Improved Robustness Analysis of Reinforcement Learning Embedded Control Systems

Reinforcement learning emerges as an efficient tool to design control algorithms for nonlinear systems. There are, however, few results available on how the robustness of the closed-loop dynamics with reinforcement learning is performed. While μ

-analysis is well established as the robustness analysis tool for linear systems, there is also a limitation caused by ignoring the equilibrium shift by the uncertain parameters. An improved linearisation method for μ

-analysis is presented and the method is applied to the inverted-pendulum system with the reinforcement learning control. The resulting robustness analysis provides a significantly less conservative upper bound to the smallest worst-case perturbation.