Bug in BEAUti which prevents it from emitting the updown operator in some cases, which can slow convergence

[msuchard]

Originally reported on Google Code with ID 718

In a post on low ESS values for *BEAST analyses, Heled commented "There is a bug in
BEAUti which prevents it from emitting the updown operator in some cases, which can
slow convergence." which indicates to me that the operators should be included for
*BEAST analyses, but need to be added into the xml manually.

So, for those who are interested in improving ESS values and convergence of the species
tree parameters, here is the code I have added to my xml files to add in the up/down
operators that are no longer output by BEAUti. This should be added to the operators
section of the xml.

        <upDownOperator scaleFactor="0.75" weight="30">
            <up>
                <parameter idref="species.yule.birthRate"/>
            </up>
            <down></down>
        </upDownOperator>
        <upDownOperator scaleFactor="0.75" weight="3">
            <up></up>
            <down><speciesTree idref="sptree"/></down>
        </upDownOperator>
        <upDownOperator scaleFactor="0.75" weight="3">
            <up></up>
            <down><parameter idref="species.popMean"/></down>
        </upDownOperator>
        <upDownOperator scaleFactor="0.75" weight="3">
            <up></up>
            <down><parameter idref="speciesTree.splitPopSize"/></down>
        </upDownOperator>

You may need to change the parameter idrefs depending on your particular xml code.
This is for the Yule species tree prior, if you use the birth death prior the idref
would be <parameter idref="species.birthDeath.meanGrowthRate"/>.

 I have split the parameters so they each have their own operator, I don't know if
this is better or different than doing the following:

<upDownOperator scaleFactor="0.75" weight="3">
            <up>
                <parameter idref="species.yule.birthRate"/>
            </up>
            <down>
   <speciesTree idref="sptree"/>
   <parameter idref="species.popMean"/>
   <parameter idref="speciesTree.splitPopSize"/> 
   </down>
        </upDownOperator>

I decided to split them because other parameters that were in a single up/down operator
in previous releases of BEAUti (i.e. geneX.treeModel.allInternalNodeHeights) have their
own operator in newer version of BEAUti. The transition seemed to be around BEAUti
1.7.2 where the allinternalNodeHeights up/down operators were written to the xml file
twice in each format, grouped into one up or down and split into their own upDownOperators.

I hope that adding these operators helps others with mixing in their analyses.

Reported by rambaut on 2013-12-17 23:56:34

[msuchard]

Adding all the <upDownOperators> with only one parameter is the equivalent of just adding
extra scale operators for these parameters. This suggests the weights of these operators
should be increased by default. Need to check if the updown interation of these parameters
would also help.

Reported by rambaut on 2013-12-18 00:10:28

[msuchard]

The issue is not with the "<upDownOperators> with only one parameter".

What is missing is the *combined*  upDownOperator, which includes the species tree,
all gene trees and all the relevant rates (birth death, pop sizes and gene clock rates).

Reported by jheled on 2013-12-18 00:16:13

[msuchard]

But the single-parameter operators are what the user claimed improved mixing. It may
be worth increasing the weights of these too? I think these will mostly be cheap operators
to do (in terms of what needs to be reevaluated)? No harm in adding the full updown
back in (it isn't clear how it got lost in the first place).

Reported by rambaut on 2013-12-18 00:22:18

[msuchard]

Increasing the weight of the "species-tree" operators may hurt analysis with a smaller
number of loci. Ideally the overall weights should be a some fixed ratio (say 20%)
of the total operators weight. We are trying to resolve this in BEAST2 by introducing
an "operators block" which can have a relative weight to the other "blocks".

Reported by jheled on 2013-12-18 00:45:48

[msuchard]

Reported by rambaut on 2013-12-18 07:12:01

• Status changed: WontFix