Federated Reinforcement Learning for Distributed Dependent Task Offloading with Batch Processing in CPU-GPU Heterogeneous MEC

By offloading tasks to proximal edge servers, mobile edge computing (MEC) has emerged as a promising paradigm to enable computation-intensive and latency-sensitive applications on mobile devices. This paper investigates the challenging problem of joint computation offloading and resource scheduling on dependent tasks in a heterogeneous CPU-GPU MEC environment with batch processing. The optimization problem is modeled as a Markov decision process (MDP), and a distributed Transformer based Federated Soft Actor-Critic (TFSAC) framework is proposed to minimize the overall task latency and energy consumption. Graph attention networks (GATs) are leveraged to extract high-dimensional features from the task dependency graph, and transformer encoders are applied to learn contextual relationships between agents. Rather than simply averaging weights, TFSAC helps selective aggregation of relevant knowledge during federated model training to preserve agents' privacy. Extensive experiments on real-world trace data demonstrate that TFSAC is superior over other benchmarks in maximizing the quality of-service (QoS) across multiple configurations and enables the seamless onboarding of new agents into the federation.