Subject:
Backpropagation: getting better results
Author:
anonymous
Date:
11/20/2005
I'm trying to implement a neural network algorithm in C. Do not ask me why I didn't use an already coded library such as "fast neural network" under Linux because the answer is simple: learning.
Basicaly, the learning procedure I choosed for the network is of type "online" but I plan to use "bath mode" soon. Also, I don't use momentum value for instant because I will have to reimplement a new buffer to do that.
I trained the network with a few patterns in order to learn XOR rule but I can't get good results. Here's a graph showing the MSE calculated over time for a same pattern.
http://pierreluc.aqra.ca/pierre-luc/images/graph1.gif
I choosed 0.85 as step constant for that example and randomly presented 4 predetermined patterns during 30 epochs. That's what I got... Yes, the curve is droping but that's not enought to get goods results and it's seems to be staled there.
What should I improve ?
Could the problem be coming from my code ?
If you wanna take a look on it, here's the link:
http://pierreluc.aqra.ca/projet/BackPropNN1/
Thanks
Pierre-Luc
*Oh, I almost forgot the translation for a few words in my code:
cachee: hidden
sortie: output
entree: input
|