genann_run() always return constant value about genann

[codeplea]

You need to loop through your training data many, many times. Not just once.

Each time, you should randomize the order that you train on your inputs.

After each training, you should call genann_run() and subtract that output from your expected output. Square this difference so that it's always positive, and then add it to a running sum for that training epoch. That's your error.

You should monitor the error for each loop through all the training data. It should start high, and then go down as you train.

You should also try 1 layer and 10 neurons to start. Get that working, then increase it and see if you get better results. 40 x 40 is huge, and you would probably need deep learning techniques for that.

Also, are you sure your outputs are correlated with your inputs? A neural network isn't psychic, it can only calculate something that is possible to calculate.

Anyway, there's a ton I left out, but that's a starting point.

Good luck, and let me know what you find out.

[korovkin]

i didn't try randomizing the order of the training set for each epoh yet, or calculating the err, as i just let it run for many epohs for about an hour or so the result is the same, at the end of the training session, the network output is always the same for any input.

…

On Mon, Jul 9, 2018 at 7:29 PM Lewis Van Winkle ***@***.***> wrote: You need to loop through your training data many, many times. Not just once. Each time, you should randomize the order that you train on your inputs. After each training, you should call genann_run() and subtract that output from your expected output. Square this difference so that it's always positive, and then add it to a running sum for that training epoch. That's your error. You should monitor the error for each loop through all the training data. It should start high, and then go down as you train. You should also try 1 layer and 10 neurons to start. Get that working, then increase it and see if you get better results. 40 x 40 is huge, and you would probably need deep learning techniques for that. Also, are you sure your outputs are correlated with your inputs? A neural network isn't psychic, it can only calculate something that is possible to calculate. Anyway, there's a ton I left out, but that's a starting point. Good luck, and let me know what you find out. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AECFGYresu7xOvOstnawc2GBjxtNa78dks5uFBGfgaJpZM4VIW91> .

-- Best Regards, H

[codeplea]

If you take my earlier advice, I'm sure you'll have better luck.

You can post the complete example with training data if you need more help.