ACTIVE SUSPENSION CONTROL BASED ON DEEP REINFORCEMENT LEARNING COMBINED WITH METAHEURISTIC TECHNIQUES

Abstract

The Active Suspension System (ASS) can effectively generate active control forces to suppress vehicle body vibrations. This article presents a Deep Reinforcement Learning (DRL)-based control strategy for the ASS, where a Deep Q-Network (DQN) is integrated with the Particle Swarm Optimization (PSO) algorithm to optimize the DQN hyperparameters. The control objectives are to enhance ride comfort by minimizing the body acceleration |
sz
| and its Root Mean Square (RMS) value, while maintaining good road holding and ensuring feasible, energy-efficient control forces. The DQN employs a discrete action space with 11 control force levels and a deep neural architecture to learn adaptive control policies. The PSO algorithm optimizes critical DQN parameters such as the learning rate and exploration decay coefficient. To accelerate convergence, 12 PSO particles are executed in parallel using Python’s multiprocessing.Pool framework. Simulation results demonstrate that the proposed DRL–PSO-based controller achieves superior ride comfort and road-holding performance under complex road disturbances, outperforming conventional control schemes including PID, Linear Quadratic Regulator (LQR), and modern Model Predictive Control (MPC).