Risk-Sensitive and/or Adaptive Identification of Dynamic Systems by Neural Networks

通过神经网络对动态系统进行风险敏感和/或自适应识别

• 批准号: 9707206
• 负责人: James Lo
• 金额: $ 19.45万
• 依托单位: University of Maryland Baltimore County
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 1997
• 资助国家: 美国
• 起止时间: 1997-07-15 至 2001-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9707206&HistoricalAwards=false
• 关键词: Risk; Sensitive; Adaptive; Identification; Dynamic

This project will address the fundamental issues of robustness and adaptiveness in the identification of dynamic systems in both series-parallel and parallel formulations. These two issues have been the topics of major concentrated research activities in system identification, control and filtering in the past 20 years. generally speaking, a system identifier should be adaptive to adaptation-worthy environmental parameters and robust to these which are adptative-unworthy. Taking a synthetic approach inspired by the development of the artificial neural networks (ANNs) it will lay the mathematical foundations, develop the methodologies, and test their feasibility's for robust and/or adaptive identification of dynamic systems in the project. The main idea for adaptive identification is to use the nonlinear and linear weights of an adaptive neural identifier as the long and short-term memory respectively, the former being determined in a prior off-line training and the latter adjusted on-line by an LMS or RLS algorithm. Such an adaptive neural identifier is expected to have the advantages of minimizing computation, focusing on learning about and adapting to the adaptation-worthy environmental parameters, and eliminating poor local extreme of the performance surface during the operation of the adaptive neural identifier. This idea was inspired by the way a brain performs adaptation.

本项目将解决鲁棒性和适应性的基本问题，在识别动态系统的串并联和并联配方。 这两个问题是近20年来系统辨识、控制和滤波领域主要集中研究的课题。 一般来说，系统标识符应当适应值得适应的环境参数，并且对于不值得适应的环境参数是鲁棒的。 采用人工神经网络（ANN）的发展启发的综合方法，它将奠定数学基础，开发方法，并测试其可行性的鲁棒性和/或自适应识别的动态系统的项目。 自适应辨识的主要思想是利用自适应神经元辨识器的非线性权值作为长期记忆，线性权值作为短期记忆，非线性权值在离线训练时预先确定，线性权值通过LMS或RLS算法在线调整。 这样的自适应神经识别器被期望具有最小化计算、专注于学习和适应值得适应的环境参数以及在自适应神经识别器的操作期间消除性能表面的不良局部极值的优点。 这个想法的灵感来自于大脑执行适应的方式。