Forums

ARMA modeling

Started by Tom June 13, 2007
On Jun 17, 7:35 pm, John_W_Her...@yahoo.com (John Herman) wrote:
> If you are looking for Levenburg-Marquardt, I would suggest levmar. > > http://www.ics.forth.gr/~lourakis/levmar/
I tried the version in the link above. It wasn't as robust as the lmfit.c version. When starting out the sum of squared errors can be very big and cause floaiting point overflow problems. This occured in our application when there are 2000+ data points. lmfit.c didn't have this problem as it had extra code to check the size of the squared error and kept a couple accumulators to keep them from over flowing. Finding this problem is how I learned what the algorithm was doing. I recommend https://sourceforge.net/projects/lmfit/ The code is ugly c code converted from fortran but it works well and is easy to implement. If there are other versions that work equally well then it would be good to hear. This post has saved the world 100s if not 1000s of man hours debugging code. I don't recommend writing this from scratch if you don't have too. One's numerical methods and programming techniques better be good. That is why a bypassed suggesting the books and directed Tom directly to the code. Peter Nachtwey