Improving the efficiency of algorithms for fundamental computational tasks such as matrix multiplication can have widespread impact, as it affects the overall speed of a large amount of computations. In this talk I will present AlphaTensor, our reinforcement learning agent based on AlphaZero for discovering efficient and provably correct algorithms for the multiplication of matrices. Using this approach, we find new algorithms outperforming the best ranks for many matrix sizes. I will describe how to formulate this problem as a single-player game, and the key machine learning ingredients that enable tackling difficult mathematical problems using reinforcement learning.
This talk is based on this Nature paper.
This video was produced by the University of Warwick as part of the 7th Workshop on Algebraic Complexity Theory (WACT) 2023.
