On-line rule generation for robotic behavior controller based on a neural-fuzzy inference network

Fuzzy logic method presents an efficient way for the control of behavior modules in robot navigation. However, it is difficult to construct and maintain correct and complete fuzzy rule base by a human expert. In this paper, considering the drawbacks of the existing neural-network-based approaches for rule generation, we present an on-line learning scheme for the adaptive and dynamic rule generation of a fuzzy controller based on a developed neural-fuzzy inference network. The proposed learning scheme consists of the rule construction and the rule examination. The former is used to locate the parameters of membership functions and create if-then rules simultaneously for each incoming training data. The latter is for the optimization of the existing rule base. The proposed scheme is applied to the wall-following control of an omnidirectional mobile robot equipped with an ultrasonic sensor ring. The simulation results show the designed fuzzy controller has a satisfying performance.