282
F. Gottsche, F.-S. Olesen
judged by observing
the error for the training data which shows the general ability of the network to
produce the correct outputs for the given inputs and
• the error for the validation data which allows to detect over-fitting of the data,
which would diminish the ability of the networks to generalise.
After 19500 epochs the training was stopped and the network was saved
("Manual-NN" in Table 2). The error for the validation data is 0.16K, which is
small compared to the intrinsic accuracy of MODTRAN-3 (0.7K). The Manual-NN
showed
Table 1. Input and output layers as determined by the structure of the data
Input layer (45 neurons)
14 temperatures
14 dew-point temperatures
14 altitude levels
1 land-surface temperature
1 scan-angle
1 surface elevation
Table 2. The NNs determined manually (Vollmer et aI., 2000) and using ENZO
Layer
Manual-NN
ENZO-NN
Input
45 neurons
31 neurons
Hidden 1
44 neurons
22 neurons
Hidden 2
18 neurons
4 neurons
Output
1 neuron
1 neuron
2790 weights
286 weights
no signs of over-fitting. However, the large number of neurons and weights
indicated that better solutions might exist. Therefore, evolution was utilised to
search for NNs, which are closer to the global optimum. The Manual-NN served as