[j23185]

What is the meaning of those little green numbers that appear on a hand diagram after buttoning "GIB" when replaying a game?

[pescetom]

j23185, on 2024-April-03, 11:16, said:

What is the meaning of those little green numbers that appear on a hand diagram after buttoning "GIB" when replaying a game?

The (confusingly named) "GIB" button shows the double dummy result corresponding to the play of each card. IIRC, a green number is overtricks and a red number is undertricks.

[mycroft]

And specifically, a green equal sign "=" means contract makes exactly.

Know that "double dummy" means "assuming after this card played, everyone makes the exact best play *seeing everybody's hand*." That means that frequently, double-dummy results are not possible in real life - because nobody would finesse the 7 at trick 2, or find the underlead of the AKJ to get the ruff, or play for the double squeeze that is 10% when there's a 50% finesse available (or that loses to a ruff on 60% of the other possible layouts), or...

[thepossum]

Correct me if I am wrong but I think DD is only seeing partner's hand
It's not that unrealistic - so the defence are making best defence like declarer is trying to make best play
That is why it requires simulations rather than some complex attempt at trying to find the best play - which may well be computationally impossible or infeasible in realistic time
I will go even further never having studied all the code of one, or tried writing one - it is possible that what the expert player who provided the intelligence on how to play to the simulation of a declarer or defence play is not optimal - better than what most of us would do I imagine
But its a form of model where it averages out a random sample of hands from the information available at that stage - not realistic at all
From what I remember of chess experts have almost every game inn their heads - so maybe expert-level bridge play is not that bad when used to inform said model
I am assuming most just use fairly simple averaging type model - I imagine there are other statistics/parameters/estimators ( will lose my licence lol) that could be used - mean, median, something else - I have never looked at enough code to find out - median makes more sense than mean to me

I did say correct me if I am wrong lol - but I think I am confident that it would be computationally impossible/infeasible to try to optimise the play of all four hands - what would an expert play here makes more sense to me - and maybe it does need to see all the hands - but its a model -after making a decision and seeing what happens next maybe there is a limited rollback. Oh that didn't work. But to do a full optimisation I don't think so. You can tell how much I care about the details. It's a strange idea really how they work. Just magic in a box

[thepossum]

PS - can I add I think people are mixing up two different things - one is when you see the result of a sim based on best play on each individual hand - who knows exactly how they do it but of course either knowing whether the finesses are offside or not is a moot point with limited rollback on a finesse lol

PPS I reckon they could put any number they liked in those little boxes

[mycroft]

You are wrong.

Double Dummy Analysis of a hand looks at all 52 cards (or the tricks taken and all the cards remaining in the 4 hands), and determines the result that will happen if all 4 players, seeing all the cards, make the correct play.

It comes from a game where two players will hold a bridge hand, but both their partners are face up on the table - therefore "double dummy". Note that knowledge of 26 cards face up, plus the 13 in your hand, means you know the 13 remaining - so it's equivalent to playing with all cards face up.

In the case asked about by the OP, we can add to that explanation "...if this card is played now". Giving you the (useful?) information that "this hand makes unless the defence switches to a small diamond RIGHT NOW" or "if they play correctly [guess which squeeze is operating, or the finesse, or the endplay, or...], there's nothing you can do here". You can watch during the replay as different choices from different players that differ (or not) from DD best change the potential outcome of the hand. Quite interesting, actually.

What you're talking about is "Single Dummy play using Double Dummy Simulations". This is a tool used by computer programs (who can do this quickly) to determine their best call or play in a simulation where there "are no rules" set.

• First, it takes its hand, the auction (and meaning thereof) so far, and the cards shown so far in play, and treats this as "known information".(*)
  
• Then it generates however many hands it has time for that fit the "known information".
  
• Then it does a double dummy analysis on each hand, determining the results of all contracts (in the auction) or this contract (in the play) for each potential action at this point. (this is the Simulation.)
  
• From those results, it works out the Expected Value of each option, and takes the option with the highest EV for its "single dummy" play (which may, or may not, work well on the specific lie of the cards in the hand. It's a "percentage action").

It's actually a good way to emulate "thinking" that bridge players do - sometimes "theory percentage play", or "placing cards"; sometimes "innate experience from thousands of times hitting this problem".

Spoiler

The difference between "double dummy play" and "single dummy best percentage play" is so strong (in both style and frequency) that many of the "self-kibitzing" malefactors were caught (sometimes by their partner!) because their play "looked wrong/different, but worked". And in fact, one of the metrics for (at least) one of the anti-cheating tools includes "loss against DD on opening lead" and "loss against DD play in general" - and too close, not too "bad", is the flag. IIRC, world-class experts hover around 83% on OL - about one in five hands they pitch a trick on the opening lead!

GIB using DD Sims is a good thing to know about, if you play with/against robots. But it's NOT what the OP is talking about (looking at those little green and red numbers in play history, or as kibitzer waiting for declarer to just make up his mind again).

(*) This is why a great strategy for robot games is to lie - in specific ways, at least. Bidding a "15-17 Balanced" 1NT with 13-18, maybe a small singleton club, if you win the auction means that the hands simulated will always put that Q in your hand that you "have to" have that is partner's setting trick. Sure, partner will also believe you (which is why *some* lies are immediately and irretrievably fatal) but again, there are two opponents to one partner. It's even better when you're playing a non-GIB system against a pair of fill-in bots. It's also why Bridge, as opposed to robot games, treats violations of declared agreements (whether innocent or deliberate, gross or fine, explicit vs implicit,...) as such an issue.

[mycroft]

Back In The Day, DD simulation was hard, and cracking the problem (and getting sufficient computer resources to do such a trivial thing) was a Big Thing. I remember the first DD simulator that would work on "my computer" - and because I knew the guy and knew how to test, I got to play with it For Free!

But Back In The Day, there were about 1000 people on OKBridge, and 40-bit encryption was still considered (by all but the cryptography community) to be sufficient for bank transactions.

Now, it's okay to get Bridgemate Control Software to *redo* all the DD simulations on all the hand records because it's easier than either clicking the checkbox off or moving your mouse to the "No". Not faster, but not by that much.

[pescetom]

mycroft, on 2024-April-05, 11:49, said:

Back In The Day, DD simulation was hard, and cracking the problem (and getting sufficient computer resources to do such a trivial thing) was a Big Thing. I remember the first DD simulator that would work on "my computer" - and because I knew the guy and knew how to test, I got to play with it For Free!

Was it really?
I was pretty good at programming back then... had yet to encounter bridge, but would not have expected DD to be difficult in terms of analysis ("but you may fade, recursion will always come true") or impossible in terms of (then very limited) resources. Understanding resource limitations and being able to conceive and implement optimizations would have been fundamental, of course. Plus maybe an honest warning if resources were not sufficient for the situation.

[smerriman]

pescetom, on 2024-April-06, 14:52, said:

Was it really?
I was pretty good at programming back then... had yet to encounter bridge, but would not have expected DD to be difficult in terms of analysis ("but you may fade, recursion will always come true") or impossible in terms of (then very limited) resources. Understanding resource limitations and being able to conceive and implement optimizations would have been fundamental, of course. Plus maybe an honest warning if resources were not sufficient for the situation.

When Ginsberg wrote his paper on Partition Search, a fundamental and extremely non-trivial part of modern DD solvers, his tables only compared the performance of positions with 12-48 cards remaining with standard algorithms, because:

Quote

hands generated using a full deck of 52 cards were not considered because the conventional method was in general incapable of solving them

[pescetom]

smerriman, on 2024-April-06, 15:45, said:

When Ginsberg wrote his paper on Partition Search, a fundamental and extremely non-trivial part of modern DD solvers, his tables only compared the performance of positions with 12-48 cards remaining with standard algorithms, because:

I can imagine that with all cards remaining, the resource problem is immense, fair enough - see my last comment about warning.
I also humbly concede that their may be some non-trivial part of the analysis
Wish I had been invited at the time.