Evolutionary design of a fuzzy knowledge base for a mobile robot

Abstract

This paper presents a learning method which automatically designs fuzzy logic controllers (FLCs) by means of a genetic algorithm (GA). A messy coding scheme is proposed which allows a compact and flexible representation of the fuzzy rules in the genotype. It reduces the complexity and size of the rule base, through which the GA is able to solve the design task even for FLCs with a large number of input variables. A dynamically weighted objective function is proposed for control problems with multiple conflicting goals, which prevents the GA from premature convergence on FLCs that are specialized exclusively in the easier subtasks. In order to achieve a robust control behavior for a broad spectrum of control states, a second GA coevolves a set of training situations to evaluate the performance of the FLCs. We employed the method to train an FLC which implements a behavior of a mobile robot. The robot obtains the task of reaching an aiming point and avoiding collisions with obstacles on its way. It perceives its environment by means of ultrasonic sensors, which provide the measured distances to objects as input to the FLC. The knowledge base of the FLC is learnt in a simulation based on a simplified model of the sensors and the environment. The adapted control behavior is tested afterwards in real world experiments with the mobile robot.