Policy gradient based Reinforcement learning control design of an electro-pneumatic gearbox actuator ^⁎

Abstract

The paper presents a reinforcement learning based solution for the control design problem of a gearbox actuator. The system is operated by an electro-pneumatic, three-state, floating piston cylinder. Besides the primary goals of positioning the piston, the nonlinear system’s quality objectives are to minimize switching time and overshoot. The control strategy based on the measurable parameters of the system is realized by a dense feedforward neural network. With the utilization of the policy based reinforcement learning architecture, the learning agent develops the optimal strategy for fast and smooth switching, under different and changing conditions.