[Bbradley62]

I know... we can't do anything about GIB's poor defense... but I don't understand how this works...

---

After cashing two high spades, it should be clear for West to cash a diamond before switching to hearts, looking for the big payoff if partner has either round ace. I don't see how GIB could simulate hands and not have any of them include declarer holding ♥AQ♣A while East has ♥K.

[cloa513]

This seems to suggest everything we have told about GIB's defence is total rubbish. I don't think GIB defenders can see the dummy- only infer what's in it and certainly don't play as through opponents play DD.
Its 100% obvious how to play the hearts but GIB still does it wrong- the lead can only be singleton since with any honour + ten, partner would lead the higher one- assumes opponents can see through the card

[Stephen Tu]

I don't see why what you've been told is rubbish. From West's point of view, he can get extra undertricks leading stiff if east has HA+DK, and breaks even as the cards actually lie (because he is assuming east won't blunder by playing the HK, since he is assuming other people at the table are playing double dummy). From East's point of view, I think the problem is that lead conventions only apply on *opening lead*, and not in the middle of the hand? So it doesn't understand that putting up the K can't be right. Also it might be that the lead convention only specifies which card GIB *plays*, for human readability, but GIB *partner* doesn't bias his sample based on this info. Not 100% sure about these things, barmar would be better able to clarify.

Of course the defense can see dummy; the problem is they can't see partner's hand, and they can't make inferences about partner's hand based on his carding. Signalling/lead conventions are huge part of bridge defense, and GIB doesn't understand them at all. It's playing with a partner who is not completely random, trying to set the contract, but very unreliable about what order card he plays when "it doesn't make a difference if everyone else is peeking".

[cloa513]

Stephen Tu, on 2012-May-24, 10:40, said:

I don't see why what you've been told is rubbish. From West's point of view, he can get extra undertricks leading stiff if east has HA+DK, and breaks even as the cards actually lie (because he is assuming east won't blunder by playing the HK, since he is assuming other people at the table are playing double dummy). From East's point of view, I think the problem is that lead conventions only apply on *opening lead*, and not in the middle of the hand? So it doesn't understand that putting up the K can't be right. Also it might be that the lead convention only specifies which card GIB *plays*, for human readability, but GIB *partner* doesn't bias his sample based on this info. Not 100% sure about these things, barmar would be better able to clarify.

Of course the defense can see dummy; the problem is they can't see partner's hand, and they can't make inferences about partner's hand based on his carding. Signalling/lead conventions are huge part of bridge defense, and GIB doesn't understand them at all. It's playing with a partner who is not completely random, trying to set the contract, but very unreliable about what order card he plays when "it doesn't make a difference if everyone else is peeking".

But East does put down the KH which according to GIB DD is only way that N-S can make the contract other assuming a failure to cash all the winners.
If Defence can see dummy then cashing the ace according to DD should be straightforward.

[Stephen Tu]

I don't think you have a good understanding of what double dummy Monte Carlo simulation actually entails; your reasoning is flawed.

It doesn't mean that West actually sees all the cards and makes the right double dummy play. Doesn't mean East does either.

It *can* see dummy. It *cannot* see partner's or declarer's hand. It can only deal *possible* hands for partner and declarer, and double dummy solve the *possible* hands, and play in such a way to best defeat the *possible* hands on average. This may or not work out against the *actual* hand. But it is using dummy's actual hand as a fixed thing; it *does* "see the dummy".

This is (roughly) what West is thinking, after it cashes the second spade:
At this point in time, I can cash DA for down 1, down 3 if partner have HA. But if I switch to the heart ten, if partner has HA and DK, I can get heart ruff, diamond to K, and another heart ruff. And if partner has the K of hearts, since I am assuming he is defending double dummy and knows I have stiff HT, he won't do something silly like put up the K, so we still beat this contract. It's not thinking deep enough that "oh partner *doesn't* know for sure that my HT is stiff, and might f*k up, so I better just settle for down 3 instead of down 4 if partner has HA, to make sure they don't make when he doesn't". The assumption on deciding on a card is based on *everyone else* playing double-dummy. But GIB choosing the card is not playing double-dummy. But you can see from West's point of view, if everyone else is omniscient, (and it has guts to omnisciently underlead DA later) clearly heart is better than diamond, so heart is chosen.

What East is thinking, apparently:
-Partner from hqt or HAT might be leading the T. So I better put up the K, since these cases are more common than stiff T. And even if stiff T, it won't matter if partner has club ace and declarer diamond ace instead of the actual holding.

The defence seeing *dummy* only, does NOT make the defense straightforward. It also has to be able to use leading conventions and deeper reasoning in order to get this hand right. Not seeing dummy isn't the reason it goes wrong. Now, human is not going to make this error, because we know partner is not going to be diabolical and lead t from QT or AT in this situation. We have lead convention and inference on our side and assumption partner trying to help us out. Bot does not.

Ginsberg's idea of how to improve bot defense was to get them to think more "recursively", get in the head of declarer and partner, have declarer playing single dummy, figure out why partner doing what they did rather than something else. But he never got around to implement because this just gets to be massively more computation, way too slow for the PCs when he was still working on the program.

[Bbradley62]

Stephen Tu, on 2012-May-24, 16:07, said:

This is (roughly) what West is thinking, after it cashes the second spade:
At this point in time, I can cash DA for down 1, down 3 if partner have HA

or CA.