In this paper, we present a neuro-fuzzy approach to design a controller directly from numerical data. The proposed neuro-fuzzy system is implemented as a two-layer Fuzzy Degraded HyperEllipsoidal Composite Neural Network(FDHECNN). We used a real-valued genetic algorithm to adjust weights of the composite neural networks. After sufficient training, the synaptic weights of the trained FDHECNN can be utilized to extract a set of fuzzy if-then rules. The performance of a trained FDHECNN is shown to be computationally identical to a fuzzy logic controller. The effectiveness and feasibility of the neuro-fuzzy system are tested on the truck backer-upper control problem.
Relation:
一九九六自動控制研討會暨兩岸機電及控制技術交流學術研討會論文集=Proceedings of 1996 Automatic Control Conference,頁289-294