Optimizing ride comfort through deep reinforcement learning for autonomous vehicle control

Autonomous vehicles have been a research trend over the past two decades, and both industrial and academic institutions have exerted considerable effort to achieve the highest level of autonomy. All these efforts have resulted in the use of deep reinforcement learning for autonomous driving and autonomous vehicles, as it provides great flexibility in achieving autonomy, especially in end-to-end control. One challenge when using deep reinforcement learning for end-to-end control of autonomous vehicles is the reward function design, which is the basis on which the behaviour of the vehicle is designed. Many efforts have been made by researchers and engineers to achieve an ideal reward function design, but to the best of our knowledge, this has still not been achieved. Reward function design has many challenges, one of which is the missing attributes that pose a great deal for end users, such as comfort. This study presents a novel reward function specifically designed to enhance ride comfort in autonomous vehicles. The proposed design process surpasses other methods by prioritizing passenger comfort as a core objective. The efficacy of the proposed reward function is demonstrated by the increased total accumulated rewards per episode and the acceleration profiles proved by a 44.34% reduction compared to the baseline model.