Leela Chess Zero, akin to its proprietary predecessor AlphaZero, is an engine with a fundamentally different approach to search than that of the contemporary modern engines that are ruling the chess stratosphere. As a chess player, trainer and free software programmer I have been following the reddit thread and the lczero mailing list with great interest for a while. She is pretty new in the scene really but Leela is approaching the ELO rating of stockfish and friends slowly and firmly. As such it started make an impact on the general trend of chess programming already. Komodo-12.1.1 released earlier this week (2018.06.04) comes with a new feature to change its search algorithm: Use MCTS option may be enabled to use Monte Carlo Tree Search rather than Alpha–beta pruning which has been the state-of-art search algorithm in chess engine development and has been pushed to its limits over the last decade.

My main point of interest at this point surrounds around two questions:

  1. How different is LeelaChessZero for human vs. engine matches compared to her alpha-beta counterparts?
  2. How beneficial is LeelaChessZero analytically and didactically compared to e.g. Stockfish?

The second question may be too soon to answer yet but taking a sneak peek at the first question is fortunately possible. Simply browse through play.lczero.org and stand for a challenge. You may also use Leela as an UCI engine with your preferred chess interface. The preferred way for chess professional should be to use the engine via the UCI interface, as this public server presumably would have less horsepower than your personal computer, however the web interface is a quick and practical way to give this new member of the party a try. Driven by the curiousity of the aforementioned riddle about human vs. engine games in their arguably future shape and feeling motivated to contribute my gray matter as I do not own a GPU or much idle CPU, I have played a couple of games against it. I have analyzed the final game which was a memorable success for me. I managed to liquidate the game into an endgame which I think was slightly favourable for me. There were a few moments in the endgame where I think I could play better and gain substantial winning chances. Admittedly, having been using engines for chess analysis before and during tournaments for preparation for more than a decade now I had some educated guesses up my sleeve. In the analysis below I explained how that knowledge enabled me to make some concrete decisions to let Leela enter an endgame where I think she stands slightly worse. Leela, however, defended strongly in the endgame and the game resulted in a draw.

For those who want to delve deeper into the topic of computer chess analysis, I recommend the book Modern Chess Analysis by Robin Smith who is an International Correspondence Chess Grand Master since 2004 and as such a prime authority on the matter. To list further classics on the matter, John Nunn has written three books based analysis with endgame tablebases which are enlightening and takes a mathematician and international grandmaster's rational and artistic peek on the topic of Playing Chess with God, that is Retrograde Analysis of Certain Endgames, pioneered by Ken Thompson.

So here are my 4 plies on Leela. In the iframe below you may find the lichess study with my analysis of the game. I have not included many variations at this point but I have elaborated on the critical moments which I spotted as the turning points of the game preferring natural language over concrete variations. Click here if the iframe does not work for you. Note, the lichess study has chat open for everyone so feel free to share your comments and variations. Cloning is also open for everyone so you can base your analysis on this one. Please drop a comment with your study link if you clone so people reading the main study will be aware of your version as well.

Admittedly picking this game out of 4 games, 3 of which I lost with the Black pieces. Unfortunately I did not record them and was sad to see I could not get an export of them from the server. I do hope to follow up with more games with either colour using Leela as an UCI engine using SCIDvsPC though so stay tuned.

tl;dr I like Leela. She is a fairly strong opponent who does not make too much of those "computerish" moves, a concept that the chess players have been using in one form or the other about distinguishing some extremely nonintuitive engine moves from general human style of play.

Now, is Leela more human?

Add a comment