Hi!
I have a dataset with about 10000 data points and I am trying to
construct an ANN with 2 inputs and 1 output.

I am using the Levenberg-Marquadt training algorithm with early
stopping technique and I initially set the training goal at 1e-6

and the max number of
epochs at 2000 all other parameters are set to default values
(in matlab).

I started with 1 hidden layer varying the number of neurons
from 1 to 30.
The Training and Validation MSE drops from about 0.0020
in 2-1-1 configuration to 1e-6 in to 2-10-1 configuration
and remains 1e-6 in all network configurations with more than
10 neurons.

When setting the training goal to 1e-9 I get MSE=1e-9 in
2-15-1 configuration and this figure remains
to all ANNs with more than 15 neurons in the hidden layer.

Questions!:
1) If I choose training goal of 1e-6, does the 2-10-1
configuration is the optimal one?

2) Do I have to try ANNs with more than 1 hidden layer?

3) What training goal should I set?

3) Since the results are more than excellent Is there any
problem with overfitting?

Hi Greg,
I really appreciate your help!
Could you please elaborate a little more on choosing the MSEgoal with
regard to variance?

E.g. the variance of my target data is about 180. According to var/
100, I should use a MSEgoal
of 1.8. Isnt this number a little odd compared to usual targets found
in journal papers 1e-6,1e-9 etc...??