Steady-State Error Compensation for Reinforcement Learning with Quadratic Rewards

Quadratic reward functions in Reinforcement Learning (RL) lead to steady-state errors, while absolute value reward functions, although alleviating this issue, induce substantial fluctuations in specific system states, causing abrupt changes. In response to this challenge, this study proposed an approach that integrates an integral term into quadratic type reward functions. Through experiments and performance evaluations on the Adaptive Cruise Control (ACC) and lane change models, we validate that the proposed method effectively diminishes steady-state errors and does not cause significant spikes in some system states.