Nash Equilibrium Solution Based on Safety-Guarding Reinforcement Learning in Nonzero-Sum Game
Published in 2023 International Conference on Advanced Robotics and Mechatronics (ICARM), 2023
In this paper, a safety-guarding controller is introduced to keep the safety of exploration in constrained state space. The controller is utilized to obtain the nonzero-sum game Nash equilibrium solution via a model-based reinforcement learning architecture. To deal with the uncertainty of persistent excitation, a concurrent learning approach is applied and both historical and transient data are employed in the learning process. In order to reduce the computational load, a single-critic network is utilized for approximation. To demonstrate the effectiveness of the proposed method, a two-player nonzero-sum game is developed, toward both convex/non-convex safe state-space constraints.
Recommended citation: Tan, Junkai and Xue, Shuangsi and Cao, Hui and Li, Huan (2023). Nash Equilibrium Solution Based on Safety-Guarding Reinforcement Learning in Nonzero-Sum Game. 2023 International Conference on Advanced Robotics and Mechatronics (ICARM).
Download Paper