Optimization of Tensor Network Codes with Reinforcement Learning

Tensor network codes enable structured construction and manipulation of stabilizer codes out of small seed codes. Here, we apply reinforcement learning to tensor network code geometries and demonstrate how optimal stabilizer codes can be found. Using the projective simulation framework, our reinforcement learning agent consistently finds the best possible codes given an environment and set of allowed actions, including for codes with more than one logical qubit. The agent also consistently outperforms a random search, for example finding an optimal code with a $10\%$ frequency after 1000 trials, vs a theoretical $0.16\%$ from random search, an improvement by a factor of 65.