Actor-Critic Convolutional Neural Network-Based Centralized Computation Offloading for Vehicular Edge Networks

Abstract

Efficient computation offloading is essential for addressing the growing demand for low-latency, computation-intensive applications in vehicular edge networks. Existing approaches, including traditional optimization techniques and reinforcement learning-based methods, often struggle to cope with the highly dynamic network environments and the complex task dependencies characterized by Directed Acyclic Graphs (DAGs). To address these challenges, this paper proposes a novel centralized Actor-Critic Convolutional Neural Network-based Offloading Computation (ACCOC) algorithm. By leveraging the Actor-Critic framework enhanced with convolutional layers, ACCOC efficiently handles high-dimensional task offloading decisions, balancing delay and energy consumption under dynamic network conditions. The proposed algorithm is evaluated in a realistic vehicular edge computing environment with varying numbers of users, bandwidth availability, and edge server computational capabilities. Simulation results reveal that ACCOC achieves superior performance compared to baseline methods, including a non-convolutional Actor-Critic algorithm, Deep Q-Network (DQN), and random or fully local execution strategies. Specifically, ACCOC demonstrates faster convergence and higher system utility, achieving an average reward of 0.78 compared to 0.72 for the non-convolutional baseline. Furthermore, the algorithm maintains robust performance across diverse parameter settings, highlighting its scalability and adaptability.