Speaker
Description
Training variational quantum circuits (QAOA, VQE, and quantum neural-network (QNN) models) is a noisy, shot-expensive stochastic optimization problem. The de-facto standard optimizer, SPSA, is powerful but performs well only once its hyperparameters are tuned per problem, and that tuning is itself a shot-hungry optimization.
We present a scalable local Gaussian-process-model (GPM) optimizer that is tuning-free by construction. Because a circuit's expectation value is a trigonometric polynomial in each rotation angle, the structure of the reproducing kernel of the cost function is known in advance: we equip a GP surrogate with a product kernel whose factors are read directly off the circuit ansatz, encoding the known frequencies, periodicity, or smoothness of each parameter when applicable.
At every iteration, GPM uses the accumulated noisy observations to drive all internal decisions from the posterior alone: the current argmin estimate, the optimal number of measurements in the next iteration, their placement, and the number of shots spent on them. Crucially, the number of probes per iteration does not grow with the parameter count, making the approach scalable, in contrast to previous proposals of global GPM optimizers.
Across four benchmarks (QAOA, a hardware-efficient VQE, a unitary-coupled-cluster VQE, and a QNN classifier), a single untuned configuration of our GPM optimizer matches or exceeds SPSA tuned per problem, at markedly better shot efficiency and without any per-problem tuning.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |