Hikaru vs. bots

[M ANSARI]

Let's face it ... humans are completely unpredictable tactically. With engines, even a 2000 ELO engine will not let you off the hook if you made a major tactical oversight. I was in Bahrain watching live when Kramnik missed an incredibly simple and very obvious mate in 1 after spending quite a long time thinking. I mean the Knight was right next to the Queen with an obvious glaring mate in one ... and he just completely overlooked it. That is humans for your! In the same match I think he did what Kasparov did against Deep Blue and resigned in what was probably a drawn endgame. The thing is that for a long time, humans could comfort themselves that if tactics are not missed they will outplay an engine. With NNUE ... this is absolutely no longer the case whatsoever.

[Fritz 0]

Yes, but here we are talking about crippled engines, with limited search depth and randomized evaluation, which do make both tactical and positional mistakes. It goes for NNUE and non-NNUE engines alike.

[M ANSARI]

Yes, but here we are talking about crippled engines, with limited search depth and randomized evaluation, which do make both tactical and positional mistakes. It goes for NNUE and non-NNUE engines alike.

Yes true but the problem that for an engine an obvious tactical or positional blunder is not the same for a human. So the game becomes very weird ... sometimes a one move positional mistake such as allowing opposition in pawn endgame ... which is easily winning for even a 1600 player ... that for an engine is considered the same blunder as some deep piece sacrifice that somehow allows a piece to be trapped and become useless and thus give a winning advantage. Both are considered equally bad by the engine, but a human it is totally different. That is why I find play against these stunted engines very inconsistent. They can play at a 3400 ELO level and then suddenly play to 1400 ELO level. I guess for an engine blunders are all seen as equal while for a human they are not.

[jkominek]

Garry Kasparov once put it very well, I thought. Speaking of near-2700 strength engines from the mid-90s he said (paraphrasing) "they play like 2500 level Grandmasters who never blunder."

That is to say, chess playing entities of equal strength can have dramatically different error profiles. I presume what Larry does in creating his suite of strength levels for Komodo is weaken the engine by increasing degrees and calibrate against chess.com games.

There is an alternative approach. That is to let the engine think for as long as the time control allows, and then make a move choice based on the full gathering of information, not necessarily selecting the best move. This approach involves player modeling. It might be easier to model human player Elo classes using an lczero-style search, rather than alpha-beta. I don't know.

[Fritz 0]

Garry Kasparov once put it very well, I thought. Speaking of near-2700 strength engines from the mid-90s he said (paraphrasing) "they play like 2500 level Grandmasters who never blunder."

That is to say, chess playing entities of equal strength can have dramatically different error profiles. I presume what Larry does in creating his suite of strength levels for Komodo is weaken the engine by increasing degrees and calibrate against chess.com games.

There is an alternative approach. That is to let the engine think for as long as the time control allows, and then make a move choice based on the full gathering of information, not necessarily selecting the best move. This approach involves player modeling. It might be easier to model human player Elo classes using an lczero-style search, rather than alpha-beta. I don't know.

Yes, it is probably the most natural way to decrease the engine's strength, to limit it's thinking time. But it has a big drawback - it is hardware dependant, so it can not be standardized. What they did in Komodo/Dragon is the next best thing I think - to limit and fix the search depth and nodes per move, so it doesn't depend on the hardware. I think it gives a reasonably human-like play (because the engine itself is intended to "behave" that way) and fixed-strength levels. If those levels are realistic or not, is another question. I personally think they are overestimated by about 100 Elo points (in Dragon 2.6), but it certainly has to be tested yet. I have no doubt that it will be fixed in future realeases. But regardless of that, I think that Larry and other gentlemen in Komodo/Dragon have made an unprecedented effort to make an engine weaker in a "human-like" and natural way, which is, as we all know I believe, a much harder task than making the absolutely strongest engine.

[M ANSARI]

Garry Kasparov once put it very well, I thought. Speaking of near-2700 strength engines from the mid-90s he said (paraphrasing) "they play like 2500 level Grandmasters who never blunder."

That is to say, chess playing entities of equal strength can have dramatically different error profiles. I presume what Larry does in creating his suite of strength levels for Komodo is weaken the engine by increasing degrees and calibrate against chess.com games.

There is an alternative approach. That is to let the engine think for as long as the time control allows, and then make a move choice based on the full gathering of information, not necessarily selecting the best move. This approach involves player modeling. It might be easier to model human player Elo classes using an lczero-style search, rather than alpha-beta. I don't know.

I think what Kasparov said is pretty accurate but I would not say it plays at 2500 that doesn't blunder as at that time engines did many positional blunders. I think more accurate would be that they play tactically at 3200 ELO level and positionally at around a 2500 ELO level. But today with NNUE they play both tactically and positionally at a 3400 ELO level. I just can't imagine a human ever outplaying one of the top engines today ... maybe if he has a prepared forced line that the engine falls into ... but even that preparation was not made by a human alone ... that preparation came with engine help.

[Fritz 0]

Does anybody know, what are the bots on chess.com now? Are they based on the levels of Komodo or Dragon? I've just played two games against a bot named Danya, rated 2650 (I suppose it is blitz rating). It is in the same group as Nakamura bot, for example. I lost the first game and drew the second, spending about 15-20 minutes per game. Is it 7 ply? Perhaps 8?

[lkaufman]

Does anybody know, what are the bots on chess.com now? Are they based on the levels of Komodo or Dragon? I've just played two games against a bot named Danya, rated 2650 (I suppose it is blitz rating). It is in the same group as Nakamura bot, for example. I lost the first game and drew the second, spending about 15-20 minutes per game. Is it 7 ply? Perhaps 8?

Chess.com bots use Komodo levels (no NNUE). I think that the ratings are supposed to be about right for slow blitz or fast rapid (i.e. 5' + 5" or 10' with no increment). I don't recall the exact formula for converting level to estimated bot elo, but I think that a 2650 bot is probably either 7 or 8 ply. The bots are probably over-rated, but not way overrated. You must be a strong player to draw against that bot even with 15-20 minutes taken.

[Fritz 0]

Thanks. I guess I got used to playing against the engines over the years.

[Marcus9]

Thanks. I guess I got used to playing against the engines over the years.

Some time ago ccc made a match between bots, in the description of the bots you can see the level and personality of the bot, of what should be komodo 14.
Danya is skill 21 with "active" personality