GAPPO-AVNs: GA-Bounded PPO for DAG-Aware Resource Optimization in MEC-Enabled Vehicular Networks

Abstract

Autonomous vehicular networks (AVNs) supported by multi-access edge computing (MEC) require reliable execution of latency-critical workloads under dynamic traffic, wireless, and computing conditions. Existing resource-allocation methods often lack the ability to jointly provide system-level coordination, real-time adaptability, and dependency-aware execution for directed acyclic graph (DAG)-based autonomous-driving workloads. This paper proposes GAPPO-AVNs, a hierarchical GA-bounded Proximal Policy Optimization (PPO) framework for delay-constrained energy optimisation in MEC-enabled AVNs. The key novelty lies in a feasibility-guided GA–PPO coordination mechanism, where a genetic algorithm (GA) periodically generates global resource bounds for CPU frequency, transmission power, bandwidth allocation, and task offloading, while PPO performs online refinement within the GA-bounded feasible region through a Monitor–Analyse–Plan–Execute (MAPE) control loop. A DAG-based task model is also incorporated to capture subtask precedence constraints and support dependency-aware parallel execution. Simulation results show that GAPPO-AVNs improves the energy–delay trade-off, task-completion reliability, and policy stability compared with GA-only, PPO-only, PSO-based, DE-based, and heuristic baselines. These results confirm the effectiveness of GA-bounded action control for scalable resource allocation in MEC-enabled AVNs.