I have an OCR program where I'm using an Encog NN. I started out with Encog 2.5. When I trained the NN down to an error of 0.01, I had a good recognition of the training set. The problem was that Resilient propagation didn't work. The error rate would go down a bit (to 0.03-0.08) and would just get stuck there and not reduce further. I had to use BP which was very slow.
Now I upgraded to Encog 3.1. RP works, gets the error down to 0.01, but in trying to recognize the training set, the recognition is less than 50%. Sometimes much less. I tried to improve it, fiddled with the network, but no good. Then I reduced the training error to 0.001 and suddenly it recognized perfectly. Same as 2.5 did with a 0.01 error. Is this an expected change between 2.5 and 3.1?