Vibration control using self-organizing look-up tables

The self-organizing map neural network is used in a supervised way to represent a sensor–actuator mapping. The learning of the controller assumes no prior information, but only reward/failure signals that are produced by an evaluation criterion. The evaluation criterion used is based on the low-pass filtering of the gradient of a reward function and the local storing of the filtered gradient value. The control method is tested in vibration isolation of a flexible spray boom used in agriculture for pesticide application. The neural network learns to stabilize the boom on-line without any prior information and with a very high performance.