Control of nonlinear dynamic systems using neural networks with incremental learning

Abstract

Nonlinear dynamic systems present complex behavior that is not easy to control using conventional techniques. Even more, neural networks cannot always be trained in a straightforward learning scheme for solving dynamic control problems. This paper proposes incremental learning methods for training neural networks for the control of nonlinear dynamic systems using the Dynamic Back Propagation algorithm. By analyzing the complexity of the control problem, learning strategies are formulated in an incremental scheme similar to human learning: starting from easy and simple tasks and continuing with increasingly complex and difficult tasks. The results obtained in the control of highly unstable nonlinear systems, and the positioning control of mobile robots verify the effectiveness of the proposed incremental learning strategies.