Embedding fuzzy knowledge into neural networks for control applications

Fuzzy control has recently emerged as a new technique of knowledge-based intelligent control in which precise knowledge of control algorithms is not required. The control knowledge is expressed in terms of membership functions for control parameters and a given rule set which defines the relationship among various parameters. Although this technique is robust, it cannot learn and adapt as parameters change over time. Neural network control uses learning for defining mapping between input and output data. By using fuzzy logic rule/membership knowledge and neural network learning capability, this work proposes a new method for combining the two intelligent control methods. The proposed neuro-fuzzy method embeds fuzzy rule and membership knowledge into a neural network for training via a backpropagation algorithm. Results based on fuzzy control, the Iwata-Machida-Toda method, and this proposed method are given for cart-pole problems. The proposed method has the best response time and the smallest magnitude of oscillations near the setpoint.