Possible quick easy GIB improvement

[armantt2k]

When I play with one of my regular partners, our convention card is loaded and available, even against two robots.

 

Do the GIB robots assume we play GIB 2/1 and all other GIB card conventions?? Or do they check human opponents' card for exceptions to the usual conventions that give critical info useful in later play of the hand?

 

I first thought only of UDCA or O/E during play, but being aware of basic alertable bids by the opponents such as opening 1NT weak or 1!C that "could be short" would certainly have an impact on the random hand generation algorithm GIB uses to chose competitive bids during the auction and later when choosing a lead or lines of play. Another example: checking which of the myriad 1NT overcall systems (DONT, Cappelletti, Meckwell, Landy, ...) the opponents play and using that info to restrict the suit distributions in the random hands under consideration.

[steve2005]

Gib uses 2/1 Gib CC. See GIB System Notes (bridgebase.com).

The card is essentially available to you in that when you click on potential or actual bids an explanation for the bid is shown.

Not only that but Gib assumes opponents are using the same 2/1 GIB. So it can cause strange things to happen when opps use a different bidding system lol.

 

GIB comes pretty close to using no card signaling. Though it does sometimes show count but even that you can't count on.

[mycroft]

This "simple" fix is about as simple as teaching GIB to understand Spanish. But it's German next round.

 

No, really, this is *incredibly hard*, and one of the reasons why the computer championships use such a restrictive convention card - because otherwise one team will walk in with a totally different (even if seriously suboptimal) system and totally bamboozle the opposition.

 

which, of course, is how "anti-GIB", the "system" used by the best robot players, works.

 

But seriously, I play:

• K/S with Keri, 1NT overcall for takeout, and Woolsey with one partner
  
• except when we play 10-12 NT, incredibly wide-ranging 2openers and overcalls, and very sound 1-suit bids
  
• well, except when we're bored and play Phantom Club
  
• but with another partner, it's Toronto-standard 2/1, with Hello
  
• or the next, Calgary-style 2/1 with DONT (and odd two-bids, and odd invm responses)
  
• or the next, Méxican 2/1 with Meckwell, Bergen, and no invm
  
• and if we're both bored, what we can remember of the BM Precision we played in 2008. I think we played Mod Capp against NT...

 

Oh, and our carding could be:

• UD everything *including suit preference*, frequent suit preference, Reverse Smith and some special A/K agreements
  
• UDCA and basically "signal what partner needs, when partner needs, but usually attitude to partner/count to declarer"
  
• Standard, but attitude leads to NT, and "middle-only encourage to T1"
  
• UDCA, with coded 10s and 9s (when we remember) and odd-even discards (when we remember)

 

And that's just one player. Good luck.

[smerriman]

Yep, this is about as far from 'possible', 'quick', and 'easy' as you can get :) If you want to play against robots, you'll need to play their system or accept very weird things will happen.

[pilowsky]

GIB comes pretty close to using no card signaling. Though it does sometimes show count but even that you can't count on.

 

I know it says in the system notes that they "sometimes show count" but what does that mean? When do they not show count?

The statement has no heuristic meaning at all.

OTOH, I have the impression that GIB plays high encourage. So if it discards a high honour it implies that leading that suit might result in a trick.

But I don't know how to formally test this idea.

[thepossum]

How much of GiB signalling is genuinely programmmed in versus just being confirmation bias from time to time

 

Occasionally it shows attitude or count, but it cannot be relied upon :)

 

As for reading other players CCs I am not sure it has read its own that well

[armantt2k]

Sorry, but it seems everyone has missed the point of my posted question. So I will simplify it:

 

Do BBO robots look at human opponents' posted convention card?

 

At the risk of confusing everyone again, here's an example:

 

My partner and I play 1NT weak (12-14 hcp) - does GIB use this info to restrict the random hands it uses to analyze how to bid/play?if m

[pescetom]

Sorry, but it seems everyone has missed the point of my posted question. So I will simplify it:

 

Do BBO robots look at human opponents' posted convention card?

 

At the risk of confusing everyone again, here's an example:

 

My partner and I play 1NT weak (12-14 hcp) - does GIB use this info to restrict the random hands it uses to analyze how to bid/play?if m

 

I think everyone understood the first time.

The answer is no.

 

(Having said that, I do think that people overstated the difficulty of explaining basic agreements to a robot (future robots, not GIB). Ok we can't explain the nuances and interpreting the current free text card is not practical, but it is quite possible to structure the card to have fields for things like shape/number of cards and HCP range of each opening and for robots to read this, for example. Half a cake is better than none, and not just for robots either).

[armantt2k]

I know it says in the system notes that they "sometimes show count" but what does that mean? When do they not show count?

The statement has no heuristic meaning at all.

OTOH, I have the impression that GIB plays high encourage. So if it discards a high honour it implies that leading that suit might result in a trick.

But I don't know how to formally test this idea.

 

I doubt the GIBots play high to encourage. They simply play the highest of equals when intervening cards have been played. For example, holding Q97 the GIBot will play the Q from the remaining Q9 when the K is led after the first round was AJT7. This 'false carding' not only disguises the distribution (is he void?) but also makes their ability to 'remember' ALL the cards played a bigger advantage against most humans who fail to keep track of high pip cards.

[smerriman]

(Having said that, I do think that people overstated the difficulty of explaining basic agreements to a robot (future robots, not GIB). Ok we can't explain the nuances and interpreting the current free text card is not practical, but it is quite possible to structure the card to have fields for things like shape/number of cards and HCP range of each opening and for robots to read this, for example. Half a cake is better than none, and not just for robots either).

GIB doesn't deal hands based on HCP and suit lengths; it deals hands based on whether they would have resulted in a matching auction sequence. And to match the auction, it needs to have an entire database built around that bid.

 

And the database doesn't have HCP and suit lengths 'variables' that it can simply change and have the rest adapt automatically; it's all around sequence matching. So you're basically starting that chunk from scratch.

 

Making a simple bug fix seems extremely hard already, so adding an entire new branch of the database with every followup and defense - and not just that, but adjusting all the sequences where that bid *isn't* made anymore - is basically impossible.

[pescetom]

GIB doesn't deal hands based on HCP and suit lengths; it deals hands based on whether they would have resulted in a matching auction sequence. And to match the auction, it needs to have an entire database built around that bid.

 

And the database doesn't have HCP and suit lengths 'variables' that it can simply change and have the rest adapt automatically; it's all around sequence matching. So you're basically starting that chunk from scratch.

 

Making a simple bug fix seems extremely hard already, so adding an entire new branch of the database with every followup and defense - and not just that, but adjusting all the sequences where that bid *isn't* made anymore - is basically impossible.

 

Sure, but as I said, future robots not GIB. BBO has a clear interest to replace GIB and the parent company already owns at least two better robots.

[smerriman]

Oops, missed that. But I still disagree. OK, suppose you allowed someone to enter solely an HCP range and suit lengths ranges for a 1♣ opening. I enter 10-20, 5+ spades, 4+ diamonds. No robot in the world could ever be designed to have a dynamic bidding system that could compete against that and we're only at the second call in the auction..

 

Let alone more complex stuff like a bid which shows length in unspecified suits.

[pilowsky]

I doubt the GIBots play high to encourage. They simply play the highest of equals when intervening cards have been played. For example, holding Q97 the GIBot will play the Q from the remaining Q9 when the K is led after the first round was AJT7. This 'false carding' not only disguises the distribution (is he void?) but also makes their ability to 'remember' ALL the cards played a bigger advantage against most humans who fail to keep track of high pip cards.

 

Well, obviously they don't play high encourage in the sense that they are informing a sentient being that they "want" something.

It's a computer program not a spouse.

It's a problem with anthropomorphising machines. They don't "remember" (as in random access memory), they assign values to things and then use an algorithm to calculate the optimal answer to a problem based on the assigned values.

 

Having said that, when the program seems to consistently play a high card if it is beneficial to have that suit returned (and I repeat, I may be entirely wrong here) humans might be tempted to label this "behaviour" high encourage.

At other times they randomise (so do I) to avoid giving anything away.

[pescetom]

Oops, missed that. But I still disagree. OK, suppose you allowed someone to enter solely an HCP range and suit lengths ranges for a 1♣ opening. I enter 10-20, 5+ spades, 4+ diamonds. No robot in the world could ever be designed to have a dynamic bidding system that could compete against that and we're only at the second call in the auction..

 

Let alone more complex stuff like a bid which shows length in unspecified suits.

 

The problems start when an opening shows more than one suit, possibly unspecified: but there are and/or operators and precedence, this is more a problem of making things human readable than machine readable. Yes we're getting closer to a markup language than a static form, but it is still doeable at low cost.

As for robots in the world... Argine can more or less effectively switch agreements about length and strength (and even aggressiveness) at partner's whim, I would not be surprised if there was some flexibility about opponent agreements already present or easy to insert. And that is just an algorithm that you or I could program with techniques of decades ago: point a leading edge AI like AlphaZero at bridge and all this will be blown out of the water, with the real problem becoming how do robots disclose their mutually formed agreements to humans in a format the poor things might understand B-)

[johnu]

Oops, missed that. But I still disagree. OK, suppose you allowed someone to enter solely an HCP range and suit lengths ranges for a 1♣ opening. I enter 10-20, 5+ spades, 4+ diamonds. No robot in the world could ever be designed to have a dynamic bidding system that could compete against that and we're only at the second call in the auction..

 

Let alone more complex stuff like a bid which shows length in unspecified suits.

Not just robots, but nearly all human partnerships would not have a system to handle that opening bid. As far as just handling NT HCP ranges, and maybe 4/5 card majors, etc, I thought the original standalone GIB allowed you to configure your own bidding system to some extent. If so, GIB should be able to simulate those differences in the play also based on the original coding if that coding hasn't been stripped out of the program.

[armantt2k]

Well, obviously they don't play high encourage in the sense that they are informing a sentient being that they "want" something.

It's a computer program not a spouse.

It's a problem with anthropomorphising machines. They don't "remember" (as in random access memory), they assign values to things and then use an algorithm to calculate the optimal answer to a problem based on the assigned values.

 

Having said that, when the program seems to consistently play a high card if it is beneficial to have that suit returned (and I repeat, I may be entirely wrong here) humans might be tempted to label this "behaviour" high encourage.

At other times they randomise (so do I) to avoid giving anything away.

 

By 'remember' I meant perfectly track all 52 cards, played and remaining to be played, a very easy task for a computer program, and beyond the ability of most humans. I frequently see GIBots discard an honor on trick 12, and win trick 13 in the same suit with a low pip card (e.g. 7 over a 6). A human wouldn't do that; it's too risky, and illogical. Why not just keep the honor and toss the pip card? Unless, of course, you're trying to win with the beer card! 😉

[thepossum]

Sure, but as I said, future robots not GIB. BBO has a clear interest to replace GIB and the parent company already owns at least two better robots.

 

This comment made me very sad

 

[armantt2k]

If BBO intends to replace the GIBots with one of the other robots they own, then there is no point in us submitting robot reports or suggestions, is there?