[bobby]

http://www.bridgebase.com/tools/handviewer...3487-1285872301

GIB played the hand above in 6NT -1 when there were 12 tricks from top and 13 with squeeze.
It got a club lead and played a diamond to ten and later had to overtake jack of diamonds for only four diamond tricks.
It "missguessed" the end position for only 11 tricks.

Other tables played 13 tricks in 7NT or 6NT by GIB so it was -16 imps

I have found that GIB plays differently when bidding is different in other tables even when opponents are silent.

[BunnyGo]

I particularly like that even after blocking the diamond trick, with 3 tricks to go the robot just had to endplay West; instead took a line which seems to only work if West has exactly K ♠ T ♣ and J ♥. Not terrible, but seems against the odds.

[Bbradley62]

bobby, on Sep 30 2010, 03:49 PM, said:

I have found that GIB plays differently when bidding is different in other tables even when opponents are silent.

GIB will not always make the same plays because he sometimes does simulations to determine which line of play is best, and might come to different conclusions if his sample size is small. (Just repeating an answer I was given to similar questions.)

All 5 Norths who played this hand similarly blocked the diamond suit. However, this North was in a slightly different situation than the others, because the opening lead was different (low club instead of high spade) after this NS had a different auction because this South opened differently (1♥ instead of 2♣). Because of the spade lead, none of the other Norths could find their 12th trick in spades; therefore, they all had to try the successful heart finesse instead.

[jamegumb]

GIB would be much more powerful if, in addition to the double dummy simulations it runs to determine plays, it utilized proper techniques as "tie breakers" for determining play lines.

When in doubt, GIB should:

1) Not block suits
2) Not cash stoppers assuming other suits will break
3) When drawing trumps or cashing long suits, do so in a manner to allow later transportation between hands

etc.

[Stephen Tu]

GIB doesn't work by learning "proper techniques". It just calculates what will work on what layouts. The problem I think is that the web version is set to some lightning fast calculation setting, so it doesn't consider enough layouts, it doesn't happen to deal any 5-1 splits in its tiny sample so it gets "careless" quite often. My home GIB for example doesn't have issues with making this contract on the slower settings, though I saw it randomly make same mistake when set to really fast.

I just wish that they put option in the web version for time control on GIB like they do on Windows client so you can put it on a setting to play better. And maybe slow it down a tad for the robot tourneys also.

[barmar]

The robot tourneys do use a slower setting than they do in the MBC.

Would you be OK if they charged more to rent slow robots than fast ones?

[jamegumb]

Stephen Tu, on Oct 1 2010, 01:38 PM, said:

GIB doesn't work by learning "proper techniques".

No, it doesn't. Is it impossible to add techniques to its repertoire?

[barmar]

Nothing is impossible, but solving this problem in general is probably hard. Especially while keeping its basic strategy of dealing a bunch of hands and doing double dummy analysis to plan the play.

[cloa513]

Couldn't GIB workout its top winners in NT contracts and then if that's enough for the contract just play them- isn't just a most basic bridge technique?

[uday]

There are a couple of complications


1. How much time do we give GIB to think?
Generally, more time makes it better, but also slower.

2. How many tricks should GIB try to take?
At IMP the answer is almost clear. At MP, the ans is much harder to figure out, even for humans, and GIB usually goes for the line that works most frequently. If GIB had a target of 12 tricks this hand would not be here.

3. Trick 1 and trick 2
Gib tends to get smarter after trick 2 ( the smarter code is seemingly too slow for tricks 1 & 2 )

fwiw, i think the right line is either:

Unblock clubs, cash SA, run D. making 12, and the OT if any of the following are true: 3-3 clubs OR HJ dropping OR LongHJ_with_long_club_or_SKJ

or perhaps

Unblock clubs, top hearts, run D. Making 12, and the OT if any of the following are true: 3-3 clubs OR HJ Dropping OR long_club_with_sK or long_club_with_west_and_HJ_with_east or rarely_e_known_to_have_sk_due_to_count


I'm pretty sure that what happened was that Gib ran into the 'Trick 2 not as smart as trick 3' issue but I am curious, and will find out. If anyone cares, i'll report back.

[Bbradley62]

uday, on Oct 2 2010, 10:17 AM, said:

I'm pretty sure that what happened was that Gib ran into the 'Trick 2 not as smart as trick 3' issue but I am curious, and will find out. If anyone cares, i'll report back.

Thanks, Uday. Please do let us know what was up here, because many of us would like to know.

[uday]

This is what I learned


1. A rental bot was in use here
2. The rental bot ( which are faster and dumber that their Tournament counterparts ) made the assumption that the Diamonds would break and went down
3. The tournament bot made 7 as expected ( W got squeezed as declarer cashed out ) at imps, but went down at MP, playing W for 4414 rather than 5314 by taking the hook after running clubs, hearts and diamonds. The tournament bot did not misplay the diamond suit.


I don't mind that the bot goes down in 6 trying for 7 at mp. I'm not wild about the diamond play at trick two but the rental bot does have to play quickly so I'm not shocked.


We've been discussing offering the tourney class bot as a (pricier) alternative to the normal fast rental bot. We'll try that in a week or so and see if people feel strongly about this .

[Bbradley62]

Shouldn't even the weaker bot be able to count 12 top tricks and cash them at IMPs? In fact, if he's not clever enough to think about overtricks, isn't this exactly what he should do?

It appears to me that this hand was played during a free duplicate robot game. Do those games use the faster/weaker bot?

Btw: which version were we playing with during Friday's "play with a robot for free" day?

[rwbarton]

Bbradley62, on Oct 2 2010, 08:46 PM, said:

Shouldn't even the weaker bot be able to count 12 top tricks and cash them at IMPs? In fact, if he's not clever enough to think about overtricks, isn't this exactly what he should do?

Yes, but we can also make 12 top tricks by playing the ♦10 at trick 2 and later overtaking the ♦J and running the diamond suit.

Of course what I wrote above is false, but if GIB never simulates a layout with diamonds 5-1, then it won't know the difference between winning the 10 and winning the J at trick 2. The scoring is immaterial. As far as I know GIB has no notion of "plan" that would allow it to see that it has 12 top tricks here regardless of the opponents' hands or plays.

[barmar]

Does GIB take form of scoring into account? I thought it always tried to take the maximum number of tricks, regardless.

In my experience, it's pretty easy to tell whether you're playing with fast or slow robots, if you ever become dummy. With fast bots, these hands usually get played in the blink of an eye, except maybe the first 2-3 tricks. With slow bots, you can usually see what's happening for most of the play.

[calm01]

If the fast GIB was free to compensate for its problems, I for one would pay more -
namely a $1 a week for the slower better version.