possible gib bug?

[hoodwinked]

Today when playing in a money bridge tournament came up with an interesting deal - GIB had a hand something like this

 

xxx xx xxx xxxxx about 6 hcp and I had opened 1 club. Instead of the expected 1NT or 3c inverted response he responded 1 diamond!!! He then bid 2 clubs over a 1 heart rebid which left me in a bit of a quandary with my strongish nt hand with both majors. I had no clue whether this was a strong hand with both minors or a weak one. I made the error of bidding 2nt and Gib now rebids 3 spades and we play in a very unmakeable 3nt when 2 or 3 clubs was the max for the hand.

 

Another hand, GIB decided to open 1c and I passed holding 5 clubs. The e/w pair eventually played in 2 hearts when 5 clubs made in N/S. You guessed it - GIB opened a short club!!! I've NEVER seen that one before. Did Gib just decide to do something outside partnership agreement? I find that hard to believe.

[fred]

GIB does indeed sometimes make bids that it "knows" violate its partnership's agreements. As you might expect, sometimes this works well and sometimes it works badly.

 

This is an interesting subject that I would be happy to discuss further at some point, but it will have to wait for another day.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[barmar]

I've seen GIB open a 3-card minor when it had 4 in the other minor, and the 4-card suit also had better texture (I think it was something like KQxx versus Jxx).

 

Fred posted a while ago asking people to mail him the hand records when GIB makes weird bids like these, so he can find and fix the bugs in the bidding database.

[fred]

I've seen GIB open a 3-card minor when it had 4 in the other minor, and the 4-card suit also had better texture (I think it was something like KQxx versus Jxx).

I've seen several excellent players do the same thing (obviously intentionally).

 

Please do send me hands in which GIB does strange things, but it not the "minor psychs" that GIB sometimes makes that concern us so much. We are more concerned about "impossible bids" (for example: 1NT-3NT-6NT) or bids that demonstrate a clear lack of knowledge of the meanings of bids (obvious penalty doubles that are interpretted as takeout or 3NT overcalls that are interpretted as showing 25+ HCP for example).

 

I am not expecting people to weed out which strange bids are of interest to us. Send us examples of anything you consider strange and we can do this ourselves. The only reason for the previous paragraph is that I thought some of you might find it interesting.

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[barmar]

Do you consider transfers to 4-card suits "clear lack of knowledge" or "minor psyches" (it's just 1 card short of the requirement, not a "gross misstatement")? How about when it uses Stayman, discovers a fit, and then bids 3NT anyway?

[fred]

Do you consider transfers to 4-card suits "clear lack of knowledge" or "minor psyches" (it's just 1 card short of the requirement, not a "gross misstatement")?  How about when it uses Stayman, discovers a fit, and then bids 3NT anyway?

Neither.

 

I consider them to be valid bridge strategies with certain types of hands for both technical and psychological reasons.

 

GIB looks at these things from a technical point of view. It uses simulations to try to "judge" how various actions will turn out. Sometimes the results of these simulations are not accurate for a variety of reasons. This may cause GIB to judge poorly - perhaps GIB will transfer to a 4-card suit with a hand for which this strategy is inappropriate. Sometimes GIB misjudges because it is unlucky in the sample of random hands that it generated to do its simulation, deciding that an action that worked well on this (biased) set of deals is the action it should take.

 

GIB chooses the bid that it thinks will work best in the long run. I will be the first to admit that sometimes it does ridiculous things, but some of the things it does that generate ridiculous results are not so ridiculous. They fall into the category of "it could have been right".

 

For GIB the only aspect of the psychological aspect of bidding that exists falls out of the technical aspect - GIB's simulations make predictions about the future actions of the other players. GIB does not consider who the human is or what his mannerisms were.

 

Even a psych (be it minor or major) is a technical thing for GIB. When GIB psychs its simulation is predicting that the opponents will misjudge thereby giving GIB a good result. GIB is not explicitly programmed to psych. A psych it just another bid to GIB. It choices that bid if it thinks it will be the long run winner.

 

GIB does understand the concept of the "book bid", but in most auctions it has the freedom to depart from the book bid if its simulation suggests that this is a good idea.

 

GIB can be configured to make the book bid whenever there is an entry in it that covers the situation it is in. Unfortunately GIB's bidding book has some problems including, to borrow your phrase, a "clear lack of knowlege" about some aspects of bidding. We have found that letting GIB simulate tends to produce better results (though the problems in its bidding book manifest themselves in its simulations as well).

 

Fred Gitelman

Bridge Base Inc.

www.bridgebase.com

[barmar]

OK, I thought GIB had mostly switched over to book bidding. I remember when it first came out, people found that its bidding was totally inscrutible because it was done almost entirely using Monte Carlo simulation. If you didn't give it enough thinking time, it would make crazy bids because it didn't try enough simulations and statistical flukes would make the bid seem best. I was under the impression (apparently mistaken) that simulation had mostly been removed from bidding, and was just used during play.

[uday]

It is an option. We have a choice of whether to make the robots use "book bidding" or a combination of "book bidding" and a simulation.

 

Neither approach is perfect, imo, but the latter seems to be less flawed. Using book bidding would (i think) stop it making as many odd 4-card xfers etc but there is little room for recovery if the "book" is wrong.

 

Simulations are slower and weirder sometimes but can also recover from flaws in the "book".

 

The first robots rolled out on BBO with book bidding only. I switched a few weeks later point to book+simulations

 

The more time we give it to think (also an option) the better it usually gets (to a point, and the slower it gets as well).

 

We poke gently at the bidding database (the heart of most of the issues w/the robot) but I dont know enough about the syntax to make a big dent in it myself. Someday we'll find someone with both knowledge and cycles to spare and we'll be more effective at making GIB better.

[MickyB]

I suspect that GIB's liking of 4-3 fits reflects that they are easier to handle double dummy.