A quantum-inspired neural model for forecasting cloud workloads reported substantially lower prediction error than five comparison methods in benchmark tests, with average reductions ranging from 22.83% to 75.12%. The evidence came from simulated conditions, and the paper's deployment discussion acknowledges practical challenges in real-world integration.
How the model works
The proposed system, called QB-HNN, applies controlled-Hadamard activation at hidden and output qubit neurons. It provides two versions: QB-HNN-SAP for single-resource prediction and QB-HNN-MAP for multiple-resource prediction.
Both variants use QB-BiO for offline training on prior workload data, while prediction is described as occurring concurrently in real time. The comparison assessed normalized mean squared error and mean absolute error, two measures of prediction mistakes, along with prediction accuracy, average training time, computational complexity, convergence and the frequency of absolute errors.
Reported average error reductions were 22.83% against EQNN, 72.15% against SaDE, 72.16% against BaDE, 36.36% against LSTM and 75.12% against NN-BP. No confidence intervals or formal uncertainty estimates were reported for these comparisons.
The benchmark evidence
A separate RMSE comparison, using a measure of prediction error, reported QB-HNN as the lowest across the tested windows for the GC traces. At the 30-minute window, its reported RMSE was 2.42 versus 11.6 for MCT-AQNN, while on GM it was 0.51 at five minutes.
The evaluation used six benchmark datasets spanning three workload categories: Cluster, Web server and High Performance Computing. For the cluster inputs, logs were recorded every five minutes over 29 days. The first 10 days yielded 2 million CPU entries and 2 million memory entries across two channels, with prediction windows of 5, 10, 30 and 60 minutes and visible segments of 60 minutes.
The other traces were reported as follows: NASA-HTTP covered 31 days and 1.8 million records; Saskatchewan covered seven months and 2.5 million records; AuverGrid covered 365 days and 2.3 million records; and SHARCNet covered 11 days and 188,000 records.
The reported configuration used 10 input nodes, 7 hidden nodes and 1 output node, with a maximum of 50 epochs and a 75% training-data ratio. Its qubit-star population ranged from 15 to 36, and the model used four qubit clusters.
Accuracy came with a cost
The reported convergence comparison favored QB-BiO in the cited cluster analysis: it was reported to converge before 20 epochs, versus more than 45 epochs for SaDE and BaDE, which had slightly higher prediction error.
The multi-resource QB-HNN-MAP variant was reported to have accuracy equivalent to the single-resource QB-HNN-SAP variant with similar epoch usage, but higher average training time and memory consumption. Simultaneous prediction used almost two times the memory required for a single resource. No formal equivalence test or uncertainty estimate was reported for that comparison. The authors also report computational complexity equivalent to SaDE and BaDE and greater than NN-BP.
A sensitivity analysis showed that the results changed with the tested qubit-cluster and population settings. K = 5 and Z = 25 produced the lowest reported testing MSE, 0.00412; Z = 15 with K = 5 produced 0.00551, while K = 6 with Z = 36 produced a training MSE of 0.06404.
What the test cannot answer
Those figures describe a forecasting benchmark, not a field trial. The study's deployment discussion acknowledges that the convergence and learning-efficiency evidence came from simulated conditions and that integrating the method into real-world systems poses practical challenges. The reported metrics concern prediction performance and computational use under the tested setup, so they do not establish an operational gain in resource reservation.
The results therefore do not establish that the reported advantage would generalize beyond the six benchmark datasets and tested parameter settings, or that it would persist in a live resource-management system.
The visible front matter carries arXiv:2608.25754v1 dated 26 August 2026 and states that the article has been accepted in IEEE Transactions on Systems, Man, and Cybernetics Journal, with a 2025 IEEE copyright notice.
Paper data and sources
Original title: Quantum Blackhole Learning-Optimized Hadamard Neural Network Model for Dynamic Resource Reservation in Industry Clouds
Authors: Deepika Saxena, Hari Mohan Gaur, Ashutosh Kumar Singh, Anand Mohan
Journal/Repository: IEEE Transactions on Systems, Man, and Cybernetics: Systems 2025
Status: Peer-reviewed
First online: 2026-08-26
DOI: 10.1109/tsmc.2025.3621341
Original paper · Full text