In this chapter the most important results from the tests performed are
presented. A full description of these tests is given in
Appendix 
.
Some of the tests have been performed several times with different
data sets, but the conclusions may still not be strictly statistically valid.
Still, we believe the following
results are correct:
Four Equally Good Methods:
The three methods using voting and the weight of evidence measure
method all are equally good. Even though they did not produce the same
results in the tests, these four methods had about the same
classification accuracy for the test sets, and seems to be just as
good overall. The ``best rule'' method is significantly worse than
these four methods, and is not recommended for use.
Since the exponential, linear and squared voting give almost the same
results in all tests, the use could be restricted to the ordinary
linear one.
Top Node Rules Most Important:
The classification results were best overall when the rules from the
top node was among the rules used. If these rules were the only rules
used, no significant worse classification resulted.
The Methods Performed Differently on Subsets of Rules:
The weight of evidence measure method was better than voting when
applied to rules from the bottom nodes of the lattice. At the top of
the lattice there were no real differences.
Noise has Some Impact:
The classification is degraded when some degree of noise is added to
the system, but not very much. The systems ability to handle noise
seems satisfactory.
Learning Time Considerable:
The time it takes to create rules with RGEN is considerable when
compared to the time it takes to use these rules in RCLASS on a
training set of the same size.