Biomedical Engineering Reference
In-Depth Information
Fig. 6.10
The average values
of quality functions versus the
number of the “generation”
Table 6.1
N
N of generation
N of individuals
Values of quality functions
Error number
1
10
19
160
18
2
10
7
160
18
3
10
5
160
18
4
9
3
160
18
5
7
24
160
18
6
6
21
160
18
7
3
7
160
18
8
1
8
160
18
9
0
2
198
19
10
0
3
198
19
features corresponding to the best “individuals” was considered the best to use for
subsequent work.
In the second experiment, we tested 11 generations, each having ten “indivi-
duals.” Figure
6.11
represents the dependences of the average values of quality
functions for each generation. The ten best “individuals” from all generations were
selected. Their parameters are given in Table
6.2
.
Results of recognition with these ten best “individuals” were tested on an
associative-projective neural network that contains 8,192 neurons. The quantity
of errors on the new network is given in Table
6.3
.
A neurocomputer (to be described in detail in Chapter 7) with high productivity
developed by Ukrainian and Japanese scientists from the company WACOM was
used in the experiments. The complete cycle of training and recognition required
several minutes. The results show that the evolutionary model makes it possible to
improve the feature set, but neurocomputer productivity was insufficient for a
detailed study of the properties of the evolutionary optimization algorithm. For
this purpose, neurocomputer productivity must be increased in such a way that the
cycle of training and recognition is approximately 1 second (or less). There are
technical capabilities for the realization of such a neurocomputer [
21
,
22
].
