Task Scheduling in Edge Computing Environments: a Hierarchical Cluster-based Federated Deep Reinforcement Learning Approach

Edge computing (EC) reduces latency and response times, thereby enhancing the quality of service (QoS) for internet of things (IoT) applications and services. However, scheduling heterogeneous tasks with low latency can be challenging for EC in distributed and resource-constrained infrastructures. Thus, this paper offers a hierarchical federated deep reinforcement learning (HFLDRL) scheduling architecture with feedback-driven closed-loop learning between scheduling levels. A global agent (federated deep Q-network, FL-DQNc) selects clusters for incoming tasks, while each cluster uses a locally pre-trained global agent (FL-DQNn) to assign tasks to appropriate nodes. Experimental findings show that the hierarchical federated deep Q-network framework (HFL-DQN) outperforms HDQN scheduling under various simulated workloads. The proposed approach results in 50% more cumulative incentives for the global agent than the single agent, proving its scheduling effectiveness. Across nodes, it reduces average task delay by 87.80% and execution time by 85.27%. It also reduces the makespan by up to 40.4% across varying workloads and network latencies. These findings show that HFLDRL enhances distributed edge learning latency, scheduling, and scalability.