24 June 2019

TCEC Not-so-Rapid Bonus; CCC9 Stage Two

So where was I in reviewing the two ongoing engine-to-engine competitions? To summarize last week's post, More TCEC Bonus Events; CCC9 Starts:-

TCEC: The 'Champions Bonus' ended as expected. The Stockfish version that won S14 finished first, the version that won S13 finished second, and so on. The TCEC is currently running a 'Rapid Bonus' with 14 engines. • CCC: The site is currently running 'CCC9: The Gauntlet Qualification', with 12 engines, including Leela and Stockfish.

Both competitions have advanced enough that I couldn't skip a post this week.

TCEC: The 'Rapid Bonus' is still running with Stockfish and Leela currently tied for first. The event should finish later this week. The TCEC published a blow-by-blow report of season 15 with TCEC15: the 15th Top Chess Engine Championship (chessdom.com). Although the TCEC is the front runner in organizing engine-to-engine competitions, the report struggles with some basic concepts. It doesn't specify which engines run on the two configurations -- CPU & GPU -- and it speaks of ‘Shannon AB’, ‘AB conventional’, 'neural-network', and 'non-Shannon' engines. Insiders understand, outsiders don't.

CCC: The 'Gauntlet Qualification' finished with the results shown in the following chart, where Leela finished ahead of Stockfish by a significant margin. Stockfish lost one game because of a bug handling a tablebase, but it had no impact on the final standing.

The red and green bars to the left of the chart show which engines were seeded into the next stage, the 'Gauntlet Quarterfinals'. According to CCC planning - CCC9 The Gauntlet (docs.google.com), Stockfish and Leela play at every stage. In the quarterfinal they are joined by four qualifying engines from the previous stage (the green bar in the chart) and 'strong finishers from CCC8'. Along with Stockfish and Leela, two engines will qualify into the semifinal, where they will be joined by Komodo and Houdini. Got it? The engine Dark Queen, the last to qualify from the first stage, is worth a note:-

!dq: Dark Queen is a neural network that focuses on using q-learning for its value head, still in its early development stages. It is currently trained completely on lichess games. It uses LC0 binary.

What's q-learning and how does it relate to reinforcement / deep learning? Wikipedia's Q-learning page is a typical Wikipedia science article, where you have to understand the subject to understand the article, so: Pass! (for now).

[For further information from the various stakeholders in the engine-to-engine events, see the tab 'TCEC/CCC Links' at the top of this page. • NB: Leela = LC0 = LCzero]

No comments: