Adaptive Identification and Control of Dynamical Systems Using Neural Networks

控制动态系统的基本过程包括系统的数学建模、基于实验数据的辨识、对输出的处理，以及利用它们依次合成控制输入以实现期望的行为。自动控制的目的是快速、准确、稳定地实现后者。控制一个已知的具有多输入多输出的线性对象是一个困难的问题。当这类系统的参数未知时，问题就更加困难，需要进行自适应控制。当对象是已知的但又是非线性的时，难度要大得多。当被控对象是具有未知特性的非线性对象时，我们就会遇到一个真正令人敬畏的非线性自适应控制问题。在空间技术、制造和机器人等新领域的许多应用中，以及在过程控制和飞机控制等已建立的领域中，此类问题正在以越来越频繁的方式出现。该提案涉及使用神经网络控制这类系统的理论和实践方面。这项工作的第一部分将讨论非线性系统的表示问题和控制器的存在性。在第二部分中，将讨论与选择神经网络作为识别器和控制器有关的问题。第三部分将讨论稳定这个重要问题。第四部分对多模型控制进行了详细的研究。过去三年进行的经验研究表明，这种基于多种模型的控制方法是有效应对环境快速变化的需要。该方案的四个部分代表了使用神经网络的非线性自适应控制的四个重要方面。

The fundamental processes involved in controlling a dynamical system include the mathematical modeling of the system, identification based on experimental data, processing of the outputs, and using them in turn to synthesize control inputs to achieve desired behavior. The aim of automatic control is to achieve the latter rapidly, accurately, and in a stable fashion. The problem of controlling a known linear plant with multiple inputs and multiple outputs is a difficult one. When the parameters of such systems are unknown, the problem is considerably more difficult, and adaptive control is needed. The difficulty is substantially grater when the plant is known but nonlinear. When the plant is nonlinear with unknown characteristics, we have a nonlinear adaptive control problem which is truly formidable. In numerous application in new areas such as space technology, manufacturing, and robotics, as well as established areas such as process control and aircraft control, such problems are arising with increasing frequency. The proposal deals with theoretical and practical aspects of the control of such systems using neural networks. The first part of this work will deal with the representation problem of non linear systems and the existence of controller for them. In the second part, questions related to the choice of neural networks as identifiers and controllers will be addressed. The third part will deal with the important question of stability. In the fourth part, a detailed study of control using multiple models will be carried on. Empirical studies carried out in the past three years suggest that such a control methodology based on multiple models is needed to cope efficiently with rapid changes in the environment. The four part of the proposal represents four of the important aspects of nonlinear adaptive control using neural networks.