# Re: how to change convergence criteria

From: <arthur.gilmour_at_DPI.NSW.GOV.AU>
Date: Tue, 28 Oct 2008 09:42:01 +1100

Dear Arild

re: I'm having some problems getting convergence of models. LogL does not
change for the last 1000 (!) iterations and there are (apparently) no
changes in variance component estimates either (% change are all 0), but
yet the model does not converge. I suppose this is due to a very flat
likelihood profile so that the model can't find a optimum that satisfies a
change in the REML LogL of 0.002*current iteration number and for the
individual variance parameter estimates to change less than 1%.
Would it make sense to change these criteria to obtain convergence and how
would one code asreml to change these (i.e. one or both the LogL and the
var comp change)?

I make several observations.
First, the convergence criterion in is ASReml is just a rule it uses to
say whan to stop. You are free to choose another rule.
It is not uncommon to accept a result when the ASReml criterion has not
been fully met.

Second, there will be something that is stopping ASReml from deciding the
model has converged if you have done 1000
9iterations and the last 990 all have the same likelihood. Maybe two
parameters are oscillating (likelihood is flat).
Otherwise, the residual variance might be trying to go negative. In Luan
Shengs query, fitting the model '!r sire animal', the convergence
'failed'
but the sequence was

47 LogL= 557.211 S2= 0.40566E-03 1292 df : 2 components
restrained
48 LogL= 557.215 S2= 0.25724E-04 1292 df : 2 components
restrained
49 LogL= 557.211 S2= 0.40566E-03 1292 df : 2 components
restrained
50 LogL= 557.215 S2= 0.25724E-04 1292 df : 2 components
restrained
Final parameter values 292.46 692.01 1.0000

or from the .res file

Iteration 1 42 43 44 45 46
LogL 460.122 557.215 557.211 557.215 557.211 557.215
Change % 64 0 0 0 0 0
Adjusted 0 2 2 2 2 2
StepSz 0.141 0.500 0.500 0.500 0.500 0.500
9 G 0.10 4624.20 292.46 4624.20 292.46 4624.20
1481.1
10 G 0.10 10941.70 692.01 10941.70 692.01 10941.70
1481.1

Notice the LogL and parameters are oscillating near the boundary of the
parameter space because ASReml
keeps them from crossing the boundary.
I suspect this is happening for you as well.

May Jesus Christ continue to be gracious to you,

Arthur Gilmour, His servant .

Mixed model regression mapping for QTL detection in experimental crosses.
Computational Statistics and Data Analysis 51:3749-3764 at
http://dx.doi.org/10.1016/j.csda.2006.12.031

Skype: arthur.gilmour
mailto:Arthur.Gilmour_at_dpi.nsw.gov.au, arthur_at_cargovale.com.au
Principal Research Scientist (Biometrics)
NSW Department of Primary Industries
Orange Agricultural Institute, Forest Rd, ORANGE, 2800, AUSTRALIA

fax: 02 6391 3899; 02 6391 3922 Australia +61
telephone work: 02 6391 3815; home: 02 6364 3288; mobile: 0438 251 426

ASREML 2 is available from http://www.VSNi.co.uk/products/asreml
Register on the ASReml forum at http://www.vsni.co.uk/forum
ASReml-L Archives are at
https://listserv.dpi.nsw.gov.au/cgi-bin/wa.exe
Cookbook: http://uncronopio.org/ASReml