MoNiF: a modular neuro-fuzzy controller for race car navigation

This paper characterizes how well the modular neuro-fuzzy controller, MoNiF, can perform in robotic applications by testing it in a race-car simulator. The performance of this controller is compared with that of a neural network controller (NNC) and the best programmed car (the ideal car). During the phase of learning from the ideal car, MoNiF reaches its stable performance (when it demonstrates good control) at lap 300; while the NNC requires 700 laps. Without further learning and when competing in the same learning track, MoNiF completes 20 laps of the track about 24 seconds later than the ideal car but 29 second ahead of the NNC car. However, when they all are competing in different unfamiliar tracks, MoNiF performs comparably to the ideal car, while NNC has completely lost its ability to drive. Observing the behaviors of NNC and MoNiF, we conclude that MoNiF learns to refine the output of expert rules and thus can drive in most tracks. Meanwhile NNC learns to drive in a particular track but loses its generality for other tracks. It is shown that this modular control method, equipped with the pretrained knowledge of only few simple expert rules, learns much faster than a NNC without any apriori knowledge.