Progressive optimization of a fuzzy inference system

The paper describes an automatically on-line tuned fuzzy navigation system for an autonomous robot using modular structure to generate the angular speed in function of the sensor data. The goal is to obtain a reactive behavior such as wall-following with the adaptivity necessary for coping with large modifications in the physical characteristics of the robot. For this behavior, the building of the navigation controller is done entirely on-line by the optimization of a zero order Takagi-Sugeno fuzzy inference system (FIS) by a back-propagation-like algorithm. It is used to minimize a cost function that is made up of a quadratic error term and a weight decay term that prevents an excessive growth of parameters of the consequent part. The procedure is performed entirely on-line, but in two steps. The first one is done on a miniature robot or on its dedicated simulator. Then the obtained controller is carried on the real robot and a further optimization step is performed. At the end of the procedure, it is possible to extract the knowledge by interpreting the result parameters in a symbolic form. One can notice that the two tables deduced on the miniature robot and on the real robot are very close with respect to the linguistic concepts. Moreover, these two automatically extracted tables of rules are quite close to those empirically written. But, we can observe that some human expertise rules work wrongly because the expert doesn't expect particular situation. In fact, the main advantage of this procedure is the optimization of the controller with respect to the actual characteristics of the robot. That means that for example the rough manual tuning of the global gains acting on the width of the universes of discourse is replaced by a fine local automatic tuning and this improves very significantly the performances. This method is simple, economical and safe since it is done on a miniature robot. It leads to a very quick and efficient optimization technique.