Dear colleague,

Maybe it is interesting and helpful for You:http://penerbit.uthm.edu.my/ojs/index.php/JST/article/viewFile/262/141 as well as following http://www.mathworks.com/help/nnet/ref/trainlm.html . It is not my primary interest but hope I could suggest You some good readings.

Good luck. Sincerely,

Ljubomir Jacic

I see You want to apply levenberg-marquardt approach in order to get fast solution! Am I right? I do think this will be useful for You: http://crsouza.blogspot.com/2009/11/neural-network-learning-by-levenberg_18.html

Regards,

Ljubomir Jacic

Levenberg-marquardt method is computationally expensive. I would suggest some alternate approximations such as the one given by Schraudolph in "Fast Curvature Matrix-Vector Products for Second-Order Gradient Descent"

Although LM is computationally expensive, it is the most widely used optimization algorithm (It is the default algorithm in Matlab's ANN Toolbox). In most cases LM converges very fast (in few epochs) and outperforms other training algorithms (GD, etc.) by many orders. Also, LM uses an approximation to the Hessian matrix in the following Newton-like update.

http://www.codeproject.com/Articles/55691/Neural-Network-Learning-by-the-Levenberg-Marquardt