文献类型：期刊文献

英文题名：Single-critic reinforcement learning based optimal control for multi-player game systems with coupled unknown dynamics

作者：Zhang, Zhixiang[1];Wang, Mengling[1];Yang, Wen[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China

年份：2026

外文期刊名：INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE

收录：;EI(收录号：20260720084827);WOS:【SCI-EXPANDED(收录号:WOS:001687759600001)】；

语种：英文

外文关键词：Non-linear NZS systems; integral reinforcement learning; single critic design; off-policy; optimal control

摘要：This paper proposes a single-critic reinforcement learning (SCRL) for estimating coupled unknown dynamics to solve the optimal control problem of a multi-player non-zero-sum (NZS) game. A data-driven approach is employed to approximate the Nash equilibrium of the game system with input constraints. Firstly, the multi-player optimal control problem is converted to a data-iterative solution form, where the derived coupled unknown dynamics are handled by off-policy integral reinforcement learning (OPIRL) with complementary polynomial fitting. In contrast to existing IRL methods, the SCRL method proposed in this paper learns from real system data without knowing any model information, and can handle complex heterogeneous input dynamic. Subsequently, the single-critic structure is proposed to learn optimal control law, in which the structure parameters are integrated with determined coupled unknown dynamics. To enhance convergence efficiency, an off-policy method is applied to learn the parameters of the critic network, where off-line measurements are utilised as an empirical dataset. Theoretical proofs are provided, along with simulation experiments of the proposed SCRL framework. The experimental results demonstrate the accuracy and computational efficiency of the proposed algorithm.