What is Darkforest?

What is Darkforest?

Darkforest is a computer Go program developed by Facebook's AI Research team, based on deep learning techniques using a convolutional neural network. It combines these techniques with Monte Carlo tree search (MCTS), a method commonly seen in computer chess programs.

Darkforest improved move prediction for Go compared to earlier MCTS only systems by using a policy network to guide search. Public evaluations reported strong amateur to professional level play depending on the version and time period, though reported ranks vary across servers and evaluation setups.

The Darkforest Go engine is open source, and its source code and pre-trained models have been publicly available on GitHub. This has been beneficial to the community, as it allows other researchers and developers to use and build upon the work done by the Facebook AI Research team.

Note: the phrase "dark forest" is also used, unrelated to this Go engine, as a metaphor in discussions about AI generated content and, separately, as the name of a theory in Liu Cixin's science fiction about civilizations staying hidden to avoid detection. These are coincidentally-named concepts with no connection to the Darkforest Go program.

How does darkforest use deep learning techniques?

Darkforest's convolutional neural network (CNN) is designed to predict the next k moves based on the current state of the game and each player's territory. The architecture is a 12-layer full convolutional network with a width of 384 nodes, with each convolutional layer followed by a rectified linear unit (ReLU) activation.

The network is trained on game positions paired with corresponding best moves, allowing it to learn to predict strong moves given a game position. During play, this policy network guides a Monte Carlo Tree Search (MCTS), letting Darkforest evaluate the game state and make strategic decisions more efficiently than MCTS alone.

Some Darkforest variants experimented with architectural changes and different search integrations to improve efficiency and strength. Reported innovations include deeper networks and tighter coupling between policy evaluation and tree search.