HTBM: A Hybrid-Transformer-Based Method with Double Deep Q-Learning and Reptile Meta-Learning for Dynamic MEC Task Offloading

Abstract

Mobile edge computing (MEC) enables computation-intensive applications to be processed close to end users. In dynamic MEC environments, task-offloading decisions must account for changing communication conditions, heterogeneous computing resources, and dependencies among fine-grained subtasks. Conventional reinforcement-learning approaches may require substantial interaction data and may suffer from unstable value estimation. This study proposes a meta-reinforcement-learning framework, termed HTBM (Hybrid-Transformer-Based Method), for latency-aware task offloading in dynamic MEC systems. HTBM combines a hybrid-transformer sequential representation mechanism, Double Deep Q-Network (DDQN) for discrete offloading decisions, and the first-order Reptile algorithm for rapid adaptation across related MEC environments. The system model follows a fine-grained directed acyclic graph (DAG) representation in which applications are decomposed into dependent subtasks that can be executed locally or at an edge server. Experimental evaluation is conducted under a representative MEC simulation scenario. Across transmission-rate experiments, a latency reduction of approximately 2.00% was observed under the evaluated run configuration compared with GASTO, an existing graph-based meta-reinforcement-learning offloading method. Additional experiments examine rapid adaptation, robustness across DAG topologies, generalization to an unseen topology, inference time, and an ablation study isolating the contribution of each HTBM component.