr.util.ErrorLogHandler.publish(Unknown Source)
        at java.util.logging.Logger.log(Unknown Source)
        at java.util.logging.Logger.doLog(Unknown Source)
        at java.util.logging.Logger.log(Unknown Source)
        at java.util.logging.Logger.severe(Unknown Source)
        at dr.app.beast.BeastMain.<init>(Unknown Source)
        at dr.app.beast.BeastMain.main(Unknown Source)

Original comment by dong.w.xie@gmail.com on 29 Sep 2009 at 4:08

[GoogleCodeExporter]

coalescentLikelihood -INF when using UPGMA starting tree, where tree prior is
constant, exponential.

Original comment by dong.w.xie@gmail.com on 5 Oct 2009 at 11:38

• Changed title: coalescentLikelihood -INF when using UPGMA starting tree

[GoogleCodeExporter]

This seems to be an issue with UPGMA producing zero internal branch lengths. 
There are two solutions to this - perturb the initial tree a bit (when 
sampling, zero internal branch lengths won't happen), or put a hack in the Tree 
priors to be robust to this (doesn't really matter how - just don't return a 
-Inf log likelihood). My preference is for the former. In fact the initial tree 
should be checked for zero internal branches and perturbed irrespective of how 
it is provided (i.e., for user trees too).

Original comment by ramb...@gmail.com on 10 Sep 2010 at 1:39

[GoogleCodeExporter]

Fixed by perturbing zero branches when loading the initial tree into TreeModel. 
This will also fix issues with user specified starting trees.

Original comment by ramb...@gmail.com on 10 Sep 2010 at 1:57

[GoogleCodeExporter]

Original comment by ramb...@gmail.com on 10 Sep 2010 at 1:57

• Changed state: Fixed