Incremental Tuning of Fuzzy Controllers by Means of an Evolution Strategy

This paper presents a novel evolutionary tuning method that adapts the knowledge base of a fuzzy logic controller. The evolution strategy adopted in this paper employs a variable size genome, which it adapts to the complexity of the control problem. The incremental learning scheme starts with a single control rule, which is re ned through superposition of additional rules, variables and fuzzy sets in later generations. Additional regulator genes control the activation of the genome segments that represent the fuzzy rules and sets. The incremental development of the fuzzy knowledge base enables the evolution strategy to partition the controller design problem into smaller steps that become more feasible. The proposed method is applied to adapt a wall-following behavior of a mobile robot. The evolutionary learning process takes place by a simulation of the robot, its sensors and the training environment. The adapted control behavior is subsequently evaluated on the physical robot situated in a realworld environment.