Robust reinforcement learning with augmented state for leveling control of multi-cylinder hydraulic system

Robust reinforcement learning with augmented state for leveling control of multi-cylinder hydraulic system

연구 분야: Infrastructure

학회: The Journal of Supercomputing

In this article, a deep reinforcement learning (DRL) robust control method based on maximum entropy framework is proposed to solve the leveling problem of multi-cylinder hydraulic press machines (MCHPM). The problem of steady-state errors in the model free reinforcement learning (RL) was addressed by adding augmented states, and the robustness to both internal parameter perturbations and external disturbances was improved. Meanwhile, by introducing a robust regularizer, the adverse effects of disturbance on state observations caused by sensor measurement errors on the system have been greatly alleviated. The experimental results show that the method proposed in this article significantly improves the control accuracy and system robustness.