284
....
~
....
w
(J)
::2:
CI:
'ä)
(J)
1ij
(I)
Ioe _
H. Wilson . F. Recknagel
o hidden nodes
2 hidden nodes
5 hidden nodes
o
I
I
I
I
I : I
I
I
I
I : I
I
I
I
I
0.5 0
2 1.5 1 0.5 0: 2 1.5 1 0.5 0: 2 1.5 1
Error at which Training Stopped
= Bootstrap aggregate model RMS error
I
~
= Distribution of RMS errors
.L
o
= Outlier
Figure 14.11. Example of the effect of training time and the number of hidden
layer nodes on training set RMS error. Lake Kasumigaura same day model.
The effect of hidden layer neurons on the 30 day ahead Myponga model can be
clearly illustrated by comparison of the time series plots of Figures 14.9b and
14.10b. Compared to the 5 hidden node model, the 0 hidden node model fails to
predict the magnitudes, timing of onset, or the duration of bloom events. On the
other hand, comparison of the time series plots of predictions for the 0 and 5
hidden node Lake Kasumigaura same day models CFigures 14.5c and 14.12
respectively) shows subtler differences. Close examination shows that the 5
hidden node model better predicts peak biomass in some years Ce.g. 1983, 1984),
although is more prone to slight overestimations Ce.g. 1987, 1992). In some years,
the 5 hidden node model is also somewhat better at following the patterns in
chlorophyll a dynamics Ce.g. 1986, 1987). These types of distinctions between the
....
~
....
w
(J)
::2:
CI:
'ä)
(J)
1ij
(I)
Ioe _
H. Wilson . F. Recknagel
o hidden nodes
2 hidden nodes
5 hidden nodes
o
I
I
I
I
I : I
I
I
I
I : I
I
I
I
I
0.5 0
2 1.5 1 0.5 0: 2 1.5 1 0.5 0: 2 1.5 1
Error at which Training Stopped
= Bootstrap aggregate model RMS error
I
~
= Distribution of RMS errors
.L
o
= Outlier
Figure 14.11. Example of the effect of training time and the number of hidden
layer nodes on training set RMS error. Lake Kasumigaura same day model.
The effect of hidden layer neurons on the 30 day ahead Myponga model can be
clearly illustrated by comparison of the time series plots of Figures 14.9b and
14.10b. Compared to the 5 hidden node model, the 0 hidden node model fails to
predict the magnitudes, timing of onset, or the duration of bloom events. On the
other hand, comparison of the time series plots of predictions for the 0 and 5
hidden node Lake Kasumigaura same day models CFigures 14.5c and 14.12
respectively) shows subtler differences. Close examination shows that the 5
hidden node model better predicts peak biomass in some years Ce.g. 1983, 1984),
although is more prone to slight overestimations Ce.g. 1987, 1992). In some years,
the 5 hidden node model is also somewhat better at following the patterns in
chlorophyll a dynamics Ce.g. 1986, 1987). These types of distinctions between the