The "expert" cant make 3NT

[Wackojack]

[hv=d=n&v=n&n=s94hkqdkj972c10764&w=sk862hj3d843cq532&e=saqjha87daq65cakj&s=s10753h1096542d10c98]399|300|Scoring: Total Points[/hv]

 

I have played money bridge fairly regularly and understand that I have to take the rough with the smooth when experiencing GIB's many ideosynchrasies. Nonetheless, this latest one has made me feel quite dispirited about partnering GIB again. Would you believe GIB went down in this contract.

 

Bidding:

- 2♣

2♦ 2NT

3NT

 

The lead was J♥ to Q and ducked. Then K♥, A♥ 2, 3. GIB then played out the clubs and spades ending in dummy ♠8,♦843 with ♥7, ♦AQ6 in hand. 8 tricks taken. Did GIB settle for the contract by taking the A♦? No! The idiot robot played a spade "knowing" it would be taken by south, with surely the long hearts.

 

31 combined HCP. GIB somewhat underbidding the east hand with a rebid of 2NT and then risking +400 for the (one would have thought) remote possibility of an overtrick. Does anyone understand why GIB would do this?

[helene_t]

The only explanation I can make up is that the hearts blocked in all the simulations it made. Maybe it made some assumptions on South's systematic heart lead.

[barmar]

I don't think that's it. If it assumes South led a 4th-best ♥, the Rule of 11 tells it that North has no more high ♥, so there's no blockage.

 

I think it was playing for ♥ originally 5-3. South has presumably discarded a ♦ and ♥ on the ♣, so now it only has 2 ♥ left. It will take them and then be end-played into giving the free finesse into the ♦AQ for an overtrick.

 

As Helene points out, GIB doesn't use bridge logic, it just uses simulations. If this generates an overtrick in 90% of its simulations, it goes for it.

[Wackojack]

ok thanks for shedding some light on unravelling GIB's logic. Incidentally in order to get a better score in the long run, assuming the 2 alternatives are -50 and +430, (using school maths) the odds on making an overtrick must be better than 93.75%. Thus unless it does more than 26 simulations all generatiing an overtrick (1/1-.9375) then it should not go for the overtrick.

[cherdano]

Maybe GIB thought North would play K, then Q, from KQ tight.

[helene_t]

the odds on making an overtrick must be better than 93.75%.  Thus unless it does more than 26 simulations all generating an overtrick (1/1-.9375) then it should not go for the overtrick.

I don´t know how GIB would asses the probability of making an overtrick if all his, say, 26 simulations makes an overtrick. My suspicion is that he simply says 26/26=100%. From a Bayesian point of view this can't be quite correct but it may be a reasonable rule in practice. I mean, the estimator for the best strategy may still be correct even if the involved probabilities are not.

 

But maybe the algorithm could be improved by making a distinction between plays that are theoretically 100% safe and plays that might theoretically fail but just always succeed in the simulations.

[barmar]

I suspect Helene's explanation is what happened. Figuring out the truly 100% plays requires single-dummy analysis, which GIB doesn't have (it's much more complex, except for some simple cases of just counting top tricks). It only knows how to do double-dummy analysis, so it deals a bunch of cards into the opponents' hands (consistent with the bidding and play so far) and analyzes this. Unfortunately, this only provides accurate results if you deal lots of hands.

[Robert]

Hi everyone

 

Isn't this a case of GIB will return the favor(money?) since it will go down when it plays another hand 'against' you?

 

Regards,

Robert

[ashokg]

This is not 100 % pertinent to the issue here, but I'll mention it as a curiosity. When there are exactly two possible outcomes for a random experiment (e.g. dealing a random hand from a constrained set of possibilities, where a certain line either makes the contract or goes down one--the two possibilites), and you seek to estimate the probability of it making, making the a priori assumption that it may equally well be any number in the range 0 to 1, then the probability after simulations that resulted in m successes and n failures [corrected now; thanks barmar] is (m + 1) / (m + n + 2); e.g. if in 26 simulations (randomly dealt hands, of course), there were 26 successes, the p of success in the actual hand is 26/27 = 96.3 %. This assumes, of course, that the sampling is perfect, i.e. the hand generator depicts the actual situation perfectly.

[barmar]

Your example doesn't seem to match the formula you gave.

 

n = 26

m = 0

(m+1)/(m+n+2) = 1/28 = 3.6%

 

I suspect you got m and n mixed up, and meant that the formula is

 

(n+1)/(m+n+2) = 27/28 = 96.4%

 

However, this formula is only useful when there are two possible outcomes, but in bridge you can make, make with overtricks, go down 1, go down 2, etc. Also, GIB doesn't work by choosing a line and then dealing hands to see if the line makes. Rather, it deals hands and then performs a double-dummy analysis to determine the number of tricks that can be taken and the line necessary to get them. Whatever the next play is that appears the most in these analyses is then chosen.

 

I wonder if part of the problem is that it always looks only at the lines that maximize the number of tricks, rather than all the lines that make the contract (when playing IMPs or total points). If it looks at all the making lines, I'd expect it to play more safely, although it's still probably the case that it won't protect against rare layouts like 6-0 breaks unless something in the auction or earlier play has made them more likely.