Deep Reinforcement Learning of Semi-Active Suspension Controller for Vehicle Ride Comfort

Among the controllable suspension systems, the control of the semi-active suspension is mostly based on optimal control. Recently, deep reinforcement learning is widely used as a method to solve the optimal control problem. Control strategies developed using reinforcement learning have shown performance beyond conventional control algorithms in some fields. In the current study, we have proposed a near optimal semi-active suspension ride comfort controller using deep reinforcement learning. An algorithm suitable for a semi-active suspension control environment was selected based on deep reinforcement learning theory to increase convergence in training. Furthermore, a state normalization filter was designed to improve the generalization performance. When compared with the ride comfort oriented classical control algorithms, our trained controller showed the best performance in terms of ride comfort. Policy map comparison with mixed SH-ADD (Skyhook-Acceleration Driven Damping) algorithm suggested the direction to the design of the semi-active suspension control algorithm.