Efficient neuro-fuzzy rule generation by parametrized gradient descent for seismic event discrimination

We show that parametrized gradient descent is very efficient to train fuzzy expert systems with examples. We first present how fuzzy expert systems work and explain their relevance compared to neural classifiers. Then, we describe the proposed learning algorithm. We further explain in more detail its application in each parameter of the fuzzy expert system: the position and the width of the premise fuzzy sets, the rule weights and the conclusion activation levels. Finally, we show the results obtained on real-world problems using several databases and compare them to other classification methods.