18 February 2019

TCEC S14 Final, CCC5/-6; Part 2/2

Following the TCEC and CCC computer chess competitions presents two challenges. The first challenge is to keep track of their progress without being glued 24/7 to their GUIs. I've addressed that by preparing a weekly report in the form of a Monday blog post.

The second challenge is what to call each weekly post. Last week's post was TCEC S14 Final, CCC5/-6, where the title summarized the content of the post. For this week's post I could use exactly the same title as last week, but that's obviously not satisfactory. Tacking on 'Part 2/2' should keep things clear.

TCEC: The TCEC S14 final match ('superfinal' aka 'sufi' in TCEC jargon) between Stockfish and Leela ('LC0' in Leela jargon, meaning 'Leela Chess Zero') has seen 72 games completed. Leela leads by one game, +9-8=55. If Leela ('she' in Leela jargon) can maintain this slim lead through the full 100 games, it will signal a major shift in chess engine dominance. The traditional alpha-beta engines with their handcrafted evaluations are quickly giving way to the AI/NN engines which work out their own evaluations from scratch.

Leela also won the preceding TCEC competition with a little help from her friends. For a final report, see TCEC Cup 2 report (chessdom.com).

CCC: Last week's 'CCC5/-6' post started with a preliminary report on the finish of CCC5:-

Stockfish then played Lc0 in a 100 game match, which finished yesterday. The result hasn't been published yet, although the PGN is available.

The result was published in a short note on the 'CCC #tournament-results' page (see the tab above this post for a collection of 'TCEC/CCC Links'):-

Ccc5 finals SF vs LCO dev 60-40

I loaded the corresponding PGN file into PGN software and determined that Stockfish (SF) beat Leela (LC0) by a score of +23-3=74. The games were played at a time control of '10|5' (chess jargon for 10' per game plus 5" per move). As for the next tournament, the 'CCC5/-6' post said,

A new event titled 'Battle of the Leelas' started with Stockfish, Komodo MCTS, and four Leela variants.

The final results are summarized in the following chart.

'CCC #tournament-results'

Stockfish finished fourth in the field of six, losing only one mini-match, against Antifish. I'll leave it to others more knowledgeable to interpret the complete results. I assume the six engines were chosen for a reason.

CCC6 is currently underway. The rules say,

Qualifiers (5|2) - escalation, 3 rounds, 16 engines, top 8 qualify, SF, Lc0, Komodo & Houdini get buys [sic; 'byes'] • Playoffs (10|10) - 12 engines, 3 rounds, top 8 plus SF, Lc0, Komodo & Houdini, book ON • Finals (10|10) - 2 engines, 200 games, book ON

The escalation format, where the lowest ranked engines play mini-matches against each other at the beginning and the stronger engines enter the competition at each successive stage, makes it impossible to predict which engines will eventually prevail. For a preliminary report, see CCC 6: The Winter Classic (preview).

No comments: