[pigpenz]

---

please post other hands, the thread speaks for itself
comments dont do these type of hands justice

[dkham]

I wonder what the logic was that lead to leading queen of trumps, maybe it thought AKQ were equals?

[wynsten]

dkham, on 2012-May-01, 09:38, said:

I wonder what the logic was that lead to leading queen of trumps, maybe it thought AKQ were equals?

It doesn't "think". Computing strong single-dummy play is not as easy as it seems.

[Stephen Tu]

wynsten, on 2012-May-01, 10:47, said:

It doesn't "think". Computing strong single-dummy play is not as easy as it seems.

At least not for the first few tricks. After gibson (true single dummy reasoning engine, instead of monte carlo DD sim) kicks in, around trick 4 or 5 or so on the advanced bots, GIB is actually quite a bit better than most humans. Given enough thinking time it should be able to do better on this one though. The basic robots are just set to ultra-fast though, so they make more blunders than usual. Basic bots are kind of bad at the time settings BBO is using currently. And the advanced bots are more prone to making blunders the first few tricks when it is not doing real single-dummy analysis.

My home GIB doesn't blunder this hand. Perhaps it is dealing so few samples that it didn't have any that the trumps didn't split ??

[pigpenz]

Stephen Tu, on 2012-May-01, 12:26, said:

At least not for the first few tricks. After gibson (true single dummy reasoning engine, instead of monte carlo DD sim) kicks in, around trick 4 or 5 or so on the advanced bots, GIB is actually quite a bit better than most humans. Given enough thinking time it should be able to do better on this one though. The basic robots are just set to ultra-fast though, so they make more blunders than usual. Basic bots are kind of bad at the time settings BBO is using currently. And the advanced bots are more prone to making blunders the first few tricks when it is not doing real single-dummy analysis.

My home GIB doesn't blunder this hand. Perhaps it is dealing so few samples that it didn't have any that the trumps didn't split ??

when using your home GIB
are you just loading it into the apps as lin or pbn or
are you using the bridge.exe command line?

[wynsten]

Stephen: Does the single dummy logic care about the contract, or about the scoring method (IMPs vs MPs)? None of this enters into double dummy calculations, but can be critical for single dummy.

[Stephen Tu]

I use the GUI on my home GIB and just hand enter the deal. Or use the command line and hand enter.

The logic does care about scoring method, there are flag for IMP and MP.

[pigpenz]

Stephen Tu, on 2012-May-01, 19:09, said:

I use the GUI on my home GIB and just hand enter the deal. Or use the command line and hand enter.

The logic does care about scoring method, there are flag for IMP and MP.

on didnt realize on the gui there was an option for imps or mp
i knew on the bridge.exe command line there was
but on this hand it looks like GIB is playing against like 8% that trumps are 4-0

[Bbradley62]

pigpenz, on 2012-May-01, 19:44, said:

but on this hand it looks like GIB is playing against like 8% that trumps are 4-0

No; the way to defend against the possibility of 4-0 trumps is to cash the K first so you can lead to the 10 if East has no trumps. You cannot pick up the suit if East has all four.

[pigpenz]

Bbradley62, on 2012-May-01, 21:37, said:

No; the way to defend against the possibility of 4-0 trumps is to cash the K first so you can lead to the 10 if East has no trumps. You cannot pick up the suit if East has all four.

yes
I am only assuming it had some reason for the play it made
what it is i dont know

[wynsten]

The reason is that in all its simulations the Queen was good enough to maximize double dummy results (so was low to the King, of course - it is always good enough). The only time that Queen is worse than King is when East holds tripleton Jack, (which as luck would have it, is how the cards lay). Of course if West has three to the Jack, the Queen must be overtaken (double dummy, remember), and the Jack taken by finesse. But by the time it comes to make that play, he is no longer playing double dummy, so it would really never happen.

This suggests an improvement to the algorithm that approximates single dummy play: Prefer the line that has fewest forks. Leading out the King, Ace, and (if necessary) the Queen has no forks. Leading Queen requires overtaking against some distributions, but playing low against others. As humans, we always prefer the branchless trees. Maybe the robots should too.

[Oops - the Queen is also worse when West has all four, see Bbradley62 above, so presumably that distribution was not simulated either]

[Stephen Tu]

It has no idea what a "line with fewest fork" is. In double dummy mode, early tricks, it's not thinking in terms of lines yet; it's just picking the best card, *right now*, not the *best line*. The hope is most of the time picking the best card of the moment will stumble onto the best line. This fails when
- sample space has too few hands to consider the relevant possibilities (lack of thinking time, other parameters that control how many deals considered)
- sample space inaccurate due to bidding database bugs giving it a bad picture of opps probable holdings
- inability to foresee possibility of misguessing a position single-dummy, thus randomly failing to find "lines" that avoid this guess

It would be far simpler just to give it enough time to deal enough samples to find hands where it has to play low to the K like a normal person.

Only after the single dummy engine kicks in on later tricks on the advanced bots, or downloaded versions, does it start thinking in terms of lines, and thinking like "I'm making if west has 1-3 diamonds, or 4 diamonds with the J", and playing appropriately.

[wynsten]

OK - well I have no idea how the engines work - just making some uneducated guesses. Time to push on to a different topic. Thanks Stephen.

[ahydra]

What about telling it to make 100% plays? So here playing to the K is 100% whereas playing the Q first is only 28%. If they're 2-2 there's no difference; 3-1 or 4-0 then you need to unblock; so unblock.

I guess how do you define a 100% play. What might be a 100% play in one suit could be wrong for the hand. But here it's not: diamond to K followed by diamond to (10 or Q depending on whether E showed out) followed by drawing trumps if necessary followed by cashing all the clubs is a 100% line. It's these kinds of lines that should never be deviated from because everything else has a chance to go wrong.

Oh, and tell GiB holding AK in a suit to (nearly) always lead the ace or king from that suit. This would have set this hand, not to mention that 6NT from a while back

ahydra

[Antrax]

Playing the Q is a 100% play - in all the simulations GIB did, playing the Q won.

[ahydra]

Antrax, on 2012-May-03, 04:35, said:

Playing the Q is a 100% play - in all the simulations GIB did, playing the Q won.

I'm talking about a new algorithm, not just simulations. For example

AQ

xx

Consider all the possible combinations of EW's cards in this suit. Playing small to the Q gains 50% of the time whereas playing the A only gains a very small percentage of the time and any other line doesn't gain at all. Thus, all other things being equal, I will play small to the Q.

Writing an algorithm to figure these kinds of plays out should be trivial. It's using it that's the hard bit, since you need to consider this play in the context of the whole hand (so need to consider (nearly) all 26C13 combinations of EW cards). But some simple triggers could at least be put in, e.g. if I need 2 tricks to break the (NT) contract and I have two winners then cashing them will defeat the contract, so that's what I will do.

From a programmer's point of view, I find this really quite interesting, and is an excellent example of how things that are obvious to humans are sometimes hard to code into computers.

ahydra

[Stephen Tu]

I love how you guys who have obviously very little clue how GIB works, some of whom have no idea what altering a program would involve (how many of you have CS degrees and/or software dev work experience?), just toss out ideas to "oh just tack on a new algorithm for certain situations" like this is a totally feasible thing to do. Why don't you go out and write a bridge program then?

You are asking GIB to throw out its whole approach to the game, think about how to play cards in a totally different way. A long time ago people wrote bridge programs that tried to work using human-like thinking techniques. They were horrible. GIB came out, and despite the limitations of Monte Carlo DD approach, it pretty much obliterated them, and programmers switched to brute force statistical techniques.

Throwing out suggestions like "oh GIB should think this way" are totally useless, what BBO developer going to rewrite whole bridge engine? You do it if you think so easy. How is it going to recognize when to switch between Monte Carlo and your "find the 100% play" algorithm? When to use which? Don't you realize that it basically has to deal the losing case(s) for the Monte Carlo approach to recognize this, and then the Monte carlo algorithm itself would handle it?

The only things I think are feasible (until they can get Ginsberg back on project, or Ginsberg-level AI programmer on board), are
- increase thinking time of basic bots as the power of their server farm increases
- update download client more frequently (so people can use thinking time on their home machine instead of GIB servers); download client (on slow setting) plays better, but being older has more uncorrected bidding bugs. Maybe they could release a small patch that just overwrites the bidding DB in the BBO directory? I just really don't want to pay $1 a day to use GIB at its best. I only payed $80 for my home GIB, and that was a one time payment! There's probably some happy medium between the two that is reasonable cost to the user and supports continued bidding fixes.
- if things get fast enough try to cut in the superior Gibson single-dummy algorithm trick or two earlier

[Antrax]

a) You're right, but why so angry?
b) I think BBO could invest in distributing GIB simulations to people using the BBO server. It'll be difficult to write, but there's a potential to allow GIB a lot more thinking power without investing in new servers.

[Stephen Tu]

Not really angry. Just trying to stop people from posting their silly suggestions about GIB's card play, get them to understand a bit better why GIB makes play blunders and how it's not easy to fix them, and not going to happen anytime soon. That they pretty much have to shrug their shoulders at plays like these, and concentrate on reporting bugs that *can* be fixed, like silly point ranges and distributions in the bidding database.

I think upgrading the download client is the fastest way to get GIB thinking time. It's just frustrating playing with bidding bugs that are years out of date, I guess it depends on which you find more frustrating, GIB's bidding or its play.

Even as it is, I find GIB less frustrating on average than random humans in the MBC; I gave up playing with randoms years ago.

[pigpenz]

so holding
AQXXX vs XXX

most players play to the Q or A first then lead up to Q
most club players play to the Q first
better players will play the %'s and play A first then lead to Q

back to original diamond holding in this problem its 96% to play to K♦'s then to AQ♦