A Learning Controller for Decentralized Nonlinear Systems

This paper describes a learning controller which uses both a general model and a local model to produce good generalization and fast learning. The controller also differs from past learning controllers by generating control commands without using the inverse dynamics model. The controller was applied to the decentralized control of a highly coupled nonlinear system - a three link manipulator. Simulations show that the controller maintains a high tracking performance when the payload was varied from 0 to 4 times the normal payload. The tracking accuracy in terms of rms errors is 10⁵ radians, which is a two orders of magnitude better than a conventional decentralized adaptive controller previously reported for a similar system.