EDGECLOUD-DRL: A DEEP REINFORCEMENT LEARNING-BASED TASK SCHEDULING FRAMEWORK FOR EDGE-CLOUD COMPUTING

Efficient task scheduling is essential in cloud and edge computing environments to minimize execution latency, energy usage, and resource utilization. Although optimization problems can be approached through heuristics and metaheuristics, like Particle Swarm Optimization (PSO), Multi-Objective Bat Algorithm (MBO), and Multi-Objective Particle Swarm Optimization (MOPSO), which are often time-consuming, cannot address dynamic workloads well and are non-adaptive. Recent advances in Deep Reinforcement Learning (DRL)-based methods hold promise for addressing these issues. Yet recent advances in state-of-the-art DRL work, e.g., Deep Q-Networks (DQN), Deep Deterministic Policy Gradient (DDPG) and Multi-Agent DRL encounter scalability bottlenecks, inferior learning efficiency, and challenges of real-time scheduling in large-scale hybrid cloud-edge settings. To address these drawbacks, we propose EdgeCloud-DRL, a delivery framework based on Deep Q-Network (DQN) that uses target network stabilization and experience replay to improve decision-making efficiency. Our proposed framework is based on Q-learning, where at each step, it learns an optimal scheduling policy by updating its set of Q-values to maximize the task success rate while minimizing the latency and energy cost. Numerous experiments were performed using a customized edge-cloud simulation environment, and the developed model was compared with PSO, MBO, and MOPSO for various workload intensities (1000 to 20,000 tasks). Experimental results indicate that EdgeCloud-DRL outperforms baseline methods by 28% in execution latency, 19% in task success rate, and 23% in energy efficiency. The proposed EdgeCloud-DRL framework also provides adaptive, scalable, and learning-driven schedulings, which are very demonstrably applicable in real-time cloud-edge applications, IoT task schedulings, and distributed AI workloads. These findings further manifest the impact of workload management based on reinforcement learning optimization approaches within intelligent workloads in the current computing infrastructures.