Abstract
Recently, intelligent control approaches through reinforcement learning have achieved robust performance. The practical complication of reinforcement learning in approximating complex functions is resolved through the introduction of deep learning into the framework. We developed such a deep reinforcement learning-based optimal vehicle cruise control. The most commonly used algorithm, deep deterministic policy gradient (DDPG), overestimates Q-functions that can lead to sub-optimal agent policy. We proposed the twin-delayed DDPG (TD3) algorithm, the extension of the DDPG algorithm. We have compared our proposed method with model-based control approaches and non-self-learning as physics-based car-following models to see the effectiveness of the proposed approach. The TD3 agent has achieved a minimum spacing error than MPC and PID. Moreover, the TD3 agent has shown effective tracking performance compared to physics-based models and risk assessments.
