[daveharty]

I knew I was dialing it in a bit when I decided not to bid game on this hand, but GIB played very oddly:

http://tinyurl.com/6prawrr

[barmar]

I assume the most odd play you're talking about is ducking the ♦ on trick 2. I took a look at GIB's simulations on this trick. Of the 25 hands it dealt, it only made a difference whether it won or ducked on 10 of them. Of those 10, there were 2 where it needed to play the A to make the contract, 3 where it needed to play the 10 to make the contract, and 5 where playing A just makes while the 10 makes an overtrick (there were no hands where playing the 8 gained). So the average score when playing the 10 was 112, but the average for the A was only 100. Here's an example hand it dealt where it has to play the 10:

---

The other 2 were similar; in all of them, East has ♠AQ6 ♥Qxxxx ♦Kx ♣xxx and West has ♣QTxx.

I haven't figured out why it thinks ducking is necessary to make the contract, but I'm assuming its double dummy analysis is correct. Maybe one of the readers here can work it out.

[Stephen Tu]

barmar, on 2012-April-11, 17:27, said:

I haven't figured out why it thinks ducking is necessary to make the contract, but I'm assuming its double dummy analysis is correct. Maybe one of the readers here can work it out.

I don't see it. My double dummy program (bridgify 104) says winning the ace makes (better lead heart up next), but ducking is down 2. I'll experiment with my PC-based GIB later.

[Bbradley62]

GIB guessed the spade finesse wrong. I thought he played the hand double-dummy and always got those right.

[barmar]

He's playing double dummy with the simulated hands, not the actual hands.

Which reminds me of one of my pet peeves. Players rarely underlead aces on opening lead against suits, and even GIB avoids it. So human players know that when there's a low opening lead through KJ in dummy, it's almost always correct to play the J rather than K. Yet GIB frequently (it seems like always, but that could be selective memory) plays the K. I'm not sure how to fix this, other than to hard-code an override. I think the reason this happens is that when it's doing its simulations, it doesn't try to figure out what the opponent would do single dummy, it assumes he's also playing double dummy. So if there are hands in the simulation where it's best (or maybe just safe) to underlead the A, it assumes the opening leader will do it, so it doesn't exclude this from the possibilities the way a human does.

[Stephen Tu]

Bbradley62, on 2012-April-11, 19:03, said:

GIB guessed the spade finesse wrong. I thought he played the hand double-dummy and always got those right.

GIB doesn't cheat. It plays the hand double-dummy in its simulations, when deciding what to bid and what to play in earlier tricks. But that's double-dummy with randomized hands for the opponents, biased by the bidding, picking the best card on average, not necessarily the best card on the exact hand. On the real hands it will guess wrong randomly. The advanced GIBs actually have the GIBson algorithm so actually kick in real single-dummy reasoning after a few tricks, not just "best play statistically double dummy", so it can do stuff like find 100% endplays instead of assuming it will always guess AJT vs. kxx suits. (A single-dummy aware program will do the throw-in when possible and capture all layouts, while a double-dummy sim-only program will randomly choose between the throw-in and playing the suit itself, since it doesn't see the throw-in as superior, since it can always just guess the two-way finesse suit and always guess right.)

But on this hand it seems its double-dummy sims went haywire, I find it hard to construct a hand where ducking the DK gets you more tricks.

[barmar]

It's possible my assumption was wrong.

This was a basic bot, and it has a very limited amount of time allowed for thinking. This might be causing it to cut the DD analysis short, resulting in incorrect results. Will investigate further.

[Stephen Tu]

On my PC, GIB 6.1.5:

- single dummy, GIB places ace without thinking very long at all
- double dummy, it agrees with Bridgify that ace is necessary to make contract (followed by heart up).