Information Technology Reference
In-Depth Information
...acdcbddbbadbcbdaddcdcdbddcdbbaab...
...bdbacabccbaacabadbacadbaabdbcccb...
...dcbacbcdacdccbcabadabc158358772736301
C = {1, 3, 0, 7, 2, 3, 8, 1, 2, 5} (9.8)
It has a fitness of 95 on the training set and 38 on the testing set, which
corresponds to an accuracy of 91.35% on the training set and 86.36% on the
testing set and, therefore, is better than the best results previously reported
for this dataset (Cestnik et al. 1987, Clark and Niblett 1987, Michalski et al.
1986). For instance, a Bayesian classifier achieved 89% accuracy on the
training set and 83% on the testing set, whereas the CN2 algorithm achieved
91% accuracy on training and 82% on testing (Clark and Niblett 1987).
But even better rules than the model (9.8) above can be learned from this
dataset by letting the system evolve for a larger number of generations. For
instance, the decision tree below, also with 28 nodes, was created after 146,833
generations and has a fitness of 95 on the training set (91.35% accuracy) and
39 on the testing set (88.64% accuracy):
MGOFRAdDBHbBBcabbcbdcbdccbcbbcadddb...
...dbcbddacabcbccbadcdcaccbabcbbcdd...
...dcbbaacbcdacadbbcbaaccbadbcdbaca...
...cadcdacaddbabcdcacdada458381830727089
C = {3, 10, 6, 5, 2, 9, 8, 9, 1, 8} (9.9)
And obviously even better solutions can be created if (larger) populations
are allowed to adapt for even longer periods of time. It all depends on how
accurate one wishes the model to be and how much time one wants to spend
on its design.
Let's now analyze other results obtained with a dataset containing only
nominal attributes.
9.3.4 The Postoperative Patient Problem
The postoperative patient problem is a complex real-world problem with
three different classes (admitted to general hospital floor “A”, intensive care
unit “I”, and safe to go home “S”) and eight nominal attributes (Table 9.9).
The original dataset contains a total of 90 instances distributed very asym-
metrically between the three classes (64 instances for class “A”, two for
class “I”, and 24 for class “S”). Previous results on this dataset report very
low accuracies of just 48% (Budihardjo et al. 1991).
