Research on Cloud Task Scheduling Problem Based on Pareto Optimization Strategy for Deep Reinforcement Learning

Abstract Cloud data centers face highly dynamic cluster loads and sharply heterogeneous task demands, which place heavy pressure on task scheduling. Most existing schedulers rely on local state features, adapt poorly to load fluctuations, and struggle to balance competing objectives. This paper proposes ITS-D3QN (Interactive Two-Stage D3QN), a deep reinforcement learning scheduler built on D3QN (Dueling Double Deep Q-Network) and driven by a Pareto optimization strategy. The method has three components. First, a two-layer scheduling model is formulated in which the lower layer optimizes task completion time, system energy consumption, and resource utilization. Second, the upper layer uses the Pareto hypervolume (HV) to capture cluster-wide load and resource conditions, and the scheduling problem is cast as a Markov decision process. Third, an incremental sampling mechanism is introduced to improve convergence and adaptability to non-stationary load. Simulations on the Alibaba Cluster Trace v2018 show that, relative to the compared algorithms, ITS-D3QN reduces average task response time by approximately 4.75% and average energy consumption per task by approximately 4.32%, while raising average resource utilization by approximately 5.62%. These results indicate that the method achieves balanced scheduling in complex cloud environments.