Self-learning AI beats the world’s most famous chess engine

Chess engine Bc0 has won the Computer Chess Championship. It is the first time that a chess engine that has taught itself to play chess through machine learning has won the tournament.

Bc0 defeated the famous chess engine Stockfish in the final of CCC: Blitz Bonanza. It is not only the first time that a machine learning chess engine has won the tournament, but also the first time that Stockfish has not won. That engine took first place in the previous seven editions of the Computer Chess Championship.

In third and fourth place at CCC: Blitz Bonanza were the engines Leelenstein and Antifish. These are variants of Bc0. After the tournament, the championship organizers changed the rules so that no more than two finalists with a similar codebase are allowed.

Chess.com’s Computer Chess Championship is an annual tournament where chess engines compete against each other. There are more such tournaments, such as Chessdom’s TCEC. Bc0 stands for Leela Chess Zero. It is an open source project based on Leela Zero Go, which in turn builds on AlphaGo Zero from Google’s DeepMind.

The engine teaches chess based on a set of ground rules and trains based on a network for distributed computing. The intention is for the engine to improve through reinforcement learning, whereby decisions that work out well are rewarded.

Users can help train the engine or play against the engine themselves using the Lc0 client. The developers advise to use a powerful GPU.