DDPG based self-learning active and model-constrained semi-active suspension control

In classical suspension control, parameter calibration is one of the most critical and time-consuming progress in algorithm development. One set of parameters is hardly possible to adapt to different working conditions. The ideal algorithm should be adaptive, cope with any road conditions and learn to apply the most appropriate control. Reinforcement learning is well suited to this scenario, giving adapted output in interaction with the environment. The DDPG network is adopted to implement active and semi-active suspension control. In active suspension control, the output forces of the network are used to exert control directly, while in semi-active suspension control, experimentally obtained damper constraints are introduced. The algorithm can reduce the RMS of acceleration by 66% in active suspension control and 15% in constrained semi-active control, adapted to various road conditions, vehicle velocities and system parameters.