A modular neural-fuzzy controller for autonomous reactive navigation

Fuzzy systems are modular by definition but lack the trainability of neural networks. We will introduce a modular, neural-fuzzy system called threshold fuzzy systems (TFS). A TFS has two unique features that distinguish it from traditional fuzzy logic and neural network systems; (1) the rulebase of a TFS contains only single antecedent, single consequence rules (called a behaviorist fuzzy rulebase (BFR)) and (2) the fuzzy inference mechanism is modified to incorporate a highly structured adaptive node network (called a rule dominance network - RDN). Each rule in the BFR is a direct mapping of an input sensor to a system output. Connection nodes in the DN occur when rules in the BFR are conflicting. The nodes of the DN contain functions that are used to suppress the output of other conflicting rules in the BFR. Several different approaches to tuning the unknown parameters of the dominance function can be used; including supervisory training methods, self-organizing and evolutionary-based exploration. For the supervisory training approach, a unique gradient-matrix error back-propagation algorithm (GMEBA) has been developed and will be discussed.