The arboretum procedure

 

The SCORE statement creates output data sets.  The DATA= option specifies

the data set you want to score. The  OUTFIT= option creates a data set containing

fit statistics.

       score data=sampsio.dmtbase outfit=netfit   

---

 

The OUT=option identifies the output data for predicted values and residuals.

The RENAME= option renames the variables in the OUT= data set containing predicted

values and residuals.

    out=netout(rename=(p_logsal=predict r_logsal=residual));

    title 'NN:1 Hidden Unit, Direct Connections, 

           and Reduced Input Set';

run;

---

 

PROC PRINT lists selected variables from the OUTFIT=  data set.

proc print data=netfit noobs label;

   where _name_ = 'LOGSALAR';

   var _iter_ _pname_ _tmse_  _trmse_ _tmax_;

   title 'Partial Listing of the Score OUTFIT= Data Set';

run;

---

 

PROC GPLOT plots diagnostic plots for the scored data set.

proc gplot data=netout;

    plot logsalar*predict / haxis=axis1 vaxis=axis2;

       symbol c=black i=none v=dot h=3 pct;

    axis1 c=black width=2.5;

    axis2 c=black width=2.5;

    title 'Diagnostic Plots for the Scored Test Baseball Data';

    plot residual*predict / haxis=axis1 vaxis=axis2;

run;

quit;

---

The NEURAL Procedure

Example 3: Neural Network Hill-and-Plateau Example (Surf Data)

Features

Specifying input, hidden, and output layers

q   

Scoring Data with the Score Statement

q   

Outputting Fit Statistics

q   

Plotting the Posterior Probabilities

q   

This example demonstrates how to develop a neural network model for a continuous target. A multilayer perceptron architecture is employed

with 3 and then 30 hidden units. The example test data set is named SAMPSIO.DMDSURF (Surf Data). It contains the interval target HIPL,

and two interval inputs X1 and X2. The data set was artificially generated as a surface containing a hill and a plateau. The hill is easily learned

by an RBF architecture. The plateau is easily learned by an MLP architecture.

The SAMPSIO.DMTSURF data set is a test data set that is scored using the scoring formula from the trained model. The

SAMPSIO.DMDSURF and SAMPSIO.DMTSURF data sets and the SAMPIO.DMDSURF catalog are stored in the sample library.

Program: 3 Hidden Units

proc g3d data=sampsio.dmdsurf;plot x2*x1=hipl 

         / grid side ctop=blue caxis=green 

           ctext=black zmin=-1.5 zmax=1.5;

  title 'Plot of the Surf Training Data';

  footnote 'Hill Plateau Response Surface';

 run;

 

title 'Hill & Plateau Data';

%let hidden=3;

proc neural data=sampsio.dmdsurf 

           dmdbcat=sampsio.dmdsurf 

           random=789;

   input x1 x2 / id=i;

   target hipl / id=o;

 

   hidden &hidden / id=h;

   prelim 10;

   train maxiter=1000 outest=mlpest;

   score data=sampsio.dmtsurf out=mlpout outfit=mlpfit;

   title2 "MLP with &hidden Hidden Units";

run;

 

proc print data=mlpfit noobs label;

   var _tase_  _tasel_ _taseu_;

   where _name_ ='HIPL';

   title3 'Fit Statistics for the Test Data';

 run;

 

proc gcontour data=mlpout;

   plot x2*x1=p_hipl / pattern ctext=black coutline=gray;

   pattern v=msolid;

   legend frame;

   title3 'Predicted Values';

   footnote;

run;

---

 

proc g3d data=mlpout;

   plot x2*x1=p_hipl / grid side ctop=blue 

                       caxis=green ctext=black 

                       zmin=-1.5 zmax=1.5;

 run;

Output: 3 Hidden Units

PROC GCONTOUR Plot of the Surf Training Data

PROC Neural Output

                              Hill & Plateau Data

                                  MLP with 3 Hidden Units

    Iteration     Pseudo-random     Objective

     number       number seed        function

          0               789         0.04023

          1         761237432         0.03673

          2        1092694980         0.04187

          3         577625332         0.05216

          4         261548896         0.03593

          5         616485149         0.04030

          6         692687363         0.03973

          7        1510804008         0.04141

          8        1385020003         0.03583

          9        1070679467         0.04167

                              Hill & Plateau Data

                                 MLP with 3 Hidden Units

---

                               Optimization Start

                               Parameter Estimates

 ------------------------------------------------------------------------------

   Parameter       Estimate   Gradient Label

 ------------------------------------------------------------------------------

    1 X1_H1        6.064004 -2.6744E-6 X1 -> H1                                

    2 X2_H1        0.880274 5.28411E-6 X2 -> H1                                

    3 X1_H2        0.048809 5.37355E-7 X1 -> H2                                

    4 X2_H2       -4.988958 6.03873E-7 X2 -> H2                                

    5 X1_H3       -5.916343 -6.0486E-6 X1 -> H3                                

    6 X2_H3        0.730854 -0.0000207 X2 -> H3                                

    7 BIAS_H1     -3.004936 -0.0000105 BIAS -> H1                              

    8 BIAS_H2      1.791982 2.17127E-6 BIAS -> H2                              

    9 BIAS_H3      0.864474 -0.0000126 BIAS -> H3                              

   10 H1_HIPL     -0.261095  0.0000179 H1 -> HIPL                              

   11 H2_HIPL     -0.484358 -0.0000107 H2 -> HIPL                              

   12 H3_HIPL     -0.265490  0.0000123 H3 -> HIPL                              

   13 BIAS_HIP    -0.490112 -0.0000182 BIAS -> HIPL                            

                   Value of Objective Function = 0.0358271766

                              Hill & Plateau Data

                                 MLP with 3 Hidden Units

                        Levenberg-Marquardt Optimization

        *** Termination Criteria *** 

        Minimum Iterations  . . . . . . . . . . . . . . .         -1 

        Maximum Iterations  . . . . . . . . . . . . . . .       1000 

        Maximum Function Calls. . . . . . . . . . . . . . 2147483647

        Maximum CPU Time  . . . . . . . . . . . . . . . .     604800

        ABSGCONV Gradient Criterion . . . . . . . . . . .  0.0000100 

        GCONV Gradient Criterion  . . . . . . . . . . . .       1E-8

        GCONV2 Gradient Criterion . . . . . . . . . . . .          0

        ABSFCONV Function Criterion . . . . . . . . . . .          0 

        FCONV Function Criterion  . . . . . . . . . . . .  0.0001000 

        FCONV2 Function Criterion . . . . . . . . . . . .          0 

        FSIZE Parameter . . . . . . . . . . . . . . . . .          0

        ABSXCONV Parameter Change Criterion . . . . . . .          0 

        XCONV Parameter Change Criterion  . . . . . . . .          0 

        XSIZE Parameter . . . . . . . . . . . . . . . . .          0

        ABSCONV Function Criterion  . . . . . . . . . . . -1.341E154

        *** Other Control Parameters ***

        Trust Region Initial Radius Factor  . . . . . . .    1.00000

        Singularity Tolerance (SINGULAR)  . . . . . . . .       1E-8

                              Hill & Plateau Data

                                 MLP with 3 Hidden Units

                        Levenberg-Marquardt Optimization

                         Scaling Update of More (1978)

                        Number of Parameter Estimates 13

Optimization Start: Active Constraints= 0  Criterion= 0.036 

Maximum Gradient Element= 0.000 Radius= 1.000

         Iter rest nfun act   optcrit  difcrit maxgrad  lambda     rho

            1    0    2   0    0.0358 1.989E-7 0.00002       0   0.568

            2    0    3   0    0.0358 1.764E-7 0.00002       0   0.698

---