Results of "appcr1" application demo

Hi,

I used "appcr1" application demo built in MATLAB to test how NN behave for character recognition example. I noticed that there are unstable and high error in recognising alphabet letters based on the noise provided.

I tried modifying some parameters like for example number of neurons in the input layer, momentum constant, required goal value in additoin to the number of iteration for training the network with noise, however, still the results are unstable with high error.

There are not too much discussion about this issue. I read the manual for the NN and I did some recommended modificatin (as mentioned above) based on the manual, but still with same results.

Is this example missing something? If I really need to modify the script further, then where should be the modification?

What if I want to compare the performance in term of error detection for two trainined network, how to represent this as plot?

Best Regards Nayef