Autostructuration of fuzzy systems by rules sensitivity analysis

We present a destructive (pruning) method aiming at gradually finding the appropriate number of rules in the case of fuzzy models. A particular attention has been paid to Takagi-Sugeno fuzzy systems (TS) for the problem of functions approximation. The proposed system can be seen as a generalization of the conventional TS system and allows to evaluate the importance of one particular rule in the inference process. The advantage of our approach has been put in light on two well-known benchmarks related to the field of chaotic time series forecasting. We also study and compare possible local and global learning strategies for these systems in terms of readability and performance. (C) 2001 Elsevier Science B.V. All rights reserved.