BBO Discussion Forums: Shouldn't GIB see that it needs to unblock? - BBO Discussion Forums

http://tinyurl.com/73cks57

Its not too difficult to figure out even with Easts return.

Regards,

Simon

http://www.bridgebas...sername=pigpenz

you would think so but look at these results,
3NT-1
1NT -2
both by North/South

where theoretically 3NT could be down 5
and 1NT down 3

on both hands at these contracts GIB refuses to take setting tricks(3nt) or
lets go of good tricks to keep worthless cards(1nt) at trick 11. Now I am one
who likes to go back and look at results at other tables, and if one does this
they will see that there is alot of discrepancies in what happens with different
GIBs.

I've recently been studing GIB's playing code, and it turns out to be much dumber than I ever thought. It doesn't really know how to make inferences from the play. It doesn't remember that you led 4th best (and since it often doesn't lead 4th best, I guess that makes sense). If a defender wins a trick with an unnecessarily high card, it doesn't realize that this denies holding the lower cards that could have won. These are why it does some really strange things, like not repeating finesses, or not giving partner a ruff when his singleton King drops under your Ace.

Basically, what it uses when doing its simulations during declarer play and defense is expected suit distributions learned from the bidding, and inferences from standard honor leads to trick 1.

We asked Ginsberg about this, and he said trying to include all the inferences from the play when selecting hands for simulations was too slow.

barmar, on 2012-March-13, 02:55, said:

Why does GIB lead back diamonds? Anyway simulations should tell it the correct play- obvious GIB ignores simulations.

cloa513, on 2012-March-13, 06:47, said:

It can't ignore simulations, since that's all it has to go on.

The problem is with the hands in the simulations. If it doesn't deal the right hands, it won't come to the right conclusion about the best play.

I haven't actually done the analysis, but I think it's pretty obvious that East's diamond continuation made it think that its diamonds were significantly better than Kx.

barmar, on 2012-March-13, 14:57, said:

You didn't answer about declarer's play- it assumes opponents play DD so by leading back diamonds it must lose 4 diamond tricks at least plus what other loser it has.

Oops, I misread your question, I thought you asked "Why doesn't GIB lead back diamonds?", referring to when North switched to clubs.

I'm going to have to investigate that.

I tested this with 200 different random seeds, and it always played a high ♥ or ♠ at trick 2.

My guess is this was a basic robot (and so were all the other tables that played the board, since they all returned a ♦), so I'm not going to investigate any further. I've said it over and over: if you want more reasonable play, rent the advanced robots.

The hand is actually still makable after the ♦ exit -- South is squeezed on the 4th ♠. But not after East finesses the ♣Q, as long as North unblocks; there were some tables where it did this.

May I add my two cents worth? GIB is programmed in Prolog, which uses database technology. If you have modded a wargame like Waterloo you,ve encountered Prolog. Alternetively you could get a feel for it from Oxford Bridge's bidding editor.

Gib's bidding is based on a truncated version of Meadowlark's bidding database. Simulation takes over from convention at a fairly early stage. Improvements in bidding may only entail extension of the database although any extension/change may require heavy debugging.

I do not find Giorgi's replies defensive, sometimes laconic. Here is the thing, it is frustrating to report an appalling action by GIB and have it rationalised as a permissable error but if you had to fix software you would first check that any report was a genuine bug.

I suspect the best solution is to follow Fred's approach and play against GIB with a human partner, it is easier to tolerate silly plays by opponents than by partner.

And no sane person would rate GIB as their favourite partner.

Scarabin, on 2012-March-15, 15:29, said:

Reading all these replies should make one wonder wether it is even worth reporting all the weird things that happen with GIB, cause it sounds like its no easy task to take care of.....when comparing results especially if you play with GIBs in open room against GIBs you are at the vagrancies of the settings that other people have for the old versions of GIB.

Scarabin, on 2012-March-15, 15:29, said:

No it isn't, it's programmed in C. I suspect he originally wrote it in Lisp, then rewrote it in C -- he has some C macros that are patterned after Lisp (e.g. dotimes for looping, T and NIL to represent true and false). And simulations don't "take over" from conventions, they work in tandem (after finding the book bid, it optionally looks for rules in the DB for hands similar to the one it holds, then uses simulations to choose among them).

I agree with most of the rest of what you said. I find GIB enjoyable for random play -- it's as good or better than many of the partners I've gotten when playing in EFAF tourneys (last week I had a partner open 1NT with more than 20 HCP). But I wouldn't ask GIB to partner with me in any serious tournament.

We all agonize over the shortcomings of GIB, and I've spent several hours skyping with Uday, discussing ways that we might improve it beyond the individual rule tweaks that Georgi does. It's a difficult problem, since the programming is so tangled.

Sorry for the mistake about prolog - I could have sworn I got it from one of Ginsberg's many papers on GIB. On the point about "simulations taking over from convevtions at a fairly early stage" I meant that my impression is that GIB and some other robots only follow bidding rules for the first couple of rounds. I originally bought GIB as a way of learning Kaplan-Sheinwold and Moscito, but was sadly disappointed.

Anyhow apologies for the slip.

Early versions of GIB made much more extensive use of simulations during the bidding than the current version does. It used to make lots of really random bids because of this, as well as many unwarranted leaps to slams (far more than it currently does). I think this was before the full Meadowlark bidding DB was incorporated.

barmar, on 2012-March-16, 01:06, said:

so are we talking about the commercial versions of GIBs?
In the commercial version there were alot more programable parameters that could be set.
I dont think I ever noticed the commercial version doing some of the things done by the bots
here( probably old ones)....but the BBO interface is so much easier to use than the commerical
GIB, that I hardly every use the commericial GIB, if it could be changed to a more GI like BBO i would use it.

I think he's just too lazy to do that work. Wbridge5 does a lot of that kind of inferences and is fast. Here comes the topic I raised several years ago, GIB isn't a very sound choice for the money bridge purpose because it is badly programmed.

barmar, on 2012-March-13, 02:55, said:

xxhong, on 2012-March-16, 11:53, said:

Are you talking about Ginsberg? It's not his problem any more. BBO bought the program from him years ago, he's no longer involved in it at all. We occasionally send him emails when we're confused by something in the code.

So if anyone is lazy, it's me and Uday. I just don't feel competent to tackle some of the tricky AI problems in this program.

pigpenz, on 2012-March-16, 10:48, said:

The BBO advanced bots are essentially the same as the commercial version, but we don't make the parameters available -- everyone uses the same parameters. BBO basic bots are also mostly the same code, but with different options (e.g. less thinking time, and disabling simulations during the auction) and some of the code is commented out (like the code that switches from Monte Carlo simulations to GIBson after several tricks of play).

What I meant was that when GIB was first written, it made much more extensive use of simulations during the auction. It had a database for the first couple of rounds, and I think it also handled conventions like Blackwood, but it frequently went off on its own, just bidding what it thought it could make. Sometimes it would bid 4NT intending it as natural, not realizing that partner would respond with Blackwood. I don't think I ever used GIB when it was like this, I just remember reading about it on rec.games.bridge. This was many years ago, perhaps before BBO acquired GIBware.

I know the story. What I meant is that it is often very difficult to modify the code because of the early inflexible programming designs. Also, it is often like the refinement of the biddings, you can probably start from those easy and important problems, especially defensive cardings and inferences. Also, if you really work hard to allow the human player to declare, you would already get rid of half of the problems, which definitely worth the efforts I think.

barmar, on 2012-March-16, 12:16, said: