[1175]

smerriman, on 2024-June-27, 00:50, said:

But even if such a hand didn't exist, while everyone is in agreement about poor assumptions (arguably GIB's greatest flaw), to accept your theory that GIB's shift is a mistake even given its assumptions, you need to find examples where the the spade shift scores strictly less than the club when South has Ace 6th in spades. What hand layout are you thinking of there?

I just prefer to concentrate on one defensive task at a time. On the original hand, South actually could hold a stronger hand (say AJxxxx,x,xxx,Axx). In this case, a Robot holding the South hand might (after a spade shift) lead a low club ("assuming" that partner led from KQ). Granted, with the opponents at the 5-level, that would only cost an overtrick.

I wish I had kept track of the number of Souths that bid 2♠. Yes, that bid has "flaws," but (by making the opponents play at the 5-level), it gave North-South a chance for a huge pickup. I have quickly learned that to do well against the Robots, one must often bid non-existent values.

[1175]

I want to get back to the original topic of "obvious" mistakes with this hand (which I had nothing to do with, other than suffering a lower score because of the Robot's misdefense):

---

The less said about the bidding, the better (although one other table had an identical auction). In total, five tables took on 3NT, one with South opening 1♦, one with South opening 1♣!, and one where South passed initially, bid 1♠ in response to North's 1♦ opening bid, and then jumped to 3NT after North's 2♦ rebid.

Of course (as so many of the Robot's errors do), the incident took place at trick 11. Surely the Robot places the ♠K in the South hand. Why would it play the ace instead of the queen?

[smerriman]

1175, on 2024-June-27, 04:53, said:

I just prefer to concentrate on one defensive task at a time. On the original hand, South actually could hold a stronger hand (say AJxxxx,x,xxx,Axx). In this case, a Robot holding the South hand might (after a spade shift) lead a low club ("assuming" that partner led from KQ). Granted, with the opponents at the 5-level, that would only cost an overtrick.

Of course. But none of that is really relevant; my sole point here was that if North believes South has A sixth, then a *double-dummy* analysis will show that a spade switch is correct. It sounded like you were disagreeing with that, but I guess not. Of course, the assumptions are flawed, and double dummy analysis isn't realistic, but those are an absolute given.

1175, on 2024-June-27, 05:13, said:

Surely the Robot places the ♠K in the South hand. Why would it play the ace instead of the queen?

Same answer as last time. It doesn't consider any logic about why past plays were made. Yes, this is bad.

The very original (pre-BBO) version of GIB used to try to handle this - from a 1999 paper:

Quote

To conform to the card play thus far, it is impractical to test each hypothetical decision against the cardplay module itself. Instead, GIB uses its existing analyses to identify mistakes that the opponents might make. As an example, suppose GIB plays the ♠5. The analysis indicates that 80% of the time that the next player (say West) holds the ♠K it is a mistake for West not to play it. If West in fact does not play the ♠K, Bayes' rule is used to adjust the probability that West holds the ♠K at all. The probabilities are then modified further to include information revealed by defensive signalling (if any), and the adjusted probabilities are finally used to bias the Monte Carlo sample, replacing the evaluation sum_d s(m,d) with sum_d w_d*s(m,d) where w_d is the weight assigned to deal d. More heavily weighted deals thus have a larger impact on GIB's eventual decision.

But this part of the algorithm was deleted sometime before or when BBO took over, as it was too slow / buggy - considering how badly GIB messes up hands by assuming the bidding is accurate, consider how much more it would mess up when weighting the accuracy of every single card played to date (and how trivial it would then be to fool it into going down in cold contracts as defender).

[1175]

smerriman, on 2024-June-27, 15:45, said:

Same answer as last time. It doesn't consider any logic about why past plays were made. Yes, this is bad.

The particular problem doesn't need any fancy algorithm on simulations. Just a "win the trick with the cheapest possible card" will do (I assume that the Robots, when defending, can "see" dummy's cards).

I know you have described this as the "same" answer, but (as this case does seem "slightly" different), what does the Robot look at when deciding to play the ♠A? Does it conclude that East has the ♠K (at that point, the unseen hands have only one spade - the king)? Presumably, East would have cashed the missing heart if it had it, although one could claim that East, "knowing" that the unseen (from its view) hands had only one club (the queen), should have cashed both clubs before playing a spade).

The large number of "this can't gain and can only lose" plays probably ranks as my biggest surprise as I learn about the Robot's capabilities.

[smerriman]

Oops, so sorry - somehow was completely looking at the wrong trick (or hand diagram?!) there. My comment doesn't even make sense in that context :/

From GIB's perspective, South's 3NT bid denies a 4 card major, so no matter how baffling the play may have seemed to date, South has precisely KT of clubs left, and GIB must win the Ace in order to put partner in with the King to cash the setting heart trick.

Intentionally lying during the bidding like that is a relatively common tactic if you're in want of a top score, as the defense will bend over backwards to believe you, giving up tricks in the process.. if you open 1NT with a 5 card major, try responding 2♦ to Stayman, accepting an invite with even a minimum, and watch as the opponents lead into your 5 card suit repeatedly with another suit wide open due to the lack of signalling..

[1175]

smerriman, on 2024-June-27, 19:47, said:

Oops, so sorry - somehow was completely looking at the wrong trick (or hand diagram?!) there. My comment doesn't even make sense in that context :/

I think I have just one hand diagram where I show the entire play.

smerriman, on 2024-June-27, 19:47, said:

From GIB's perspective, South's 3NT bid denies a 4 card major, so no matter how baffling the play may have seemed to date, South has precisely KT of clubs left, and GIB must win the Ace in order to put partner in with the King to cash the setting heart trick.

That makes some sense, although does the 3NT bid by South even exist by a passed hand? If it doesn't, then every aspect of that bid deserves some doubt. So I guess (on this hand) the blame has to go with East not cashing the two clubs at trick 11, which would give West no chance to go wrong.

smerriman, on 2024-June-27, 19:47, said:

Intentionally lying during the bidding like that is a relatively common tactic if you're in want of a top score, as the defense will bend over backwards to believe you, giving up tricks in the process.. if you open 1NT with a 5 card major, try responding 2♦ to Stayman, accepting an invite with even a minimum, and watch as the opponents lead into your 5 card suit repeatedly with another suit wide open due to the lack of signalling..

I believe I have seen some instances of that when examining results at other tables. Some bids seem so bad that I have assumed a lack of knowledge or skill, but I guess some people have learned to steer the Robot wrong by deliberately making abnormal bids.

[smerriman]

1175, on 2024-June-28, 06:46, said:

That makes some sense, although does the 3NT bid by South even exist by a passed hand? If it doesn't, then every aspect of that bid deserves some doubt. So I guess (on this hand) the blame has to go with East not cashing the two clubs at trick 11, which would give West no chance to go wrong.

It's not a bid that GIB would ever make, correct; the rule for 3NT shows 13-15 HCP and the initial pass shows <= 11 HCP and <= 12 total points. But it has to assign a definition to every bid; they're combined to result in exactly 11 HCP and exactly 12 total points. (When you see a definition specify an exact number of points, that's a sign things are likely to go wrong..)

And however the definition is assigned, it still assumes you have it, for better or for worse.

And yes, East cashing two clubs would be the simple human logic, but as usual, GIB's double dummy logic isn't capable of telling it that one 100% play involves a guess / some assumptions and another 100% play doesn't.

[1175]

I just finished a hand that shows the Robot making a "poor shift," that fortunately didn't frustrate me this time since I benefited from it.

---

At most tables, nothing much of interest happened.

Ten tables had this auction (down 1, +100):

---

At three tables, South competed the 3♦ (making, +110):

---

At one table, South didn't understand inverted minors:

---

After the opening lead of the ♣2, the defense can no longer beat the contract, however this South didn't play the hand any better than he bid it (cashing the ♥A before setting up either the third diamond trick or third club trick), and ended up going down 3 (-150).

Finally, at two tables (including mine), South (at least speaking for myself) decided that the favorable vulnerability made "preempting" one-level balancing calls worthwhile:

---

At the other table with this auction, West led the ♥3. East won the ♥K, and continued hearts, leaving South with no play for the contract. This South should have held the contract to down 1, but he also managed to go down 3 (-150).

At my table, West led the ♥Q, which I, of course, ducked. West then shifted to the ♠Q. I won in dummy with the ace, and then, wanting to put West back on lead, played the ♦A and a small diamond, West winning the queen. Reverting to hearts still beats the contract, but West persevered with the ♠J, ultimately giving me ten tricks (+430).

I guess signalling would help if East had three hearts, but even if East had just two (possibly either Kx or Tx), playing a second heart looks like a better option to me.

[smerriman]

I just wrote a sim that deals 1000 hands with South holding the promised 6-10 HCP and no 4 card major, and no other constraints on South / East other than the cards played to the first 4 tricks.

Double dummy results:

- on 573 deals, the spade and heart score equally
- on 246 deals, a spade beats 3NT while a heart does not
- on 29 deals, a heart beats 3NT while a spade does not
- of the other 152, the differences were overtricks/undertricks, but a spade was still better in 102 of these

After 1 trick, it's a little closer, but spade is still comfortably better:

- 730 equal
- 24 only spade beats contract
- 12 only heart beats contract
- of the other 234, 177 have spade a better result

GIB runs smaller samples but comes up with the same conclusions.

[1175]

smerriman, on 2024-June-29, 02:37, said:

I just wrote a sim that deals 1000 hands with South holding the promised 6-10 HCP and no 4 card major, and no other constraints on South / East other than the cards played to the first 4 tricks.

[...]

That surprises me some (but maybe part of that comes from the fact that West has no clear entry to cash hearts, particularly if South has the ♠K).

Just to make sure I understand, does "double dummy results" mean on opening lead or does that mean with West on lead at trick five (and does "after 1 trick" meat with West on lead at trick two)?

[smerriman]

On lead at tricks 5/2. And yeah, I think that's probably it; noting that if East has the ♥K, then South *must* have the ♠K and ♣Q for the 1NT bid.

[1175]

I have to ask about the Robot's "thought processes" on this hand:

---

Three tables (including mine) played 4♥ on this auction:

---

With the help of a defensive error, I managed to hold this to down 1 (still an IMP loss), while the other tables in this contract went down 2 and 4.

The other thirteen Souths played either 4♦ or 5♦ on various auctions. Those in 4♦ made the contract six times (three of those with an overtrick), and went down five times. The game contract failed by two tricks at both tables where attempted.

At a table with the auction shown in the diagram, the low club by East at trick five (the topic of my curiosity, of course) turned a two-trick set into two potential overtricks (declarer did not play optimally after that, and ended up with only one overtrick). It seems bizarre to me (South can't have the ♥A after showing up with the ♠K and KQJT... of diamonds). I assume that East wanted to get to West's hand for a spade ruff, but maybe some other reason exists.

[smerriman]

This one seems pretty straightforward - South can't hold either J (4♦ is a max of 9 HCP). If they have at least one club and heart, it doesn't matter which suit you play. But if they have a heart void, a low club is required to beat the contract (while if they have a club void, the difference is just an overtrick).

[1175]

Not my table, so painful to watch (the Robot only dropped one trick against me - at this table, it dropped FIVE).

---

I still say that bad shifts rank as the Robot's top defensive problem.

[1175]

Why does the Robot keep making this type of mistake (which I see fairly frequently)? This took place on board 16, during a set which I had fifteen plus-IMP scores going into the last board. I needed the Robot to drop two tricks on defense to make my contract. It didn't disappoint:

---

[smerriman]

1175, on 2024-July-09, 11:41, said:

Why does the Robot keep making this type of mistake (which I see fairly frequently)? This took place on board 16, during a set which I had fifteen plus-IMP scores going into the last board. I needed the Robot to drop two tricks on defense to make my contract. It didn't disappoint:

What do you think is the mistake here? Ducking the diamond at trick 3?

Ace is needed if North has the QJ, but GIB finds a lot more hands where low is needed..

---

---

---

[1175]

smerriman, on 2024-July-09, 23:30, said:

...GIB finds a lot more hands where low is needed.

How did you simulate those (and how many hands did you look at?)?

[1175]

I have a hand where I came out on the good side of a mistake by the Robot. Let me post this first as a defensive problem - the solution(s) of which appear obvious to me, and I will update it later:

---

Hmmmmm. Looks like I can't select cards for the hidden hands.

You (West) lead the ♣T, and the first trick goes ♣A, ♣7, and ♣4. Declarer plays the ♥6 from dummy. Partner follows with the ♥7, declarer plays the ♥J, and you win your king. What do you play now?

[smerriman]

1175, on 2024-July-13, 04:31, said:

How did you simulate those (and how many hands did you look at?)?

Initially via the generally undocumented command line parameters of the old Windows version of GIB, to see firstly whether it played the same way (it did), and whether that was consistent if I repeated it multiple times (it was), and then whether those samples it prints in debugging information were flawed in some way (not that I could tell). Those were just a handful of examples I picked out - obviously there were plenty of others where both high and low were equal as well.

I didn't do a full custom simulation which may show something else, but I just didn't see any clear-cut bug here that warranted further investigation - what was your logic as to why it was wrong?

[1175]

smerriman, on 2024-July-13, 22:45, said:

...what was your logic as to why it was wrong?

I probably latched onto it as a (in retrospect, flawed) example of more obvious errors made by the Robot on the same theme (inappropriately ducking), which I didn't save at the time I played those hands.