Next: Classification of Sonar Data Up: Problems with a Small Previous: Prediction on Credit Card

The Ionosphere database contains data of radar signals, used to find structures in the ionosphere. It is a two-class prediction problem. The signal is given as 350 instances of 34 continuous values normalized between -1 and 1. The first 200 records were used for training the remaining 150 for testing (this is according to the previous usage of the data).

Training a single network $M=(34,1,[8,4])$ over 49 cycles (time needed 18'') led to a performance of 97.3%. The very low number of training cycles shows that the data is very easy to memorize and generalization is simple on the domain.

To compare all three network types the data was converted into 340 binary inputs, using the coding:

```[-1,-0.8)   => 0 0 0 0 0 0 0 0 0 1
[-0.8,-0.6) => 0 0 0 0 0 0 0 0 1 1
[-0.6,-0.4) => 0 0 0 0 0 0 0 1 1 1
[-0.4,-0.2) => 0 0 0 0 0 0 1 1 1 1
[-0.2,0)    => 0 0 0 0 0 1 1 1 1 1
[0,0.2)     => 0 0 0 0 1 1 1 1 1 1
[0.2,0.4)   => 0 0 0 1 1 1 1 1 1 1
[0.4,0.6)   => 0 0 1 1 1 1 1 1 1 1
[0.6,0.8)   => 0 1 1 1 1 1 1 1 1 1
[0.8,1)     => 1 1 1 1 1 1 1 1 1 1```

Table 7.2 shows the best results achieved with each type of network on the recoded data set:

 Network $P$G Training Time RAM based Logical network 92.5% 5'' $M=(340,1,[10,6])$ 97.3% 9'01'' $N=(340,2,7,small,π,(50,1,[4]),(7,2,[2]))$ 97.3% 2'21''

On the recoded data set the time difference is significant, but it must be recalled that the single network using the original data set was much faster. For this data set the recoding brought no advantage, in fact a significant disadvantage.

Next: Classification of Sonar Data Up: Problems with a Small Previous: Prediction on Credit Card

Albrecht Schmidt
Mit Okt 4 16:45:34 CEST 2000