[eagles123]

it seems so often GIB finds the killing lead, I encourage and for no reason at all GIB switches. This is much more common against NT than suit contracts, sorry I have no specific board to show.
others found the same?

[ArtK78]

I would rephrase your comment. It seems so often GIB finds the killing lead against me. However, for whatever reason, GIB often finds a reason to switch and let me sneak through with my contract anyway.

When GIB is my partner and is on lead, the chances of GIB leading a (or the) suit that I bid is less than 50-50. And the chances that GIB is right not to lead my suit is much less than 50-50.

[msmaple]

I sometimes play Bingo which in itself is a challenge trying to get the GIB to bid a part score, especially in a major, however, what I find most interesting is the percentage of finesses that are off side. Are the hands pre-dealt, I cannot imagine that what i see as at least 95% of offside finesses being
the result of random deals? Any explanation?

[Bbradley62]

msmaple, on 2015-June-13, 08:34, said:

... what I find most interesting is the percentage of finesses that are off side. Are the hands pre-dealt, I cannot imagine that what i see as at least 95% of offside finesses being the result of random deals? Any explanation?

There have been several threads about this, some with very detailed analysis of data. If you think that anywhere near 95% of finesses are offside, you have a very selective memory.

[msmaple]

Bbradley62, on 2015-June-13, 10:32, said:

There have been several threads about this, some with very detailed analysis of data. If you think that anywhere near 95% of finesses are offside, you have a very selective memory.

I am specifically referring to Bingo games played, I have not kept track of the actual percentage of offside finesses (in fact I think 95% is a low estimate) but I will in future

[ArtK78]

It is my experience that approximately 50% of the finesses are offside. And this is a distressingly large number. I would much prefer that 80-90% of my finesses worked.

[1eyedjack]

ArtK78, on 2015-June-13, 12:37, said:

I would much prefer that 80-90% of my finesses worked ...

... and only at my table

[barmar]

eagles123, on 2015-June-11, 11:38, said:

it seems so often GIB finds the killing lead, I encourage and for no reason at all GIB switches. This is much more common against NT than suit contracts, sorry I have no specific board to show.
others found the same?

GIB is not able to interpret signals, so it doesn't care if you encourage or not.

[1eyedjack]

barmar, on 2015-June-17, 12:51, said:

GIB is not able to interpret signals, so it doesn't care if you encourage or not.

Not (always) strictly true I think:

Quote

When it's following to partner's opening lead, it will give an attitude signal:

• High spot card with an Ace or King
• High spot card with a Queen behind dummy's Ace or King
• Low in any other situation

from

https://www.bridgeba...ystem_notes.php

[Bbradley62]

I didn't look for the context of that quoted passage, but it seems to indicate that GIB doesn't high-low holding a doubleton when CHO leads AK.

[barmar]

1eyedjack, on 2015-June-17, 13:30, said:

Not (always) strictly true I think:

That passage explains that GIB gives signals. But it doesn't know how to interpret the signals that its partner sends.

I think it does know how to recognize standard honor leads, so if you lead from QJx, which is covered by the king and won by the robot's ace, it knows it's safe to return the suit.

[1eyedjack]

barmar, on 2015-June-18, 10:18, said:

That passage explains that GIB gives signals. But it doesn't know how to interpret the signals that its partner sends.

I have to take your word for this, but I find it very strange.
Each GIB is identical to each other GIB (of same version etc etc).
What you are saying is that when GIB is partnering GIB, one GIB will make a signal but the other GIB is in ignorance of the plays that its partner might make. And yet that same GIB on lead could just as easily have been on the other side of the table, with exactly the same "brain", but because it happens to be on the side of the table making the lead it must act in ignorance of what it would have done in the other seat.

I find this hard to swallow.

[Stephen Tu]

1eyedjack, on 2015-June-18, 11:11, said:

I find this hard to swallow.

What's hard to swallow? To teach the computer signalling, you need to both do
1. program it to echo from certain holdings
2. on subsequent tricks, bias its deal algorithm appropriately based on partner's signal, basically throwing out the right deals "partner wouldn't have signaled high from that".

Computers are really stupid. Teaching it the first thing doesn't mean it can figure out itself how to do the second. A human has to teach it how to do both things. The second part is apparently a lot harder to do than just the first. So no one got around to it.

Eventually AI probably advances enough that computers can figure stuff out by themselves given a little knowledge. But for now it is basically science fiction, you have to lead them by the nose for every single logical step.

[1eyedjack]

I always try to keep an open mind, but really for now I do not accept this.

GIB is on lead opposite partner GIB and leads (say) Spade "x" from rubbish. Dummy plays low from Kxx and partner (GIB) contributes (1) Q or (2) J, in either case won by declarer's Ace.

Opening leader GIB is later put back on lead and has to decide on its defence, the options depending on whether partner-GIB has the missing Q or J or Spades.

It is my understanding, and I am open to persuasion otherwise but would be surprised, that if partner-GIB played the Q at trick 1 then opening leader-GIB would place declarer with the J and plan its defence accordingly. If partner played J then the location of Q is up for grabs.

In other words, after seeing the Q appear, GIB in its later simulations eliminates possible holdings in which GIB-partner holds the J.

How is this any different from its eliminating the possibility of partner holding the King when GIB-partner plays the 2 on lead of Ace?

If there is one thing that computers are *good* at it is counting cards and eliminating impossible holdings. "Throwing out deals", in your words, is its bread and butter.

Now, had you said that GIB attitude signals are based on an evaluation of the entire hand rather than based on a specific holding in the signal suit, I would understand that further extrapolation becomes complex. But that is not my understanding of the crude signalling that is currently available in GIB.

[Stephen Tu]

1eyedjack, on 2015-June-18, 17:37, said:

It is my understanding, and I am open to persuasion otherwise but would be surprised, that if partner-GIB played the Q at trick 1 then opening leader-GIB would place declarer with the J and plan its defence accordingly. If partner played J then the location of Q is up for grabs.

I really don't know whether GIB can do this to begin with.

Quote

How is this any different from its eliminating the possibility of partner holding the King when GIB-partner plays the 2 on lead of Ace?

Even if it could do the first thing, why would it be able to figure out how to do the second, without careful instruction from a human on how to do this? Humans are good at inferring from previous experience how to think in analogous but not 100% identical situations. Computers are not, they basically have to be taught every single situation separately.

Quote

If there is one thing that computers are *good* at it is counting cards and eliminating impossible holdings. "Throwing out deals", in your words, is its bread and butter.

Computers aren't really inherently good at anything other than making lots of fast number calculations then deciding which subsequent instructions to execute depending on comparing the results of such calculations. They "know" basically nothing, knowledge is solely in the program itself, which is the product of human minds. If the human didn't specifically put instructions in on how to deal with something, the computer isn't going to figure it out and come out with the "similar but not exactly the same" code to handle it. The code either knows how to do something or it doesn't. It doesn't adapt, doesn't extrapolate, doesn't learn. Maybe decades in the future AI advances to the point where computer programs can adapt & self-program, but this stuff is really in its infancy research stages and is mostly still in the realm of science fiction.

It only knows how to throw out deals if a human teaches it very carefully how to throw out deals. And if the human tells it how to throw out deals in one situation, it can't work out how to extrapolate that to other situations. AFAIK it may only know how to bias the deals by the bidding database and actual cards played, and not know how to bias the deals based on signals and inferences like "lowest from equals by third hand".

Your surprise to me is mainly a factor of you not being a computer programmer, not realizing how incredibly stupid machines they are at the core. Programs do amazing things these days, but that's only because of tons of effort by humans to tell them exactly what to do, building on many millions of lines of code written by earlier generations of human programmers. But they still don't think like a human at all, all this effort is basically to create the *illusion* of human thought. Computer code basically boils down to add these numbers, subtract these others, compare, if a>b do this next, if a<=b do something else. They know nothing of suits, finesses, anything of bridge, it's all a bunch of numbers and calculations to do in a particular order exactly as some human programmer wrote down for it.

Computers are stupid. Don't expect too much out of them.

[iandayre]

To Stephen Tu or anyone else knowledgeable on the subject:

Does anyone have any experience with other bridge-playing programs currently on the market? I know that there is an annual computer championship and that GIB participated in the early years, but not since. I am very curious to know if GIB is comparable to the best programs currently available. Is so, our expectations of GIB obviously have to be toned down. But if not...

[Stephen Tu]

I haven't spent a lot of time playing against competitors. Haven't played a lot against GIB lately either. Wbridge5 is available for free download, you can also play against it at bridgez.net . Why don't you try it out yourself and give a report on how strong you think it is in comparison?

[barmar]

1eyedjack, on 2015-June-18, 17:37, said:

I always try to keep an open mind, but really for now I do not accept this.

GIB is on lead opposite partner GIB and leads (say) Spade "x" from rubbish. Dummy plays low from Kxx and partner (GIB) contributes (1) Q or (2) J, in either case won by declarer's Ace.

Opening leader GIB is later put back on lead and has to decide on its defence, the options depending on whether partner-GIB has the missing Q or J or Spades.

It is my understanding, and I am open to persuasion otherwise but would be surprised, that if partner-GIB played the Q at trick 1 then opening leader-GIB would place declarer with the J and plan its defence accordingly. If partner played J then the location of Q is up for grabs.

There's a difference between encouragement signals and standard plays from sequences.

GIB's defense is almost entirely based on simulations. It deals a bunch of hands for partner and declarer consistent with what it knows from the auction and play so far, determines the double dummy result of each possible play, and chooses the one with the best average result.

GIB knows that third hand will play the lowest of equals. So if it follows with the Q it has denied the J, and subsequent simulations will not include hands where partner has the J.

But attitude signals are harder for it to deal with. The hands consistent with a positive/negative attitude are harder to characterize when constraining the simulated hands.

With enough computer power, it could be done by dealing hands for partner, and for each of them figuring out what that hand would have done on the previous tricks, and whether that's consistent with the plays partner has made so far. But this results in a combinatorial explosion of simulations. When Ginsberg was first developing GIB he experimented with this type of "what if" simulation, and it didn't produce enough benefit and slowed the robot down too much.

So GIB is limited to very simple inferences when it can place specific cards somewhere, but not anything sophisticated.

[GreenMan]

Stephen Tu, on 2015-June-18, 18:51, said:

Programs do amazing things these days, but that's only because of tons of effort by humans to tell them exactly what to do, building on many millions of lines of code written by earlier generations of human programmers. But they still don't think like a human at all, all this effort is basically to create the *illusion* of human thought.

I think this illustrates the situation reasonably well:

[1eyedjack]

Not a good analogy I think. Unless the cartoon was written 5 years ago, in which case the reference to 5 years may be valid. These days if you take a picture of a bird, you expect software to identify it as a bird. Maybe 5 years ago you might have had a problem.

It is a theme really - most of what we do now with computers was 5 years ago considered to be beyond imagination.