Komodo Dragon 2.6 released

[lkaufman]

Komodo Dragon 2.6 Released

KomodoChess has released Dragon 2.6 at komodochess.com. It is a significant upgrade from Dragon 2.5.1 (or 2.5). Estimated elo gains from 2.5.1 on one thread in blitz (2' + 1") using a normal opening book in standard mode are 18 elo (based on 20,000 direct games), in Monte Carlo mode 24 elo, and in standard mode playing Fischerandom (chess960) 26 elo (based on about 5500 direct games in those cases). Elo gains for similar tests against Dragon 2 showed +83 Elo for a normal opening book in standard mode, +113 elo in MCTS mode, and a remarkable +126 elo in FRC (standard mode). Compared to the final Komodo version before NNUE (Dragon), Komodo 14.1, the gains in direct matches were 300 elo in standard chess and 351 elo in FRC. Note that all gains in standard chess will be significantly larger with unbalanced opening books, but with normal books should be smaller with longer time controls or more threads. The elo gains from 2.5.1 are due to newer nets, small speedups, and some search improvements.
One new feature added this time is UCI Elo, which replaces the Skill levels. To activate it set UCI LimitStrength to true. There is no real difference between setting Elo to 1800 or setting Skill to 18 in older versions (for example), but the advantage is that you can use intermediate values like 1725 Elo for example. The strength of the Elo levels is not much different than the previous Skill levels except in or near the grandmaster range where they are stricter now. The intent is that the Elo ratings should be about fair against humans with similar FIDE ratings at Rapid (15' + 10") time control, although Dragon will move instantly with any Elo set in the human range. Elo levels should be used with just one thread; using more raises the strength by varying amounts, but there is no reason to do this. We would appreciate any feedback on the accuracy of the Elo levels in human terms.
As usual Dragon 2.6 is free for subscribers, is priced the same as 2.5, and is eligible for a 20% discount if you bought Komodo 14 or any later version.

[bastiball]

Congrats for the release!

[Werner]

Thanks for the nice
Christmas Present!

[Cornfed]

I do not recall if I have seen anyone ask this before, but can you kind of indicate how you arrive at the engine playing at a general UCI elo (Skill Levels)?

I mean...how does the engine get to those - is it set to make say 4 positionally dubious moves when the rating is set to 2400 and 5 when set to 2200...or are there random tactical mistakes thrown in...or some combination or does it randomly pick 3 or 4 moves a game and opt for the 4th or 5th 'best' pv as long as the difference is not too far out of some set tolerance?

One other thought (and it may not even apply to Dragon) is when playing in one of these lesser modes, does the engine throw out it's move after day a 5-10 sec think? I ask because it is always disconcerting when one is trying to play a G/30 and you move only to have the engine move instantly...kind of detracts from the whole experience if you know what I mean.

[mjlef]

I do not recall if I have seen anyone ask this before, but can you kind of indicate how you arrive at the engine playing at a general UCI elo (Skill Levels)?

I mean...how does the engine get to those - is it set to make say 4 positionally dubious moves when the rating is set to 2400 and 5 when set to 2200...or are there random tactical mistakes thrown in...or some combination or does it randomly pick 3 or 4 moves a game and opt for the 4th or 5th 'best' pv as long as the difference is not too far out of some set tolerance?

One other thought (and it may not even apply to Dragon) is when playing in one of these lesser modes, does the engine throw out it's move after day a 5-10 sec think? I ask because it is always disconcerting when one is trying to play a G/30 and you move only to have the engine move instantly...kind of detracts from the whole experience if you know what I mean.

In the past we used a combination of search depth limitation and applying some randomness to the eval to get to playing levels to match a needed elo. We do something a little better now (an idea from Larry) which I do not want to reveal until someone figures it out. But anyway, the depth change is kind of the big knob, and the randomness the smaller tuning know. Since Komodo is one of the engines on chess.com we have looked at hsitroical results of our old Skill levels to help tune these things.

As for the "instant moving", if enough people want a time delay added, I can do that. The bots on chess.com do add in a little time for this reason, so if we change Dragon, we would need to get that server side feature removed. But adding a delay does complicate things a little. Anyway, opinions? We could do something like take 1/3 the allocated time or something, so it does not slow down games too much.

Mark

[lkaufman]

I do not recall if I have seen anyone ask this before, but can you kind of indicate how you arrive at the engine playing at a general UCI elo (Skill Levels)?

I mean...how does the engine get to those - is it set to make say 4 positionally dubious moves when the rating is set to 2400 and 5 when set to 2200...or are there random tactical mistakes thrown in...or some combination or does it randomly pick 3 or 4 moves a game and opt for the 4th or 5th 'best' pv as long as the difference is not too far out of some set tolerance?

One other thought (and it may not even apply to Dragon) is when playing in one of these lesser modes, does the engine throw out it's move after day a 5-10 sec think? I ask because it is always disconcerting when one is trying to play a G/30 and you move only to have the engine move instantly...kind of detracts from the whole experience if you know what I mean.

Although search depth and randomness (as well as opening book depth) do both play a role in the Elo/Skill settings, 1500 elo is a dividing line in that above that elo search depth is the dominant factor while below that randomness becomes the dominant factor. One datapoint I can give is that our Elo level is an almost perfect match in Rapid (15' + 10") when set to match Rybka 2.3.2a (the last free Rybka) at its CCRL Rapid rating on one thread. Whether this means it is correct in human terms is debatable; Frank Q. would say we are overrating it, others might say the opposite. Based on limited human feedback we may be a bit strict in the 1600 to 1800 range and probably too generous with our ratings in the under 1200 range, but we don't have nearly enough human feedback to say much. As for an artificial delay, chess.com uses one second (which would also be my personal preference), some users want it to take the full allowed time, others like you want a few seconds. Hard to please everyone without a separate option for this.

[Odd Gunnar Malin]

I have observed at ICC (thanks for letting me running it there) that players play alot faster against Dragon than they normaly do, in classical game (90 + 30) they could have 90 min. left when the game is finnish. This is also true for 15+10 games an in a less extend for 5+3 games.
I was thinking of creating an (easy move) engine to only add some timedeley with a configuration like: fast, normal, slow and letting the timedelay be depended on ho easy the move is ex. with only one possible recapture it should still do it fast. This easymove engine could sit between Dragon and the gui (like InBetween). I personally like an engine to move fast, not so fast as Dragon do, but more like a fast human.
If you could add something like this into Dragon in LimitedStrength mode It would even be better.

[Paloma]

Not Found

The requested URL was not found on this server.

[Cornfed]

I have observed at ICC (thanks for letting me running it there) that players play alot faster against Dragon than they normaly do, in classical game (90 + 30) they could have 90 min. left when the game is finnish. This is also true for 15+10 games an in a less extend for 5+3 games.
I was thinking of creating an (easy move) engine to only add some timedeley with a configuration like: fast, normal, slow and letting the timedelay be depended on ho easy the move is ex. with only one possible recapture it should still do it fast. This easymove engine could sit between Dragon and the gui (like InBetween). I personally like an engine to move fast, not so fast as Dragon do, but more like a fast human.
If you could add something like this into Dragon in LimitedStrength mode It would even be better.

I absolutely never (knowingly...) play against engines on ICC or Chess.com (Cornfed on both) as there are certain old timers I like to play on the former and due to sheer volume, always someone slightly stronger to available to play on the later. But yes, people generally play WAY too fast online even with slower time controls.

It's...disconcerting frankly, when an engine immediately puts you 'on move' in non-forcing situations as one does not get to think in strategic generalities or draw up 'to do lists' on your opponents time (as I was tought)..it disturbs the whole experience.

Even taking a given TC and dividing it by 1/4 and letting an engine space it's moves out based on that would be an improvement.
Say G/30 on an avg game might be 40 moves and so = 45 sec per move and for an engine letting it equal a move roughly every 11 sec, that would at least make the experience a bit more 'normal' and pleasant.

[Rowen]

I do not recall if I have seen anyone ask this before, but can you kind of indicate how you arrive at the engine playing at a general UCI elo (Skill Levels)?

I mean...how does the engine get to those - is it set to make say 4 positionally dubious moves when the rating is set to 2400 and 5 when set to 2200...or are there random tactical mistakes thrown in...or some combination or does it randomly pick 3 or 4 moves a game and opt for the 4th or 5th 'best' pv as long as the difference is not too far out of some set tolerance?

One other thought (and it may not even apply to Dragon) is when playing in one of these lesser modes, does the engine throw out it's move after day a 5-10 sec think? I ask because it is always disconcerting when one is trying to play a G/30 and you move only to have the engine move instantly...kind of detracts from the whole experience if you know what I mean.

Although search depth and randomness (as well as opening book depth) do both play a role in the Elo/Skill settings, 1500 elo is a dividing line in that above that elo search depth is the dominant factor while below that randomness becomes the dominant factor. One datapoint I can give is that our Elo level is an almost perfect match in Rapid (15' + 10") when set to match Rybka 2.3.2a (the last free Rybka) at its CCRL Rapid rating on one thread. Whether this means it is correct in human terms is debatable; Frank Q. would say we are overrating it, others might say the opposite. Based on limited human feedback we may be a bit strict in the 1600 to 1800 range and probably too generous with our ratings in the under 1200 range, but we don't have nearly enough human feedback to say much. As for an artificial delay, chess.com uses one second (which would also be my personal preference), some users want it to take the full allowed time, others like you want a few seconds. Hard to please everyone without a separate option for this.

When playing against engines I use them to train for my long OTB Human games. so I prefer them to simulate a Human chess game so clearly that means that I prefer human thinking time intervals for the engine.
What would be ideal would be an option to control the computers thinking time to simulate a slow or quick opponent, Not seen that before.