| A linear plant can track an input command signal if it is driven by a controller whose transfer function approximates the inverse of the plant transfer function. A stable inverse can be obtained even if the plant is nonminimum-phase. No direct feedback is used, except that the plant output is monitored and utilized to adapt the parameters of the controller. A model-reference inverse control system can learn to approximate a desired reference-model dynamics.; In this thesis, we demonstrate that similar principles can be utilized to control nonlinear systems. Two key features of neural networks, namely universal approximation and adaptivity make them useful for this extension. Recurrent neural networks are used to identify a model of the plant and to construct its inverse. An on-line algorithm is derived to adapt the internal weights of the inverse of the plant model.; Inverses of nonlinear systems can be used as linearizing pre-compensators. The combined system can then be controlled using standard linear feedback methods. Alternatively, the plant inverse can be utilized as part of the Internal Model Control scheme. |
