search

jeffheaton / encog-java-core

http://www.heatonresearch.com/encog
Other
743 stars 268 forks source link

Issue with NWR Weight Init in Encog3.1 #41

Closed seemasingh closed 12 years ago

seemasingh commented 12 years ago

There were some changes to the NWR weight initialization in 3.1. I do not believe either the current or original code really have it right. However, the original 3.0 code does seem to work better on some training sets. I can see that the current NWR is not touching every neuron, which it should.

This was discovered looking at #38

jeffheaton commented 12 years ago

DId an overhaul of the way NWR weight init works. I now do not get this error. I did about 10 analyst based trains in a row, one did settle into a local minima, but the rest did just fine.

PetrToman commented 12 years ago

Thank you, it appears to be working for NNs with my data again! :-)

seemasingh commented 12 years ago

Okay, sounds good!