Closed-Loop Decision-Focused Learning for User-Aware Cloud Orchestration under Uncertainty

2026-08-06

AI Summary

This work addresses the inefficiencies and contention during peak periods caused by time-varying cloud workloads by proposing CL-DFL, an end-to-end closed-loop decision-focused learning framework. The approach formulates heterogeneous task scheduling as a multi-objective combinatorial optimization problem under uncertainty, enabling joint optimization of resource awareness and scheduling decisions. CL-DFL integrates spatiotemporal forecasting based on MTGNN, zeroth-order decision-focused learning via Tree-structured Parzen Estimator (TPE), and GRPO-driven cooperative local search, further enhanced by the GNeuro-PLS strategy to improve robustness under heterogeneous loads. Experimental results on four real-world datasets demonstrate that the proposed method significantly outperforms existing approaches, achieving a superior trade-off among SLA violation rate, user satisfaction, and resource utilization, while maintaining robust performance even under highly saturated conditions.

Abstract

Time-varying cloud workloads often cause resource under-utilization during off-peak periods and resource contention during peak periods. Existing prediction-then-optimization (PTO) frameworks suffer from two-stage decoupling, hindering the balance among violation rate, user satisfaction, and resource utilization. We formulate heterogeneous job scheduling as a multi-objective combinatorial optimization problem (MOCOP) under uncertain constraints and propose a closed-loop decision-focused learning (CL-DFL) framework for cloud orchestration. CL-DFL integrates a Multivariate Time-series Graph Neural Network (MTGNN)-based spatio-temporal predictor with a zeroth-order decision-focused learning (DFL) mechanism based on the tree-structured Parzen estimator (TPE). This integration establishes an end-to-end (E2E) feedback pathway between resource perception and scheduling decisions. Furthermore, we develop the GNeuro-PLS strategy by incorporating group relative policy optimization (GRPO) into cooperative local search to improve robustness under heterogeneous workloads. Extensive experiments on four real-world datasets demonstrate that CL-DFL achieves superior trade-offs among violation rate, user satisfaction, and resource utilization. It effectively controls overload risks under regular workloads and maintains resilience under highly saturated scenarios compared with state-of-the-art baselines.

Problem

• cloud orchestration
• uncertainty
• user satisfaction
• resource utilization
• violation rate

Innovation

• Closed-loop Decision-Focused Learning
• Multi-objective Combinatorial Optimization
• MTGNN
• Zeroth-order Optimization
• User-aware Orchestration