In the circle in the square problem the network must learn to classify points that are inside a circle and those that are not. In the sample files included with yawn the network must predict a "1" if the point is inside de circle and a "0" if not.
The figures bellow show the original test set and the predictions made by fuzzy ARTMAP and AppART. These figures were created with plot_cs.m, a matlab .m file included in the <yawn-dir>/data/circle-in-square directory. These results, in particular those of AppART are not optimal. For example, by increasing the <trainMatchVigilance> parameter a more precise but less generalizing approximation is obtained.
 |
 |
|
C-in-S test set
|
C-in-S test set contour
|
|
 |
|
Fuzzy ARTMAP prediction of the test set
|
Contour of the fuzzy ARTMAP prediction of the test
set
|
 | |
| AppART prediction of the test set | Contour of the AppART prediction of the test set |