Mapping and Implementation of Reinforcement Learning Algorithms for Quarter-Car Semi-Active Suspension Systems: An Analogy-Based Approach

Abstract

This paper presents an analogy-based framework for integrating Reinforcement Learning (RL) algorithms into a Quarter-Car Semi-Active Suspension System to improve overall ride comfort and handling stability. The framework establishes a direct mapping between suspension parameters and RL components - state, action, reward, and policy enabling learning-driven control design. Multiple RL algorithms are investigated, including value-based (DQN), policy-based (PPO, A3C), actor–critic, and model-based approaches (DDPG, TD3, SAC). Parameter variation and performance analysis reveal that RL-based controllers effectively adapt to nonlinear suspension behaviour and varying road excitations. Simulation results show significant improvements in ride comfort, reduced sprung-mass acceleration, and enhanced tire–road contact stability compared to conventional semi-active control techniques. Among all methods, DDPG and SAC demonstrated superior adaptability and convergence. The proposed analogy-based RL framework provides a systematic pathway for developing intelligent, self-optimizing vehicle suspension systems suitable for next-generation adaptive vehicle dynamics control.