An alternative default strategy, namely using Minimum OSR search to generate new rules for uncovered observations before applying the default rule, has been studied. Our experimental results indicate:
Minimum OSR appears to be a desirable default strategy for decision list learners, particularly early in training. It is possible that a more search-intensive OSR default strategy than Minimum OSR would yield better results still, though as training increases we are liable to experience diminishing returns in any case. However, this strategy requires a new search for each uncovered observation and therefore is more expensive than the default rule strategy. Since the design of Minimum OSR has taken efficiency issue into account, it takes almost the same amount of time as the original BruteDL does on the small training sets. On the large data sets, BruteDL learns many homogeneous rules and the Minimum OSR is not applied very often. Therefore, although it would take some time on the large training sets for each Minimum OSR search, it is still reasonable for the learning system as a whole.
However, it still may be preferable to find other strategies that provide the same improvement without space and time costs higher than the default rule. Thus, we have studied two alternative strategies, which are described in the later parts of this thesis. Unfortunately, these strategies did not turn out to be competitive with the Minimum-OSR default strategy in terms of accuracy. These alternatives are comparable in accuracy to BruteDL's simple default rule strategy -- no considerable improvement has been revealed. Detailed descriptions of these alternative strategies are given in the following three chapters.
Jing Lin