Hybrid Learning Algorithm for Fuzzy Neuro
Ching‐Hung Lee, Yu-Ching Lin · 2004
In this paper, we propose a hybrid learning algorithm for fuzzy neural network (FNN) systems, which combining the hack-propagation and the genetic algorithms. Without any pre-training, the algorithm achieves high accuracy performance. Here, we make a breakthrough of the restriction of membership function to be some specific shape (e.g., triangular form, trapezoid form and shape of bell). The membership functions of the FNN are constructed by a group of line segment and then are fine tuned by genetic algorithm (CA) for achieving the mapping accuracy. The proposed training algorithm can he described as: (a) Firstly, we construct and train the FNN using the back-propagation algorithm to obtain membership functions and consequent weight vector. (b) Membership functions with a group of line segment by partitioning and sampling themselves are constructed. Thus we can represent membership functions in a string form- chromosome for genetic algorithm (GA). (e) Finally, for every partition point, we use the GA to search the optimal value and obtain the optimal membership functions. Simulation results show that the mapping capability of the FNN trained by the proposed method is much better. In addition, the application on the fuzzy rules reduction is presented to show the effectiveness of the approach.