Wrong explanation

[Yu18772]

http://tinyurl.com/d5vh677

http://www.bridgebase.com/forums/public/style_emoticons/default/rolleyes.gifYu

[cloa513]

http://tinyurl.com/d5vh677

http://www.bridgebase.com/forums/public/style_emoticons/default/rolleyes.gifYu

Why North couldn't shade his hand to call it 1NT- its 15 TP which would have made it much easier.

[Stephen Tu]

GIB plays 16-18 1nt overcalls. upgrading to 16 is a bit much.

 

No, I think the real problem is the various meanings of 2nt + 3nt by the doubling side. If my home GIB still matches BBO on this auction, over 2h, GIB wants 16 pts to bid 3nt, and 15 points to bid 2nt. 16 pts opposite balancing passed hand takeout double might be about right. But probably 2nt can be dropped to 14-15. Penalty doubling should also be considered. With this sort of shape in this 12-14 range I am considering 2nt/dbl/pass depending on my major suit honor cards/spots, and not really cue bid.

 

But GIB feel stuck for a bid and cue bids. Then South bids 2nt, which is "invitational to 3nt"? My home GIB also describes this as forcing, which is inconsistent. If GIB is going to cue with 13-14, opposite 9-11 2nt it should be able to pass this.

 

Then again over 2nt, it wants 16 pts to try 3nt, again feels stuck and thus invents the 3d bid. If some of these point ranges were relaxed a bit then GIB wouldn't need to make these weird maneuvers.

[Bbradley62]

GIB plays 16-18 1nt overcalls. upgrading to 16 is a bit much.

GIB's convention card ( http://webutil.bridgebase.com/v2/acblcc/acblCC.html?x=656469743d6e266b65793d4143424c2f4e2f7261696e2f313239383931343233395f313038365f667265642f3132393839313432333926706172746e65723d7261696e ) indicates that he plays 15-18 1NT overcalls.

[cloa513]

http://tinyurl.com/d5vh677

http://www.bridgebase.com/forums/public/style_emoticons/default/rolleyes.gifYu

Surely you should have realised that logically 3D showing 5 and 13TP+ is impossible- GIB didn't overcall afterall (the bidding description can't keep up with GIB's freewheeling simulation bids).

[barmar]

This was a simulation bid -- the book bid is 3NT.

 

The problem is that it was unable to find hands in the simulation that seemed consistent with the bidding. You showed stoppers in ♠ and ♥, yet West is bidding like it has good majors, and North can see good cards in his hand those suits.

 

In a little more than half the simulations, it got to 3NT whether it bid it directly or bid 3♦ along the way (the rest of the auction was always 3♥-3♠-3NT).

 

But in many of the rest, East came to life over 3♦, and often NS were able to double him for a huge number. As a result, the average score from bidding 3♦ was much higher than just bidding 3NT, so it chose it.

 

I'm not sure what can be done about situations like this. One thought I had was that it should ignore simulations if it can't find hands that match the auction, and just go with the book bids. But that basically means that when someone psyches, GIB may stop thinking, and that doesn't seem right.

 

In this case, you were the one who confused it with your 2NT bid. GIB's book bid with your hand is 3♣, and if you bid this North will bid 3NT.

[Bbradley62]

In this case, you were the one who confused it with your 2NT bid. GIB's book bid with your hand is 3♣, and if you bid this North will bid 3NT.

It's not really clear that 3♣ is the "book bid", since it shows 5+C.

[hv=lin=pn|pcassidy,~~R30945bt,~~R4465zeg,~~R4468enf|st%7C%7Cmd%7C4S47H5TQD467AC2TJQ%2CS259TAH24KAD259CK%2CS68JKH679JDTQKC9A%2C%7Crh%7C%7Cah%7CBoard%202%7Csv%7Cn%7Cmb%7Cp%7Cmb%7Cp%7Cmb%7C1S%7Can%7CMajor%20suit%20opening%20--%205%2B%20S%3B%2011-21%20HCP%3B%2012-22%20total%20points%7Cmb%7Cp%7Cmb%7Cp%7Cmb%7Cd%7Can%7C3%2B%20C%3B%203%2B%20D%3B%203%2B%20H%3B%209-11%20HCP%3B%2012-%20total%20points%7Cmb%7C2H%7Can%7C4%2B%20H%3B%205%2B%20S%3B%2021-%20HCP%3B%2015-22%20total%20points%7Cmb%7C2S%7Can%7C13-16%20total%20points%3B%20forcing%7Cmb%7Cp%7Cmb%7C3C%7Can%7C5%2B%20C%3B%203%2B%20D%3B%203%2B%20H%3B%209-11%20HCP%3B%2012-%20total%20points%7Cmb%7Cp%7Cmb%7C3N%7Can%7C5-%20H%3B%205-%20S%3B%2014%2B%20HCP%3B%2016-%20total%20points%3B%20likely%20stop%20in%20H%3B%20likely%20stop%20in%20S%7Cmb%7Cp%7Cmb%7Cp%7Cmb%7Cp%7Cpc%7C]360|270[/hv]

[barmar]

It's not really clear that 3♣ is the "book bid", since it shows 5+C.

True, but it's the closest lie, so that's what GIB would bid.

[Stephen Tu]

But in many of the rest, East came to life over 3♦, and often NS were able to double him for a huge number. As a result, the average score from bidding 3♦ was much higher than just bidding 3NT, so it chose it.

 

I'm not sure what can be done about situations like this. One thought I had was that it should ignore simulations if it can't find hands that match the auction, and just go with the book bids. But that basically means that when someone psyches, GIB may stop thinking, and that doesn't seem right.

 

- Fix initial bid so it can bid a more sensible 2nt or double, rather than torture partner with a cue bid unlikely to illicit useful information?

 

- Why should it expect East to come to life over 3 diamonds, having passed originally and over 2 spades? What bids would he come to life with, major suit big fits it can't have since North has so many major cards? Perhaps the definitions of these bids should be looked at, to make East not at all likely to choose them, then it doesn't see that unlikely upside. This used to be a bigger problem for GIB, it always made these weird assumptions about what the opps were going to do. I don't think it should simulate if it anticipates further competition, it should stick to book bids. Simulation should be reserved for decisions likely to be relatively final or limited to very few possible continuations.

 

- if there are too many points in the deck, trust partner to have his hand, but assume opps have fewer HCP, more shape, adjust their lower limit down a couple points and redeal until it can get some matches?

[barmar]

- Fix initial bid so it can bid a more sensible 2nt or double, rather than torture partner with a cue bid unlikely to illicit useful information?

It's too strong, 2NT shows 12-13 HCP (although it will actually bid it with 11-13).

- Why should it expect East to come to life over 3 diamonds, having passed originally and over 2 spades? What bids would he come to life with, major suit big fits it can't have since North has so many major cards? Perhaps the definitions of these bids should be looked at, to make East not at all likely to choose them, then it doesn't see that unlikely upside. This used to be a bigger problem for GIB, it always made these weird assumptions about what the opps were going to do. I don't think it should simulate if it anticipates further competition, it should stick to book bids. Simulation should be reserved for decisions likely to be relatively final or limited to very few possible continuations.

Sometimes East bid 3♠, sometimes it jumped to 4♠, holding just a doubleton. It generally had strange distribution.

 

Not simulating when you expect competition is not an answer. It's really hard to write bidding rules that cover complex, competitive auctions, there are just too many cases to program. So we depend MORE on simulations to decide what to do in situations like this.

 

I realize it's got imperfections, since you can't depend on the opponents to do what you predict. But we can't write book bids that cover these cases well, so sims are the best tool we have.

- if there are too many points in the deck, trust partner to have his hand, but assume opps have fewer HCP, more shape, adjust their lower limit down a couple points and redeal until it can get some matches?

The way it currently works is that it deals several hundred hands (using approximate dealer parameters), ranks them by how closely they match the auction, then runs the simulations using the 25 hands with the fewest mismatches. In this case, the closest it could get were 2 mismatches.

 

Something like what you suggest could potentially be done, but I think it would slow things down too much. After noticing that every hand has too many mismatches, it would have to try lots of different adjustments (length of each suit, reducing HCP) and go through the same redeal and analysis for each of them until it found a set of acceptable hands.

 

And you say that it should trust partner, but that's one of the things that got it into trouble on this hand: it expected you to have a spade stopper because of your 2NT bid.

[Stephen Tu]

It's too strong, 2NT shows 12-13 HCP (although it will actually bid it with 11-13).

So fix the range for 3nt??? If it's too strong for 2nt, then let it bid 3nt. Or widen the range for 2nt. There shouldn't be a gap between 2nt & 3nt.

 

Sometimes East bid 3♠, sometimes it jumped to 4♠, holding just a doubleton. It generally had strange distribution.

Why would East do this, is it because it thinks that if *it* does something weird, NS will bid more and go down? NS bids weird because it thinks EW will do something crazy? And EW does something weird because NS will then do something more crazy? Setting up infinite loop of craziness here.

 

Fix the GIBs so East doesn't support spades without say 4+ card support and a heart card or something. Then maybe the problem goes away.

 

Not simulating when you expect competition is not an answer. It's really hard to write bidding rules that cover complex, competitive auctions, there are just too many cases to program. So we depend MORE on simulations to decide what to do in situations like this.

 

Don't at all agree why it's not the answer. Make the book bid, then if there is more competition deal with that later. In complex, competitive auctions, you want to do the "normal thing" as much as possible, early in the auction, because you can't possibly know for sure what the enemy is going to do, the number of possible continuations is too high that you can't manage all the possibilities. If you make the book bid and there's further bidding you or partner can do something sensible later. If simulations cause you to deviate from book, and there is further bidding, it's harder to recover. Simulations aka judgment should come in on the possible *final* call of the auction. Bid game or not. Pass or balance. Not if you think there is likely to be further bidding.

 

I can't see how making the book bid of 3nt after cue bid-pass-2nt-pass is going to make GIB worse? It has a reasonable book bid here. If on some auctions there isn't a sensible book bid, then we fix the book!

 

I realize it's got imperfections, since you can't depend on the opponents to do what you predict. But we can't write book bids that cover these cases well, so sims are the best tool we have.

You have a book bid here already that's sensible and covers this case, bid 3nt, the sim made it worse not better.

 

And you say that it should trust partner, but that's one of the things that got it into trouble on this hand: it expected you to have a spade stopper because of your 2NT bid.

 

Why should this throw it off course of the normal 3nt book bid? So what if we have spades extra-well stopped? I don't agree with your supposition that making book bids if it can't match the auction is a bad thing. If the default book bids are defined sensibly, it should generally land in a reasonable spot. GIB gets in more trouble in weird auctions when it starts speculating IMO, not when it follows book when the book is sensible. The times it gets in trouble in weird auctions following book is where the book is crazy, promising 22+ or 25+ points.

 

I'd argue, that if it can't deal hands that match the auction, that is precisely the case where simulation is *least* reliable and accurate, so you want to rely on sim less, not more.

[Bbradley62]

True, but it's the closest lie, so that's what GIB would bid.

Ok. Of course, it's hard for most of us (those who weren't at the table at the time) to see this since the description of 2N is truncated. You identified the truncation problem six months ago; is there any prospect of this being fixed?

[barmar]

Here's one of the simulated hands where it believed it could get a huge score by bidding 3♦:

[hv=pc=n&s=sa7hkq52d9542ckj8&w=st9542hat84d63c75&n=skj86hj976dkqtca9&e=sq3h3daj87cqt6432&d=e&v=n&b=2&a=pp1sppd2h2sp2np3d4sppdppp]399|300[/hv]

I know, West doesn't really have an opening hand, let alone one that would bid a second time, and East wouldn't have passed the opening. Remember, I said that it couldn't find any hands that were consistent with the bidding -- this is just one of the many mismatching hands.

 

Why is GIB bidding 4♠ with that East hand? West's bidding shows 15+ total points, and East, with 11 total points, thinks they have enough for game. And in a pinch, it will play in a 7-card fit.

[Stephen Tu]

So isn't this perfect example of why if it can only find grossly mismatched hands to the auction, it should just fall back and make the book bid and hope it works out?

 

Basing strategic decisions on partner passing opening bids, and assuming opponents opened 4 counts can't be good bridge. If you are relying on your opponents to be nut cases you will often be disappointed. Making the book bid may not always be the *optimal* thing to do, but at least it rates to be a *reasonable* thing to do, and doesn't create disasters if the book has reasonable defaults. If you can't figure out what's going on, bidding normal will usually be OK, maybe partner can figure it out.

[barmar]

Don't at all agree why it's not the answer. Make the book bid, then if there is more competition deal with that later. In complex, competitive auctions, you want to do the "normal thing" as much as possible, early in the auction, because you can't possibly know for sure what the enemy is going to do, the number of possible continuations is too high that you can't manage all the possibilities. If you make the book bid and there's further bidding you or partner can do something sensible later. If simulations cause you to deviate from book, and there is further bidding, it's harder to recover. Simulations aka judgment should come in on the possible *final* call of the auction. Bid game or not. Pass or balance. Not if you think there is likely to be further bidding.

But you need to use simulations to deal with the competition in the first place.

 

Think about how you behave in competitive auctions. You upgrade or downgrade honors and shortness in the opponent's suit, depending on where they sit in relationship to the opponent. For instance, if you have Kx or AQ in RHO's suit, that's much better than having them in LHO's suit. Singletons and voids in a suit that the opponents have bid and raised are more valuable than other shortness. You upgrade your hand if you have a double fit. These are the types of things that it's really hard to write comprehensive rules for. To fill in these gaps in the bidding rules, humans use judgement, GIB uses simulations.

 

What I think you may actually be trying to say is that when GIB is performing its simulations, it shouldn't try to predict what the opponents will do. But is that really reasonable? Do we just assume that the opponents will get out of the way and let us bid comfortably? That's hardly realistic.

 

We know this isn't a perfect system. For instance, when GIB is simulating, it assumes that everyone will make book bids during the ensuing auction, but they would actually often simulate as well. But simulating the simulations would result in combinatorial explosion, so it's not feasible.

 

And as hands like this demonstrate, there are times when it gets things totally wrong. But we feel that it works well enough, and none of us thinks we know better than Ginsberg in this regard that we could change it properly. Maybe if Matt were still involved we could have a good technical discussion about it, but none of us has the AI expertise to make such a fundamental design change. I have no idea how to specify precisely when GIB should or shouldn't simulate, other than the existing mechanism (each bidding rule says whether it prohibits simulations from overriding it, requires simulation before choosing it, allows it to be selected by a simulation, or is agnostic about simulations).

[Stephen Tu]

Think about how you behave in competitive auctions. You upgrade or downgrade honors and shortness in the opponent's suit, depending on where they sit in relationship to the opponent. For instance, if you have Kx or AQ in RHO's suit, that's much better than having them in LHO's suit. Singletons and voids in a suit that the opponents have bid and raised are more valuable than other shortness. You upgrade your hand if you have a double fit. These are the types of things that it's really hard to write comprehensive rules for. To fill in these gaps in the bidding rules, humans use judgement, GIB uses simulations.

 

But it shouldn't go to simulations too early. If you do the anti-book bid too early in the auction things can spiral out of control, because partner starts playing you for a hand you don't actually have. Ideally GIB would be able to reevaluate honors based on the auction, adding/subtracting from its raw total point count, then still make book bids based on adjusted point count, but it can't. Fine. But don't let it go off on wild flights of fancy predicting the next few bids by the opponents or partner when not making auction ending decisions.

 

What I think you may actually be trying to say is that when GIB is performing its simulations, it shouldn't try to predict what the opponents will do. But is that really reasonable? Do we just assume that the opponents will get out of the way and let us bid comfortably? That's hardly realistic.

 

Why is that unrealistic? If you make the book bid, you pass max accurate information to partner. This gives a solid platform to react further, no matter how the opponents continue. Propose some auction where making the book bid puts you in some untenable spot, where the book bid is reasonable.

 

I'm saying:

- If it can't deal hands that match the auction, revert to book, because its simulation is likely to go haywire in these cases.

- If it's not making a "final decision", stick to book.

- If you are making some decision that is likely final, simulate, but simulate only against other "final decisions", not against auctions where partner or opps keep on bidding.

 

I strongly feel that GIB's bidding can be improved by fixing the book bid definitions in places where there are holes and inconsistencies, and having it rely on the book as much as possible, not simulate more. Just like chess engines rely on extensive opening book libraries, GIB should rely on a huge DB to keep it on a reasonable footing in the beginning part of an auction. If the book were fixed for 2nt/3nt instead of cue bid, then it avoids the problem of 3nt vs. 3d after 2s - 2nt. And if it can't figure out what people have after 2nt, since supposedly someone misbid at this point, I don't think sticking to book and bidding 3nt is at all a bad thing. Blame partner if it goes wrong.