Realizing a Reinforcement Learning Agent on a Field-Programmable Gate Array for Real-time Control of Superconducting Qubits

Real-time controllers of quantum systems process the outcome of intermediate measurements to determine which subsequent actions to apply to the quantum system. Realizing such adaptive control, on timescales much shorter than the coherence time, has a wide range of potential applications, such as in quantum error correction and in quantum state preparation. Here, we implement a deep neural network on a field-programmable gate array (FPGA) and investigate its use as a real-time reinforcement learning agent to efficiently initialize a transmon qubit into its ground state. The agent repeatedly measures the qubit and chooses after each cycle whether to idle, to apply a bit-flip gate, or to terminate. After the agent chooses to terminate the initialization process, we perform a validation measurement to infer the probability of having successfully initialized the ground state. To train the agent, we use model-free reinforcement learning that is based solely on measurement data.