Abstract
In multi-scenario Mobile Edge Computing (MEC), offloading and resource allocation for topologically heterogeneous Directed Acyclic Graph (DAG) tasks are NP-hard, while Federated Meta-Learning (FML) suffers from generalization degradation under non-independent and identically distributed data. To address both challenges, this paper proposes a Difficulty-Aware FML (DAFML) algorithm. We construct a three-dimensional difficulty metric—integrating task load, topological complexity, and resource supply–demand tension—and determine its weights via the Analytic Hierarchy Process. Building on this metric, a difficulty-guided non-uniform meta-task sampling strategy prioritizes complex tasks, while Proximal Policy Optimization (PPO) handles inner-loop adaptation. The offloading problem is formulated as minimizing a normalized composite cost of latency, energy, and deadline violations. Simulation results show that DAFML achieves a composite cost of 0.8096 and a completion rate of 82.61%, outperforming Greedy, centralized PPO, FedAvg-PPO, and Fed-MAML. Compared with Fed-MAML, DAFML reduces the composite cost by 2.43% and improves the completion rate by 2.26 percentage points, with greater gains on complex subsets. Ablation studies validate the contributions of the sampling strategy and the three difficulty dimensions, and convergence analysis reveals that DAFML has the most stable trajectory among all algorithms.