Last week Deepmind announced AlphaTensor, a mathematics package that follows AlphaFold and continues the tradition of using AI to expand the horizons of science.
In this case, the problem is matrix multiplication, known to most people from their high school days.
The issue is not just the actual multiplication but the fastest method to perform the multiplication. The speeding up of matrix multiplication calculations has a high impact because matrix multiplication is a part of many applications – especially in deep learning and image processing.
AlphaTensor models matrix multiplication problems as games and trains the AlphaTensor agent using reinforcement learning.
We then trained an AlphaTensor agent using reinforcement learning to play the game, starting without any knowledge about existing matrix multiplication algorithms. Through learning, AlphaTensor gradually improves over time, re-discovering historical fast matrix multiplication algorithms such as Strassen’s, eventually surpassing the realm of human intuition and discovering algorithms faster than previously known.Source: https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor
Reinforcement learning, when used to model real-world problems, always had limitations. Hence, as a technique, it excelled in situations like games (AlphaGo) or virtual simulations(used for autonomous training vehicles). But reinforcement learning to address scientific problems is a new way to do science and a game changer. For more details on this approach and philosophy see this (older) youtube video from the AlphaTensor lead researcher pushmeet kohli.