6
§
PROC DMNEURL: Approximation to PROC NEURAL
------------ Optimization Cycle (Stage=1) --------------
------------ Activation= SQUARE (Stage=1) --------------
NOTE: ABSGCONV convergence criterion satisfied.
SQUARE: Iter=1 Crit=0.06280042: SSE=741.517369 Acc= 83.1376
------------ Activation= TANH (Stage=1) ----------------
NOTE: ABSGCONV convergence criterion satisfied.
TANH: Iter=4 Crit=0.06299539: SSE=748.120566 Acc= 83.0705
------------ Activation= ARCTAN (Stage=1) --------------
NOTE: ABSGCONV convergence criterion satisfied.
ARCTAN: Iter=5 Crit=0.06300879: SSE=748.255575 Acc= 83.0705
------------ Activation= LOGIST (Stage=1) --------------
NOTE: ABSGCONV convergence criterion satisfied.
LOGIST: Iter=7 Crit=0.06302038: SSE=749.059872 Acc= 83.1208
------------ Activation= GAUSS (Stage=1) ---------------
NOTE: ABSGCONV convergence criterion satisfied.
GAUSS: Iter=37 Crit=0.06583974: SSE=783.332834 Acc= 82.2987
------------ Activation= SIN (Stage=1) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
SIN: Iter=5 Crit=0.06294376: SSE=747.863259 Acc= 83.0201
------------ Activation= COS (Stage=1) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
COS: Iter=6 Crit=0.06575231: SSE=782.280952 Acc= 82.2819
------------ Activation= EXP (Stage=1) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
EXP: Iter=5 Crit=0.06318155: SSE=746.822973 Acc= 83.1208
The best accuracy went up from 81.61 to 83.28, the (1,1) entry from 229 to 319 counts
(which is still less than that of TANH in the first stage):
Classification Table for CUTOFF = 0.5000
Predicted
Activation
Accuracy
Observed
1
0
SQUARE
83.288591
1
319.0
870.0
741.977606
0
126.0
4645.0
EXP
83.187919
1
304.0
885.0
747.249371
0
117.0
4654.0
SIN
83.120805
1
306.0
883.0
748.284252
0
123.0
4648.0
TANH
83.221477
1
309.0
880.0
748.436926
0
120.0
4651.0
ARCTAN
83.255034
1
310.0
879.0
748.614897
0
119.0
4652.0
LOGIST
83.171141
1
305.0
884.0
749.309756
0
119.0
4652.0
COS
82.583893
1
309.0
880.0
781.342717
0
158.0
4613.0
GAUSS
82.567114
1
309.0
880.0
Purpose of PROC DMNEURL
§
7
782.484989
0
159.0
4612.0
Goodness-of-Fit Criteria (Ordered by SSE, Stage 1)
Run
Activation
SSE
RMSE
Accuracy
1
SQUARE
741.97761
0.352835
83.288591
8
EXP
747.24937
0.354087
83.187919
6
SIN
748.28425
0.354332
83.120805
2
TANH
748.43693
0.354368
83.221477
3
ARCTAN
748.61490
0.354410
83.255034
4
LOGIST
749.30976
0.354575
83.171141
7
COS
781.34272
0.362074
82.583893
5
GAUSS
782.48499
0.362339
82.567114
Here starts the third stage (stage=2):
Component Selection: SS(y) and R2 (Stage=2)
Comp
Eigval
R-Square
F Value
p-Value
1
15337
0.006514
39.068994
<.0001
5
8117.555354
0.005566
33.564983
<.0001
7
7371.205837
0.005429
32.918782
<.0001
When fitting the second order residuals the average value of the objective function
dropped from 0.063 to 0.061.
------------ Optimization Cycle (Stage=2) --------------
------------ Activation= SQUARE (Stage=2) --------------
NOTE: ABSGCONV convergence criterion satisfied.
SQUARE: Iter=1 Crit=0.06128736: SSE=720.996529 Acc= 83.4228
------------ Activation= TANH (Stage=2) ----------------
NOTE: ABSGCONV convergence criterion satisfied.
TANH: Iter=5 Crit=0.06164516: SSE= 726.12791 Acc= 83.6242
------------ Activation= ARCTAN (Stage=2) --------------
NOTE: ABSGCONV convergence criterion satisfied.
ARCTAN: Iter=5 Crit=0.06167716: SSE=726.468943 Acc= 83.6074
------------ Activation= LOGIST (Stage=2) --------------
NOTE: ABSGCONV convergence criterion satisfied.
LOGIST: Iter=8 Crit=0.06157175: SSE=725.393203 Acc= 83.5403
------------ Activation= GAUSS (Stage=2) ---------------
NOTE: ABSGCONV convergence criterion satisfied.
GAUSS: Iter=2 Crit=0.06272854: SSE=740.745518 Acc= 83.3054
------------ Activation= SIN (Stage=2) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
SIN: Iter=6 Crit=0.06164909: SSE=726.185402 Acc= 83.6074
------------ Activation= COS (Stage=2) -----------------
NOTE: ABSGCONV convergence criterion satisfied.