⚛️ quantum physics

QCOEM: Quantum Cloud Orchestration with Evolutionary Multi-Objective Optimization

The paper proposes QCOEM, a quantum cloud orchestration framework that utilizes evolutionary multi-objective optimization (specifically NSGA-II/III) combined with an Augmented Achievement Scalarization Function to dynamically schedule tasks across heterogeneous backends, achieving significantly higher execution fidelity and zero rescheduling compared to existing noise-agnostic heuristics.

Original authors: Tam N. Pham, Hoa T. Nguyen, Quan Le-Trung

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tam N. Pham, Hoa T. Nguyen, Quan Le-Trung

Original paper licensed under CC BY 4.0 (http://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

The Quantum Traffic Jam

Imagine a world where computers don't just count numbers; they dance with probabilities. This is the realm of quantum computing, a field where machines use tiny particles called "qubits" to solve problems that would take supercomputers thousands of years. But right now, these machines are like fragile, temperamental artists. They live in a phase scientists call the "Noisy Intermediate-Scale Quantum" (NISQ) era. Think of them as high-performance race cars that are also incredibly sensitive to the weather; a little bit of static electricity or a slight temperature change can ruin their performance.

Because these machines are so rare and expensive, we can't just buy one for every person. Instead, they are offered as a service over the internet, much like renting a supercomputer in the cloud. This is called Quantum-as-a-Service. However, managing a fleet of these fragile, different-looking quantum computers is a nightmare. Some are fast but make mistakes often; others are slow but very accurate. Some have broken parts (called "noise"), and their queues (waiting lines) change every second. If you send a complex math problem to the wrong machine at the wrong time, the answer might be garbage. The big question for scientists is: How do we automatically decide which job goes to which machine to get the best answer as fast as possible, without crashing the system?

The Smart Scheduler: QCOEM

Enter QCOEM, a new "traffic cop" for the quantum cloud, proposed by researchers Tam N. Pham, Hoa T. Nguyen, and Quan Le-Trung. Instead of using simple, old-fashioned rules to assign jobs (like "send the next job to the first free machine"), QCOEM uses a clever, nature-inspired strategy called Evolutionary Multi-Objective Optimization.

To understand why this is special, imagine you are organizing a massive school talent show with three different judges who hate each other. Judge A wants the show to end as fast as possible. Judge B wants every act to be perfect, even if it takes longer. Judge C wants to make sure no single stage is overcrowded while another sits empty. If you just try to please Judge A, the show might be fast but full of bad acts. If you only please Judge B, the show might drag on forever.

Old methods tried to combine these judges into one score, but they often failed because the "perfect" balance isn't a straight line; it's a messy, curved shape that's hard to find. QCOEM, however, uses an algorithm that acts like a swarm of evolving ants. It tries thousands of different ways to assign the talent show acts to the stages. It doesn't just look for one "best" answer; it finds a whole menu of "good enough" options that balance speed, accuracy, and fairness. Then, it uses a special decision tool (called AASF) to pick the single best schedule based on what the user cares about most at that moment.

The researchers tested this system in a simulated quantum cloud with five different types of "virtual" quantum computers. They fed it a batch of real-world quantum problems (like those used in chemistry and finance) and watched how it performed compared to the old, simple methods.

The results were promising. In these simulations, QCOEM managed to eliminate the need to reschedule tasks entirely, whereas the old methods often had to send jobs back to the line because they were assigned to broken machines. More importantly, the average quality of the answers (called "fidelity") was about 30% higher than the noise-ignoring methods. The system also showed that by tweaking the "weights" (telling the system to care more about speed or more about accuracy), users could get different results:

If you wanted speed, the system focused on the fastest machines, but this sometimes created long waiting lines and lower accuracy.
If you wanted quality, the system picked the most reliable machines, boosting accuracy to about 0.505 (compared to 0.376 for the old methods), though the total time increased.
If you wanted balance, the system found a sweet spot, keeping the load spread out evenly across all machines.

While the system does take a bit more time to make the decision (about 94.58 seconds for a large batch of 50 tasks, compared to less than a second for the simple methods), the researchers suggest this is a fair trade-off. In the quantum world, where waiting in line for a real machine can take hours or days, spending a minute or two to ensure the job is assigned perfectly is worth it.

The paper concludes that QCOEM suggests a way to make quantum cloud computing more stable and reliable. It doesn't claim to have solved every problem in the universe, but it offers a much smarter way to navigate the messy, noisy reality of today's quantum hardware. By treating the scheduling problem as a complex balancing act rather than a simple race, QCOEM helps ensure that when we finally harness the full power of quantum computers, we won't be wasting time on broken answers.

Technical Summary: QCOEM – Quantum Cloud Orchestration with Evolutionary Multi-Objective Optimization

1. Problem Statement

The practical deployment of quantum computing has shifted toward cloud-based Quantum-as-a-Service (QaaS) platforms. However, current hardware operates in the Noisy Intermediate-Scale Quantum (NISQ) era, characterized by limited qubit counts, gate/readout errors, short coherence times, and restricted connectivity. Crucially, these noise profiles and topologies vary across devices and drift over time, meaning identical circuits can yield significantly different fidelities depending on the execution backend.

Existing cloud orchestrators largely rely on noise-agnostic heuristics (e.g., Greedy, Round Robin) designed to maximize throughput and minimize latency in classical clusters. When applied to quantum workloads, these strategies are systematically suboptimal because they ignore backend-specific error characteristics. This leads to:

Reduced execution fidelity.
Load imbalances (queueing hotspots on high-quality devices).
Frequent task rescheduling due to infeasible assignments or stale state information.

The core challenge is formulated as a Multi-Objective Optimization Problem (MOOP) where a scheduler must simultaneously navigate three competing goals: minimizing mean completion time, minimizing execution error rates, and minimizing load imbalance. The problem space is non-convex due to the multiplicative accumulation of per-operation errors with circuit depth, rendering classical weighted-sum scalarization methods ineffective as they cannot recover optimal solutions in concave regions of the Pareto front.

2. Methodology: The QCOEM Framework

The authors propose QCOEM, a quantum cloud orchestration framework that leverages Evolutionary Algorithms for multi-objective task scheduling. The system is designed to be noise- and delay-aware, utilizing real-time hardware data rather than idealized analytical models.

System Architecture

QCOEM is deployed on a Kubernetes-based infrastructure comprising:

Centralized Scheduler: Manages a global queue, batches incoming tasks, and acts as the control plane. It maintains a cached state of the cluster derived from periodic heartbeats.
QNodes (Quantum Nodes): Data-plane workers that emulate specific quantum backends using calibrated snapshots (e.g., coupling maps, gate durations, error rates). They execute circuits sequentially via a First-In-First-Out (FIFO) queue and report availability and queue states via heartbeats.

Mathematical Formulation

The orchestration task is defined as minimizing an objective vector $F(X) = [f_1, f_2, f_3]$ subject to hardware constraints (indivisible tasks, unique assignment, and qubit capacity):

Mean Completion Time ( $f_1$ ): Minimizes the average time tasks spend in the system, calculated as the sum of the current queueing delay (derived from heartbeat telemetry) and the estimated execution time based on the circuit's critical path.
Mean Error Rate ( $f_2$ ): Minimizes the average probability of failure, estimated by multiplying the success probabilities of individual gates and readout operations on the target QNode.
Load Imbalance ( $f_3$ ): Minimizes the standard deviation of completion times across all QNodes to prevent queueing hotspots and ensure fair utilization.

Optimization Strategy

To solve this NP-hard combinatorial problem, QCOEM employs Non-dominated Sorting Genetic Algorithms (NSGA-II and NSGA-III) to approximate the Pareto-optimal front.

Handling Non-Convexity: Unlike weighted-sum approaches, the evolutionary algorithms can navigate the non-convex regions of the objective space.
Decision Making: To select a single dispatchable schedule from the non-dominated set, the framework applies an Augmented Achievement Scalarization Function (AASF). This function maps the Pareto set to a single solution based on user-defined preference weights ( $w$ ), allowing the system to adapt to different Service Level Agreements (e.g., prioritizing speed vs. fidelity).

3. Key Contributions

The paper outlines three primary contributions:

Framework Proposal: The introduction of QCOEM, a noise- and delay-aware orchestration framework that integrates calibrated hardware snapshots and heartbeat-based queueing estimates to jointly optimize completion time, error rate, and load imbalance.
Algorithmic Innovation: The application of NSGA-II and NSGA-III to approximate task-to-QNode assignments, coupled with AASF for preference-based selection. The authors note this is among the first frameworks to combine AASF-based decision-making with an explicit load-imbalance objective in quantum scheduling.
Empirical Evaluation: A comprehensive evaluation using MQTBench workloads in a heterogeneous quantum cloud environment, demonstrating the trade-offs between latency, fidelity, and stability.

4. Performance Results

The framework was evaluated in a heterogeneous environment with 5 QNodes emulating IBM Quantum backends, processing workloads from MQTBench (including VQE and QAOA algorithms).

Fidelity vs. Baselines: Compared to noise-agnostic baselines (Greedy, Round Robin, Random), QCOEM achieved approximately 30% higher mean fidelity. For instance, while Greedy achieved a mean fidelity of ~0.376, NSGA-II/III raised this to ~0.49.
Rescheduling: The framework achieved zero task rescheduling, whereas noise-agnostic heuristics often triggered infeasible assignments requiring re-execution.
Trade-offs:
- Load-Focus: Configurations prioritizing load balancing reduced total completion time by mitigating queueing delays (e.g., ~5.68s for NSGA-II).
- Quality-Focus: Prioritizing fidelity increased mean fidelity to ~0.505 but extended total completion time and load imbalance.
- Speed-Focus: Aggressive routing for speed minimized execution time but significantly increased waiting times and load imbalance, ultimately degrading end-to-end latency and fidelity.
Algorithm Robustness: NSGA-II demonstrated greater robustness and lower variance than NSGA-III across repeated runs, particularly in the "Balanced" configuration.
Overhead: The scheduling overhead for MOEAs was higher than simple heuristics (e.g., ~94s for NSGA-II with batch size 50 vs. <1s for Greedy). However, the authors argue that in QaaS settings where QPU queueing dominates end-to-end latency, the gains in feasibility and fidelity outweigh the control-plane latency for moderate-to-large batches.

5. Significance and Claims

The paper claims that QCOEM addresses a critical gap in quantum cloud orchestration: the lack of end-to-end, cloud-native workflows that jointly model noise, queueing dynamics, and load imbalance.

Stability and Fidelity: The framework provides a mechanism to deliver stable, high-fidelity execution by avoiding the pitfalls of noise-agnostic routing.
User-Centric Control: By utilizing AASF, the system allows operators to dynamically adjust routing behavior based on specific priorities (speed, quality, or balance) without re-engineering the scheduler.
Scalability: While acknowledging the computational cost of evolutionary search, the results suggest that for batched workloads typical in cloud environments, the benefits of optimized resource allocation justify the scheduling overhead.

The authors conclude that noise-aware scheduling is a necessary evolution for quantum cloud computing to move from theoretical investigations to reliable, high-performance services. Future work is proposed to include broader benchmarks against deep reinforcement learning techniques and validation on real quantum hardware.