Researchers have brought the famous AlphaZero algorithm to the quantum realm.
AlphaZero is a computer algorithm that was developed 2 years ago by DeepMind company. This extraordinary algorithm defeated chess computer programs such as Stoke Fish with only 5000 games via self-play.
Although many algorithms have been developed to optimize quantum dynamics such as quantum variational eigensolvers, annealers, simulators, circuit optimization, optimal control theory, and Boltzmann machines, they all have the limitation of reliance on good initial guesses. Initial guesses can be random or based on heuristics and intuitions. Researchers at the University of Aarhus have now implemented a quantum version of the AlphaZero algorithm to eliminate this limitation. AlphaZero can learn things on its own without any form of human expertise.
AlphaZero uses a deep neural network with deep predication guided tree search. This method can be used to estimate the predictive hidden-variable approximation in the quantum parameter landscape, as the researchers applied the algorithm to three sets of control problems. Using AlphaZero greatly improved both the quality and quantity of results. In fact, AlphaZero can find hidden symmetry and structure in solutions.
Although AlphaZero works very well on its own, combining it with a quantum algorithm improves the results. The researchers provided free access to the code to speed up the development of the field. Their work has just been published in Nature Quantum Information.