GIB's decision

[PetteriLem]

6 clubs/ 2S - 4C - pass - 6C; all pass.

A

92

AJ32

AKT432

 

J432

AQJT43

-

J65

 

The lead was K of diamonds, GIB ruffs and plays a club to the ace.

GIB ruffs again and plays a spade to the ace. Removes trumps and has to give

a trick to the east. A heart comes back and T wins! GIB ruffs a spade. A heart to the ace and the contract goes down.

 

Let's assume that GIB takes a finesse early, which sets the contract if it fails, but it makes this time. Later GIB discovers a line that makes the contract more often than 50%, but requires the same finesse to fail. How does GIB handle situations like this?

What if the early finesse did not yet set the contract, but a hold up was a possibility?

[xcurt]

To my knowledge GIB never solved this problem.

 

I will simplify your example. You are playing a small slam with Kxxx vs Qxxx in trumps and solid outside. How to you play? Of course you pick someone and play them for Ax. GIB plays low to one honor and then low to the other, losing to all combinations.

 

Why does this happen? It's an artifact of how GIB attacks the problem of figuring out which card to play. As far as I know Matt Ginsberg tried to solve this (Google on 'gibson' 'single dummy') a while back but the computational complexity grows much too quickly to work on 13-card play problems. I believe he got up to something like 7-card endings, with each card adding 100x to the computing time needed. With today's computers and threading the solver (I don't know how well gibson could be threaded but GIB could be threaded easily enough) that means doing 8 cards endings today. You're still a factor of 10^10 away from doing 13 card endings though.

 

In your example once GIB discovers that east has certain cards, the sample may start to include a majority of hands where west has HKx. GIB lacks the inferential rule that says "the opponents almost certainly didn't duck the setting trick against a slam," so it tries to drop out the HK.

[hotShot]

GIB simulates deals/partial deals that fit the knowledge it has. This knowledge is basically restrictions from bidding and played cards.

 

Now lets say it simulates 10 partial deals (including sidesuits) and it finds the card that produces the most tricks in most of the deals.

 

The next time GIB is to play a card it generates 10 new partial deals (with less cards) and finds the card that that produces the most tricks in most of the new deals.

 

These 2 steps do not have to go to the same (right) direction.

[awm]

One of GIB's big weaknesses is ideas of the form:

 

"If the layout were such-and-such, then the opponents choice of play to a previous trick makes no sense. Thus I will assume that the layout is not such-and-such."

 

In fact this kind of reasoning is often one of the things that separates decent bridge players from really good ones. The problem is that GIB simply generates hands consistent with the play thus far and tries to take the percentage action based on those hands. It ignores the fact that some of those hands would imply that opponents made a ridiculous lead or play earlier in the defense.

 

Of course, this kind of reasoning by good players often leads to the so-called "Grosvenor Coup" where a defender in fact does make a ridiculous lead or play that potentially gives away the contract, but the declarer proceeds to go down anyway because he assumes a layout that does not require the ridiculous defensive play!

[Stephen Tu]

I think he eventually got Gibson to kick in earlier, maybe 10 or 11 tricks to go.

 

I believe the problem with this hand is that GIB thinks that KQxxxx spades and out is an "impossible" weak 2 bid, being too weak for its standards. East has shown up with everything, so West "has to" have the HK to have his bid, so it plays for the defense to have pulled the wrong card. It just doesn't think it can make it, and winning the hook the first time doesn't change its mind.

[CarlRitner]

There is a thread that indicates 11 and 12 tricks to go were achieved,

at least in alpha:

 

GIB vs GIBson on RGB 6-12-2000

 

 

It may not kick in that early in the commercial version due to speed.

[xcurt]

There is a thread that indicates 11 and 12 tricks to go were achieved,

at least in alpha:

 

GIB vs GIBson on RGB 6-12-2000

 

 

It may not kick in that early in the commercial version due to speed.

GIB is a fantastic achievement and Matt gets full credit for showing a new way to play computer bridge, but infortunately I'm very skeptical of Matt's claims in this area given how many times he overpromised and underdelivered.

[Stephen Tu]

GIB's declarer play is pretty damn good if it understands the auction, I'd pit it vs. all but the very best humans. Unfortunately the bidding database is full of holes so it gets rather confused at times in both bidding & play because of those gaps. And some years ago Ginsberg abandoned fixing the holes and any other GIB improvements for his "day job". Maybe someday he gets back to it.

 

Also the format of the database is rather obtuse so it's rather difficult even for some other seasoned programmer to decipher and fix.

[CarlRitner]

Also the format of the database is rather obtuse so it's rather difficult even for some other seasoned programmer to decipher and fix.

This is probably 50% due to the need to compact the database size, and 50% to hide the source. Although I was given the "source" and it's every bit as mystical.

 

One would have to disassemble the evaluation routines that this "source code" invokes to reverse-engineer the original source code, or rule set. Although Matt did not want to see GIB as a rule-based program, if you include the simulations and subroutines as tools to the rules, that's essentially what it (and all other commercial bridge programs) are.

 

GIB continually improved for a while through human intervention. To my knowledge it never "learned" to play bridge. Had Matt continued full time on the project, he would have eventually encountered "the wall". That said, it still was the breakthrough in techniques Matt engineered that led to the very (or at least more) successful programs of today.