Environmental Engineering Reference
In-Depth Information
accuracies were obtainedwithmoderate training threshold values
(0.3-0.5).
(Fig. 7.3D), and number of iterations (Fig. 7.3E). Overall, neural
network models were highly sensitive to the learning rate used,
as indicated by the large standard deviation. The models with the
learning rate ranging from 0.005 to 0.01 yielded higher classifica-
tion accuracies. As the learning rate increased, the classification
accuracy plunged by more than 20%, although another peak
did occur when the learning rate increased to 0.2 (Fig. 7.3C).
Overall, increasing the momentum value helped boost the classi-
fication accuracy (Fig. 7.3D), especially when this value was larger
than 0.6; by adjusting the value of momentum, the classification
7.3.5.3
Training parameters and
classification accuracy
Table7.4(Nos26-53)andFig.7.3(C,D,andE)showthe
overall classification accuracies in relation to the three train-
ing parameters, namely, learning rate (Fig. 7.3C), momentum
90
90
(a)
(b)
80
85
70
80
60
50
75
40
SD
=
6.24
70
30
20
Log-sigmoid function (Average
=
79.00, SD
=
2.65)
Tan-sigmoid function (Average
65
38.34, SD
5.79)
10
=
=
60
0
0
1
2
3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Number of Hidden Layers
Training Threshold
90
85
(c)
(d)
85
84
80
83
75
70
82
65
81
60
Raw Data
Average
=
71.04
SD
=
9.63
Linear Trend
y
=
−
2.8995x
+
85.534
R
2
=
0.6802
Raw Data
Average
Linear Trend
y
80
=
82.17
=
0.2122x
+
80.998
55
R
2
SD
=
0.94
=
0.4731
5
0.001
79
0
0.005
0.01
0.05
0.1
Learning Rate
0.15
0.2
0.25
0.3
0.1
0.2
0.3
0.4
Momentum
0.5
0.6
0.7
0.8
0.9
85
(e)
84
83
82
81
80
Raw Data
Average
Linear Trend
y
=
82.81
=
−
0.3945x
+
84.787
R
2
79
SD
=
1.48
=
0.5307
78
400
700
1000
1300
1600
1900
2200
2500
2800
Number of Iterations
FIGURE 7.3
Classification accuracies (
y
-axis) as related to different internal settings (
x
-axis): (A) number of hidden layers, (B)
activation function and training threshold, (C) learning rate, (D) momentum, and (E) number of iterations. For each figure, the
average accuracy and the standard deviation (SD) are provided, and for C-E, both raw data (solid lines) and linear trends
(dash lines) with linear regression equations and R square values are shown.
Search WWH ::
Custom Search