Comparative Study of DQN and DDQN in Reinforcement Learning for Autonomous Vehicle Collision Avoidance Systems

This paper evaluates the performance of Deep Q-Network (DQN) and Double Deep Q-Network (DDQN) agents for collision avoidance in autonomous vehicles using the CARLA simulator. By integrating a hybrid sensor system, comprising radar, collision, and lane invasion sensors and utilizing a prioritized experience replay (PER) buffer, both agents were able to achieve 100% accuracy within 1,000 training episodes. Experimental results demonstrate that DDQN outperforms DQN in terms of average reward, training stability, and responsiveness to dynamic traffic conditions. Its architecture mitigates common reinforcement learning challenges, such as value overestimation, leading to safer and more reliable autonomous driving behaviors. These findings offer valuable insights for enhancing autonomous driving safety and warrant further real-world validation.