Classical Reinforcement Learning for Experimental Quantum Error Correction

Quantum feedback control framed as a partially observable Markov decision process poses a significant challenge for reinforcement learning methods due to the minimalistic observability of quantum states through projective measurements. To alleviate this, all applications of reinforcement learning to quantum control to date rely on the simulation of the system which gives the learning agent direct access to the wavefunction. In contrast, using model-free reinforcement learning, we solve a challenging task of learning the parameters of quantum error correction protocol for bosonic grid-state logical encoding directly from ancilla qubit binary measurement outcomes. Such approach, applied to a quantum system whose wavefunction is not directly accessible in a real experiment, will completely eliminate the model bias and allow to learn quantum control policies tailored to the specific analog error channels present in the experimental system.