[A2003]

GIB Declarer play
Why GIB cashing out winners and not looking ways to make the contract

[olegru]

I may be wrong, but according to my understanding declarer play by GIB implemented by the following algorithm.
1. Generate a finite number of possible (according the current knowledge) boards.
2. Play each of generated boards double-dummy.
3. Select the play that gives you a better results averaged on the set of generated boards and play it in the current board.
4. Repeat the procedure on the next round.

The main difference between the basic and advanced GIB is the number of boards program generates and analyses before choosing a play.
I don't know the number of generated boards, but I guess it is relatively small especially for basic version, so it is possible that generated set of boards will not be fair representation of actual probability, so GIB will choose very strange line.

[zhasbeen]

A2003, on 2018-February-01, 13:46, said:

GIB Declarer play
Why GIB cashing out winners and not looking ways to make the contract

GIB has only 11 tricks without the club finesse (or stiff K). Note he didn't cash majors until after cashing the AD, giving him an extra chance if the KD was doubleton. The only other combination that would allow him to take 5 diamonds would be Kxx. That would mean he'd need only 2 club tricks, but he'd have to get them without losing one first (he'd have to use his last heart to reach dummy to finesse clubs.) His other entry would be used by crossing to dummy's QS to repeat diamond finesse.
In a nutshell it looks like GIB gave himself the best chance.

Declarer play is only part of GIB's game that is usually good, and at times it is great. Defense is a distant 2nd and bidding is horrific.

[steve2005]

zhasbeen, on 2018-February-01, 15:58, said:

GIB has only 11 tricks without the club finesse (or stiff K).

No, ♦ finesse worked, why Gib didn't lead Q to get rid of all entry problems is bizarre.
So 3 diamonds 2 hearts 3 spades 4 clubs just giving up the King while suits are stopped.

[zhasbeen]

steve2005, on 2018-February-01, 16:37, said:

No, ♦ finesse worked, why Gib didn't lead Q to get rid of all entry problems is bizarre.
So 3 diamonds 2 hearts 3 spades 4 clubs just giving up the King while suits are stopped.

I suppose that could help in terms of damage control if the club loses, but GIB's line was pretty decent in terms of making. He's combining chances of Kx of diamonds with club finesse and taking only 2 club tricks if they happened to be 5-0

Compared to some of the stuff GIB does it was wonderful. Either way you'll need more than one club to make.

[steve2005]

zhasbeen, on 2018-February-01, 20:37, said:

I suppose that could help in terms of damage control if the club loses, but GIB's line was pretty decent in terms of making. He's combining chances of Kx of diamonds with club finesse and taking only 2 club tricks if they happened to be 5-0

Compared to some of the stuff GIB does it was wonderful. Either way you'll need more than one club to make.

No if the diamond finesse works that's all you need, you have 4 clubs (you have all the high spots) you don't need the club finesse.
The club finesse you only have 11 tricks. You need to find the K♦ still.

[johnu]

zhasbeen, on 2018-February-01, 15:58, said:

GIB has only 11 tricks without the club finesse (or stiff K). Note he didn't cash majors until after cashing the AD, giving him an extra chance if the KD was doubleton. The only other combination that would allow him to take 5 diamonds would be Kxx. That would mean he'd need only 2 club tricks, but he'd have to get them without losing one first (he'd have to use his last heart to reach dummy to finesse clubs.) His other entry would be used by crossing to dummy's QS to repeat diamond finesse.
In a nutshell it looks like GIB gave himself the best chance.

Declarer play is only part of GIB's game that is usually good, and at times it is great. Defense is a distant 2nd and bidding is horrific.

Maybe this was the best line of play in a nut house, not a nutshell Count your tricks. 3 spades, no more, no less. 2 hearts, no more, no less. That's 5 tricks. If the club finesse is on, you can take 5 club tricks except against a 5-0 split, so you still need 2 (or 3 diamonds) diamonds. Win the heart lead and take club finesses. If it loses, lead ♦Q for a finesse. You need 1 of 2 finesses to make.

Before the human declare robot tournaments, I used to dread having GIB declare a hand. There were usually at least a hand a tournament when GIB would kick tricks and sometimes the contract by taking a hopeless line of play.

[steve2005]

johnu, on 2018-February-02, 08:50, said:

Win the heart lead and take club finesses. If it loses, lead ♦Q for a finesse. You need 1 of 2 finesses to make.

Does nobody read? take the diamond finesse twice. if it works you can take club finesse for 7 but you can just duck a club for 4 club tricks and make.
3♠ 2♥ 3♦ 4♣ = 12 tricks

If you finesse club first you need both finesses to make 12 as you still need 2 ♦ tricks even with 5 ♣ tricks.

[zhasbeen]

johnu, on 2018-February-02, 08:50, said:

Maybe this was the best line of play in a nut house, not a nutshell Count your tricks. 3 spades, no more, no less. 2 hearts, no more, no less. That's 5 tricks. If the club finesse is on, you can take 5 club tricks except against a 5-0 split, so you still need 2 (or 3 diamonds) diamonds. Win the heart lead and take club finesses. If it loses, lead ♦Q for a finesse. You need 1 of 2 finesses to make.

Before the human declare robot tournaments, I used to dread having GIB declare a hand. There were usually at least a hand a tournament when GIB would kick tricks and sometimes the contract by taking a hopeless line of play.

Sorry, I'm still not counting 12 tricks without both finesses working. Also, there is no diamond holding that will yield 5 tricks that wouldn't work with GIB's line except if west held exactly 3. In that case you would STILL need at least one club finesse to work. For making the contract I'm not seeing a better line.

That being said, your line is better in terms of damage control. You would be down 1 instead of 5 like GIB was, although it looks to me that you reduce your chances of making with 5 diamonds and only 2 clubs when west holds all 5 clubs.

"I used to dread having GIB declare a hand..."

I have a friend who said the same thing but am wondering if this isn't one area of GIB's game that has improved in the last year or so. I played my first robot tournament in August.

[smerriman]

steve2005, on 2018-February-02, 09:26, said:

If you finesse club first you need both finesses to make 12 as you still need 2 ♦ tricks even with 5 ♣ tricks.

Uh, if you finesse the club first, exactly the same principle as your line applies. If it holds, you have 12 tricks after knocking out the king of diamonds.

Whichever suit you finesse first, you have 12 tricks as long as one finesse works (not sure why zhasbeen thinks you need both - maybe the same reason you thought you needed both when playing clubs first ). Far better than GIB's line.

[zhasbeen]

smerriman, on 2018-February-02, 13:41, said:

Uh, if you finesse the club first, exactly the same principle as your line applies. If it holds, you have 12 tricks after knocking out the king of diamonds.

Whichever suit you finesse first, you have 12 tricks as long as one finesse works (not sure why zhasbeen thinks you need both - maybe the same reason you thought you needed both when playing clubs first ). Far better than GIB's line.

I see the light. "...after knocking out..." were the words you used that woke me up. You don't need the diamond finesse to work in order to make 2 tricks in the suit. You can afford to give one up if you still have entries and the clubs come in.

[smerriman]

This could be an issue with limited simulations as Olegru says.

I suspect it may however be due to the double-dummy nature of GIB. It probably can't see anything wrong with leading a low diamond - if the queen is onside, you can return with a spade and finesse again; while if the queen is offside, you can take the club finesse. And maybe a low diamond helps with rarer cases (eg a singleton King onside?)

The problem is, after the first finesse holds, it's no longer playing double dummy, and doesn't know where the queen is anymore. Now that it's messed up entries has to change tack.

GIB is meant to have some logic built in which avoids the issue of always putting decisions off because it will 'know' the right thing to do later, but maybe it didn't hold up well here.

[johnu]

zhasbeen, on 2018-February-02, 11:29, said:

"I used to dread having GIB declare a hand..."

I have a friend who said the same thing but am wondering if this isn't one area of GIB's game that has improved in the last year or so. I played my first robot tournament in August.

BBO has said many times that the play engine is so complex and the original programmer is no longer associated with the company that they basically don't do any major fixes to fix play problems.

For something very delineated and (probably?) simple like leading an ace against 7NT, IIRC that BBO implemented a change (or was it leading AK against 6NT?).

[johnu]

steve2005, on 2018-February-02, 09:26, said:

Does nobody read? take the diamond finesse twice. if it works you can take club finesse for 7 but you can just duck a club for 4 club tricks and make.
3♠ 2♥ 3♦ 4♣ = 12 tricks

If you finesse club first you need both finesses to make 12 as you still need 2 ♦ tricks even with 5 ♣ tricks.

I can read just fine Think about what I said. You need 1 of 2 finesses to make 6. I don't know why you would think about ducking a club if the diamond finesse wins. Why give up an overtrick if the finesse works?

[zhasbeen]

johnu, on 2018-February-02, 22:15, said:

BBO has said many times that the play engine is so complex and the original programmer is no longer associated with the company that they basically don't do any major fixes to fix play problems.

For something very delineated and (probably?) simple like leading an ace against 7NT, IIRC that BBO implemented a change (or was it leading AK against 6NT?).

I was talking about declarer play only. However, my assessment of GIB's declarer play could be distorted since I only see it while defending, which is about 25% of the time. I've been end played or squeezed many a time--I play only the tournaments where the human declares all the hands. I don't know if you agree on this, but it seems that a high percentage are set ups when GIB plays the hand, such as when 2 passed hands bid up to game and a magic dummy comes down. Typically there are 4 trumps or more, a long running side suit, and/or void. You usually have an impossible lead. Actually, it seems that all hands are set ups when I think about it. Have you ever noticed how often that there are 2 or 3 small in your hand and a suit headed by KJ in the dummy, or vice versa? It seems like more than half the time.

I see GIB making a lot more errors on defense. On that topic I'll never understand why it would be programmed to lead a stiff trump so often, which is often the K or Q. Another common situation is when it gets off to killing lead or switch but doesn't continue. To it's credit, however, it lays down quite a few slick defenses and all the false cards make it more difficult to keep track of the spot cards, such as discarding the 7 first, then the 8, and saving the 6. I'm sure you've noticed that when it comes down to 2 key spot cards, with you having 1 and defense the other, GIB usually has you by one, e.g. your 5 vs his 6.

[Stephen Tu]

smerriman, on 2018-February-02, 16:16, said:

I suspect it may however be due to the double-dummy nature of GIB. It probably can't see anything wrong with leading a low diamond - if the queen is onside, you can return with a spade and finesse again; while if the queen is offside, you can take the club finesse. And maybe a low diamond helps with rarer cases (eg a singleton King onside?)

The problem is, after the first finesse holds, it's no longer playing double dummy, and doesn't know where the queen is anymore. Now that it's messed up entries has to change tack.

I'm pretty sure it's this issue on this particular problem. GIB has two declarer play engines, the double-dummy Monte Carlo approach (deal randomly that matches bidding/play so far, randomly pick among ties for double dummy best card), and a true single-dummy engine "GibSon" that kicks in after a few tricks on the advanced GIBs (but is disabled on the basic GIBs to minimize resource consumption). The single-dummy engine plays for particular layouts and card locations, it does real "planning" that avoids a lot of these double dummy reasoning errors (card plays that are equivalent/free, not really errors, if you are going to guess everything right subsequently, but aren't as good as some other play when the possibility of guessing wrong is taken into account).

Here it's making the decision at trick two before Gibson has cut in.

On my PC version of GIB, with the normal startup flags, it does the same error at trick two not leading DQ so it can repeat finesse. But it recovers later crossing to SQ and taking club finesse, and maintaining entries to take diamond finesse the 2nd time. If the "par/Bernasconi" mode is turned on, which kicks in Gibson right away I think, it runs the DQ.

[Stephen Tu]

zhasbeen, on 2018-February-03, 09:37, said:

I see GIB making a lot more errors on defense. On that topic I'll never understand why it would be programmed to lead a stiff trump so often, which is often the K or Q.

It's not difficult to understand when you know how it thinks. Remember it is assuming double dummy declaring and defense. So it is assuming you will always drop the stiff K/Q offside or whatever. It is assuming you will always pick it up, so from its point of view it is not giving up anything by leading it, and leads it if it thinks it doesn't want to break any other suit because those can give up tricks by force.

That's one of the perils of relying on double dummy reasoning. To fix it, they'd either have to do something on the lines of
1) Special programming to make it not want to lead stiff honors (but pretty hard to figure out exactly how much to weigh against this without banning it completely since sometimes you actually do want to lead stiff honors)
2) calculate what the best card is assuming the opponent proceeds *single dummy* (takes way too much resources at the moment to simulate a single dummy opp over a useful number of hands)

Quote

Another common situation is when it gets off to killing lead or switch but doesn't continue.

That's because it doesn't have any concept of reading defensive signalling. It is programmed to give signals in some particular cases to help humans, but doesn't understand how to read a signal given by its partner to continue.

[zhasbeen]

Stephen Tu, on 2018-February-03, 13:02, said:

It's not difficult to understand when you know how it thinks. Remember it is assuming double dummy declaring and defense. So it is assuming you will always drop the stiff K/Q offside or whatever. It is assuming you will always pick it up, so from its point of view it is not giving up anything by leading it, and leads it if it thinks it doesn't want to break any other suit because those can give up tricks by force.

That's one of the perils of relying on double dummy reasoning. To fix it, they'd either have to do something on the lines of
1) Special programming to make it not want to lead stiff honors (but pretty hard to figure out exactly how much to weigh against this without banning it completely since sometimes you actually do want to lead stiff honors)
2) calculate what the best card is assuming the opponent proceeds *single dummy* (takes way too much resources at the moment to simulate a single dummy opp over a useful number of hands)


That's because it doesn't have any concept of reading defensive signalling. It is programmed to give signals in some particular cases to help humans, but doesn't understand how to read a signal given by its partner to continue.

Hi Stephen-good to see you are still coming around. I saw a couple of your Monterey results--nice