Biomedical Engineering Reference
In-Depth Information
Table 5. Classification with fixed search: percent accuracy
Low Concentrations
Low Concentrations
Training (134)
Training (134)
Validation (39)
Validation (39)
Comercial (2)
Comercial (2)
Run1: Selected Variables
[12 159]
Run1: Selected Variables
[12 159]
125 (93.28%)
125 (93.28%)
23 (58.97%)
23 (58.97%)
0 (0%)
0 (0%)
Run2: Selected Variables
[23 67]
Run2: Selected Variables
[23 67]
129 (96.27%)
129 (96.27%)
26 (66.66%)
26 (66.66%)
1 (50%)
1 (50%)
Run3: Selected Variables
[102 129]
Run3: Selected Variables
[102 129]
129 (96.27%)
129 (96.27%)
25 (64.10%)
25 (64.10%)
0 (0%)
0 (0%)
ANN Configuration for low concentrations
Topology: 2/ 10 / 60 / 7
ANN Configuration for low concentrations
Topology: 2/ 10 / 60 / 7
learning rate: 0.0001 stop criterion: mse=5 or epochs=500.000
learning rate: 0.0001 stop criterion: mse=5 or epochs=500.000
High Concentrations
High Concentrations
86
86
43
43
21
21
Run1: Selected Variables
[12 159]
Run1: Selected Variables
[12 159]
82 (95.35%)
82 (95.35%)
31 (72.09%)
31 (72.09%)
19 (90.47%)
19 (90.47%)
Run2: Selected Variables
[23 67]
Run2: Selected Variables
[23 67]
81 (94.18%)
81 (94.18%)
31 (72.09%)
31 (72.09%)
19 (90.47%)
19 (90.47%)
Run3: Selected Variables
[102 129]
Run3: Selected Variables
[102 129]
81 (94.18%)
81 (94.18%)
33 (76.74%)
33 (76.74%)
21 (100%)
21 (100%)
ANN Configuration
Topology: 2 / 10 / 60 / 5 learning rate: 0.0001 stop criterion: mse=1 or epochs=500.000
ANN Configuration
Topology: 2 / 10 / 60 / 5 learning rate: 0.0001 stop criterion: mse=1 or epochs=500.000
one solution provided valid classification rates.
As it can be noted, classification results are very
similar, although the variables used for perform-
ing the classification are different.
ANNs models obtained are slightly worse
than those obtained using 176 wave numbers,
although the generalisation capabilities of the
best ANNs model are quite satisfactory, as there
is only one error when commercial beverages are
classified.
population consists of individuals with n genes,
where n is the amount of wave numbers that are
a priori considered enough for the classification
according to some external criterion. Each gene
represents one of the 176 wave numbers considered
in the IR spectra. The GA tries to find the subsets
of variables that perform the best classification
model using only the number of variables pre-
defined in the genotype.
In this case, the final number of variables is
given in advance, so some external criterion will
be needed. In order to simplify the comparison
of the results, in this work the final number of
variables is defined by the minimum number of
Fixed Search
This approach uses a real codification in the chro-
mosome of the genetic individuals. The genetic
Search WWH ::
Custom Search