Federated learning-based approach for heterogeneous task scheduling in edge computing environments

Edge computing (EC) aims to facilitate internet of things (IoT) applications and services with low latency at the edge, thereby reducing response time and providing quality of service (QoS). However, increasing user demand for low-latency applications has highlighted the need to reduce task completion time delay in EC environments. Therefore, this paper introduces an approach for heterogeneous task scheduling in heterogeneous EC environments to minimise task delay for time-sensitive applications. A combination of deep reinforcement learning (DRL) and federated learning (FL) techniques is used to build the scheduling framework. Initially, the deep Q network (DQN)-based scheduling framework is employed to reduce the delay of tasks generated on an edge cluster within heterogeneous nodes. For collaborative learning, DQN agents are trained in different edge clusters for multiple FL rounds. The federated averaging model (FedAvg) is applied to calculate the average of the parameters of each trained agent in every FL round to generate a global agent that improves task completion time across the entire system. Compared to the standard DQN-based scheduling model, simulation results show that using the FL technique improves the learning curve over training time, reducing task delay and speeding up processing by about 50% in scheduling tests. Furthermore, collaborative learning by all trained agents confirms the global agent’s stability and improvement, in contrast to the individual agent’s fluctuating performance.