RE: [R] optimisation procedure with flat log-likelihood

From: Gabor Grothendieck <ggrothendieck_at_myway.com>
Date: Sat 31 Jul 2004 - 01:06:21 EST

Two things to try:

  1. Try transforming the parameters. It may be that one or more parameters transformed by log or reciprocal, say, will improve the objective function from the optimizer's viewpoint.
  2. specify the gradient explicitly. If its complicated but not too complicated you might try a computer algebra package such as the free one, yacas, to get the derivative.
Date:   	Fri, 30 Jul 2004 15:41:27 +0200
From:   	Camarda, Carlo Giovanni <Camarda@demogr.mpg.de>
To:   	'r-help@stat.math.ethz.ch' <r-help@stat.math.ethz.ch>
Subject:   	[R] optimisation procedure with flat log-likelihood

Dear R-friends,
I use

optim(par=c(mystartingpoints), fn=myloglikelihoodfunction, gr=NULL, method=c("L-BFGS-B"), ## I would like to do not use any bounds
control=list(trace=6, ## just to see what it's going on maxit=c(20000)), ## to be sure the it
doesn't stop reaching the max iterations data=mydataset)

to optimize some demographic model. I assume that the log-likelihood is relatively flat because the estimated results are very similar to my starting values. In addition, I know the "real" parameters as I have used simulated data (which have been also found by using GAUSS and replicated by it).
I already tried various methods and also various starting values but it did not help. Can maybe anyone give me some suggestion what I could do?

Thanks,
Carlo Giovanni Camarda



R-help@stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html Received on Sat Jul 31 01:11:56 2004

This archive was generated by hypermail 2.1.8 : Wed 03 Nov 2004 - 22:55:22 EST