[zzxjoanw]

Here's the deal and the auction:

---

At trick 1, east leads the ♣3, won in dummy by the Q. At trick 2, the ♦3 is led from dummy and won in hand with the A.

This play at trick 2 seems very strange to me. It guarantees that should the finesse in hearts lose, the defenders will be able to immediately cash the ♦K, setting the contract.

Instead, declarer should attempt to pitch the ♦10 on a club. If this loses, the contract still makes whenever the trump queen is onside, so it seems like a strictly superior line.

Am I missing anything?

This is the advanced GIB btw.

[zzxjoanw]

Forgot to mention, GIB cashes the ♥A at trick 3 and finesses the ♥J at trick 4, going down. It wasn't simply a matter of delaying the club finesse.

[johnu]

Once GIB decided to take the heart finesse, it doesn't matter to GIB that it created an immediate diamond loser. By taking the heart finesse, GIB has determined that it is going to win so there will be time to get rid of a diamond loser later. Of course, this is double dummy type of pseudo logic which makes no sense when you are playing single dummy.

[smerriman]

To be specific, double dummy it can never cost to cash the ♦A at trick 2, because if the ♣K is onside, it can finesse clubs at trick 3.

The algorithm that uses logic like "but you won't know what to do at trick 3" and plans accordingly doesn't kick in until trick 3 sadly.

[wbartley]

Unless I'm misunderstanding the responses to OP, none of them explain GIB's behavior satisfactorily. It seems more likely to me that the set of hands on which GIB ran its double dummy analysis did not contain any hands where the queen of hearts was in the East hand. Because GIB is required to play quickly, its sample size is limited. Had it considered hands where the queen was offside, it would have taken the finesse immediately rather than cashing the ace of diamonds.

I don't know how practical this suggestion is, or even if I'm in the realm of reality when it comes to how GIB's play engine works, but if I were writing a play engine, I would generate a large number of layouts consistent with the bidding and play up to the current point and group them according to the distribution and/or missing honors, then randomly choose one hand from the top n most frequently occurring groups to run DD analysis, weighing the correct play from each group by its relative frequency. I say this because generating hands is orders of magnitude faster than running double dummy analysis.

[Stephen Tu]

Advanced GIB has two engines, one is the advanced "GibSon" engine which has single dummy reasoning, which is kicked in after a few tricks (trick 3 usually? tweakable by the programmers), one is the older, dumber double-dummy Monte Carlo reasoning engine. Using double dummy reasoning, for the matter of cashing DA, it doesn't matter whether it is dealing sample hands with the HQ offside or not, because from GIB's point of view it will simply not finesse hearts when the HQ is offside and take club hook again. It only matters if sample is containing hands where it actually matters, i.e. void diamond. The double dummy approach is very susceptible first few tricks of doing kind of nullo random plays for no purpose that are super unlikely to affect the final result one way or the other if subsequent guesses are going to be successful, when the samples aren't big enough to capture deals with like the void diamond. Only single dummy reasoning realizes that the heart hook may go wrong and it better not create the loser in diamonds so it can take both chances.

I kind of feel like it isn't kicking in GibSon soon enough though, my home GIB if I force it into the line for first few plays will eschew the heart finesse, go for heart drop then repeat the club hook.

This is an interesting hand from human perspective also. Is the best play to take the safety play in hearts? It would be quite diabolical and spectacular for West to find the duck from CKxx holding HQx/HQTx, winning 2nd heart and shifting to diamond. I think I am going to play duck is impossible to find for mortals so I would likely take safety play in real life. Or maybe should cross and take 2nd club hook right away.

[wbartley]

I see. Double dummy analysis ignores the obvious fact that you can only make one play. In a sense, the play of the ace of diamonds is decoupled from the subsequent play even though in reality they are linked.

[TylerE]

Stephen Tu, on 2018-August-26, 12:56, said:

Advanced GIB has two engines, one is the advanced "GibSon" engine which has single dummy reasoning, which is kicked in after a few tricks (trick 3 usually? tweakable by the programmers), one is the older, dumber double-dummy Monte Carlo reasoning engine. Using double dummy reasoning, for the matter of cashing DA, it doesn't matter whether it is dealing sample hands with the HQ offside or not, because from GIB's point of view it will simply not finesse hearts when the HQ is offside and take club hook again.

Given how much cheaper CPU time is than 10-15 years ago, I don't see why the single-dummy engine isn't kicking in at trick 1, for paying customers at least.

[zzxjoanw]

johnu, on 2018-August-25, 20:34, said:

Once GIB decided to take the heart finesse, it doesn't matter to GIB that it created an immediate diamond loser. By taking the heart finesse, GIB has determined that it is going to win so there will be time to get rid of a diamond loser later. Of course, this is double dummy type of pseudo logic which makes no sense when you are playing single dummy.

I don't understand. You're saying GIB is proceeding under the assumption that the heart finesse is going to win? Why does that make sense? It should take the club finesse first so that even if the heart finesse loses it can still make the contract. Unless, of course, you assume that the heart finesse and the club finesse are equally likely to work, in which case GIB's play does make sense. Is that it?

[Stephen Tu]

GIB, first few tricks, is simulating *double dummy* play, meaning it can see whether or not the heart finesse is going to work. It is playing sample deals looking at ALL the cards. Therefore (barring rare void diamond which may not show up given its limited sample size of analyzed deals to maintain performance), the diamond ace is a free play (neither hurting nor helping), because if the heart hook is on it will take it while if it is off it will refuse it (or take the backwards finesse for that matter, if Txx/tx with west) and take the club hook. So from GIB's point of view playing diamond ace is equal to playing a top heart honor preparing for a finesse, and among equal plays it will randomly choose one or the other. It is not until the point of no return (heart hook) that it can realize that DA was a mistake, since earlier it was assuming it was infallible at future guessing, seeing all the cards. If you are infallible at guessing, and no diamond void, basically why not cash the DA since it makes no difference? There's no logic in there, looking at all the cards, for it to decide the DA is bad, because it takes an equal # of tricks on all the deals compared to alternative plays.

It takes the single dummy reasoning advanced engine, which realizes it has to play for cards in particular spots later, and might get some of these wrong, to avoid these kind of errors, but it isn't kicking in soon enough, which is what we have been discussing. Whether advances in computer speed/memory are enough for it to kick in earlier and maintain performance for the multitude of users, is something for BBO developers to measure.

It also doesn't have the human reasoning that the club hook is very likely on, very hard for west to duck trick 1.