8
§
PROC DMNEURL: Approximation to PROC NEURAL
COS: Iter=3 Crit=0.06277476: SSE=741.349891 Acc= 83.2886
------------ Activation= EXP (Stage=2) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
EXP: Iter=4 Crit=0.06131052: SSE=721.679089 Acc= 83.4396
The best accuracy increased from 83.29 to 83.47 and the (1,1) entry from 319 to 343
counts:
Classification Table for CUTOFF = 0.5000
Predicted
Activation
Accuracy
Observed
1
0
SQUARE
83.473154
1
343.0
846.0
721.118524
0
139.0
4632.0
EXP
83.422819
1
337.0
852.0
721.301910
0
136.0
4635.0
LOGIST
83.607383
1
337.0
852.0
724.687746
0
125.0
4646.0
TANH
83.607383
1
340.0
849.0
725.553808
0
128.0
4643.0
ARCTAN
83.607383
1
341.0
848.0
725.724668
0
129.0
4642.0
SIN
83.607383
1
340.0
849.0
725.889780
0
128.0
4643.0
GAUSS
83.372483
1
317.0
872.0
741.471407
0
119.0
4652.0
COS
83.322148
1
316.0
873.0
741.668156
0
121.0
4650.0
Even though SQUARE shows the best SSE, the accuracy rates for some other func-
tions (e.g. LOGIST) are slightly better:
Goodness-of-Fit Criteria (Ordered by SSE, Stage 2)
Run
Activation
SSE
RMSE
Accuracy
1
SQUARE
721.11852
0.347841
83.473154
8
EXP
721.30191
0.347885
83.422819
4
LOGIST
724.68775
0.348700
83.607383
2
TANH
725.55381
0.348909
83.607383
3
ARCTAN
725.72467
0.348950
83.607383
6
SIN
725.88978
0.348989
83.607383
5
GAUSS
741.47141
0.352715
83.372483
7
COS
741.66816
0.352762
83.322148
Component selection w.r.t. the residuals of the stage 2 starts the estimation of stage
3. Note, that the
Ê
¾
values become smaller and smaller.
Purpose of PROC DMNEURL
§
9
Component Selection: SS(y) and R2 (Stage=3)
Comp
Eigval
R-Square
F Value
p-Value
8
6938.083228
0.005571
33.383374
<.0001
20
5345.603436
0.004223
25.409312
<.0001
12
6136.575271
0.004059
24.517995
<.0001
Also the size of the objective function at the optimization results decreases:
------------ Optimization Cycle (Stage=3) --------------
------------ Activation= SQUARE (Stage=3) --------------
NOTE: ABSGCONV convergence criterion satisfied.
SQUARE: Iter=1 Crit=0.06049339: SSE=710.516275 Acc= 83.7081
------------ Activation= TANH (Stage=3) ----------------
NOTE: ABSGCONV convergence criterion satisfied.
TANH: Iter=4 Crit=0.06052425: SSE=710.396136 Acc= 83.7752
------------ Activation= ARCTAN (Stage=3) --------------
NOTE: ABSGCONV convergence criterion satisfied.
ARCTAN: Iter=3 Crit=0.06052607: SSE=710.489715 Acc= 83.7081
------------ Activation= LOGIST (Stage=3) --------------
NOTE: ABSGCONV convergence criterion satisfied.
LOGIST: Iter=6 Crit=0.06055936: SSE=711.054572 Acc= 83.6577
------------ Activation= GAUSS (Stage=3) ---------------
NOTE: ABSGCONV convergence criterion satisfied.
GAUSS: Iter=6 Crit=0.06111674: SSE= 719.41694 Acc= 83.3725
------------ Activation= SIN (Stage=3) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
SIN: Iter=3 Crit=0.06051959: SSE=710.308709 Acc= 83.8087
------------ Activation= COS (Stage=3) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
COS: Iter=6 Crit=0.06117044: SSE=719.262211 Acc= 83.3725
------------ Activation= EXP (Stage=3) -----------------
NOTE: ABSGCONV convergence criterion satisfied.
EXP: Iter=2 Crit=0.06051088: SSE=710.810558 Acc= 83.7081
The accuracy of the best fit impoves slightly from 83.47 to 83.79 and the size of the
(1,1) entry inceases from 343 to 364.
Classification Table for CUTOFF = 0.5000
Predicted
Activation
Accuracy
Observed
1
0
SIN
83.791946
1
364.0
825.0
709.778632
0
141.0
4630.0
TANH
83.758389
1
363.0
826.0