approach was performed according to the flow chart shown in figure
54, whereby 10 neural networks were grown in each loop cycle. The complete
calibration data set was split into training (60 %), monitor (20 %) and selection
(20 %) subsets for each loop cycle. For R22 the framework stopped after 8 loop
cycles with a network topology consisting of 16 input neurons, 56 links and
15 hidden neurons organized in 4 hidden layers shown in figure
57. For R134a the framework stopped after 7 loop cycles with a network topology
consisting of 13 input neurons, 35 links and 8 hidden neurons organized in 3
hidden layers shown in figure 58. The predictions
of the validation data by these network topologies show the best results of
all multivariate calibration methods used for this data set with relative errors
of 1.50% for R22 and 2.37% for R134a (see table
4). The true-predicted plots show no bias and very low standard deviations
for all concentration levels (see figure 59). Compared
with the parallel approach the loop-based network topologies use rather many
input variables. It is also remarkable that the number of 3 respectively 4 layers
of hidden neurons is unusually high. Yet, the non-uniform network design helps
to keep the number of adjustable parameters low by building a sparse network
topology with only few links. The topologies of the grown neural networks show
that the common recommendation ,- to use only 1 or at the furthest 2 hidden
layers for fully connected networks is only a vague rule, as the growing neural
network algorithm automatically decides, how many hidden layers are optimal.
The good generalization ability demonstrates that the non-uniform topology efficiently
uses small networks and is superior to fully connected networks.
same test for chance correlation and reproducibility was performed for the
loop-based approach as already described for the genetic algorithm framework.
Thereby the network topologies are by far more reproducible than using single
runs of the growing neural network algorithm. The network for R22 of the second
run uses the same variables than the network of the first run except of one
variable being not used. The network for R134a uses the same variables of the
first run except of one variable, which was exchanged by another one. Both
networks did not use a random variable, whereby inside a loop cycle some networks
used a random variable but these networks were not selected for the next loop
cycle due to worse predictions of the selection data sets.
figure 57: Neural network with
4 hidden layers built by the loop-based framework for R22.
figure 58: Neural network with
3 hidden layers built by the loop-based framework for R134a.
figure 59: Predictions of the
validation data by neural networks optimized by the loop-based growing network