CAREER: Artificial Learning Control Systems for Performance Critical Applications

许多研究人员都坚持类似人类(或更好的)智能控制的目标，即将人类操作员和机器的最佳元素结合在一起以实现更有能力的整体的系统。具有中等智能和适应性(介于当前系统和类人控制之间)的控制器具有重要意义，至少有两个原因：(1)它们是通向更有能力的系统的垫脚石；(2)它们本身具有重要的实际用途，作为半智能系统的一部分，或作为更有能力的人/机器系统的一个元件。本项目将不同于以前重要和有用的人工神经网络(NN)自适应控制方法，基于动态逆或反馈线性化控制器和在线训练的神经网络。这类自适应控制最近被首席研究人员扩展到包括不能反馈线性化(即使是近似的)的对象。由于对新的不确定性和重要的理论稳定性结果的快速响应，这些方法对于性能关键应用(即安全、经济压力或其他因素意味着必须将系统故障的概率降至最低的应用)具有明显的优势。将采用一种新的方法，它保留了上述优点，但利用神经网络自适应控制器的未开发潜力，使系统能够在所有可获得信息的状态和控制空间上学习对象，包括先验信息、记录数据和当前状态。它将探索这一想法，即同时进行在线培训，并在其他自适应控制框架内利用记录的过去数据进行培训。这些方法的发展有望导致许多重要的关键性能控制器的新应用，这些应用通常需要人的“触摸”才能获得可接受的性能(还需要稳定性和健壮性保证)，例如：工厂自动化(那些不确定性不可忽视的方面)、地面车辆控制、建筑设备、集成航空电子设备、航天器对接和飞行控制。这种预测的影响是基于提供了人类作为控制器可以做的一个重要元素：在很大一部分状态和控制空间内高度熟练地控制植物，同时还能够对植物动态的突然变化做出反应。

The goal of human-like (or better) intelligent control, and systems that combine the best elements of human operators and machines to achieve a more capable whole, is held by many researchers. Controllers with intermediate levels of intelligence and adaptability (between current systems and human-like control) are significant for at least two reasons: (1) They are a stepping-stone to the more capable systems envisioned; and (2) They have important practical uses in their own right, as part of semi-intelligent systems or utilized as an element of more capable human/machine systems.This project will depart from previous important and useful artificial neural network (NN) adaptive control methods, based on a dynamic inverse or feedback linearization controller and a neural network trained online. This class of adaptive control has recently been extended by the principal investigator to include plants that are not feedback linearizable (even approximately). These methods have distinct advantages for performance critical applications (i.e., those where safety, economic pressures, or other factors imply that the probability of system failure must be minimized), because of rapid response to new uncertainties and important theoretical stability results. A new approach will be taken, which retains the advantages described above, but utilizes untapped potential of NN adaptive controllers to enable the system to learn the plant over all state and control space for which information is available, including information given a priori, recorded data, and current state. It will explore this idea of simultaneous online training and training on recorded past data within other adaptive control frameworks. The development of these approaches is hoped to lead to many important new applications of performance critical controllers, applications that generally required a human "touch" to achieve acceptable performance (and also require stability and robustness guarantees), examples include: aspects of factory automation (those where uncertainty is not negligible), ground vehicle control, construction equipment, integrated avionics, spacecraft docking, and flight control. This predicted impact is based on providing an important element of what humans can do as controllers: become highly proficient at controlling a plant over a large segment of state and control space, yet also able to respond to a sudden change in plant dynamics.