[iandayre]

First time I have posted about a GIB defensive error.

http://tinyurl.com/l5omnbg

One can only guess at what bizarre form of electronic "thought" would tell GIB to follow with the ♠10 at Trick 1. Declarer must have a long strong ♠ suit on the auction. Most auctions were the same through 4H. I was the only one to raise to 5♥, two others jumped to 6 and received the same defense. The larger number who bid Blackwood got a ♦ lead and GIB did not throw the ♠10 under the K, so these declarers went down.

[johnu]

Who said GIB doesn't give signals of defense? Obviously a reverse count, standard attitude signal. This may also have been an early unblocking play to avoid getting endplayed later. I have been complaining about GIB tossing random cards on defense since I started on BBO. Why can't GIB always play the lowest card when just following suit since giving count/attitude signals seems to be too difficult.

[steve2005]

According to Gib slam still makes if East Gib plays small. but yes not a good play.

I concur with Johnu just play small, unless it makes some difference by the simultions run (avoid being endplayed, give entry to p etc.)

[1eyedjack]

I suspect that against a poor and inattentive human, GIB gains more than it loses by false-carding, even accepting that it loses some on odd occasions by pitching an active card.

Additionally, against an expert and attentive human, GIB would lose more than it would gain by always, predictably and reliably playing the lowest card (when not playing high).

Certainly it could use some improvement in calculating when a card is "small" v "active", and I expect that it will never be immune from error in that assessment. But I don't see the solution to this problem as always playing the lowest card.

[johnu]

 1eyedjack, on 2014-November-01, 16:38, said:

I suspect that against a poor and inattentive human, GIB gains more than it loses by false-carding, even accepting that it loses some on odd occasions by pitching an active card.

Additionally, against an expert and attentive human, GIB would lose more than it would gain by always, predictably and reliably playing the lowest card (when not playing high).

Against players who aren't paying attention, it probably doesn't make any difference what GIB plays. If they aren't paying attention, what can you gain by false carding?

There are some false cards that are no lose (except for fooling partner). IE you have spot cards something like 765. GIB should value them exactly the same as obviously you can play any of the 3. And it is fairly rare that letting declarer know that you have followed suit with your lowest card is any help. In any case, figuring out when randomly playing a spot card could lose a trick can get pretty complicated, so I would consider just following suit with the lowest card as a significant play upgrade for GIB.

[1eyedjack]

 johnu, on 2014-November-01, 20:33, said:

Against players who aren't paying attention, it probably doesn't make any difference what GIB plays. If they aren't paying attention, what can you gain by false carding?

Just because a player is inattentive does not mean that they are blind to every card being played. A (generally) inattentive player will pay more attention to a higher card than a lower card, and on seeing a higher card played may, if not keeping proper count, conclude in error that it is the player's lowest available. It happens. I have seen it happen, so there is no denying it. Or you can deny it but I reject the denial and we shall just have to agree to differ.

Furthermore, your suggested "improvement" would reward such an inattentive player. His conclusion that the opponent's cards in a suit are (eg) exhausted by reason of the appearance of a high spot card when he has not been paying attention to the lower spot cards or otherwise counting the suit will be valid.

 johnu, on 2014-November-01, 20:33, said:

There are some false cards that are no lose (except for fooling partner). IE you have spot cards something like 765. GIB should value them exactly the same as obviously you can play any of the 3. And it is fairly rare that letting declarer know that you have followed suit with your lowest card is any help.

I absolutely and vehemently disagree, both with your assessment of frequency and assessment of importance when it happens. GIB is regularly dealt a holding such as 765. If it always plays the 5 from this, and the 6 from 76, but on the hand in question plays the 7, then I know that his partner has both 65. How can you sit there and say that this is rarely any help?

 johnu, on 2014-November-01, 20:33, said:

In any case, figuring out when randomly playing a spot card could lose a trick can get pretty complicated, so I would consider just following suit with the lowest card as a significant play upgrade for GIB.

Absolutely not.

Agreed, it can get complicated. Existing GIB tries to resolve it; largely by way of simulations although it may have other tools.

It doesn't always get it right, and sometimes when it gets it wrong it does so in a spectacular fashion that a human would never replicate. Such examples then find their way to this forum. I doubt that it would get such publicity on the many more frequent hands that it plays the 7 when lacking the 6 or 5 after the implementation of your suggested solution and a competent declarer capitalises on it.

So again we must agree to differ.

It doesn't always get it right, and there may well be programming solutions that would improve its performance and reduce (but not eliminate) the grossest of errors. But your suggested cure is worse than the disease.

[Antrax]

Would you agree that playing lowest always (3 from 345) leaks information? Then you want "one of the equal lowest" to avoid that.
Do you agree that on the actual hand (and thus potentially in the sims), all plays result in the same outcome, double-dummy? Then all spades are equally low.
That's the "electronic thought" (as in OP) in GIB's head.

[johnu]

 Antrax, on 2014-November-01, 23:46, said:

Would you agree that playing lowest always (3 from 345) leaks information? Then you want "one of the equal lowest" to avoid that.
Do you agree that on the actual hand (and thus potentially in the sims), all plays result in the same outcome, double-dummy? Then all spades are equally low.
That's the "electronic thought" (as in OP) in GIB's head.

Didn't I already say that false cards from equals are no lose plays? I'm just saying that instead of playing random spot cards which can give away the hand, I would just as soon have GIB play the lowest card. If declarer is good enough to take advantage, more power to them. Certainly human players with extensive (any?) carding agreements leak much more information. It's a moot point because based on previous responses from BBO, I don't expect any changes in card play any time soon.

Of course, GIB frequently fails to run enough simulations to give a decent projection of future play (e.g. allowing for a fairly unlikely split in a key suit), , or worse, uses a completely wrong set of assumptions about the strength and distribution of the other hands, whether from a bug in the bidding database, or from a psych or semi-psych. The biggest flaw in this double dummy analysis is that both GIB and human declarers actually play single dummy, not double dummy.

[Antrax]

Your last sentence is correct.Nevertheless, the leading Bridge engines use the same approach. I think it means this approach is at least somewhat viable.

Your point about "not enough simulations" might be true but is irrelevant to this thread, where GIB's sample size was apparently enough to determine the truth: the contract makes on any defense.

[steve2005]

 johnu, on 2014-November-02, 02:33, said:

The biggest flaw in this double dummy analysis is that both GIB and human declarers actually play single dummy, not double dummy.

Playing single dummy the T♠ could be an important card. Playing the T♠ occasionally is bigger information leakage than playing smallest ♠ 100 other times. Imho

[iandayre]

I thought for a moment that I had figured out the double dummy line to make 6H if E plays a small ♠ at T1, but no, it doesn't work. Double dummy problems are not my strong point.

In fact, no declarers made the hand unless GIB tossed the ♠10 under the K. Some programming allowance should me made for the fact that even expert declarers are not perfect, much less average players.

And why does it pitch the ♠10 if it is the opening lead, but not after a D lead with declarer playing a ♠ at T2?

I do agree that the answer is not for GIB to always follow with its smallest card.

[johnu]

 Antrax, on 2014-November-02, 10:05, said:

Your last sentence is correct.Nevertheless, the leading Bridge engines use the same approach. I think it means this approach is at least somewhat viable.

Your point about "not enough simulations" might be true but is irrelevant to this thread, where GIB's sample size was apparently enough to determine the truth: the contract makes on any defense.

Obviously on this particular hand, double dummy play will make the contract, but since the actual play isn't double dummy, the contract won't make if declarer misguesses. It wouldn't take long to browse through recent threads to find examples where someone made a serious misbid that causes GIB to make a nullo play because the actual distribution is supposed to be impossible. If so, you could run a million simulations and not get the right result.

[johnu]

 iandayre, on 2014-November-02, 13:23, said:

I thought for a moment that I had figured out the double dummy line to make 6H if E plays a small ♠ at T1, but no, it doesn't work. Double dummy problems are not my strong point.

There are probably several successful lines, but off the top of my head,

Cross to ♦A, club finesse, cash ♣A pitching spade, diamond ruff, spade ruff, diamond ruff , spade ruff with ♥A, trump to hand, draw trump, cash ♠A. That's 2 club tricks, 1 diamond trick, 2 top spade tricks, 5 trumps in hand, 2 spade ruffs in dummy, for 12 tricks. You'll have a losing spade at trick 13.

[johnu]

 johnu, on 2014-November-02, 14:21, said:

There are probably several successful lines, but off the top of my head,
.....

A more double dummy feel line would be,

low heart to ♥K
club to ♣Q
♣A pitching a spade
low club ruffing high
♠A - If west pitches, pitch a diamond, finesse ♥9, ruff a club, trump to ♥A, cash 2 long clubs and ♦A. 4 clubs, 1 diamond, 2 spades, 5 hearts makes 12 tricks.
If west ruffs the ♠A, overruff, ruff a club, heart to ♥A to draw trumps, cash 2 long clubs, ♦A, and last trump in hand. 4 clubs, 1 diamond, 1 spade, 5 hearts, spade ruff makes 12 tricks.

[Antrax]

 johnu, on 2014-November-02, 13:50, said:

Obviously on this particular hand, double dummy play will make the contract, but since the actual play isn't double dummy, the contract won't make if declarer misguesses.

I understand that, but you have to realize that GIB will never "understand" that. There's no easy way to introduce such logic to it, so we'll have to contend ourselves with the knowledge that we have to consider plays like this "okay". It would've been great had GIB had an extra layer of "the most deceptive play" or "give the opponents the most headache" or something, all other things being equal, but that's probably never happening.

[johnu]

 Antrax, on 2014-November-02, 22:44, said:

I understand that, but you have to realize that GIB will never "understand" that. There's no easy way to introduce such logic to it, so we'll have to contend ourselves with the knowledge that we have to consider plays like this "okay". It would've been great had GIB had an extra layer of "the most deceptive play" or "give the opponents the most headache" or something, all other things being equal, but that's probably never happening.

I believe that "give the opponents the most headache" is already a feature, although sometimes it is CHO who gets the headache.

[iandayre]

I think it is remarkable that that programmers have taken the time to give GIB the ability (one which no human defender possesses) to accurately double-dummy analyze a hand such as this at Trick 1. Yet they can't teach it to take simple preferences, stop introducing new 3 and 4 card suits at high levels, or to not pass cue bids. They can't even teach it that, in a recent hand posted here - I believe it was KQ, x, AKJTxxxxxx, V - that 11 tricks are nearly assured and to not stop bidding at the 3 level. Or there was that recent hand where both minors were unbid and it held 6 Diamonds and 2 Clubs, but it bid Clubs.

Also I wonder how if it can analyze this hand a Trick 1, why it misdefends so many much simpler hands.

[Bbradley62]

 iandayre, on 2014-November-03, 12:48, said:

I think it is remarkable that that programmers have taken the time to give GIB the ability (one which no human defender possesses) to accurately double-dummy analyze a hand such as this at Trick 1.

You have to remember that there have been (at least) two sets of "programmers": Matthew Ginsberg (possibly and his team) originally wrote GIB, and BBO purchased GIB a few years ago. My understanding is that BBO's programmers have not done anything with respect to GIB's cardplay; they are only dealing with patching holes in the bidding system, or revising it to accommodate human partners. You really shouldn't compare what was done by one group to what has been done by the other, and particularly shouldn't present it as though it's one person/team prioritizing one thing over another.

[1eyedjack]

I don't believe that the new owner/programmers have completely disregarded improvements in card play. I do agree that it seems to have taken a lower priority, particularly the complete absence of defensive carding signals, but they have expressly stated in these forums that the card play does receive attention, and I would be loath to call them liars.

[Bbradley62]

If you search the GIB subforum for the word "defense" in posts made by Barmar, you will find many posts indicating that GIB's defense is such that the programming staff really can't do anything about it.