Project Maia
A human-like neural network chess engine
Why Chess?
Chess is an ideal game to study when it comes to artificial intelligence because it is popular, has well-defined rules, and has not yet been fully solved.
The game emerged in the 15th century and is played between two players controlling the black and white pieces, respectively. Many people know the game and AI researchers use it as a «model system» to study new ideas or techniques.
What is Project Maia?
Maia is a deep learning framework that learns from online human games, with the goal of understanding how humans play. The larger vision of Maia is to use chess to investigate the relationship between humans and AI. Previous AI systems for chess focus on finding the optimal sequence of moves. But it’s more complex to use AI to understand what move a human should make. For example, it’s not always clear that every person will understand a specific move: suggesting an advanced move to a novice player may be dangerous, because the player may not understand the board position that results from that move.
With this vision in mind, the Maia Project aims to develop an AI engine that holistically understands human play.
How does Maia work?
We developed Maia by taking a deep reinforcement learning neural network, previously used to predict the optimal move for a given board position and retraining it to instead predict what a human player would do.
Chess players have their own playing styles, so predicting their moves is hard: most positions that people reach are unique—due to the astronomical number of possible chess positions—and even the same player may not make the same move on a position they’ve previously seen!
Each Maia is trained on games played by players at a particular skill level and can accurately predict moves made by players at that skill level. In fact, the Maias do noticeably worse at predicting moves that are higher or lower than their target skill level, which means they have captured the “playing style” of players at that level.
Beyond Chess
As artificial intelligence becomes increasingly intelligent—in some cases, achieving superhuman performance—there is growing potential for humans to learn from and collaborate with algorithms. However, the ways in which AI systems approach problems are often different from the ways people do.
A crucial step in bridging this gap between human and artificial intelligence is modeling the granular actions that constitute human behavior, rather than simply matching aggregate human performance. Maia predicts human moves at a granular and individual level, forming the foundation for a teaching tool that can help each individual understand their weaknesses and improve their chess play. The larger vision of Maia is to create a more productive relationship between humans and AI in chess, with the hope of applying these learnings to other domains.