A decentralized control architecture for nonlinear systems using multilayer feedforward neural networks

This paper presents a decentralized control architecture with feedforward neural networks for the control problem of complex nonlinear systems. In this method, the decentralized technique was used to treat several simple subsystems instead of a full complex system in order to reduce the training time of neural networks, and the neural networks´ nonlinear mapping ability is exploited to handle nonlinear interaction variables between subsystems. The decentralized control architecture is composed of local neuro-controllers, local neuro-identifiers and an overall interaction neuro-identifier. With the interaction neuro-identifier that catches interaction characteristics, a local neuro-identifier is trained to simulate a subsystem dynamics. A local neuro-controller is trained to learn how to control the subsystem by using a generalized backpropagation through time training algorithm. The proposed neural network-based decentralized regulating scheme is applied in a typical nonlinear system of two inverted pendulums connected by a spring and compared with the conventional centralized controller