Cooperative Multi-Agent Deep Reinforcement Learning for Dynamic Task Execution and Resource Allocation in Vehicular Edge Computing

Computer vision plays a crucial role in enabling connected autonomous vehicles (CAVs) to observe and comprehend their surroundings. The computer vision tasks are typically based on convolutional neural networks (CNNs). However, CNNs often require significant processing power. Techniques like early exiting and split computing enhance CNN task execution latency and adaptability to varying environmental conditions. Since the split computing introduces additional overhead for offloading of the task from the CAV to an edge servers, we incorporate multiple autoencoders within each split point to enhance the adaptability of splitting under varying environmental conditions. However, the autoencoders introduce an additional layer of complexity related to the selection of the optimal compression strategy alongside the splitting and exiting decisions. To tackle this challenge, we introduce a novel approach based on the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. This algorithm dynamically and jointly determines the most suitable exit point, split point, and autoencoder. Furthermore, the MADDPG-based approach considers other CAVs when selecting action, promoting cooperation among CAVs. Our results demonstrate that the proposed approach reduces latency up to 44.4% while maintaining at least comparable or even higher accuracy of the computed vision outcome compared to the state-of-the-art solutions.