[pescetom]

And followed up with this second bottom...
costing me the tournament.

[smerriman]

Yeah, playing the ♠J at trick 6 is bonkers.

But I don't think this one is a "bug" in the normal sense, just a result of Monte Carlo + sample size. If I tell old-GIB to sample 20 hands, about once every 10 tries it plays the J - stating that its best line is West holding the Q and East holding the T.

The backwards finesse, needing two cards in place, is of course horrible compared to the 50% line of playing East for the Q.

But if I run the numbers:

- in a 20 deal sample, the backwards finesse will work more frequently than the normal finesse about 12% of the time
- for a sample size of 30, it's 8%
- for a sample size of 50, it's 3%
- for a sample size of 100, it's 0.5%

barmar once said GIB samples "a few dozen" hands, so it's going to get this decision wrong somewhere from 3-8% of the time, enough to be noticeable. Note that with the default settings (which uses a sample size of 100), it took about 15 seconds to make a play at trick 3. I suspect the proportion of BBO users who prefer waiting that long for greater accuracy is pretty low, which is why BBO would have cut it down.

Before GIB, robots used to work based on being taught suit combinations, etc, more like humans, to solve situations like this.. but they were simply left in the dust by Monte Carlo simulations, which all future robots then switched to.

[johnu]

You didn't mention the low club to the ♣Q at trick 4 which was an excellent play to maintain communication in clubs so GIB could make the key play of ♠J

Edit: 4-1 or 5-0 club splits with ♣J onside only happens in bridge books.

[pescetom]

smerriman, on 2025-March-23, 22:44, said:

But if I run the numbers:

- in a 20 deal sample, the backwards finesse will work more frequently than the normal finesse about 12% of the time
- for a sample size of 30, it's 8%
- for a sample size of 50, it's 3%
- for a sample size of 100, it's 0.5%

barmar once said GIB samples "a few dozen" hands, so it's going to get this decision wrong somewhere from 3-8% of the time, enough to be noticeable. Note that with the default settings (which uses a sample size of 100), it took about 15 seconds to make a play at trick 3. I suspect the proportion of BBO users who prefer waiting that long for greater accuracy is pretty low, which is why BBO would have cut it down.

We were playing in a mainly human tournament and neither I nor the opponents would worry about 15 seconds (or more) by Declarer, who will finish 2 minutes ahead of the field all the same.
I do wonder if the time has been cut down even lower (despite robot rental price hikes), given that the frequency of such disasters seems to be increasing.