Reinforcement learning for toric code error correction

I will present a summary of our efforts to use deep reinforcement learning (DRL) for quantum error correction of the toric code. A convolutional neural network is trained by exploration of the state space consisting of (hidden) error configurations and corresponding (visible) syndromes. No external input to the algorithm is provided apart from the current syndrome and the final success or failure of the correction episode. The trained network outputs action values of Pauli operations on the system given an input syndrome. Initial work on uncorrelated noise has shown that the DRL agent performs on par with the standard minimum weight perfect matching (MWPM) algorithm¹. For depolarizing noise the algorithm outperforms MWPM for all error probabilities and with a higher error threshold. The progress on extending this framework to deal with arbitrarily biased noise and syndrome measurement errors, as well as the scalability of the approach will be discussed.

1. P. Andreasson, J. Johansson, S. Liljestrand, and M. Granath, Quantum 3, 183 (2019).