Model-Free Quantum Control with Reinforcement Learning

Model bias is an inherent limitation of the current dominant approach to optimal quantum control, which relies on a system simulation for optimization control policies. To overcome this limitation, we propose circuit-based training reinforcement learning agent tasks in model-free way. Given continuously parameterized circuit, learns its parameters through trial-and-error interaction with system, using measurement outcomes as only source information about state. Focusing harmonic oscillator coupled ancilla qubit, show how reward measurements experimentally available observables. We train prepare various non-classical states both unitary and adaptive measurement-based feedback, execute logical gates encoded qubits. This significantly outperforms widely used methods terms sample efficiency. Our numerical work immediate relevance superconducting circuits trapped ions platforms where such can be implemented experiment, allowing complete elimination model adaptation policies specific they are deployed.