I would like first to state that it is beyond my capability to identify whether this is a BFGS issue or a R package problem.
I've been doing some mixed logit regression using the R package "mlogit". In total I have 32 random parameters and one fixed parameter to be estimated. Correlations are allowed among random parameters. With BFGS method, the code stopped after 13 iteration. Judging from the Chi-square value, it is far from converge. With BHHH method, however, it usually converge (although very slowly). Is this a package problem? or is it BHHH method more appropriate than BFGS in my context?
I doubt if it is related to the step size, but it's only my doubt. The code and data I've used are here.