Multiobjective Genetic Fuzzy Systems - Accurate and Interpretable Fuzzy Rule-Based Classifier Design - Plenary Talk

Fuzzy rule-based systems are universal approximators of non-linear functions [1] as multilayer feedforward neural networks [2]. That is, they have a high approximation ability of non-linear functions. A large number of neural and genetic learning methods have been proposed since the early 1990s [3, 4] in order to fully utilize their approximation ability. Traditionally, fuzzy rule-based systems have been mainly applied to control problems with a few input variables. Recently, they have also been applied to approximation and classification problems with many input variables. The main advantage of fuzzy rule-based systems over black-box non-linear models such as neural networks is their linguistic interpretability. Fuzzy rules are often written in the if-then form with linguistic terms such as “If x1 is small and x2 is small then y is large” and “If x1 is large and x2 is large then Class 1”. In this case, it is easy for human users to understand fuzzy rule-based systems since each fuzzy rule is linguistically interpretable. As we have already explained, fuzzy rule-based systems have two advantages: high approximation ability and high interpretability. These advantages, however, often conflict with each other as shown in Fig. 1. For example, accuracy maximization (i.e., error minimization in Fig. 1) often leads to accurate but complicated fuzzy rule-based systems with low interpretability. On the other hand, interpretability maximization (i.e., complexity minimization in Fig. 1) often leads to interpretable but inaccurate fuzzy rule-based systems.