Performance evaluation of evolutionary multiobjective approaches to the design of fuzzy rule-based ensemble classifiers

Evolutionary multiobjective fuzzy rule selection can find a large number of non-dominated fuzzy rule-based classifiers with different tradeoffs between complexity and accuracy. Very simple fuzzy rule-based classifiers with high interpretability are usually not accurate while complicated classifiers with high accuracy are not interpretable. In this paper, fuzzy rule-based classifiers with different tradeoffs are used as an ensemble classifier. Three multiobjective formulations of fuzzy rule selection are compared with each other in terms of the generalization ability of constructed ensemble classifiers. Those ensemble classifiers are also compared with individual fuzzy rule-based classifiers obtained from the corresponding three single-objective formulations based on weighted sums of accuracy and complexity measures.