Reinforcement learning and non-zero-sum game output regulation for multi-player linear uncertain systems
Reinforcement learning and non-zero-sum game output regulation for multi-player linear uncertain systems
复制标题
DOI：
10.1016/j.automatica.2019.108672
复制
发表时间：
2020-02
期刊：
影响因子：
--
通讯作者：
Adedapo Odekunle;Weinan Gao;M. Davari;Zhong-Ping Jiang
中科院分区：
文献类型：
--
作者：
Adedapo Odekunle;Weinan Gao;M. Davari;Zhong-Ping Jiang
研究了一类连续时间多参与者线性系统的非零和博弈输出调节问题。在不了解状态矩阵和输入矩阵的情况下，通过沿系统轨迹收集的在线数据来学习反馈控制策略的n元组纳什均衡解。一个关键的策略是，首次将强化学习（RL）、微分博弈论和输出调节技术结合起来，用于数据驱动的控制设计。与已有的自适应最优输出调节文献不同，前馈矩阵被认为是非平凡的。理论分析表明闭环系统具有良好的抗干扰和跟踪能力。仿真结果验证了数据驱动控制方法的有效性。
This paper studies the non-zero-sum game output regulation problem (GORP) for a class of continuous-time multi-player linear systems. Without the knowledge of state and input matrices, the Nash equilibrium solution, N-tuple of feedback control policy, is learned through online data collected along the system trajectories. A key strategy is, for the first time, to combine techniques from reinforcement learning (RL), differential game theory, and output regulation for data-driven control design. Different from the existing literature of adaptive optimal output regulation, the feedforward matrices are considered nontrivial. Theoretical analysis shows the disturbance rejection and tracking ability of the closed-loop system. Simulation results demonstrate the efficacy of the developed data-driven control approach.