Wednesday, April 18, 2018

A final look at season 11

Season 11 is over, Stockfish is the winner. Season 12 is about to start, there is almost no pause between seasons now, and I want to have a last look at season 11 as a whole.

Season 11 changed the competition format and introduced a division system. Engines were split to 4 divisions based on season 10 results, and a fifth division was used for candidate engines that did not participate in season 10. In general I think the new format was a success. The engines' results depended on their performance against other engines of similar strength. In the previous format in early stages there were many games where the result was almost known in advance due to the huge difference in engine rating.

In the lower divisions one of the interesting engines was Laser. In season 10 it finished last and it started season 11 on division 3. Between seasons the engine was upgraded and it was known that it was much stronger, it was not a big surprise that it promoted to division 2. In division 2 Laser still played well and managed to promote again to division 1. The division 1 engines were stronger than Laser, yet it finished the division above the relegation threshold and will start season 12 in division 1. I would like to see other engines improve between seasons, this is the sort of thing that makes TCEC interesting.
In the premier division it was expected that the top-3 engines - Stockfish, Houdini and Komodo - would fight for a place in the superfinal. In season 10 the race was very close, and Stockfish lost by half a point. This season Komodo was able to keep up until the halfway point of the premier division, but then lost sight of the leaders. It finished third despite beating Stockfish once, the only Stockfish defeat in the division. Stockfish finished first, with much better results against lower ranked engines than Houdini.

Based on the 12 games Houdini and Stockfish played in the premier division I assumed the superfinal would be close with many draws. Instead, the superfinal was extremely one-sided, Houdini struggled in many openings and lost 20-2. Stockfish was upgraded between divisions on top of the upgrade between seasons, and Houdini was the same as season 10, and it showed.

Leela to play in season 12

Towards the end of season 10 Google published its Alpha Zero chess project, where it used specialized hardware and its infinite resources to train a neural net based engine using only the rules of chess and self-play. The claim was that this engine could beat Stockfish.There was a lot of controversy surrounding this claim, and Google hasn't responded to any of the arguments, probably it only used chess as a proof of concept and it has bigger plans for the technology. The chess community, on the other hand, did not ignore the claim and a new chess engine was born - Leela Chess Zero. This project tries to imitate Google's ideas using standard hardware, where everyone can participate in the ongoing training procedure. At the end of season 11 Leela got to a level where it could compete (and lose) against the weakest engines in TCEC. It played exhibition matches against Scorpio and against a weak Stockfish, losing both but not without winning a few games. Its rating was estimated to be around 2750. Just before season 12 started Anton announced that Leela will play in division 4 after an engine dropped out. There have been last minute bugfixes and improvements, this is going to be exciting to watch.

site stats
This is a graph of the all time monthly hits for this blog: 

During the season the statistics reports were promoted on the chat and used in chessdom. The last 6 months have seen at least 1000 hits each month, peaking towards the higher divisions. The total number of hits is over 22,000.

Thanks to all the readers, I hope to continue being informative as well as interesting in the next season.


No comments:

Post a Comment