Abstract
This paper proposes the application of split computing paradigms for deep reinforcement learning through distributed computation between Connected Autonomous Vehicles (CAVs) and edge servers. While this approach has been explored in computer vision, it remains largely unexplored for reinforcement learning scenarios. We introduce a novel autoencoder trained directly through Deep Q-Network (DQN) rewards, wherein we optimize autoencoder layers using the DQN reward function while maintaining all other layers frozen. Our experimental results demonstrate that the proposed approach outperforms baseline methods by reducing data offloading requirements to the edge server by up to 98.7%. Additionally, this methodology not only decreases the data transmission burden but also achieves comparable rewards. In certain configurations, it even enhancing performance by up to 9.65%. The primary objective of this research is to reduce latency in deep reinforcement learning tasks for autonomous vehicles. In this regard, proposed approach achieves up to 66.5% improvement in latency reduction compared to baseline methods. These findings indicate that partial offloading through split computing offers significant benefits over both full offloading and complete on-device computation strategies for CAVs.