Iterative Optimization of Rule Sets
Jiawei Du · 2010
Rule learning is one of the subfields in machine learning that is specialized in the generation of rules from the data. A rule set, which consists of a series of rules, can be seen as the experience that enables the system to do the same task more efficiently. Normally, the process of learning rule sets is called the building phase. It is suggested that a single rule in the building phase is optimized in most rule learning algorithms, while the improvement can also be gained for the entire rule set in a postprocessing phase. As a well-known algorithm, including the postprocessing phase, RIPPER is suitable for the benchmark. Moreover, two variations are also derived from the original RIPPER algorithm for comparison. The first one introduces a new pruning method and the second does a simplified selection criterion. At the end, all the algorithms mentioned in this thesis are implemented and validated in the simulation platform SeCo. In different parameter settings, some conclusions are made to optimize the rule set based on the simulation results.