Parameter-Efficient Quantum-Inspired Fast Weight Programmers for Traffic-Matrix Forecasting

The Core Finding

Quantum-inspired recurrent networks can forecast network traffic matrices more accurately than standard deep learning while using a fraction of the parameters. The authors' best model (G-QKANFWP) achieves superior prediction accuracy on real network data while consuming only 22.4% of the parameters required by a larger LSTM baseline—a significant win for deployment in bandwidth-constrained network controllers.

Why This Matters

Traffic matrices describe all the data flows crossing a network, from every origin to every destination. Predicting these patterns hours ahead is essential for routing optimization and network management. But network control systems operate under brutal constraints: limited memory on routers, strict energy budgets, and minimal time to train or update models. Classical approaches—especially large transformers or diffusion models—are too expensive for these settings. This paper asks whether quantum-inspired architectures, which mimic certain quantum-mechanical properties classically, can break this efficiency-accuracy tradeoff.

The prediction task itself is demanding: forecasting 144 separate origin-destination channels 20 steps ahead (a 100-minute window) using only two hours of history. Standard LSTMs struggle here without becoming prohibitively large.

How the Method Works

The authors adapt a quantum-inspired Kolmogorov-Arnold network (QKAN) into a "fast-weight programmer" framework. This is a recurrent architecture with two components: a slow "programmer" network that learns general patterns, and a fast "programmable" module that adapts quickly to shifting traffic conditions. The quantum inspiration comes from using Kolmogorov-Arnold basis functions—a mathematical tool that represents complex functions more efficiently than standard neural network layers—to approximate quantum-like superposition and entanglement classically.

The key variants tested are:

• G-QKANFWP: A gated version pairing a classical slow LSTM programmer with a quantum-inspired fast programmer
• GQKAN-FWP: Full quantum-inspired architecture in both components
• Baselines: matched-size LSTM, larger LSTM, and a classical gated fast-weight programmer

All models train under identical compute budgets to ensure fair comparison.

The Results

On the Abilene network dataset (a real backbone network used extensively in telecom research), G-QKANFWP achieved the lowest root-mean-square error (RMSE)—the standard metric for forecasting accuracy. Critically, it outperformed both the parameter-matched and larger LSTM baselines, proving the quantum-inspired advantage isn't simply about the fast-weight framework itself.

The authors also tracked convergence speed: quantum-inspired variants showed lower validation-loss area under the learning curve (AULC), meaning they learn faster and stabilize better. Channel-wise breakdowns reveal that G-QKANFWP and GQKAN-FWP win on the majority of individual OD-channel predictions, not just in aggregate.

The 22.4% parameter ratio is the headline: comparable accuracy at one-quarter the size opens deployment pathways for real network hardware that would reject larger models outright.

This work suggests that quantum-inspired classical methods—borrowing mathematical structures from quantum computing without requiring quantum hardware—may offer a pragmatic sweet spot for resource-constrained forecasting in operational networks.