Adaptive coordination of decentralized controllers using a centralized neural network

An adaptive approach that augments existing decentralized linear controllers is considered. By employing a neural network as a centralized element, the approach greatly broadens the class of system for which linear decentralized controllers can be designed. The stability proof naturally follows from the viewpoint that a set of decentralized controllers are a special class of multi-input multi-output controllers of an existing central method. The approach is illustrated using an inverted flexible pendulum in which a neural network coordinates an acceleration controller with a controller for an rigid inverted pendulum.