Projection Neuro-Fuzzy Hybrid Learning and its Forecasting Application

To address nonlinear forecasting, this paper develops a projection neuro-fuzzy learning method for the first time, which combines projection pursuit with fuzzy neurons composed by fuzzy membership functions (PPFL). In order to reducing the dimension of input variables, the linear projection can be used to convert multi-variables to single one called the projected characteristic value firstly. Then, taking the projected characteristic value as the new input, we can build the fuzzy inference learning model based on the fuzzy membership function neuro. By the way, thinking of the new hybrid model‘s’ complicity, we combine generic algorithm with gradient descent algorithm to optimize the model's parameters iteratively, which follows the idea of extreme learning. According to the assessment criterion, the application in specific prediction example shows that PPFL model is a powerful tool in multiple input with different types and single output prediction. The new method is also a new member added to the learning network family.