Reinforcement learning for error-robust control on cloud-based superconducting hardware [Part I]

The noisy nature of today's quantum hardware limits the ability to realize functioning quantum computers. Yet, a careful design of the systems' controls allows researchers to narrow the gap between current and desired hardware capabilities. In this work, we study a black-box optimization technique based on reinforcement learning for the discovery of high-performance gate-sets on a cloud quantum computer. We show that by employing RL, where intermediate information is used to optimize a long term goal, we are able to generate single-qubit gates which, when implemented on a real device, outperform existing model-based optimized pulses. We demonstrate the performance of our learner on IBM quantum hardware accessed via Qiskit Pulse programming. The entire learning process occurs on the quantum device itself, aiming to suit the low-level gate implementation to the underlying details of the specific hardware. This allows gate optimization without any prior knowledge or assumptions on the noise model, hardware limitations or any other undesired effect that exists in real devices. Our experiments demonstrate gates up to 3 times faster than the IBM default, with performance of less than 3e-4 EPG and stability up to 10 days, as compared to the single-day calibration window for IBM pulses.