An ML-based hybrid task scheduler for classical–quantum computing environments using real graph-derived workloads

Hybrid classical-quantum computing creates a scheduling challenge. Tasks need assignment over time and across different resources, each with unique execution, setup, and queue traits. This paper introduces a machine-learning-based scheduler for placing tasks between classical CPU and quantum processing resources in a simulated environment. We use the Wiki-Vote network from the Stanford Network Analysis Project as a workload source. We build a benchmark of connected Max-Cut task instances by sampling subgraphs of different sizes. Each task is profiled under both classical and quantum execution modes. The resulting dataset combines task features, runtime behaviour, and queue-state conditions. We evaluate multiple learning models for predicting task placement. Their scheduling decisions are compared to classical-only, quantum-only, and rule-based hybrid baselines. Results show that our ML-based scheduler outperforms all others. It achieves the lowest average completion time, smallest makespan, highest throughput, and best decision accuracy. Feature analysis reveals that queue pressure, system regime, and task priority are key factors in placement decisions. Graph and circuit complexity also enhance predictive power. These findings suggest that data-driven scheduling can boost resource efficiency in hybrid classical-quantum environments. This approach offers a practical path for intelligent task placement in emerging heterogeneous systems.