A computing power network task dynamic scheduling method and system

By refining and enhancing the historical scheduling trajectory dataset, it is divided into sub-policy networks with multiple scheduling strategy modes, and a meta-policy network is constructed. This solves the robustness and accuracy problems of computing power network task scheduling in dynamic heterogeneous environments, and realizes efficient and flexible scheduling decisions.

CN122111640APending Publication Date: 2026-05-29CHINESE PEOPLES LIBERATION ARMY INFORMATION SUPPORT CORPS ENGINEERING UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINESE PEOPLES LIBERATION ARMY INFORMATION SUPPORT CORPS ENGINEERING UNIVERSITY
Filing Date
2025-12-31
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Existing computing power network task scheduling strategies lack robustness in dynamic heterogeneous computing power environments, cannot guarantee scheduling accuracy and efficiency in various scenarios, and are susceptible to data pseudo-correlation interference. A single strategy is difficult to cover diverse and effective scheduling styles.

Method used

By acquiring historical scheduling trajectory datasets, trajectory purification and counterfactual trajectory data augmentation are performed to form a high-quality causal trajectory set. Based on unsupervised clustering, the set is divided into sub-policy networks of various scheduling strategy modes. A meta-policy network is constructed to dynamically generate scheduling actions, and a model-independent meta-learning framework is used for training and deployment.

Benefits of technology

It achieves high-precision, robust, and fast-adaptive scheduling in dynamic heterogeneous computing environments, supports continuous iterative optimization of strategies, and avoids strategy bias caused by spurious correlations in historical data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122111640A_ABST
    Figure CN122111640A_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of computing resource scheduling, in particular to a computing network task dynamic scheduling method and system. Historical scheduling trajectories are obtained to form a historical scheduling trajectory dataset; a high-quality causal trajectory set is formed based on the historical scheduling trajectory dataset; the high-quality causal trajectory set is divided into M trajectory subsets, and M sub-strategy networks are trained; during training, a loss function weight is configured based on a trajectory quality weight; a meta-strategy network integrating the M sub-strategy networks is constructed and trained using a model-independent meta-learning framework; a dynamic weight vector is dynamically generated according to an environment state feature vector and a task feature vector, and a scheduling action is output based on the dynamic weight vector; the meta-strategy network is deployed to a computing network scheduler to process real-time scheduling requests containing environment state feature vectors and task feature vectors. The present application can achieve high accuracy, strong robustness and fast adaptability in computing network task scheduling.
Need to check novelty before this filing date? Find Prior Art