Adaptive Critic Design and Implementations

9413120，Pi-Wunsch：自适应批评神经网络是一种很有前途的计算智能新技术，能够解决非线性动态识别、最优实时非线性控制和时间序列预测等各种应用问题。本项目将在基于不同神经网络模型的软件中实现各种自适应批评设计(ACDs)，并证明它们对于解决仿真中的各种问题的有效性。它将根据ANSI规范为ACD的所有模块开发C/C++代码，以提供通用的图形界面和跨各种平台的可移植性。这将为早期的行业合作提供工具。PI将实现复杂识别问题(例如，化学反应堆动力学的识别)、时间序列预测(例如，股市预测)和复杂最优控制问题(例如，单级到轨(SSTO)飞机的飞行控制)实例的计算机模拟解决方案。这些将来自已公布的来源。随着应用需求的增加，以及需要大量的机会和快速原型能力，他们将把这些工具移植到专门的神经协处理器上。从数值或理论上证明ACD逼近复杂非线性动态系统(即系统加基于ACD的控制器)的Lyapunov函数的能力，以及研究神经网络的实时交互作用，也是该项目的目标。这对了解非线性动力系统的稳定性分析具有重要意义。最后，这些活动将特别注重成果的传播，包括学术界、工业界和政府的教育活动；可用于ftp的软件、讲习班、加强出版以及与政府和工业界的进一步合作。这项工作将提供更好的工具并减少与使用神经网络和非线性控制相关的学习曲线，从而提高神经网络和非线性控制作为工程学科的成功。

9413120, PI-Wunsch: The Adaptive Critic Neural Network is a promising new technique of computational intelligence capable of solving large variety of applied problems in nonlinear dynamic identification, optimal real-time nonlinear control and time series forecasting. This project will implement various Adaptive Critic Designs (ACDs) in software based on different models of Neural Network and prove their efficiency for solving various problems in simulation. IT will develop C/C++ code for all modules of ACDs in accordance with ANSI specification to provide a common graphical interface and portability across various platforms. This will provide the tools for early industrial collaborations. The PI will implement computer-simulated solutions of examples of complex identification problems (e.g., identification of chemical reactor dynamics), time series predictions (e.g., stock market forecasts) and those of complex optimal control problems (e.g., flight control of a Single -Stage-To-Orbit (SSTO) plane). These will come from published sources. As application demands rise, and the volume of opportunities necessitates and rapid prototyping capability, They will port these tools to a specialized neural coprocessor. Proving numerically or theoretically the ability of ACD to approximate Lyapunov functions of a complex nonlinear dynamic system as a whole (i.e., a system plus ACD-based controller), along with a study of real-time interaction of neural networks, is also a goal of the project. This will significantly contribute to the knowledge of stability analysis of nonlinear dynamic systems. Finally, these will be a special concentration on dissemination of the results, including educational activities for academia, industry and govern ment; software available for ftp, workshops, enhanced publishing, and further collaborations with government and industry. This work will enhance the success of both neural networks and nonlinear controls as engineering disciplines by providing better tools and reducing the learning curve associated with their use.