Use Enterprise Miner to fit a neural network.

1. Issue the following commands to SAS (with appropriate
   substitutions for "U:\").

LIBNAME GFLib 'U:\dm\sasdata' ; 

LIBNAME RCLib 'U:\dm\RC' ;

2. If necessary, apply macros to convert Excel data to
SAS (EXCELSAS, see MacroInstr.txt), to divide the data 
into training/test or training/validation/test subsets
(RANSPLIT, see MacroInstr.txt), or to explore the data
(UNIVAR and FREQ/FREQUENCY, see MacroInstr2.txt).

3. If necessary, assign numerical designations to 
categories and shorten variable names.  Illustrative
code may be found in MacroInstr5.txt. 

4. In SAS, go to the "Solutions" menu.  Go to "Analysis"
and then select "Enterprise Miner".  Do not invoke the
tutorial (unless you want to do so to satisfy your own
curiosity).  From the "File" menu, choose "New" and then
"Project".  A box will appear with "Name" and "Location"
fields as well as with "Create", "Cancel", and "Browse" 
buttons.  Press the "Browse" button and select a 
subdirectory such as 'U:\dm\RC'.  In the "Name" field,
type a name like 'NeuralNet'.  Then press "Create".  

5. In the left panel of the SAS Enterprise Miner window, 
you will see a diagram with 'NeuralNet' and, immediately 
below it, 'Untitled'.  You can right-click on 'Untitled' 
to assign a name such as 'Diabetes'.

6. Near the top of the SAS Enterprise Miner window, 
click the "Input Data Source" icon (far left) and, while 
holding the mouse button down, drag it into the right 
panel.  Assuming that you have training, validation, and 
test data, repeat this process twice so that you have 
three "Input Data Source" icons in the right panel.  

7. Double click on one of the "Input Data Source" icons 
in the right panel.  Press the "Select" button and then 
choose an appropriate library such as 'RCLib'.  Choose a 
training data set like 'DIABTR'.  In the "Role" field,
change "RAW" to "TRAIN".  Then click on the "Variables"
tab near the top of the "Input Data Source" box.  You 
can alter entries in the "Model Role" column by right
clicking and then selecting "Set Model Role".  You want
the response variable to be identified as "target", the
explanatory variables to be identified as "input", and 
the ID variable (if any) to be identified as "ID".  Any 
variables that you know will not be used may be 
identified as "rejected".  Also, make any necessary 
adjustments in the "Measurement" column.  When finished, 
click the white on red X in the upper right corner of 
the "Input Data Source" box and confirm the changes.  
Assuming that you have validation and test data, repeat 
this process twice (except that "RAW" will be changed to 
"VALIDATE" and "TEST").

8. Near the top of the SAS Enterprise Miner window, 
click the "Neural Network" icon (middle-right) and 
drag it into the right panel.  By holding the left 
mouse button down, 'draw' arrows from the data set 
icons to the "Neural Network" icon.  Also, drag the 
"Reporter" icon (right) into the right panel, then 
'draw' an arrow from the "Neural Network" icon to 
the "Reporter" icon.

9. Double click on the "Neural Network" icon.  Click 
the "General" tab.  For "Model selection criteria",
choose "Average [Squared] Error" if the target is 
continuous and "Misclassification Rate" if the 
target is categorical.  Click the "Basic" tab.  For 
"Runtime limit", choose 10 minutes.  Clicking the 
arrow for "Network architecture" will yield a menu 
with a "Hidden neurons" item.  You can set the 
number of hidden neurons yourself or let Enterprise 
Miner decide based on the anticipated level of noise 
in your data; some guidelines are provided below.  
When finished, close the "Neural Network" box and 
assign a model name such as 'DiabNN'. 

Noise	Continuous target	Categorical target
High	50% reduction in 	40% misclassified
	average squared error
Medium  70% reduction in	25% misclassified
	average squared error
Low	90% reduction in	10% misclassified
	average squared error
 
10. Right-click the "Neural Network" icon and choose 
"Run".  You will be asked if you want to view the 
results.  You can say "No" since you will acquire
what you need in the next step. 

11. Right-click the "Reporter" icon and choose "Run".  
You can "Open" the report now or view it later by 
noting to which subdirectory it has been saved.  For 
a continuous target, you are interested in the 
average squared error for the training, validation, 
and test data sets.  For a categorical target, you 
are interested in the misclassification rate for the 
training, validation, and test data sets.  For either 
kind of target, you can click the "Datastep Score 
Code" link to get SAS code that will allow you to 
generate predictions for each and every individual 
in any data set that has the same explanatory 
variables as the training data set.