A new approach to adaptive control using multiple models

摘要

The demands of a rapidly growing technology for faster and more accurate controllers have always had a strong influence on advances in automatic control theory. In recent years, problems are arising in a variety of disciplines including medicine, computer vision, and financial markets where fast and accurate decisions have to be made in the presence of large uncertainties or rapidly time-varying parameters. Classical adaptive control methods based on the use of a single identification model are found to be inadequate to cope with such problems. During the past 15 years, efforts have been made to extend the general methodology of adaptive control by the use of multiple identification models. Among numerous methods that have been proposed, two approaches, referred to as "switching" and "switching and tuning", have emerged over the years as the most successful ones. In this paper, a radically new way of using multiple models for the identification and control of an unknown linear time-invariant plant is proposed. It uses the information generated by a finite number of conventional adaptive identifiers (referred to as first level) to re-parameterize and identify rapidly the unknown plant (at a second level). The stability of the procedure, and the reasons for its resulting in faster convergence in the control of time-invariant plants, are discussed and illustrated using simulation studies. The accepted philosophy among adaptive control theorists has been that if an adaptive system is fast and accurate in a time-invariant environment, it will perform satisfactorily in a time-varying environment. Simulation studies are included, at the end of the paper, of plants with rapidly time-varying parameters, to compare the performance of the new approach with currently well-established methods.