Genetic Programming for Rule Discovery

In subsection 5.4.4 we saw that standard Genetic Programming (GP) for symbolic regression — where all terminals are real-valued variables or constants and all functions have real-valued inputs and output — can be used for classification, if the numeric value output at the root of the tree is properly interpreted. However, this kind of GP does not produce high-level, comprehensible IF-THEN rules in the style of the rules discovered by rule induction and decision-tree building algorithms (chapter 3) and GAs for rule discovery (chapter 6). As discussed in subsection 1.1.1, rule comprehensibility is important whenever the discovered knowledge is used for decision making by a human user.