Extending the logical lifetime of a stabilized Gottesman-Kitaev-Preskill qubit with reinforcement learning

Gottesman-Kitaev-Preskill (GKP) encoding of a qubit in an oscillator is a promising candidate for quantum error correction (QEC) in bosonic systems. We introduce a novel QEC protocol for the GKP encoding, and experimentally implement it with a superconducting microwave cavity coupled to a transmon ancilla qubit. The protocol contains 33 tunable parameters which we optimize in-situ using a reinforcement learning agent. We demonstrate that the agent learns interpretable directions in parameter landscape, which would be difficult and time-consuming to discover and optimize for a human experimentalist. The learned QEC protocol for GKP has process fidelity on par with that of the simplest uncorrectable Fock encoding, and yields X & Z logical Pauli operator lifetimes approaching 1 ms. Our experiment clearly demonstrates the advantage of closed-loop optimization for quantum control over the limited model-based tuneup.