Neuro-fuzzy systems for function approximation (Q1299623): Difference between revisions

The idea of exploiting learning capabilities of neural networks in order to automate or support the process of developing a fuzzy system for a given task has been developed for a long time. The authors of the reviewed paper themselves used neuro-fuzzy systems already in the domain of (neuro-) fuzzy control, data analysis and classification. The same idea is used here for a function approximation problem. The NEFPROX (neuro-fuzzy approximation system) is defined as a special 3-layer fuzzy perceptron. The NEFPROX learning algorithm is described in a very sketchy way. As an example of the learning capabilities of the NEFPROX, a chaotic time series generated by the Mackey-Glass differential equation with a delay is used. That's a surprise. As one knows a long time, description of chaotic time series is impossible in principle. Commonly, one uses such chaotic time series for a problem of prediction. And in fact, results of simulations seem very vague. It is very difficult to make a reasonable conclusion to know if the nonlinear system has 105 adjustable parameters, as it is the case here. In other words, in such system one can generate artificially what one wants, but hardly the opposite. So it seems to be rational to use another example to test the NEFPROX system, e.g., the short time prediction in a time series generated by the Mackey-Glass equation.