*********** Quad.nns is a NeuNet Pro Sample File ***********

This file contains special properties that allow NeuNet Pro to recognize it
as authorized NeuNet sample data.  Anyone using the unlicensed version
of NeuNet Pro is welcome to experiment with this sample data.
Please do not modify this file, or it will lose its status as authorized sample data.

For further information about Neunet Pro and additional sample data,
please visit the NeuNet Pro website at http://www.cormactech.com/neunet

All of this data has been created by CorMac Technologies Inc.
CorMac Technologies Inc. does not guarantee the accuracy of the data.
This data is intended solely for experimental purposes.


    ***************  More about the Quad Data *******************


The Quad.nns database contains 3 tables: Quadratic, Quad_Shuff and Quad_Shuff_Noise.

Tables:
-------
Quadratic is the original unshuffled data.
Quad_Shuff is the same data shuffled randomly.
Quad_Shuff_Noise the the same shuffled data with the PlusNoise field added.


Data:
-----
The data was generated in a MS Excel97 spreadsheet based on the functions:
	
	f(x) = a + bx + cx^2

And:
	
	g(x) = f(x) ^ (1/3)

And:

	PlusNoise = g(x) * ( 0.1 * Rnd() + 0.95 )


A, B and C are 10 sets of 3 randomly chosen integers between 0 and 9 inclusive.
The sets used in this data are:
  A B C
  - - -
  0 0 0
  2 5 6
  4 6 6
  5 5 3
  6 1 4
  6 2 6
  6 6 1
  6 6 5
  7 3 0
  8 1 1


The data consists of 10 sets of 100 where x = 1..100.  For each of these 10 sets of x,
a, b and c are chosen from one of the 10 sets of 3 above.


Field Descriptions:
-------------------
X          Integers from 1 to 100
A          Integers from 0 to 9
B          Integers from 0 to 9
C          Integers from 0 to 9
Fx         Function f(x) as above
Gx         Function g(x) as above
PlusNoise  g(x) +- 5% (only in Quad_shuff_Noise table)


Training:
---------
The included example NeuNet Pro project file uses the Quad_Shuff_Noise table.
The data was configured with A, B, C And X as Inputs and PlusNoise as the prediction target.
The number of hidden nodes was set to 6 and the UserMins and UserMaxs were left at their defaults.
The training data range was set from 1 to 300 and the testing data was rows 301 - 1000.
The data was trained for approximately 3000 cycles, achieving 1.53% error.

Testing:
--------
None of the test set was used to train.  It consisted of rows 301 - 1000.
The normalized RMS error for this set was 1.57%.
Looking at the NeuNet Pro scatter graph, (Actual vs. Predicted) of this data,
we can see that larger numbers predict worse than smaller ones.
This is because the noise was added as a percentage.
(5% of 40 is greater than 5% of 4).