13 December 2021

Komodo Dragon

Not too long ago, the top-three elite chess engines were Stockfish, Komodo, and Houdini. Then Leela, incorporating AlphaZero's AI technology, joined the select group. Komodo fell behind and Houdini retired. Stockfish added an important AI component and Komodo added a similar component, becoming Dragon. Now the top three elite engines are Stockfish, Leela, and Dragon.

After I learned that Talkchess Is Talking Again (November 2021), I set out to discover what I had missed during the months after the forum went dark, plus the months before that when I hadn't been following it regularly. I was particularly taken by a series of announcements from GM Larry Kaufman, each of which generated many comments:-

I had noticed in recent posts on this blog that Komodo Dragon had been regularly finishing among the leaders in both the TCEC and the CCC. Following are several of those posts.

2021-10-25: TCEC Cup 9, CCC C960 Blitz Final : Both Underway • 'The [TCEC] 'Cup 9' has reached the semifinal round, with Stockfish, ScorpioNN, KomodoDragon, and LCZero all qualifying for that stage.' • 'In the [CCC] 'Chess960 Blitz Semifinals', Stockfish finished a point ahead of Dragon as both engines qualified for the final match. Only one game of their 40-game minimatch was decisive, with Stockfish winning.'

2021-11-08: TCEC Swiss 2, CCC16 Rapid : Both Underway • 'In the [TCEC 'Cup 9'] semifinal round, Stockfish beat ScorpioNN +1-0=21, and LCZero beat KomodoDragon +1-0=27. In the final round, Stockfish beat LCZero +1-0=3. In the consolation match ('Bronze') for 3rd/4th place, KomodoDragon beat ScorpioNN +2-0=2.' • 'In the [CCC 'Chess960 Blitz Championship'] final match Stockfish beat Dragon +10-1=589.'

2021-12-06: TCEC Swiss 2 in Final Round; Stockfish Wins CCC16 Rapid • 'The [TCEC] 'Swiss 2' event is now in the last of 11 rounds. KomodoDragon is a half point ahead of Stockfish and LCZero, which are a full point ahead of the next engine.' • 'Stockfish won the final match of the 'CCC 16 Rapid', beating Lc0 by a score of +80-32=188. In a 'Runners-up' match, Lc0 beat Dragon by a score of +60-41=199. Dragon had finished third in the 'Rapid Semifinals', well ahead of three other engines.'

Getting back to the Talkchess threads, this is how GM Kaufman introduced 'Komodo Dragon 2.5 released':-

KomodoChess has just released Dragon 2.5 at KomodoChess.com. Normally we would call this 2.1, but it is probably the largest upgrade we have had between versions in many years, with the exception of the NNUE idea first implemented in Dragon 1. It could easily be Dragon 3, but it's only been four months since Dragon 2 and we can't use that name this soon contractually.

It has a larger, "smarter", better trained net, and noticeably more search depth due to search enhancements. Estimated elo gains in standard mode over Dragon 2 range from 50 at long time controls on four threads to 60 at CCRL blitz (2' + 1") on four threads to 72 at CCRL blitz on one thread. In MCTS mode gains are even larger by about ten elo.

All this is based on direct matches of 3400 games or more with a normal opening book. Gains are even larger with unbalanced books, and especially large in Fischerandom (chess960), where on one thread at blitz Dragon 2.5 beat Dragon 2 by a full hundred elo in a 3400 game match! It also defeated Stockfish 14 in a 650 game bullet (half minute plus half second inc) FRC match by 11 elo (132 wins, 112 losses, 406 draws).

The main new feature is a major improvement to the Skill levels. There are now 35, up from 25, and they now consider node count, depth, and eval randomizing in a much more logical, gradual way than in prior versions. The levels should much better simulate the play of a human at that level playing a game with enough time to think thruout the game. The intent is that each level should be a fair match at Rapid (15' + 10") chess with a human of the same FIDE rating as the Skill level set with two zeroes added.

The rest of the post discussed issues related to calculating ratings for the various 'Skill levels'. The feature is related to an earlier post on this blog: Komodo Personalities (April 2021).

It's a positive development for all concerned to have another engine challenging the dominance of Stockfish and Leela. I'm especially interested to see how Dragon performs in the forthcoming TCEC FRC4 (chess960) event. To close this post, I'll repeat the standard footnote to my fortnighly posts on the TCEC/CCC series...

[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; Dragon = KomodoDragon]

No comments: