This weekend, GM Andrew Tang (@penguingim1) will play a match against an unusual opponent, namely a new self-taught AI-based chess engine called Leela Chess Zero. Match details can be found at the bottom of the page.

What is Leela Chess Zero?

Leela Chess Zero (also known as LCZero or Leela) is open source and based on Leela Zero, which is an open source AI-based engine for Go. Leela is inspired by DeepMind's AlphaZero, which beat Stockfish 8 in a 100-game match (28 wins, 72 draws, 0 losses). The Leela Chess Zero team is trying to replicate and expand on the success of AlphaZero.

Except for the fundamental rules of chess, everything Leela knows, she has learned from playing games against herself. Leela doesn't use an opening book or an endgame tablebase. What happens is that Leela's neural network plays thousands of games with itself, a new neural network is trained based on those games, which then is tested against the previous neural network. If you watched NM Chess-Network play Leela "ID 82" and then later "ID 102" this month, this ID is basically the version number of the neural network. New versions are often stronger, but not always.

Since the end of February this year, Leela, has played more than 4.6 million games against herself thanks to volunteers donating their CPU and GPU time.

Rough Elo estimate over time, showing Leela's progress over more than 4.6 million games

Leela's strength is currently estimated to be equivalent to 2500 FIDE, but it ultimately depends on the time control, available hardware resources and the position in question.

Anyone can play against Leela at play.lczero.org. However, the website version is significantly resource limited to avoid overloading the server. On Hard mode, Leela considers "only" about 2,000 positions before deciding which move to make. In a very recent eleven-game practice match, with a 20+5 time control, against IM Lasse Østebø Løvik (@lovlas), Leela Chess Zero ID 125 ran at full strength on a GeForce GTX 980 Ti graphics card, considering anywhere from 25,000 to 100,000 positions before making a move. This is still orders of magnitude fewer positions than Stockfish would analyze, but Leela selects which moves to analyze more carefully, using her neural network. Leela won the match with 10 wins, 2 draws and 0 losses.

Earlier versions of Leela were said to be tactically weak, sometimes missing simple tactics and losing to short forced mates. However, this weakness seems to have been "trained away". The most likely weakness one can hope to exploit now is probably its endgame technique. In one game against @lovlas, Leela failed to win the Lucena position, an easy, well-known endgame to humans.

Match Details

Who: GM Andrew Tang (@penguingim1) vs Leela Chess Zero (@LeelaChessOfficial)

When: Sunday April 22nd, 8am PDT, 11am EDT, 3pm UTC

Duration: ~2 hours

Time controls:

• Phase 1: 2 games at 15+2
• Phase 2: 4 games at 5+2
• Phase 3: 8 games at 1+0
• Phase 4 (as time allows): Andrew can challenge Leela at any time control, faster or slower.

@LeelaChessOfficial will be overseen by @careless25 who used the new Lichess bot API to develop a program that will automatically relay Leela's moves to Lichess.

Leela will start the match with neural network ID 125 which was published on April 13th. That's the same neural network that was used in the 20+5 practice match against @lovlas which was played a couple of days ago, and in the TCEC bonus match against Scorpio 2.79 which was played yesterday. Leela scored 4 wins, 3 draws and 13 losses against Scorpio.

If Andrew manages to win a game against Leela, and there is time for additional games, her neural network will quickly be upgraded to an even newer and slightly stronger one, should one be available at that time.

Leela is currently not able to resign or offer/accept draws.

Follow Andrew's livestream on Twitch for his live commentary during the match. If you only want to follow the games on Lichess, use this link.

This should be an interesting match. Let's hear what you think about Leela and Andrew's chances in the comment thread below.