Home ›

Encog 2.3(beta) adds Scaled Conjugate Gradient (Levenberg-Marquardt) Training

Submitted by jeffheaton on Mon, 12/21/2009 - 21:03
in

We just added a new training technique. I really need to test this out more, but so far it outperforms RPROP on anything I've given it. We also have it fully multithreaded, so I really think this is going to become the training algorithm of choice for Encog. It is called Scaled Conjugate Gradient (Levenberg-Marquardt) Training. So far I am really impressed with it! It is quite easy to use.

Train train = new ScaledConjugateGradient(network, trainingSet);

Java only... for now... this will be part of the official 2.3 release which should come out in the Jan/Feb timeframe.

Comments

SCG or Levenberg-Marquardt ?

sirandreus's picture
Submitted by sirandreus on Mon, 01/04/2010 - 07:01.

Have you implemented SCG according to Martin F. Moller ?

It is SCG, according to Moller

SeemaSingh's picture
Submitted by SeemaSingh on Sun, 01/17/2010 - 05:26.

My fault, this is actually SCG (scaled conjugate gradient) training, I was thinking the two were the same thing! I implemented it, and told Jeff it offered both. We will have to add Levenberg-Marquardt at some point, it seems like a powerful training method as well.

Copyright 2005 - 2010 by Heaton Research, Inc.. Heaton Research™ and Encog™ are trademarks of Heaton Research. Click here for copyright and trademark information.