Generation of Fuzzy Rules from Decision Trees
Lawrence O. Hall and Petter Lande
Department of Computer Science and Engineering University of South Florida Tampa, Fl. 33620
The paper introduces two ways to develop fuzzy rules, using decision trees, from data with continuous valued inputs and outputs. A key problem is dealing with continuous outputs. Output classes are created, then a crisp decision tree is created using a set of fuzzy output classes and letting each training example to partially belong to classes. Alternatively, a discrete set of fuzzy outputs classes is created that includes a selected group of overlaps, such as class A.75/class B.25. Training examples are then provided to a standard decision tree learning program, such as C4.5. In both cases, fuzzy rules are extracted from the resulting decision tree. Output classes must be created for a case in which examples belong to discrete but overlapping classes. We discuss tradeoffs of the two approaches to output class creation. An example of system performance uses a discrete set of overlapping classes on the Box-Jenkins gas furnace prediction problem and a function approximation problem. The learned rules provide effective control and function approximation.