After beating Atari games, the team at DeepMind set their sights on a more ambitious target: building a neural network that could play the game of Go at a very high level.
Eventually the system they built, AlphaGo, and its follow-up system, AlphaZero, were able to beat the best human players in the world.
AlphaGo works by improving on an earlier strategy for two player games called Monte Carlo Tree Search. Here is a link to a Wikipedia article on MCTS.
AlphaGo documentary
Below you will find a link to the AlphaGo documentary:
This link jumps to the section of the documentary that explains a little bit about how AlphaGo works.
The full documentary is really worth watching from start to finish.
The heart of AlphaGo was a modified Monte Carlo Tree Search algorithm. The best explanation of that algorithm appears in the paper in Nature describing AlphaGo.
The successor to Alpha Go was Alpha Zero. Alpha Go was initially trained on many human games. Alpha Zero, on the other hand, learned the game purely from playing games against itself.
Here is a link to a video of a presentation by the head of the AlphaZero team, David Silver. In the video he discusses how AlphaZero differs from AlphaGo:
Here is a link to a blog post that provides some more details on how AlphaZero works:
After beating the game of Go, the DeepMind team next set out solve the protein folding problem. This lead to the creation of a system called Alpha Fold.
Here is a link to a video giving some detail on how the system works.
The DeepMind team's work lead to the 2024 Nobel Prize in Chemistry.
One of the most recent systems that DeepMind has built is AlphaProof, which is an automated RL system for constructing mathematical proofs.
Here is their paper describing the system.