GIB bidding issue

[jdonn]

1♣ (1♦) 1♥ (2♦)

DBL

 

GIB generally plays support doubles, but when I moused over this double after I made it, GIB didn't recognize the meaning. It's hard to say how responder-GIB took the double since it passed on a 2452 hand. So I think somehow it just doesn't know support doubles on this auction, and probably ones like it. That seems like something correctable.

[TylerE]

If they can just get it to stop transferring into 4 card suits I'd be happy.

[fred]

If they can just get it to stop transferring into 4 card suits I'd be happy.

It is likely (but not certain) that we will be releasing a new version of GIB within a weak that will never transfer to 4-card suits (and never Texas transfer to 5-card suits and...).

 

We are working hard on improving GIB's bidding. If everything goes according to plan there will be more and more improvements in the months to come.

 

Thanks for the bug report, jdonn.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[TylerE]

Glad to here it Fred. This will definatly make me more interested in playing in the robot events.

[junyi_zhu]

1♣ (1♦) 1♥ (2♦)

DBL

 

GIB generally plays support doubles, but when I moused over this double after I made it, GIB didn't recognize the meaning. It's hard to say how responder-GIB took the double since it passed on a 2452 hand. So I think somehow it just doesn't know support doubles on this auction, and probably ones like it. That seems like something correctable.

those are just small issues. You may never imagine what gib bid with this hand:

Sx HQJxxx D- CQTxxxxx

over partner's 1NT opening.

either 2D, 2NT, or even 4D, 5C are all human bids. Gib chose to pass.

As I said again and again and again, gib's whole bidding design is completely wrong. Bidding should not base on simulations of a few hands, which gib is doing. And the the fundamental of gib bidding is so called "points", which is a complete disaster. For example, in this hand, gib overbid 6 "points" with xx KQxxxx Kxx xx, the bidding went 1C 3D, gib overbid 3H, which in system agreement shows at least 14 "points", there is no way in the world to count 14 points in this hand. This bidding weirdness only shows the programmer limited understanding of bridge bidding. If one doesn't fix the basic design structure of gib bidding, there is nothing much for serious players to expect in gib's bidding improvement.

[awm]

Recently I actually saw GIB open the bidding with 1♣ holding 3♣ and 5♠ and about ten points. I don't really get this.

 

It seems to me that GIB's "simulation" method is pretty good when a lot is known about the various hands. But early in the auction (like when we're opening the bidding or overcalling or replying to partner's opening) it seems a lot better to just have some sort of table/decision tree. GIB often takes a very long time deciding whether to pass in an auction where the only other calls have been pass (this especially happens in fourth chair, but I have seen it in other positions also).

[jdonn]

Recently I actually saw GIB open the bidding with 1♣ holding 3♣ and 5♠ and about ten points. I don't really get this.

 

It seems to me that GIB's "simulation" method is pretty good when a lot is known about the various hands. But early in the auction (like when we're opening the bidding or overcalling or replying to partner's opening) it seems a lot better to just have some sort of table/decision tree. GIB often takes a very long time deciding whether to pass in an auction where the only other calls have been pass (this especially happens in fourth chair, but I have seen it in other positions also).

I actually tend to agree that it would be good to have a combination of 'rules' and 'sims' that guide GIB's bidding. Make rules for common or simple situations (they can fit the alerts for GIB's bids that already exist), and feel free to adjust them or add more as time goes, but for anything that is not covered by a rule have it revert to a sample of hands. Aside from avoiding ridiculous actions, it would speed GIB up too. He tilts me when he plays real slow in those speed tourneys (yeah I know, same for everyone, but still).

[fred]

Recently I actually saw GIB open the bidding with 1♣ holding 3♣ and 5♠ and about ten points. I don't really get this.

From a bridge point of view, I don't get it either (though some of my friends like Hampson and Grue seem to like to open a Precision 1D with such hands).

 

But Uday and I do understand why GIB does this and it will stop doing it soon. During the past few weeks I have been spending about 50% of my time looking over 1000s of hands that were bid by each day's latest version of the "new and improved GIB" and comparing its performance on these hands to that of the GIB that plays on BBO. Uday has been spending an even larger % of his time learning how GIB decides what bids to make, fixing specific problems that we identify, and fixing general problems relating to things like hand evaluation. We spend hours each day discussing GIB via phone and running long matches of new GIB vs. old GIB.

 

I can confidently claim that the next version of GIB will not:

 

- open a 3-card minor before a 5-card major

- make a Jacoby transfer on a 4-card suit

- make a Texas transfer on a 5-card suit

- bid Stayman and then opt for notrump when a 4-4 major suit fit is discovered

- open 1NT with 19 HCP (or do many other similar things as it does now)

- after a Blackwood auction play in a King-showing response instead of the agreed trump suit

- make plenty of other basic mistakes that the existing GIB makes now

 

I think we have also managed to tone down GIB's tendency to massively overbid in a wide variety of situations.

 

There is still a long way to go, but GIB is getting better every day. It is kind of hard for us to stop so that we can get the new GIB ready for being released, but we are trying to force outselves to do that soon.

 

Junyi - I am afraid you don't really understand why GIB makes most of its bidding mistakes. The main cause is that it has not been properly taught what bids should mean. The more we can teach GIB to understand, the less often we will have to rely on simulations in inappropriate situations in the hope that it does something sane. Often simulations result in something insane for the same reason - it doesn't understand the bids that come up when it tries to project into the future. The earlier in the auction the simulation takes place the more likely the projection is to spiral out of control.

 

There are situations in which simulations are appropriate and work very well, by the way.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[fred]

Recently I actually saw GIB open the bidding with 1♣ holding 3♣ and 5♠ and about ten points. I don't really get this.

 

It seems to me that GIB's "simulation" method is pretty good when a lot is known about the various hands. But early in the auction (like when we're opening the bidding or overcalling or replying to partner's opening) it seems a lot better to just have some sort of table/decision tree. GIB often takes a very long time deciding whether to pass in an auction where the only other calls have been pass (this especially happens in fourth chair, but I have seen it in other positions also).

I actually tend to agree that it would be good to have a combination of 'rules' and 'sims' that guide GIB's bidding. Make rules for common or simple situations (they can fit the alerts for GIB's bids that already exist), and feel free to adjust them or add more as time goes, but for anything that is not covered by a rule have it revert to a sample of hands. Aside from avoiding ridiculous actions, it would speed GIB up too. He tilts me when he plays real slow in those speed tourneys (yeah I know, same for everyone, but still).

That is exactly how GIB works.

 

The problem is that many of the existing rules are broken and/or silly and that important rules are missing in what you call "common or simple situations".

 

One of the main focuses of our work right now is to improve and extend the rules.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[jdonn]

The work is definitely appreciated. I used to not care since I didn't use it, but now I'm completely addicted to the best hand tourneys ever since I started a few weeks ago :)

[junyi_zhu]

Recently I actually saw GIB open the bidding with 1♣ holding 3♣ and 5♠ and about ten points. I don't really get this.

From a bridge point of view, I don't get it either (though some of my friends like Hampson and Grue seem to like to open a Precision 1D with such hands).

 

But Uday and I do understand why GIB does this and it will stop doing it soon. During the past few weeks I have been spending about 50% of my time looking over 1000s of hands that were bid by each day's latest version of the "new and improved GIB" and comparing its performance on these hands to that of the GIB that plays on BBO. Uday has been spending an even larger % of his time learning how GIB decides what bids to make, fixing specific problems that we identify, and fixing general problems relating to things like hand evaluation. We spend hours each day discussing GIB via phone and running long matches of new GIB vs. old GIB.

 

I can confidently claim that the next version of GIB will not:

 

- open a 3-card minor before a 5-card major

- make a Jacoby transfer on a 4-card suit

- make a Texas transfer on a 5-card suit

- bid Stayman and then opt for notrump when a 4-4 major suit fit is discovered

- open 1NT with 19 HCP (or do many other similar things as it does now)

- after a Blackwood auction play in a King-showing response instead of the agreed trump suit

- make plenty of other basic mistakes that the existing GIB makes now

 

I think we have also managed to tone down GIB's tendency to massively overbid in a wide variety of situations.

 

There is still a long way to go, but GIB is getting better every day. It is kind of hard for us to stop so that we can get the new GIB ready for being released, but we are trying to force outselves to do that soon.

 

Junyi - I am afraid you don't really understand why GIB makes most of its bidding mistakes. The main cause is that it has not been properly taught what bids should mean. The more we can teach GIB to understand, the less often we will have to rely on simulations in inappropriate situations in the hope that it does something sane. Often simulations result in something insane for the same reason - it doesn't understand the bids that come up when it tries to project into the future. The earlier in the auction the simulation takes place the more likely the projection is to spiral out of control.

 

There are situations in which simulations are appropriate and work very well, by the way.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

Thanks for you input, Fred. Hopefully we'll see a saner version of gib soon. By the way, have you worked on a basic signaling method of gib's defense? For now, there is no signal at all, gib basically discards randomly and often pitches the highest spot card which it thinks wouldn't cost a trick. Also, gib openleads 4th best and high low from doubleton, but when it switches, it often plays low from doubleton. If those problem can be fixed, playing with gib can be more pleasant and you may certainly attract more serous money bridge players I suppose.

[barmar]

Fred, this sounds great. I'm really looking forward to the results.

 

Hmm, would it be fun to run a Vugraph of NewGIB vs. OldGIB?

[fred]

Thanks for you input, Fred. Hopefully we'll see a saner version of gib soon. By the way, have you worked on a basic signaling method of gib's defense? For now, there is no signal at all, gib basically discards randomly and often pitches the highest spot card which it thinks wouldn't cost a trick. Also, gib openleads 4th best and high low from doubleton, but when it switches, it often plays low from doubleton. If those problem can be fixed, playing with gib can be more pleasant and you may certainly attract more serous money bridge players I suppose.

So far the only aspect of GIB we have been working on is its bidding. In particular, up until now we have been trying to focus on aspects of GIB's bidding that people complain about the most. I think it is very likely that we will continue to work more or less exclusively on GIB's bidding for at least the next several months.

 

I agree 100% that there is a lot of room for improvement in various partnership-related aspects of GIB's defense (such as signalling). I am not really in a position to make this statement with any confidence, but I am guessing that improving these areas will be very difficult for us. Most likely we will need significant guidance from Matt Ginsberg, the programmer who created GIB, in order to make much headway in terms of anything related to playing cards.

 

Matt has been extremely supportive of our recent efforts, by the way, but the time he has available for GIB-related work is very limited. While I am on the subject, I should also mention that Matt was not responsible for developing the part of GIB that causes it to bid so erractically (the bidding database).

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[fred]

Fred, this sounds great.  I'm really looking forward to the results.

 

Hmm, would it be fun to run a Vugraph of NewGIB vs. OldGIB?

Thanks Barry - we are excited about this too and the work we are doing is a nice change for me and Uday. The support and help we have received from many of our members (including you!) has been much appreciated.

 

I just wanted to offer a warning that you shouldn't get your hopes up too high for the next version. It will certainly be better in many areas, but there is still a very long way to go. This is a long term project and it will likely take a while until GIB bids well enough to make everyone happy. Also, while we do expect to continue to expend considerable resources on the GIB-improvement project for the foreseeable future, there are still plenty of other aspects of BBO that require our attention. Hopefully we will be able to manage our time effectively so that all aspects of our site will continue to improve in the months and years to come. We are hoping and expecting to bring a new full time programmer on board soon - that should help a lot.

 

I like your idea of putting GIB on vugraph one day, but if this happens I expect we will wait until we are confident that we will be able to truly proud of its performance.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[junyi_zhu]

Fred, this sounds great.  I'm really looking forward to the results.

 

Hmm, would it be fun to run a Vugraph of NewGIB vs. OldGIB?

Thanks Barry - we are excited about this too and the work we are doing is a nice change for me and Uday. The support and help we have received from many of our members (including you!) has been much appreciated.

 

I just wanted to offer a warning that you shouldn't get your hopes up too high for the next version. It will certainly be better in many areas, but there is still a very long way to go. This is a long term project and it will likely take a while until GIB bids well enough to make everyone happy. Also, while we do expect to continue to expend considerable resources on the GIB-improvement project for the foreseeable future, there are still plenty of other aspects of BBO that require our attention. Hopefully we will be able to manage our time effectively so that all aspects of our site will continue to improve in the months and years to come. We are hoping and expecting to bring a new full time programmer on board soon - that should help a lot.

 

I like your idea of putting GIB on vugraph one day, but if this happens I expect we will wait until we are confident that we will be able to truly proud of its performance.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

I have tried wbridge5, the free software and winner of 05, 07 , 08 computer bridge. It seems to me that wbridge5's play is actually very strong, both in bidding and in playing. It has been a quite pleasant experience to partner with wbridge5. There are very few mistakes in bidding and the declaring and defending are also very good cause it takes both previous bidding and plays into account and often makes the right constraints and interference. In defending, it implemented a count oriented signal in first 7 tricks and also defend according to partner's signals. It has implemented three bidding systems, wbridge5, which is an extension of French standard, SEF and SAYC. To my surprise, the program is really small and the performance is also no bad in my not so new desktop. It's quite little known to the English speaking world cause it's written by Yves Costel from France. I probably can't rate wbridge5 as a top expert, but I think its bridge is certainly very close to an expert level.

[vuroth]

Wow, this is great news.

 

As a programmer and bridge player, I think that the idea of writing/tuning something like GiB is very interesting - and very intimidating. Maintaining/supporting GiB can only be a huge amount of work.

 

Thanks for the update!