You are here

Home » Neural Network Forums » Encog Project » Using Encog in DotNet

Weights not changing when training

Thomas's picture
published by Thomas on Tue, 07/24/2012 - 08:24

I have a .net neural network (in/linear/bias 209, tanh/bias 209, tanh/bias 547, out/tanh 64). I have recreated the network several times, but the error rate doesn't change through the iterations. When I later save the network, the weights on the EG file isn't changed (only the file header). Tried back, resilient and Manhattan trainer. Uncertain where to start debugging, or is there some known issues when designing NN i'm missing?

Neural Network Forums: 
Using Encog in DotNet
jeffheaton's picture

Submitted by jeffheaton on Sat, 07/28/2012 - 18:06.

Not able to reproduce this issue, also if Encog did lose the ability to update weights most unit tests would be failing.

I tried using the XORHelloWorld and added this line in the training loop.

System.out.println(((BasicNetwork)train.getMethod()).dumpWeights());

I was able to see the weights converge. You can find the hello world app here.

http://www.heatonresearch.com/wiki/Hello_World

Theme by Danetsoft and Danang Probo Sayekti inspired by Maksimer