Deep Reinforcement Learning for Semi-Active Suspension: A Feasibility Study

Semi-active suspension as an intermediate form of passive and active suspension is widely applied for the vibration isolation of passenger cars and has recently received attention in the discipline of machine learning. This paper presents how to utilize deep reinforcement learning (DRL) algorithms to generate variable damping characteristics of the semi-active suspension. A two-degree-of-freedom quarter car suspension model featuring nonlinear friction is used to design a Markovian decision process, with a linear quadratic cost function and a stochastic road profile. Furthermore, we verify the ride comfort performance of the DRL controllers and discuss observations made on its near-optimal control force.