Written by Guy Haworth and Nelson Hernandez
Reading, UK and Maryland, USA
This is the second in a new series of analytical articles on past TCEC events. The main text can be read below on this webpage, and at the bottom you will find a link to the full layouted article in pdf format, including the important tables, graphs and images.
TCEC is very grateful to the authors for their kind permission to publish these substantial and scholarly analyses of its events!
Introduction
After the successes of TCEC Season 11 (Haworth and Hernandez, 2018a), the Top Chess Engine Championship moved straight on to Season 12, starting April 18th 2018 with the same divisional structure if somewhat evolved.
Five divisions, each of eight engines, played two or more ‘DRR’ double round robin phases each, with promotions and relegations following. Classic tempi gradually lengthened and the Premier division’s top two engines played a 100-game match to determine the Grand Champion. The strategy for the selection of mandated openings was finessed from division to division. The revised TCEC engine line-up is illustrated and listed in Fig. 1 and Table 1.
Besides using FIDE’s 3x-repetition and 50-move drawing rules, TCEC terminated a game at move 40 or later if both engines had |eval| < 0.05 for ten consecutive plies in the current phase, i.e., since the last pawn-advance and/or capture. TCEC12 fell in line with most of the participating engines by adjudicating endgames using the Syzygy DTZ50″ EGTs rather than the Gaviota DTM EGTs which do not recognise the 50-move rule. 5-man EGTs were used for the divisions and 6-man EGTs were used for the Superfinal. Games which were apparently decisive were terminated by TCEC if both engines consistently agreed for the last eight plies that the evaluation is at least 6.5 or at most -6.5.
ELO ranged from 2714 to 3554, averaging 3143. Four new engines joined the fray this time:
- LEELA CHESS ZERO, a new-architecture UCT/NN engine from a large community,
- RODENT by the Polish chess programmer Pawel Koziol,
- TUCANO by the Brazilian professional software developer/programmer Alcides Schulz, and
- XIPHOS by the Serbian mathematician and computer scientist Milos Tatarevic.
The formidable 44-core server of TCEC11 was used unchanged in TCEC12.
Division 4, two DRR phases, 28 rounds, 112 games, tempo 30’+10″/m
A principal focus was the participation of the exciting LEELA CHESS ZERO, a neural-network-architecture engine inspired by the innovations of Deep Mind’s ALPHAZERO (Silver et al, 2017). The 14 most common two-move openings in the second author’s CATOBASE (Hernandez, 2018) were allocated to rounds 1-7 and 15-21, and with colours reversed for rounds 8-14 and 22-28, see Table 2.
The results, as played, are as in Table 3 though a rule introduced in TCEC11 required that the participation of SCORPIO be scratched as it had three technical crashes. The seven connection breaks with the server were thought to be caused by deadlock conditions in the engine. These losses in fact made no difference to the final ranking on this occasion. LC0 did not in fact progress but will be greatly helped by GPU assistance in a future season. The generic statistical review of TCEC12 results and terminations is given for each phase of TCEC12 in Table 11.
For division 4, all rounds have four games so game r.n is game 4r-4+n in the pgn files (Haworth and Hernandez, 2018b) and the colour-flipped pairings of engines are 28 games apart. This division had 10.8% of its wins below the diagonal of the eventual x-table. ETHEREAL alone was much improved, undefeated and a strong first while XIPHOS kept RODENT III in a distant third place, beating it 4-0.
Division 3, two DRR phases, 28 rounds, 112 games, tempo 30’+10″/m
In this division, the same fourteen openings were mandated as for Division 4 and used in the same order. This time however, the colour-reversed game followed immediately rather than 28 games later, giving an earlier balanced view on the bilateral contests. Rounds were therefore of eight games rather than four and were numbered from 1 to 14.
Draws under the 50-move rule are very rare in TCEC, less than 1% of terminations. This is because most engines seem to monitor the ply-count, zero their evaluations as they see it reaching 100, and allow TCEC draw-adjudication to take its course. Game 13.1/97, CHESSBRAINVB-WASP, was however a 50m-rule draw: CHESSBRAINVB retained hopes of a win to the end, despite being a pawn down.
CHESSBRAINVB was a clear winner but the race for the second promotion-spot was close. XIPHOS pulled off its second promotion this season, despite having lost its head-to-head matches with fellow candidates ETHEREAL and PEDONE. ETHEREAL more than justified its promotion to Division 3.
Division 2, two DRR phases, 28 rounds, 112 games, tempo 30’+10″/m
Fourteen of the most frequent 100 two-move openings in CATOBASE were allocated to rounds 1-14 with, again, the colour-reversed games being played immediately.
This time, FRITZ and TEXEL won promotion but the newly promoted XIPHOS and CHESSBRAINVB took the next two places. There were 10 technical crashes in the division, and both ARASAN and HANNIBAL were disqualified and relegated for disconnecting from the server: a pity as they had both scored wins against FRITZ. VAJOLET no doubt counted itself lucky to survive.
Division 1, two DRR phases, 28 rounds, 112 games, tempo 60’+10″/m
The normalised Sonneborn-Berger scores suggest that this was the most closely-contested division of TCEC12. Eventually, GINKGO and JONNY triumphed though FIZBO and BOOOT kept the result in doubt until almost the end. Thankfully, we did not see another rash of engine-disconnect fails, the one ultimately irrelevant incident being g8.8/64, FRITZ–JONNY. FRITZ as Black had beaten JONNY in their first, g1.7/7, encounter and hung on to its recent promotion: TEXEL did not.
Division P, four DRR phases, 56 rounds, 224 games, tempo 90’+10″/m
STOCKFISH was the only unbeaten engine but was still second with a Performance ELO of 3443, 121 down on its nominal 3554. KOMODO lost its eight-game match against STOCKFISH but had a better harvest of wins against the bottom three engines. HOUDINI made up the three engines that stretched away from the others. CHIRON crashed three times in the first round and was pulled from the event.
The TCEC12 Superfinal match: 100 games, tempo 120’+15″/m
The Superfinal between STOCKFISH and KOMODO surprised in two ways. First, it was not close and secondly, the win-rate was high. STOCKFISH won 29-9, 23/6 as White and 6/3 as Black so White also dominated Black 29-9. In terms of the fifty two-game pairs, STOCKFISH won 22 pairs 1½-½ and the g71-72 pair 2-0: KOMODO won 4 pairs 1½-½. GM Thechesspuzzler (2018) created a Youtube stream for the TCEC12 Superfinal. Wool (2018) provided some useful chessic insight on TCEC12 as a whole.
Summary
We gather together some generic statistics for the Divisions and Superfinal in Tables 11 and 12. These will help aficionados and analysts of computer chess to identify the particular games of interest to them later. The pgn files and further data are included with the e-repository version of this note (Haworth and Hernandez, 2018). Our congratulations go once again to the TCEC audience who made for a lively discussion forum, to all participants, particularly to those who gained promotion and to the TCEC12 Grand Champion, STOCKFISH and all its supporters.
REFERENCES
- CPW (2018). https://tinyurl.com/icga046. Biographies of programs and authors.
- de Man, R. (2018). http://tablebase.sesse.net/syzygy/. Site providing 5- and 6-man DTZ50 EGTs.
- ‘GM Thechesspuzzler’ (2018). https://tinyurl.com/icgaj036. TCEC video playlists.
- Haworth, G. McC. and Hernandez, N. (2018a). TCEC11: the 11th Top Chess Engine Championship. ICGA Journal 40(3). See also http://centaur.reading.ac.uk/75899/ for supporting data – results, statistics and pgn files.
- Haworth, G. McC. and Hernandez, N. (2018b). http://centaur.reading.ac.uk/76985/. TCEC12: the 12th Top Chess Engine Championship. This article plus supporting data – results, statistics and pgn files.
- Silver, D. et al (2017) https://arxiv.org/abs/1712.01815 Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm.
- TCEC (2018) http://tcec.chessdom.com. Current and past TCEC tournaments.
- Twitch (2018). https://www.twitch.tv/. A video/chat platform and community for gamers.
- Wool, A. (2018) http://mytcecexperience.blogspot.co.uk/ TCEC blog.
Full article
To read the full article in pdf, click HERE
