[Stephen Tu]

---

GIB made some incredibly nutty defensive plays against my opponent here. I feel like its too restrictive constraints on 4M opening (5-9 hcp 8+H) are frying it's ability to generate reasonable sample hands. It ought to allow 7 cd suits and up to maybe 12 hcp?

They really need to figure out how to loosen the constraints on the Monte Carlo sample when playing against humans to break ties when it "doesn't matter", cater to human be +/- a Q or K on both ends of expected high card range and a card on distributions, then some double dummy tie plays no longer are ties and it won't randomly do weird stuff when it matters.

[johnu]

Unlucky play by East on 2nd spade play. The color on color fake ruff didn't work. Good job by GIB to credit N/S with the trick. And West made 2 good plays by playing low on the first 2 trump tricks to guarantee a trump trick.

[steve2005]

It needs to loosen constraints playing against other Gibs too. Often Gib is out of range on bids.
Gib should not be playing random cards in tie. Gib should still do the right bridge play, including signaling if it knew how to do it.

Speaking of signalling, what good is 4♣ at trick 1 with spade void and AK♦?
And also even if declare doesn't have KS knows finesse is working, so danger of contract making.

[Stephen Tu]

steve2005, on 2018-March-23, 04:34, said:

It needs to loosen constraints playing against other Gibs too. Often Gib is out of range on bids.
Gib should not be playing random cards in tie. Gib should still do the right bridge play, including signaling if it knew how to do it.

Signaling of course would be awesome, but it's really super hard to implement properly. It's hard even teaching humans how to signal, let alone bot. The idea would be to bias the Monte Carlo sample based on partner's card, but how to adjust the sample is like super dependent on dummy and the auction, good signaling wouldn't be an always thing (sometimes with doubleton and partner leads A from AKx+ you DON'T want continuation and ruff, you want shift because need that first with natural trump trick etc., on flip side if partner asks continuation sometimes you are supposed to shift anyway).

But otherwise, GIB SHOULD be playing random cards in a tie. Falsecarding gives humans problems, if it always gave count or played up line or whatever it would be exploitable. GIB has no idea of what is "a right bridge play", this simply falls out of the calculation of DD analysis of its sample. What is the right play is what card works most often among the possible hands, the problem now is that it is excluding some hands it shouldn't because it is too trusting of the bidding. If it adjusted the sample to be more lenient and cover all possible remaining hands even if someone misbid, that should get rid of a large chunk of the weird ducking/non-ruffing for no particular reason plays, since then those ties would no longer be ties. It should like calculate first with more trust of the bidding to get highest candidate plays, then calculate a second time with looser/no constraints to break ties among the winners of the first round.

[jdonn]

The problem is not the definition of 4M opening, it is how GIB determines plays on defense. I don't see that changing any time soon.

[Stephen Tu]

How it determines plays on defense is heavily influenced by the samples it generates though, isn't it? Which are influenced by the definitions, no? If it is only generating < 10 HCP hands by South and only 8+ suits I think it is going to make many more errors than if it is generating samples that include hands with fewer hearts and 10-12 hcp.

I know signalling is a pipe dream but looser constraints on the samples feels like it's possible. Certainly wider range for the definition to begin with ought to be possible.