Designing quantum gates using deep reinforcement learning

The advantage of a quantum computer over its classical counterpart relies heavily on the ability to perform high-fidelity quantum logic operations. Theoretical and empirical studies of error sources have resulted in many promising designs for the standard single-qubit and two-qubit gates. While these candidates capture the majority of the dynamics, unknown error processes in a realistic hardware are not explicitly addressed but implicitly via frequent calibration. In this work, we task a deep reinforcement learning agent to interact with a simulated quantum environment of superconducting transmon qubits to directly design quantum gates suitable to the true dynamics. With a learning objective based on the worst-case fidelity, instead of the commonly used average fidelity, our agent explores the vast design landscape of piecewise-constant pulses and finds non-trivial solutions for single-qubit rotation and cross-resonance entangling operation.