In our comparative assessments we use three data bases. The first two are composed of about 2500 states, randomly split into a learning and a test set comprising each 50% of the states. The third one contains 4041 states, and is split into a learning and test set comprising respectively 3245 and 796 states.
The computational requirements are given in CPU times on a 28MIPS SUN
Sparc2 workstation. Accuracies are assessed in terms of the
percentage of classification errors in the independent test set. The
size of the test set was fixed to about 1000 states, so as to
obtain sufficiently accurate error rate estimates. Thus, with error
rates (
) varying between 1% and 10%, the standard error of its
test set estimate lies between 0.3% and 1.0%.