Preprint reports faster parameter selection for quantum optimization

Researchers report in a preprint that learned models select parameters for the Quantum Approximate Optimization Algorithm several orders of magnitude faster than conventional optimization, with near-identical quality. The algorithm draws candidate solutions to combinatorial problems from a quantum circuit whose performance depends on those parameters. The models replace a separate optimization process for each problem with a single model evaluation. Tests on quantum hardware produced samples indistinguishable from those using conventionally optimized parameters, the researchers say. The models also generalized to unseen graph structures. Generating training data and training the models impose upfront costs, which the approach spreads over repeated use on similar problems.