Reinforcement Learning for Pmsm: Effects of Choosing a Reward Function

This paper examines the design and evaluation of reinforcement learning (RL) reward functions for permanent magnet synchronous motor (PMSM) control in electric drive applications. Various reward structures are proposed to enhance current tracking in the $d q$-reference frame. Unlike model predictive control, which employs a quadratic reward function for convexity and computational tractability, discontinuous reward functions yield superior performance. A comparative analysis of quadratic, absolute value, square root, and quartic root reward functions demonstrates that the quadratic scheme exhibits the poorest tracking performance-up to 32 times worse on average. While square root and quartic root yield comparable results, the absolute value function serves as an intermediate option, outperforming the quadratic scheme but falling short of the square and quartic root approaches.