This is a simple 3 layer BP network with a single Output.

All columns are Standardized (except used 3 StDev. instead
of 1)

Somewhere I read about training until the variance of the
error was something like 5% of the variance of the output.

My focus was to see how well the training was fitting the
input cases.

What I learned was the following: Cases with large output magnitudes
were fit long before small output magnitudes.

seemingly large numbers of hidden nodes, a good fit of the small
outputs was still not achieved.

Here are the results of a training with 3000 inputs, 25 inputs, 110
hidden, and a single ouput which had a max mag of 80:

Ouput Magnitude Intervals:
00-10
10-20
20-30
30-40
40-50
50-60
60-70
70-80

# Cases in each Interval:
1642
828
327
106
31
19
4
1

Variance Ratio for each Interval: (Output Variance / Error Variance)
.246
.029
.0085
.0072
.0021
.0022
.983
undefined

It is disconcerting to see such a poor fit on the first and most
populated interval.

QUESTIONs:

Is it reasonable to conclude that Output predictions over the
magnitude interval of [0,10] should be poor?

Does the extremely close fit of large mag outputs imply
memorization?
If yes, perhaps due to the small number of large mag outs, that may be
a non-issue.
__________________

I am going to investigate the following:

1) Train and predict only cases that are contained in the 0,10
interval and see if generalization is achieved.

2) Try adding duplicate rows to the input file for the 0,10 interval
and see if error variance can be improved.

Any comment would be greatly appreciated, and of course
Much Thanks in Advance,
Tom H.

This is a simple 3 layer BP network with a single Output.
For that case I discovered/developed/what ever something like a "sine

All columns are Standardized (except used 3 StDev. instead of 1)

Somewhere I read about training until the variance of the error was
something like 5% of the variance of the output.

My focus was to see how well the training was fitting the input cases.

What I learned was the following: Cases with large output magnitudes
were fit long before small output magnitudes. In fact, with even
seemingly large numbers of hidden nodes, a good fit of the small
outputs was still not achieved.

Here are the results of a training with 3000 inputs, 25 inputs, 110
hidden, and a single ouput which had a max mag of 80:

Ouput Magnitude Intervals:
00-10
10-20
20-30
30-40
40-50
50-60
60-70
70-80

# Cases in each Interval:
1642
828
327
106
31
19
4
1

Variance Ratio for each Interval: (Output Variance / Error Variance)
.246
.029
.0085
.0072
.0021
.0022
.983
undefined

It is disconcerting to see such a poor fit on the first and most
populated interval.

QUESTIONs:

Is it reasonable to conclude that Output predictions over the
magnitude interval of [0,10] should be poor?

Does the extremely close fit of large mag outputs imply
memorization?
If yes, perhaps due to the small number of large mag outs, that may be
a non-issue.
__________________

I am going to investigate the following:

1) Train and predict only cases that are contained in the 0,10
interval and see if generalization is achieved.

2) Try adding duplicate rows to the input file for the 0,10 interval
and see if error variance can be improved.

Any comment would be greatly appreciated, and of course
Much Thanks in Advance,
Tom H.