The following are used to describe properties of the data
sets in use:
| Shorthand | Explanation |
| rel. attr. | Number of attributes of relevance for class |
| irrel. attr. | Number of attributes not relevant for class |
| % max | Percentage of objects in most common class |
| training | Number of tuples in training set |
| test | Number of tuples in test set |
The data sets used have these characteristics:
| Data set | rel. attr. | irrel. attr. | % max | training | test |
| Hayes-Roth & Hayes-Roth Database | 3 | 2 | 38.6 | 132 | 28 |
| Postoperative Patient Data | 8 | 0 | 71.1 | 45 | 45 |
| Tic-Tac-Toe Endgame database | 9 | 0 | 65.3 | 47 | 911 |
Rule Generation
When generating reducts from RSES and RGEN, the following options were
set:
For RGEN, two additional parameters were used:
Classification Results
We used simple voting as the classification strategy for RSES.
The reason for this is that simple voting usually is best when
nominal attribute values are used, which was the case.
(Refer to [Syn] for different strategies.) For RGEN,
different classification methods were used, as shown in the overview
of methods in Table 
. The results of the tests are
given in Table .
| 1 | RSES - Simple voting |
| 2 | RGEN - Simple voting, use only top node of lattice (simulating RSES) |
| 3 | RGEN - Highest accuracy, all nodes |
| 4 | RGEN - Simple voting with linear weight, all nodes |
| 5 | RGEN - Simple voting with exponential weight correction of accuracy |
| 6 | RGEN - Simple voting with squared weight correction |
| 7 | RGEN - Measurement of evidence method |
| Data set | 1 | 2 | 3 | 4 | 5 | 6 | 7 | |
| Hayes-Roth | 85 | 82.1 | 92.9 | 89.3 | 89.3 | 89.3 | 89.3 | |
| Postoperative | 57 | 64.4 | 64.4 | 66.7 | 66.7 | 66.7 | 55.6 | |
| Tic-Tac-Toe | 71.1 | 73.3 | 65.4 | 72.0 | 72.3 | 71.9 | 70.1 |