Discussion:
how to decide maximum epochs number
Sheila the angel
2012-09-05 15:19:48 UTC
Permalink
Hello All,
I am new to fann and ANN and through examples I tried to understand the
parameters of ANN.
In example file mushroom.c max_epochs is set to 300 but with my data I am
not getting good accuracy with this.
Can anyone tell me how to decide maximum epochs number?
How does it effects the performance on network?
Thanks

--
Sheila
subscribed-+
2012-09-06 02:28:49 UTC
Permalink
On Sep 5, 2012, at 8:19 AM, Sheila the angel <from.d.putto-***@public.gmane.org> wrote:

Can anyone tell me how to decide maximum epochs number?
How does it effects the performance on network?

The number of (training) epochs is definitely domain specific. You
want to be sure that you have "finished" training, but you want to
avoid over-training on the data too.

A lot of people (myself included) use a threshold on the running
error rate. As soon as it starts to flatten out, I stop the training.

You can also use a "holdout" pattern for training / testing, too.
Carve out 1/3 of the training data, train on the remaining 2/3, and
then check the error rate on the 1/3 you took out. You can keep
repeating this method until your error rate drops to an acceptable
limit.

If you can't get the (current) configuration trained as well as you'd
like, there are other parameters you will probably have to change
(hidden layer size(s), learning rate, etc.).

Continue reading on narkive:
Loading...