Abstract
Theoretical studies have shown that fuzzy models are capable of approximating any continuous function on a compact domain to any degree of accuracy. However, constructing a good fuzzy model requires finding a good tradeoff between fitting the training data and keeping the model simple. A simpler model is not only easily understood, but also less likely to overfit the training data. Even though heuristic approaches to explore such a tradeoff for fuzzy modeling have been developed, few principled approaches exist in the literature due to the lack of a well-defined optimality criterion. In this paper, we propose several information theoretic optimality criteria for fuzzy models construction by extending three statistical information criteria: 1) the Akaike information criterion (AIC); 2) the Bhansali-Downham information criterion (BDIC); and 3) the Schwarz-Rissanen information criterion (SRIC). We then describe a principled approach to explore the fitness-complexity tradeoff using these optimality criteria together with a fuzzy model reduction technique based on the singular value decomposition (SVD). The role of these optimality criteria in fuzzy modeling is discussed and their practical applicability is illustrated using a nonlinear system modeling example.
Original language | English (US) |
---|---|
Pages (from-to) | 362-372 |
Number of pages | 11 |
Journal | IEEE Transactions on Fuzzy Systems |
Volume | 6 |
Issue number | 3 |
DOIs | |
State | Published - 1998 |
All Science Journal Classification (ASJC) codes
- Control and Systems Engineering
- Computational Theory and Mathematics
- Artificial Intelligence
- Applied Mathematics