💻 computer science

An ML-Based Hybrid Task Scheduler for Classical–Quantum Computing Environments Using Real Graph-Derived Workloads

This paper presents a machine learning-based hybrid task scheduler that optimizes resource allocation between classical and quantum processors using real graph-derived workloads, demonstrating superior performance in completion time, makespan, and throughput compared to traditional and rule-based baselines.

Original authors: Peter Nimbe, Nicodemus Songose Awarayi, Vivian Akoto-Adjepong, Faiza Umar Bawah, Patrick Kwabena Mensah, Obed Appiah, Christopher Bombie Ninfaakang, Mighty Abra Ayidzoe, Augustine Takyi, Bernard Andoh

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Peter Nimbe, Nicodemus Songose Awarayi, Vivian Akoto-Adjepong, Faiza Umar Bawah, Patrick Kwabena Mensah, Obed Appiah, Christopher Bombie Ninfaakang, Mighty Abra Ayidzoe, Augustine Takyi, Bernard Andoh, Daniel Asamoah Apraku, Frank Opoku Kwarteng, Farida Osman

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the evolving landscape of modern computing, two distinct worlds are beginning to merge. On one side sits the classical computer, the familiar machine found in offices and homes, which processes information using bits that are either zero or one. On the other is the quantum computer, a specialized machine that uses quantum bits, or qubits, to explore many possibilities simultaneously. While quantum machines hold the promise of solving certain complex problems faster than their classical counterparts, they are not yet a universal replacement. They are often difficult to access, require extreme conditions to operate, and can be slow to set up for a single task. This reality has given rise to hybrid environments, where a system must decide, in real time, whether to run a specific job on a standard processor or send it to a quantum one. The challenge lies in making this choice efficiently; sending a task to the wrong machine can waste time, while sending it to the right one could unlock significant speed.

Researchers at the University of Energy and Natural Resources in Ghana have tackled this scheduling puzzle by building a smart system that learns how to make these decisions. Instead of relying on fixed rules that might fail when conditions change, they trained a machine-learning model to act as a traffic controller for computing tasks. To teach this system, they did not use made-up data. They started with a real-world network of interactions from a Wikipedia voting system, a massive graph of connections between users. From this complex web, they extracted hundreds of smaller, connected groups of users to serve as test cases. Each group represented a specific type of optimization problem known as the Max-Cut problem, which involves dividing a network into two groups to maximize the connections between them.

The team then ran every single one of these test cases through two different paths. First, they solved them using a standard classical computer to see how long it took and how good the answer was. Second, they sent the same problems through a simulated quantum workflow, which mimics the behavior of a real quantum processor, including the time it takes to prepare the machine and the time it takes to run the calculation. Crucially, they did not just look at the raw speed of the machines. They also simulated different levels of congestion, or "queue pressure," to see how waiting times affected the total cost of running a task. Sometimes the quantum machine was free and fast to set up; other times it was backed up with a long line of waiting jobs. By combining the structural details of the graph problems with these changing system conditions, they created a rich dataset that taught the computer when to choose the classical path and when to choose the quantum path.

The results showed that a simple machine-learning model, specifically one based on logistic regression, could learn to make these placement decisions with remarkable accuracy. In tests, this learned scheduler correctly identified the best resource for nearly 98 percent of the tasks. When compared to other methods, the smart scheduler outperformed a system that sent everything to the classical computer, a system that sent everything to the quantum computer, and even a system that followed a set of manual, pre-written rules. The learned scheduler achieved the fastest overall completion times and the highest number of tasks finished per second. It managed to balance the load so effectively that it reduced the total time the system spent waiting for tasks to finish by a significant margin compared to the rule-based approaches.

A closer look at what the computer learned revealed a surprising insight: the most important factor in deciding where to send a task was not the complexity of the problem itself, but the current state of the system. The model paid the most attention to how long a task would have to wait in line for the classical computer versus the quantum machine. If the classical queue was long, the system was more likely to send a task to the quantum processor, even if the quantum machine had its own overhead. Conversely, if the quantum line was backed up, the system kept the task on the classical side. This suggests that in a hybrid environment, the best strategy is not a rigid rule about which problems belong to which machine, but a flexible approach that reacts to real-time congestion. The study found that while the size of the problem mattered, it was secondary to the immediate availability of resources.

The researchers also discovered that the quantum workflow, even in simulation, produced solutions that were nearly as good as the classical ones, with an accuracy rate of almost 99.8 percent. This means that the quantum path was a viable option for these tasks, provided the system could manage the timing correctly. The study did not claim that quantum computers are now faster for all problems; in fact, the simulations showed that for small tasks, the classical computer was often much faster in raw execution time. The value of the quantum machine emerged only when the system learned to route tasks to it during moments when the classical resources were strained.

Ultimately, this work demonstrates that managing a mix of classical and quantum resources requires a dynamic, data-driven approach. By treating task placement as a learning problem rather than a static rule, systems can adapt to the ebb and flow of demand. The findings suggest that as quantum technology matures and becomes more integrated with classical infrastructure, the ability to make these split-second decisions based on current conditions will be just as important as the raw power of the machines themselves. The study provides a practical blueprint for how to build these intelligent schedulers, proving that with the right data, a computer can learn to navigate the complexities of a hybrid future.

Technical Summary: An ML-Based Hybrid Task Scheduler for Classical–Quantum Computing Environments Using Real Graph-Derived Workloads

Problem Statement
The emergence of hybrid classical–quantum computing environments introduces complex scheduling challenges that traditional homogeneous or mildly heterogeneous schedulers cannot address. In these systems, a scheduler must decide whether to execute a task on a classical CPU or offload it to a quantum processing unit (QPU). This decision is non-trivial because it must account for significant disparities in execution overhead, setup costs, and queue behaviors between the two resource types. Static allocation rules or fixed thresholds often fail because the optimal execution target depends on dynamic system conditions (such as queue pressure) and specific workload structures (such as graph complexity), rather than a static assumption that quantum execution is inherently superior for optimization tasks.

Methodology
The study proposes a supervised machine learning approach to task placement, treating the scheduling decision as a binary classification problem (CPU vs. QPU). The research follows a five-stage experimental design:

Workload Generation: The authors utilized the Wiki-Vote network from the Stanford Network Analysis Project (SNAP) as a real-world workload source. The directed graph was converted to an undirected representation, and connected subgraphs of varying sizes (6, 8, 10, 12, and 14 nodes) were sampled to create 200 Max-Cut task instances.
Profiling: Each task instance was profiled under two execution modes:
- Classical: Solved using an exact brute-force Max-Cut solver to establish ground-truth optimal cut values and runtimes.
- Quantum: Solved using a Qiskit-based QAOA (Quantum Approximate Optimization Algorithm) workflow on a simulator (Qiskit Aer) with a circuit depth of $p=1$ and 256 shots.
Dataset Construction: A unified dataset was constructed by merging structural graph features (e.g., node count, density), circuit descriptors (e.g., gate count, depth), and execution profiles. Crucially, heterogeneous queue-state regimes (CPU-favorable, Balanced, QPU-favorable) were simulated to introduce variability in queue delays and setup costs, ensuring the dataset reflected realistic placement diversity. The final dataset consisted of 200 tasks with an 80:20 split between CPU-labeled and QPU-labeled instances.
Model Training: Four machine learning models—Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting—were trained to predict the preferred execution target based on input features.
Evaluation: The models were evaluated at two levels:
- Prediction Level: Using accuracy, precision, recall, and F1-scores.
- Policy Level: Comparing the ML-based scheduler against three baselines: Classical-only, Quantum-only, and Rule-based Hybrid. Metrics included average completion time, makespan, throughput, and decision accuracy.

Key Contributions
The paper makes four primary contributions to the field of hybrid computing:

Framework Definition: It defines hybrid classical–quantum scheduling as a machine learning problem, moving beyond static offloading rules.
Realistic Benchmark: It establishes a workload benchmark derived from the Wiki-Vote dataset, generating connected Max-Cut task instances that reflect real-world graph structures rather than synthetic abstractions.
ML Scheduler Development: It develops and evaluates an ML-based scheduler capable of predicting optimal task placement under varying system regimes and queue conditions.
Comparative Analysis: It provides a rigorous comparison of the learned scheduler against classical-only, quantum-only, and rule-based hybrid baselines, demonstrating the superiority of data-driven approaches in heterogeneous environments.

Results

Profiling Insights: Classical execution was significantly faster (mean ~~0.023s) than quantum execution (mean ~3.79s) on these small graph instances. However, the quantum workflow achieved a high approximation ratio (~~0.998) relative to the classical optimum, indicating that solution quality remained high despite latency overheads.
Model Performance: Logistic Regression emerged as the best-performing model, achieving an accuracy of 0.975 and a QPU F1-score of 0.9412 on the test set. It correctly classified all QPU-favorable tasks in the test set.
System-Level Performance: The ML-based hybrid scheduler outperformed all baselines. It achieved the lowest average completion time (1.73s vs. 1.97s for Classical-only and 5.39s for Quantum-only), the smallest makespan, and the highest throughput.
Feature Importance: Analysis revealed that dynamic system-state variables were the dominant factors in decision-making. Specifically, queue_cpu_sec (CPU queue delay), system_regime, and queue_qpu_sec were the most influential features, outweighing static graph properties. This indicates that the scheduler prioritizes real-time resource congestion over static task characteristics.

Significance and Claims
The paper claims that effective hybrid resource management should be viewed as an adaptive scheduling problem rather than a static offloading rule. The findings suggest that data-driven scheduling can significantly boost resource efficiency in emerging heterogeneous systems. By leveraging machine learning, the scheduler can adapt to the interplay between graph complexity, execution behavior, and changing system conditions (such as queue pressure). The study concludes that while graph structure matters, practical hybrid scheduling depends heavily on real-time system states; a task structurally suitable for quantum execution may still be better assigned to a CPU if the quantum queue is congested, and vice versa. This approach offers a practical path for intelligent task placement in future classical–quantum environments.

Technical Summary: An ML-Based Hybrid Task Scheduler for Classical–Quantum Computing Environments Using Real Graph-Derived Workloads

More like this