TCEC10: The 10 th Top Chess Engine Championship

Of 276 games, 165 (59.78%) were won (102 by White and 63 by Black) and 111 (40.22%) were drawn. Of the 165 games, only 12 (7.27%) were won by the engine finishing lower in the standings, the extreme being game g17.6, Arasan – Hannibal, where, unusually, the loser seemed to see the loss before the winner saw the win. Chiron exceeded ELO-based expectations while Jonny and Rybka did not. Gull was edged out of the last qualification spot by just two Sonneborn–Berger points.

Texel-Fizbo (g2.5, 1–0) was one of the finely balanced but decisive games: a queen and pawn endgame actually concluded on move 139 but could have run to move 178. Despite rules to terminate games without a pulse, Ginkgo – Gaviota (g3.4) was drawn on move 157 as was clear by move 104.

2. Stage 2: 4 phases, 28 rounds, 112 games

OPEN IN VIEWER

Fig. 4.

The cross-table for TCEC10 Stage 2, a round robin of 28 rounds.

Of the 112 games, 37 (31.25%) were won (31 by White and 6 by Black) and 75 (68.75%) were drawn – less wins and relatively fewer for Black but the ELO range had narrowed from 570 to 119 for this stage. Form was consistent even more than in Stage 1. Of the 37 decisive games, only two (5.41%) were won by the engine finishing lower in the standings: Stockfish – Komodo (g19.1), and Komodo – Houdini (g14.1) where author Robert Houdart noted that his engine was ‘completely blind for some time’. These three engines were now almost equally powerful and so the three decisive games between them, also including Houdini – Komodo (g21.1, 1–0), stand out. Fire was a clear fourth, and Chiron, Ginkgo, Andscacs and Booot were tightly grouped in the lower half of the table.

The outcome of Stage 2 was uncertain to the end but behind Houdini, Komodo edged out Stockfish by just half a point. Stockfish had the higher Sonneborn–Berger score, being ‘ + 1 ’ against Houdini and Komodo, but had failed to land blows against weaker opposition, in particular Chiron. The engines’ contempt factors could only be set at the beginning of stages and perhaps Stockfish gave its weaker opposition too much credit for playing like itself. Many in the audience would have liked to see it in a three-way final: this would indeed have been fascinating, complete with an Olympian podium and flags, but the TCEC10 rules dictated a long match as the finale and so we said farewell, for the moment, to the formidable open-source Stockfish, the Season 6 and current TCEC champion.

OPEN IN VIEWER

Fig. 5.

The TCEC10 Superfinal of 100 games: the four Black wins are underlined.

In Stage 2, Houdini and Komodo had been locked together on 18.5 points, Houdini 4.75 better in Sonneborn–Berger terms but Komodo 48 better in ELO terms. All seemed set fair for a protracted, even struggle in the final but the opening games indicated otherwise. Houdini was 5–0 up before Komodo scored a win with game 39. In reply, Houdini won games 40 and 42 before Komodo scored again in game 49: 7–2 at half time. By this time, it was clear that a suboptimal compilation of the Komodo code had left it rather down on speed but the rules did not allow recompilation.

Game 52 saw another ‘first’. The TCEC9 Superfinal had seen Black unable to defeat White, but with the TCEC10 Superfinal half over, Black did win and, rather against the run of play, Komodo was the winner. Did this signal a Komodo come-back?! Not in fact. Although Komodo was much more competitive in the second half of the match, it still lost the last 50 games 8–7 to Houdini leaving the final score at 15–9. This was certainly not the rout initially threatened and all games were keenly contested: average game length was over 63 moves despite TCEC’s foreshortening adjudication rules deciding draws and wins. Perhaps more than expected, 24% of the games were decisive and Black again claimed one sixth of the wins as in Stage 2. White does seem to have an intrinsic advantage so should this be compensated for by giving White less thinking time than Black?

The most notable games are not exclusively decisive ones but significant wins include the four by Black (g26.2, g39.1, g40.2 and g49.1), Komodo’s first win (g20.1), and Houdini’s longest win emerging from an apparent deadlock (g6.2, 134 moves). Game 38.1 was the longest draw at 213 moves.

Congratulation to Robert Houdart, author of TCEC10 champion Houdini: his hard work across two years has reaped its reward. This was a remarkable return to the top level after four years by the champion of TCECs 1, 2 and 4. Robert suggested that a Houdini advantage was its greater appreciation of mobility, leading to ‘contempt-like play without the contempt factor’: something similar was seen in the concurrent AlphaZero – Stockfish games DeepMind (2017).

OPEN IN VIEWER

Fig. 6.

Some leading authors: Robert Houdart (Houdini); the late Don Dailey, Larry Kaufman and Mark Lefler (Komodo); (second row) Tord Romstad, Marco Costalba’s Github surrogate, Joona Kiiski and Gary Linscott (Stockfish).

OPEN IN VIEWER

Fig. 7.

The cross-table for TCEC10 Rapid Championship, a round robin of 46 rounds.

OPEN IN VIEWER

Fig. 8.

The cross-table for TCEC10 Blitz Championship, a round robin of 46 rounds.

The TCEC10 Rapid tournament ran 8^th–25^th December 2017 with a tempo of 15 ′ + 10 ″ /m. The Blitz event followed, 26^th–29^th December, with a tempo of 3 ′ + 2 ″ /m. Here, responding to audience requests for ‘decisive’ games to be played out further, the TCEC10 win rule was discarded and we saw 381 mates on the board, a 69.02% win-rate – significantly more than Stage 1’s 59.78%. Both were double round robins of 46 rounds.

While Houdini, Komodo and Stockfish continued to occupy the podium, the titles were neatly shared between them. Stockfish won the Rapid tournament and Komodo (having reverted to its ‘1959.00’ version) won the Blitz. Fire and Chiron dutifully fell in behind, fourth and fifth in all three tournaments. Stockfish as Black mated Houdini on move 110 of the very last blitz game, only failing to take first place by 20 Sonneborn–Berger points, a fitting and dramatic finale to the TCEC10 proceedings.

First, our congratulations to the TCEC10 organisers, to the champions, podium placers and all the competitors wherever they finished in the standings. The event was absorbing throughout and various initiatives, including the choice of server platform, the strategy on openings (Hernandez, 2018) and the use of the Twitch video/chat service, added successfully to proceedings. The work needed to make the event of 1592 games run smoothly is by definition hidden if successful – as it was, so ‘kudos’ to the TCEC10 team and their suppliers. The event attracted web commentaries from ‘GM Thechesspuzzler’ (2017) and others: they usefully shed light on many of the most interesting games.

TCEC11 was trailed, discussed and announced during TCEC10 and brought more initiatives. Some 40 engines, ranked on the basis of TCEC10 and known ELOs, are to play in five divisions with promotions and demotions after each stage. Every engine’s author can enjoy reasonable goals and closely contested games. Given their success in defining a fair allocation of interesting openings to the TCEC10 engines, Nelson Hernandez and Jeroen Noomen continue in these roles. Rules and logistics also remain unchanged though the lower divisions will have less generous time budgets.