Research Initiation Award: Real-Time Optimal Control of a Class of Nonlinear Systems Using Neural Networks

9309486作为这项建议的一部分，将针对一类非线性系统开发一种基于人工神经网络的最优控制器。对于状态变量具有未知非线性的系统，所提出的控制器将能够实时地调整和优化滚动时间二次型代价函数。完整的控制器设计将包括建立非线性模型、估计系统状态的未来值、设计成本评价器和实时学习最优控制的各个阶段。控制器设计的所有这些阶段都将涉及各种类型的神经网络，这些神经网络可以利用与控制问题集成的不同形式的监督学习方法。简而言之，这项工作的目标如下。开发并仿真了一种基于神经网络的实时最优控制器。将其性能与(离线)最优控制器和其他两种基于神经网络的变体进行了比较。通过实验和分析，得到了神经网络控制系统的一些稳定界(或吸引域)。在密苏里大学罗拉大学神经控制器实验室的两个硬件实验上实现了控制器。***

9309486 Acar As part of this proposal, an optimal controller based on artificial neural networks will be developed for a class of nonlinear systems. The Proposed controller will be capable of adapting and optimizing a receding horizon quadratic cost function in real time for a system which has unknown nonlinearities in its state variables. The complete controller design will consist of various stages of modeling the nonlinearities, estimating future values of the system states, designing a critic for the cost, and learning the optimal control in real time. All these stages of the controller design will involve various types of neural networks which may utilize different forms of supervised learning methods integrated with the control problem. Briefly the objectives of this work are as follows. developing and simulating a real- time optimal controller based on neural networks. Comparing its performance with the (off-line) optimal controller and two other neural network based variations. Obtaining some stability bounds (or attraction regions) of the neural network controlled system experimentally and analytically. Implementing the controller on the two hardware experiments in the Neuro-Controller Laboratory at the University of Missouri-Rolla. ***