[1175]

Hi,

First post here. I have registered for an account to write about this issue (and I suspect I will use it mostly to complain about the Robot ).

After playing for a couple of months (around 2000 hands) in the duplicate IMPs (where I basically live for now), I feel ready to start assigning grades to the Robot. For defense, it deserves a "D" (at best). Where do I begin? It often leads too passively against NT (I recall numerous cases where the "obvious" lead beats the contract, but the Robot led declarer's suit). The Robot often doesn't lead the suit that its partner bid (aggravating when I have made a questionable bid just to get a specific lead). The Robot seems particularly poor at giving partner ruffs (I recall one hand where the Robot had three chances to give its partner a ruff, but never did). It often plays higher cards than necessary, which ends up costing tricks for the defense. And as explained in the system notes, the Robot doesn't signal, which makes figuring out the best defense much more difficult for me.

Finally, the Robot often ducks when it shouldn't, and I have an incredible example of that.

---

At three tables, South rebid 3D over 2H, and played it there (making nine or ten tricks). At one table, South took a preference to 2S after 2H by North, which made eight tricks. At one table, South jumped directly to 3NT in response to the 1S overcall by North - certainly questionable, but the robot at that table led the HJ (oops!). At ten tables, after 2H by North, South bid either 2NT (seven times, which the Robot raised to game) or 3NT (three times). At those ten tables (3NT by South), West cashed four spades and the CA to beat the contract by a trick. I wanted to cater to six spades and four hearts by North, so I bid 3C at my second turn (as shown), which North converted to 3NT. East led the CJ, and West, looking at the cards to set the contract in his own hand (again, four spades and the CA, but in this case, with the South hand exposed as dummy), ducked to North's king, and I quickly scampered home with ten tricks (+10.9 IMPs).

I know that (as a recent arrival) I have much to catch up on (including past posts in this forum), but I have already seen a number fairly bizarre things happen playing with the robots, some of which I will write about in the future. I will try to filter those by asking "would a low-level intermediate player make this mistake?", and in this case, the answer seems clearly "no" (or at least, not more than once ).

[smerriman]

Firstly, be warned - the last time BBO updated GIB was February 2019, so everybody here has lost all hope of GIB being improved. This may have changed recently, as a new member of the BBO team has said they're picking GIB back up, in January that one bug I mentioned would be fixed in "the next version" and as of last month that next version might be out "soon". But BBO have a proven track record of saying things then disappearing forever and hoping you didn't notice, so I won't believe that until I see it.

So most of my replies are less of "yes, that should be fixed" and more "here's my understanding of why it broke".

Many of your preliminary points are part and parcel of the fact that most of GIB's logic is based on double dummy analysis of simulated hands. It will lead passively, because this is least likely to cost a trick, and it "knows" that it will be able to defend perfectly from that point onwards (and also that declarer will never misguess). If you bid a suit, it will adjust its simulations to make sure you have the length but that will still regularly come up as a suboptimal lead double dummy. In reality, much of this logic isn't really valid - a lead that's best double dummy often isn't actually best since partner won't know what the killing return is, while they'd have more information if you played more human leads. But double dummy robots tend to perform better than ones which blindly follow rules like "always lead partner's suit" or "always lead 4th highest from your longest / strongest suit", so this is something you generally have to put up with.

As for your specific example.. I can't be 100% conclusive, but I am fairly certain I know why it ducked, and it's really silly.

GIB, for a bizarre reason, has been told that the standard lead from JT8x is fourth highest, not the J (but J is correct from a suit headed KJTx).

West's logic is that if East has led from a 4+ card suit, it must be headed KJT, and thus ducking is clearly correct. If East has lead from a 2 or 3 card suit, then North with their 9 major cards has at most a singleton diamond, and the diamond suit isn't running, so it doesn't cost to duck.

I hear you say.. wait.. but East *did* lead the J. Yes, after it runs a double dummy simulation, it gives a bonus to the 'normal' lead.. but the gain of breaking its own rules and leading the J was greater.

Robots try their hardest to assume other robots (and humans) haven't broken the rules even though they can, and that's where GIB's biggest errors come from.

But I also hear you say.. wouldn't those cases meant that North had broken the rules by bidding 3NT without a club stopper? Yeah, sure. But the whole auction is broken; 3♣ is described as natural with biddable clubs!? And 3NT is described as extra values (15-17) and a "partial stop in clubs".. that point range is completely impossible to begin with. But anyway, it trusts inferences from its partner more that it trusts the opposition's bidding.

(Actually, testing this in an older version of GIB, it does play the Ace.. while the bulk of the initial sample was indeed leads from KJTx+, it then includes some non-matching hands in the simulation, and these were enough to outweigh it. But no doubt the basic version of GIB here didn't have a sufficiently large number).

[1175]

First of all, thank you for your reply (and explanations). I have a few comments (and a couple of questions).

smerriman, on 2024-May-26, 02:48, said:

But double dummy robots tend to perform better than ones which blindly follow rules like "always lead partner's suit" or "always lead 4th highest from your longest / strongest suit", so this is something you generally have to put up with.

Bridge represents the ultimate game of imperfect information. Good players don't blindly follow rules, but usually don't break them without a good reason.

smerriman, on 2024-May-26, 02:48, said:

GIB, for a bizarre reason, has been told that the standard lead from JT8x is fourth highest, not the J (but J is correct from a suit headed KJTx).

UGH!

smerriman, on 2024-May-26, 02:48, said:

West's logic is that if East has led from a 4+ card suit, it must be headed KJT, and thus ducking is clearly correct. If East
has lead from a 2 or 3 card suit, then North with their 9 major cards has at most a singleton diamond, and the diamond suit isn't running, so it doesn't cost to duck.

But shouldn't the Robot take into consideration that his partner might have led from JT8x? If East would absolutely never do that, then ducking makes some (a small amount) of sense. While I appreciate your comment about not trusting opposition bidding (I have seen humans make some, let me say "strange" bids in the duplicate IMPs, that work out incredibly well a surprising percentage of the time), I think that in this case, East doesn't have "room" for the CK (I would at least trust the opposition Robot bidding somewhat). It also appears that the Robot doesn't take the cost of errors into account. Playing the CA when partner has the CK will cost an overtrick, but ducking when declarer has the CK will likely give away a vulnerable game at IMPs (and those diamonds in dummy certainly do appear threatening). A human would obviously weigh the probability that partner led from JT8X versus from KJTx, as well as consider the consequences of making the "wrong" play in each case.

smerriman, on 2024-May-26, 02:48, said:

3♣ is described as natural with biddable clubs!? And 3NT is described as extra values (15-17) and a "partial stop in clubs".. that point range is completely impossible to begin with. But anyway, it trusts inferences from its partner more that it trusts the opposition's bidding.

I had four of them, did't I? And (very modest) extra values.


I have started to accept that (similar to playing with humans) that sometimes the Robot will hand me a huge gift (defensive error), and sometimes it will hand those gifts out to other people playing the same hand. I have also begun to take amusement from bamboozling the Robot, and while not particularly relevant to this topic, I have a humorous example of that. I arrived in an absolutely hopeless 3NT contract, and an early defensive error by the Robot got me up to eight tricks. I (South) had ♣AQT98 in my hand opposite ♣65 in dummy. The bidding indicated that West held the club honors, so I led the C8 from my hand, which held when Robot failed to cover (my ninth trick).

So let me move on to my question - does the Robot do these simulations in real time (and how does it do it?)? It really surprised me (the first time I saw it) that the Robot will take different actions (at least in defense and declarer play) in absolutely identical circumstances. This hand occurred during one of my first few sessions in IMPs, and I had written it up meaning to make it my first post (as a question), but will include it here.

---

Initially, I gave myself too much credit for good defense (and learned the truth after I reviewed what happened at the other tables).

Let me take care of the two oddities first. At one table, South overcalled the 1H opening bid with 3C (holding ♣AQJ98), which "worked," as (after a negative double by W) East bid 3NT (West didn't pull), and that failed by three tricks (+10.6 IMPs). At another table, South raised to 3D, and then to 5D over 4S. That didn't work, as West doubled, and suboptimal play led to a three-trick set.

At the other fourteen tables, West played 4S (doubled at two tables), making at every table except for mine. Ten times the bidding proceeded as shown in the diagram. At one table South bid 3D over the support double, at one table South bid 3C over the support double, and at two tables, South overcalled 2C directly, leading to this auction:

 W  N  E  S
      1H 2C
2S 3D  P  P
4S  X  P  P
 P

I hadn't realized before that the Robot can chose different actions (at different tables) in identical circumstances.

At one table in 4SX, the robot led the DA, and at the other table (same auction), the Robot led the CT.

At the table where South bid 3C over the support double, the robot led the C3 (I don't know why not the CT). At the table where South raised diamonds over the support double, the robot made the odd (and very unsuccessful) lead of the HA.

Finally, at the ten tables with identical auctions, the robot led the CT nine times. South took two clubs, and then shifted to a diamond (seven times), continued clubs (once), or shifted to a spade (once - the correct play, but a trick too late). For some reason I can't figure out, at my table, the robot led the S8, an absolutely devastating lead (but it still required the defense not to slip up in subsequent play). Declarer won in dummy and ruffed a heart, which revealed the heart position to me at trick two. The Souths at the other tables didn't know that declarer had absolutely no chance of setting up the hearts. When declarer played a club to the king at trick three, I won, and returned a second trump, also won in dummy. Declarer ruffed another heart, and then led a diamond towards dummy. Partner won the ace, and played his second club, which I overtook. I then played my third trump, and with no club ruffs available, the contract failed by two tricks. While I thought that would score well, the result exceeded my expectation (+10.3 IMPs). It turns out though, that I hadn't really done anything brilliant at all. The robot set up the defense with the trump opening lead, and I strongly suspect that had I received the lead of the CT (which occurred at all of the other tables with the same auction), I would not have found the necessary spade shift at trick two, which could still have beaten the contract one trick.

Anyway, I have since seen many cases where the Robot will make a different lead or card play at different tables in identical situations, which frequently leads to large swings, and I just wonder why that happens.

Anyway, thanks again for your reply, and I will read the posts in your signature.

[smerriman]

1175, on 2024-May-26, 06:52, said:

But shouldn't the Robot take into consideration that his partner might have led from JT8x? If East would absolutely never do that, then ducking makes some (a small amount) of sense. While I appreciate your comment about not trusting opposition bidding (I have seen humans make some, let me say "strange" bids in the duplicate IMPs, that work out incredibly well a surprising percentage of the time), I think that in this case, East doesn't have "room" for the CK (I would at least trust the opposition Robot bidding somewhat). It also appears that the Robot doesn't take the cost of errors into account. Playing the CA when partner has the CK will cost an overtrick, but ducking when declarer has the CK will likely give away a vulnerable game at IMPs (and those diamonds in dummy certainly do appear threatening). A human would obviously weigh the probability that partner led from JT8X versus from KJTx, as well as consider the consequences of making the "wrong" play in each case.

Like I said at the very end, the old version of GIB I have *does* take this into account, and causes it to play the Ace. Why didn't yours? Most likely because you are playing with the free robot, which has been severely crippled from the proper product, with one example being that the number of hands it simulates is very low. Of course, paid robots are still capable of making equally silly mistakes; I don't know the exact logic for how it determines how many mismatching hands to include in the sample, and how far they're allowed to mismatch.

If in a similar situation there was a very unusual hand East might hold that lets game make, should you play safe to ensure you gain 12 IMPs every time it occurs, even if 90% of the time you'll cost yourself 3 IMPs? How do you determine what assumptions you've made are less likely than others? Weighting possible deals is a very difficult problem..

But as usual, not saying GIB couldn't do it better, though it's trickier than it sounds.

1175, on 2024-May-26, 06:52, said:

I had four of them, did't I? And (very modest) extra values.

Yeah, I wasn't referring to your choice, more that GIB's definition is ludicrous in the first place, which doesn't help things.

1175, on 2024-May-26, 06:52, said:

So let me move on to my question - does the Robot do these simulations in real time (and how does it do it?)? It really surprised me (the first time I saw it) that the Robot will take different actions (at least in defense and declarer play) in absolutely identical circumstances.

GIB randomly generates a set of hands that it believes matches the bidding and play so far (with some leeway, as above), and uses that to make a decision. If you are playing in any form of robot *tournament* on BBO (both free and paid - Daylongs, challenges, or any other tournament setting with robots), the random seed is synchronised across tables. So if every bid and play is identical up to a certain point, that random sample will also be identical, so it will come up with the identical action at other tables.

But any minor variation will mean they're no longer synced, which is why you can get frustrating occurrences where following suit with a different small spot card can cause a contract to make at one table and not other.

I suspect none of this synchronisation applies if you're playing the anonymous IMP game, so you'll get random variation even if everything is identical.

[kontoleon]

I play with robbot by a lot, (over 2000 every month from years) and I don't even log in just because I hate to play with unknown humans. Robbot defence much deference than human, but this is not auto D. He leads usually really passive yes, and many times miss garandee and easy defence just because he put on deference spot some cards by bid.

Many times miss some great change, just take the Ace with no real reasons, and miss defence from this reason,even when he was playing attacker (dicleare)

Some times even lead singleton minor on NT, with out even i bid this one! (singleton major I do often too anyway). Many times he just miss the opportunity to go down and just duck with no real reason.

But many times he make super great defence anyway. But the real problem was too many tricks that you can easy break when you defence (with the robbot).

But is unfair to said defence D, just because he not play as human style., but I agree was real poor defence and not only defence anyway.

[thepossum]

I dislike passive leads on the whole, but if you can't think of a good lead no harm in doing anything you feel like
I am biased by not really caring about how many extra under or overtricks you can make

[1175]

The robot seems oddly random at times. What would you lead (and what do you think the robot led) on this hand?

---

I will follow up with the answer later.

[smerriman]

I assume the robot didn't make the obvious ♣K lead The fact you posted about it makes me want to lean towards one of the most unlikely leads, so I'll guess the ♠5, just because I know once in a blue moon it leads low from a doubleton and nobody knows why.

[1175]

smerriman, on 2024-June-09, 17:53, said:

I assume the robot didn't make the obvious ♣K lead

I don't know if I would describe the ♣K lead as "obvious" against a grand slam (I would certainly lead it against a small slam to try to set up a trick). It could backfire if South has a stiff club (reasonably likely) and North has AJT of clubs (close to the actual situation - South had a singleton ♣J, and North had A98, instead of AT9).

I think this auction screams for a trump lead (very unlikely to damage the defense), which the robot seems more fond of than humans in general (though, for some reason, not in this situation). Anyway, no lead can beat the contract, but (among other things) an opening lead shouldn't give declarer a map of how to play the hand. The robot led the ♥2, and I "strongly suspected" that West had the ♥Q. I therefore abandoned any thought of taking the heart finesse, and had to play for West to have four or more spades, as well (which, fortunately for me, West did). On many layouts though, the heart lead could give away the contract at trick one.

---

[smerriman]

1175, on 2024-June-09, 21:37, said:

On many layouts though, the heart lead could give away the contract at trick one.

It could resolve a guess, yes, though remember the robots assume declarer will play with all hands visible, so finding partner with Qx (or Qxx like here) will never "cost" in the robots mind, even though it will in reality. (Which is why it likes to lead singleton Ks offside too). It's probably difficult for it to find a deal where leading a heart actually costs double dummy (though point taken about the clubs).