Why there is no interest in Computer with odds Vs Humans match?

[Nordlandia]

Swindle or bluffing-mode is one way for the engine to play against the human when giving odds. Playing proper chess in odds matches is likely not the best engine strategy. Playing subpar moves like g4 early in the game put some pressure on the human.

[jp]

Swindle or bluffing-mode is one way for the engine to play against the human when giving odds. Playing proper chess in odds matches is likely not the best engine strategy.

Well, that's the question. Does that engine (or any other), which of course does not think about what is best for handicap matches, bluff? If it does, does it always bluff, even in standard chess?

[todd]

Perhaps with a sufficiently lost position, the whole concept of evaluating in terms of win probabilities (derived from games where mistakes are made sometimes) is more fruitful than the AB approach of attempting to delay checkmate for as many moves as possible.

This will look like a "bluff", but at some point, positions are lost enough ("enough" being relative to the strength of the opponent) that bluffing is all that's left - trying to resist by other means becomes nearly hopeless if the position is bad enough.

[lkaufman]

We're probably already at the point where Lc0 network 11248 (before they introduced resignation) can give N odds in 3' + 1" blitz to Nakamura with decent results. Lc0 11248 plays it like a super human GM would, looking for any way to complicate, like an early g4 for example. It is just much more effective than playing "proper" chess like any A/B engine will do.

Do you think Lc0 11248 is playing "practically" or "bluffing", rather than correctly or soundly?
If so, is that just with the handicap or when it's down material?

I would call it playing practically; it tries moves that have the highest probability of the opponent making a mistake in positions where it has almost no chance of a draw by playing "correctly". It doesn't matter if it is giving knight odds or if it just somehow loses a knight in play, all that matters is that if it judges its win probability to be very low, it will be less concerned about what is "correct".

[lkaufman]

Swindle or bluffing-mode is one way for the engine to play against the human when giving odds. Playing proper chess in odds matches is likely not the best engine strategy.

Well, that's the question. Does that engine (or any other), which of course does not think about what is best for handicap matches, bluff? If it does, does it always bluff, even in standard chess?

Well, there is a small element of "bluffing" with any MCTS engine even when the game is even, because MCTS aims to maximize win probability, not to find the "best" move. But when the game is roughly even, it won't stray too far from correct play, because it assumes that the opponent is also maximizing his win prob., not just playing random moves. The more lopsided the game becomes, the less point there is to finding best moves, at least when you are losing. I think the NN engines could do better to play more correctly when winning than when losing. We try to do that with Komodo mcts.

[lkaufman]

Perhaps with a sufficiently lost position, the whole concept of evaluating in terms of win probabilities (derived from games where mistakes are made sometimes) is more fruitful than the AB approach of attempting to delay checkmate for as many moves as possible.

This will look like a "bluff", but at some point, positions are lost enough ("enough" being relative to the strength of the opponent) that bluffing is all that's left - trying to resist by other means becomes nearly hopeless if the position is bad enough.

Yes, absolutely right, that's why Lc011248 is able to give larger handicaps to titled players than Komodo is able to give. But you shouldn't get desperate too early either. Maybe at f7 handicap, for example, proper play may be better than practical play.

[jp]

Perhaps with a sufficiently lost position, the whole concept of evaluating in terms of win probabilities (derived from games where mistakes are made sometimes) is more fruitful than the AB approach of attempting to delay checkmate for as many moves as possible.

This will look like a "bluff", but at some point, positions are lost enough ("enough" being relative to the strength of the opponent) that bluffing is all that's left - trying to resist by other means becomes nearly hopeless if the position is bad enough.

Yes, absolutely right, that's why Lc011248 is able to give larger handicaps to titled players than Komodo is able to give. But you shouldn't get desperate too early either. Maybe at f7 handicap, for example, proper play may be better than practical play.

This suggests that the weaker networks will be better at bluffing, and therefore at handicap matches, than the stronger networks, just because of the way they are trained. Once the networks get very strong, the training games, which are produced by these stronger networks, won't reward the bluffing even with some added noise, so the bluffing won't give higher win probabilities.

The other question is whether a network can really be good enough to bluff when it should, i.e. when the position is completely lost, but not when it shouldn't, i.e. in every other position. I sort of suspect the networks aren't good enough to make that distinction, so the bluffing networks will have some bluffing tendencies even in equal or good positions.

[lkaufman]

Perhaps with a sufficiently lost position, the whole concept of evaluating in terms of win probabilities (derived from games where mistakes are made sometimes) is more fruitful than the AB approach of attempting to delay checkmate for as many moves as possible.

This will look like a "bluff", but at some point, positions are lost enough ("enough" being relative to the strength of the opponent) that bluffing is all that's left - trying to resist by other means becomes nearly hopeless if the position is bad enough.

Yes, absolutely right, that's why Lc011248 is able to give larger handicaps to titled players than Komodo is able to give. But you shouldn't get desperate too early either. Maybe at f7 handicap, for example, proper play may be better than practical play.

This suggests that the weaker networks will be better at bluffing, and therefore at handicap matches, than the stronger networks, just because of the way they are trained. Once the networks get very strong, the training games, which are produced by these stronger networks, won't reward the bluffing even with some added noise, so the bluffing won't give higher win probabilities.

The other question is whether a network can really be good enough to bluff when it should, i.e. when the position is completely lost, but not when it shouldn't, i.e. in every other position. I sort of suspect the networks aren't good enough to make that distinction, so the bluffing networks will have some bluffing tendencies even in equal or good positions.

The networks are trained on very fast games, vastly faster than games that would be played at human time limits. So they will play in a way that optimizes win prob. against an opponent much weaker than the level of the network at a realistic time limit. Whether that is above or below the level of Nakamura (for example) at a realistic time limit is hard to say, but I suspect it's close enough for the purposes we are talking about. The problem with the newer networks is not their strength, but as has been explained here by dkappe it is that resignation was introduced so they stopped training on piece down positions usually. So I suspect that Lc0 11248 plays pretty close to optimally for serious games against a top-ten human player at some rapid TC regardless of the handicap, because its training happens to emulate such an opponent fairly well.

[jp]

The networks are trained on very fast games, vastly faster than games that would be played at human time limits. So they will play in a way that optimizes win prob. against an opponent much weaker than the level of the network at a realistic time limit. Whether that is above or below the level of Nakamura (for example) at a realistic time limit is hard to say, but I suspect it's close enough for the purposes we are talking about. The problem with the newer networks is not their strength, but as has been explained here by dkappe it is that resignation was introduced so they stopped training on piece down positions usually. So I suspect that Lc0 11248 plays pretty close to optimally for serious games against a top-ten human player at some rapid TC regardless of the handicap, because its training happens to emulate such an opponent fairly well.

"Optimally" there is just in the "bluffing" sense. We'd expect a network trained on an equal number of games, but of superior quality, to be a superior network in the "objectively correct" sense and then it would maybe "bluff" less, unless the networks bluffed strictly only when the position was lost. We'd expect that even if all training games are carried out till mate.

[lkaufman]

The networks are trained on very fast games, vastly faster than games that would be played at human time limits. So they will play in a way that optimizes win prob. against an opponent much weaker than the level of the network at a realistic time limit. Whether that is above or below the level of Nakamura (for example) at a realistic time limit is hard to say, but I suspect it's close enough for the purposes we are talking about. The problem with the newer networks is not their strength, but as has been explained here by dkappe it is that resignation was introduced so they stopped training on piece down positions usually. So I suspect that Lc0 11248 plays pretty close to optimally for serious games against a top-ten human player at some rapid TC regardless of the handicap, because its training happens to emulate such an opponent fairly well.

"Optimally" there is just in the "bluffing" sense. We'd expect a network trained on an equal number of games, but of superior quality, to be a superior network in the "objectively correct" sense and then it would maybe "bluff" less, unless the networks bluffed strictly only when the position was lost. We'd expect that even if all training games are carried out till mate.

I think there are two unrelated issues at work here. The NN training will tend to make the engine play like a player of the same level as the training games, so the higher the level of the training games, the more objectively correct the play will be. The other issue is that MCTS causes play to automatically "bluff" to some degree, because it doesn't assume that the best response will always be played. So Komodo MCTS does benefit from this when playing humans or weaker engines, unlike A/B engines, while Lc0 benefits from both of the above when playing weaker opposition. But oddly Lc0 performs worse (Elo-wise) against weaker engines, despite this "bluffing".