AIC is a popular model comparison measure (despite of its potential shortcomings).
I am wondering whether it is legit to compare AIC (or Akaike weights) for models that were fitted without requiring maximum likelihood. Consider e.g. that the EM algorithm for fitting Gaussian mixtures was terminated prematurely.
The R help page for AIC states: "... whereas AIC can be computed for models not fitted by maximum likelihood, their AIC values should not be compared."
Are there any other measures that can be used in this scenario?
Any references to papers/books that argue for it or against will be greatly appreciated.