One of the most powerful machine learning algorithms for autonomous game play learning is the AlphaZero algorithm developed by Google DeepMind. AlphaZero has outperformed both human expert players and highly specialized gaming software in GO, Shogi, and Chess. Now, a research team at Aarhus University led by Prof.MSO Jacob Sherson has managed to utilize AlphaZero for controlling quantum systems. Laser pulses based on analytical ansatzes typically steer quantum experiments. Developing analytical solutions is costly in human labor and often not feasible for a given problem. However, AlphaZero is able to learn the optimal control of quantum systems without the need for human expert knowledge. This is due to its combination of a deep neural network, which allows it to learn correlations between cause and effect similar to the human mind, and a Monte Carlo tree search that enables it to look-ahead and analyze the impact of future moves. Without any hyperparameter changes between the tasks, the quantum version of AlphaZero successfully solved three different control tasks within quantum computation on a circuit QED system. The research team also managed to combine AlphaZero with a quantum-specialized optimization algorithm, GRAPE. This led to an increase in both the quality and quantity of high-performing solutions. With this new hybrid algorithm at hand, physicists have a new powerful tool for optimizing the control of quantum systems. 

The scientific paper can be found at the Nature – Quantum Information, web page.