Speakers
Description
Quantum reservoir computing (QRC) offers a hardware-compatible alternative to variational quantum machine learning by using fixed quantum dynamics as a nonlinear reservoir and training only a classical readout. This work investigates Rydberg-atom QRC for next-step forecasting of insurance-relevant time series, where data are often short, sparse, nonlinear, seasonal, and affected by weather or external shocks. The reservoir is an array of interacting two-level atoms governed by an input-dependent Rydberg Hamiltonian. Time-series inputs are encoded through drive and detuning modulation, and forecasts are produced from time-stacked one- and two-body observables using ridge regression.
We first optimize Rydberg QRC on synthetic nonlinear benchmarks, including NARMA-n tasks and damped harmonic oscillators, then transfer the optimized settings to real-world actuarial forecasting. The empirical study uses STATS19 UK road-safety data, aggregated into daily auto-risk proxies: collision count, casualty count, and wet/icy-road collisions. We compare multi-step and single-step QRC configurations against classical baselines including a tuned echo-state network (ESN), ETS, ARIMA, and persistence forecasts. The ESN baseline uses a 200-unit leaky-integrator reservoir selected from 5000 random NARMA10 configurations over literature-guided hyperparameter ranges.
Preliminary results show that small simulated Rydberg reservoirs outperform the tuned ESN on all three STATS19 tasks, improving test R² by more than 0.21 absolute across collision, casualty, and wet/icy-road targets. These results suggest that many-body Rydberg dynamics can generate useful nonlinear temporal features for noisy actuarial data. Ongoing work extends the study to larger reservoirs and QuEra’s Aquila neutral-atom hardware.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum algorithms for machine learning applications |