Compensating for Sparse Data in Evolutionary Generation of Fuzzy Models

Evolutionary techniques have proven to be a successful strategy for generating fuzzy rule bases from training data. The locality of fuzzy decompositions permits a local evolutionary strategy consisting of an independent evolutionary generation of each rule. The fitness of a rule is determined by the training data within a neighborhood called the region of inclusion of the rule. When the amount of training data is limited, some local regions may not contain training data. This research examines the feasibility of adding a secondary criterion to the fitness measure to compensate for sparse data. A smoothness measure is computed for each region by comparing the approximating function within the region with those in adjacent regions. Several methods of incorporating the smoothness measure into the fitness evaluation are compared.