To see how well the Minimum OSR strategy works, we have run experiments
with our modified version of BruteDL.
The experiments are a 25-trial run of BruteDL with the Minimum OSR
strategy on the same data sets described in Chapter 2. In each trial, we
ran BruteDL with Minimum OSR on the same training and testing data that
we used in the corresponding trial in the experiments with BruteDL as
previously described. For example, in the experiments described in Chapter
2, we ran BruteDL (with default rule) on the 10% of the
original data as training data and tested it upon the 30% as testing data.
We then used Minimum OSR as a default strategy in conjunction with the same
decision list learned by BruteDL on the 10%
training set, and the result is compared to the result of the original
BruteDL to see whether any improvement is shown. The 
-test in the OSR
algorithm is applied to avoid the overfitting problem. Sometimes, this may
cause the learning system to learn rules that are too general. To examine
whether the 
-test is really needed, we have run Minimum OSR
both with and without it. When the 
-test is not used, the Minimum
OSR search terminates only when the information measure is not improved (i.e.,
the value of IM does not decrease) by expanding a node. The experimental
results are given in the following section.