Task scheduling is a fundamental challenge in edge-cloud computing systems, as it needs to address dynamic workloads, heterogeneous resources, and stringent latency requirements. While many traditional scheduling algorithms, such as First-In-First-Out (FIFO) and Round-Robin, are generally less adaptive to real-time situations, existing Deep Reinforcement Learning (DRL) methods may struggle with limited scalability and insufficient multi-objective optimisation. In this paper, we propose a new adaptive DRL (deep reinforcement learning) task scheduling framework that overcomes these limitations through adaptive decision-making and a multi-objective reward optimisation mechanism. It introduces an established Deep Q-Network (DQN) architecture with a changing state representation layer and a multi-criteria reward function that concurrently maximises latency, energy use, and SLA violations. We perform extensive experiments using synthetic workloads and real-world datasets (Google Cluster Traces and Azure Functions), demonstrating significant performance gains over state-of-the-art baselines. The framework from our proposal can reduce average latency by 33.3% (compared to FIFO), reduce SLA violations by 60% compared to these baseline methods, improve energy efficiency by 17.2% over DDQN, and maintain a 98.4% task completion rate with dynamic workloads. With statistical validation demonstrating robustness and scalability, the approach emerges as an efficient candidate for near-real-time, at-scale performance across edge-cloud deployments. The potential of DRL for next-generation edge-cloud computing systems is evident in its ability to adapt to environmental changes while jointly achieving multiple performance objectives.
Sravan et al. (Tue,) studied this question.