Neuro-Fuzzy Rule Generation for Backing up Navigation of Car-like Mobile Robots

An automatic neuro-fuzzy rule generation scheme is proposed for backing up navigation of carlike mobile robots. The proposed method is based on the Conditional Fuzzy C-Means (CFCM) and Fuzzy Equalization (FE) methods. The CFCM is adopted to render clusters, which can represent the homogeneous properties of the given input and output fuzzy data, and also the FE method is used to systematically construct the fuzzy membership functions for the ANFIS. From these, a compact size of fuzzy rules can be automatically obtained, which satisfy the given goal. The proposed method has been applied to a truck, and also to a truck-trailer backing up navigation problem, and good results have been achieved in comparison to previous work. To solve these problems, researchers have been attempting to automate rule extraction, based on numerical training data. Most methods are intrinsically based on a fusion scheme of fuzzy logic and neural networks, that is, a neuro-fuzzy approach is often used. Here, fuzzy logic permits easy incorporation of the expert knowledge with humanlike if-then rule thinking and neural networks, leading optimization abilities, learning abilities, and connectionist structures. In this way, the learning power of neural networks can be incorporated into fuzzy systems, and also the reasoning process can be provided to the neural networks [9-10]. One of the most widely used method has been the neuro-fuzzy scheme, which is the AdaptiveNetwork-based Fuzzy Inference System (ANFIS) proposed by Jang [12]. This paper is also proposed on this basic framework of the ANFIS approach. Applying the ANFIS to real-world problems requires two phases, namely, parameter identification and structure identification. The former is used to adjust the membership functions in the premise part of the rules and the linear coefficients in the consequent part are determined by using the steepest descent and least square estimation methods, respectively. The latter is related to finding a suitable number of rules and a proper partition of the input space. Related to the structure identification, the usual grid partition often encounters problems when there are a moderately large number of inputs which is usually referred to as the “curse of dimensionality” [12].