check convergence in BreedR
Victor Papin
In some cases, after a certain execution time, the remlf90() command returns an error message indicating that the model was unable to converge.
Otherwize, the command remlf90() returns the results of the model. In the latter case, I wonder whether it's still possible that the model didn't fully converge. Is it possible, in the results object, to check whether convergence has been achieved (for example, can we see the evolution of true likelihood)?
Facundo Muñoz
Dear Victor,
breedR's function remlf90() actually wraps two different REML algorithms (AI and EM), depending on the 'method' argument.
EM is guaranteed to converge, by construction. AI, instead, is typically faster, but more prone to divergences.
One practical way of checking convergence is thus, to check whether both methods yield similar results. Although this is not always possible, as some breedR models only work with EM.
You can check the evolution of the likelihood by inspecting the contents of the element 'reml.out' in the object resulting from the remlf90 call. This contains the output from the algorithm in text form and includes many details, including the log-likelihood at each step. You can parse this output to extract the values for inspection.
Best,
ƒacu.-
--
Report issues at https://github.com/famuvie/breedR/issues
---
You received this message because you are subscribed to the Google Groups "breedR" group.
To unsubscribe from this group and stop receiving emails from it, send an email to breedr+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/breedr/dcf9a59e-72e5-43eb-ae41-94afc70528e0n%40googlegroups.com.
Victor Papin
To view this discussion on the web visit https://groups.google.com/d/msgid/breedr/2836b273-a1ae-e0a6-2a3b-d9576238ecbb%40cirad.fr.