Information Technology Reference
In-Depth Information
Tabl e 6 . 25
Topologies of the modular neural networks.
Dataset
# experts
K-MNN
ArtificialF2
3
[2:18:3] [2:18:2] [2:18:2]
Breast Tissue
3
[9:10:2] [9:12:2] [9:12:2]
CTG
4
[22:22:10] [22:18:9] [22:18:9] [22:26:3]
Diabetes
4
[8:10:2] [8:12:2] [8:10:2] [8:12:3]
Olive
4
[8:6:6] [8:12:4] [8:12:8] [8:12:3]
PB12
3
[2:6:2] [2:6:2] [2:2:2]
Sonar
3
[60:12:2] [60:12:2] [60:14:2]
Dataset
# experts
S-MNN
ArtificialF2
3
[2:18:3][2:18:2][2:16:2]
Breast Tissue
3
[9:4:2][9:14:2][9:6:2]
CTG
4
[22:18:10][22:22:10][22:22:10][22:22:3]
Diabetes
3
[8:18:2][8:12:2][8:10:3]
Olive
4
[8:12:4][8:12:4][8:4:3][8:6:3]
PB12
3
[2:5:2][2:5:2][2:2:2]
Sonar
3
[60:10:2][60:16:2][60:16:2]
Dataset
# experts
EC-MNN
ArtificialF2
3
[2:20:3][2:12:2][2:14:2]
Breast Tissue
3
[9:12:2][9:12:2][9:3:2]
CTG
4
[22:20:6][22:18:2][22:26:2][22:26:3]
Diabetes
4
[8:14:2][8:18:2][8:12:2][8:16:3]
Olive
4
[8:10:4][8:4:3][8:12:2][8:8:3]
PB12
3
[2:4:2][2:5:2][2:2:2]
Sonar
3
[60:12:2][60:12:2][60:12:2]
Tabl e 6 . 26 Average error rates (with standard deviations) for MNNs using differ-
ent clustering algorithms. Best results (lowest error rate) in bold.
Dataset
K-MNN
S-MNN
EC-MNN
ArtificialF2
16.40 (2.40)
15.32 (3.55)
14.70 (3.22)
Breast Tissue
58.95 (7.54)
33.53 (4.47)
32.79 (3.72)
CTG
22.90 (0.86)
23.91 (2.91)
20.67 (2.38)
Diabetes
24.45 (1.45)
23.96 (1.76)
23.89 (1.64)
Olive
49.11 (2.89)
5.20 (1.11)
4.74 (0.89)
PB12
7.23 (1.17)
7.28 (0.95)
7.25 (0.80)
Sonar
16.14 (3.43)
23.69 (4.57)
18.57 (3.40)
Table 6.27 shows the errors obtained using single neural networks (SNNs
— MLP's with one hidden layer). For the studied datasets the SNN solution
outperformed the MNN solution in only one dataset (CTG). In three other
datasets (ArtificialF2, Olive and Sonar) the SNN performance was statisti-
cally significantly worse that the best MNN solution (
t
-test
p
values below
0.005). We remind that the MNN approach is only effective, achieving better
