离散时间系统的脱策强化学习鲁棒优化控制

For most of complicated industrial processes, it is usually difficult to establish their mathematical model. Moreover, there inevitably exist disturbances, which affect the control performance to a great extent. In this project, robust optimal control problems of discrete-time systems are considered and data-based off-policy reinforcement learning will be studied for control design. Two kinds of robust optimal control problem are considered: H∞ control and mixed H2/H∞ control. The off-policy reinforcement learning algorithms and their convergence theories will be studied based on value function and Q-function, respectively. Moreover, some issues will be analyzed rigorously, such as, the implementation based on offline and online data, the stability of closed-loop system and the influence of learning error on control performance. This project will provide some novel and effective methods and theories for data-based robust optimal control of discrete-time systems, which are extremely important for the development of intelligent control.

对于很多复杂工业过程，往往很难精确建立其精确的数学模型，而且不可避免地存在干扰因素,在很大程度上将影响控制性能。本项目针对离散时间系统的鲁棒优化控制问题，研究基于数据的脱策强化学习控制设计方法。主要考虑两类鲁棒优化控制问题：H∞控制与混合H2/H∞控制，分别研究基于状态值函数与Q-函数的脱策强化学习算法，建立算法的收敛性理论，同时探讨基于离线与在线数据的算法实现方法，系统的闭环稳定性及学习误差对控制性能的影响。本项目的研究将为基于数据的离散时间系统鲁棒优化控制提供一些新的、有效的方法和理论依据，促进智能控制的发展。

对于很多复杂工业过程，往往很难精确建立其精确的数学模型，而且不可避免地存在干扰因素,在很大程度上将影响控制性能。本项目针对离散时间系统的鲁棒优化控制问题，研究基于数据的脱策强化学习控制设计方法。考虑了两类鲁棒优化控制问题：H∞控制与混合H2/H∞控制，分别研究基于状态值函数与Q-函数的脱策强化学习算法，建立算法的收敛性理论，同时探讨基于离线与在线数据的算法实现方法，系统的闭环稳定性及学习误差对控制性能的影响。本项目吸收和借鉴了自适应动态规划控制理论、事件触发控制理论、偏微分方程系统控制理论、鲁棒控制理论、积分和深度强化学习技术的最新研究成果和经验，提出了一系列基于数据的离散时间系统鲁棒优化控制方法及理论。共发表论文17篇，其中SCI期刊论文14篇。发表的期刊包括国内外领域顶级期刊：IEEE Transactions on Cybernetics, IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Systems, Man, and Cybernetics: Systems, Neural Networks等等。此外，授权和实审发明专利4项。