Design method for neuro-fuzzy motion controllers

A four-step hybrid method for the design of neuro-fuzzy motion controllers is presented. The design method starts from a preliminary known "good" control strategy used as learning data. The aim of the method is to find a controller that reproduces as close as possible the good control strategy and ensures the accomplishment of the required motion pattern for the controlled mechanism. An improved simple genetic algorithm and the weighted errors balance algorithm were combined with the global and local learning principles within the first two steps of the method in order to design a Takagi-Sugeno type fuzzy controller. The evaluation of the local learning criterion is then used in order to reduce the number of rules used by the rulebase. Finally, the parameters of a feed-forward type neural network structure, that embeds the simplified Takagi-Sugeno fuzzy controller, are derived. As an example, the paper presents the simulated evolution of a brachiation mobile robot under the control of a neuro-fuzzy motion controller obtained using the proposed design method. The advantages of the method and the possibilities of further improvements are discussed.