[smerriman]

The GIB System Notes page is very clear how GIB follows to an opening lead:

Quote

When it's following to partner's opening lead, it will give an attitude signal:

• High spot card with an Ace or King
  
• High spot card with a Queen behind dummy's Ace or King
  
• Low in any other situation

Anyone who has played with GIB knows it doesn't quite work like this.

Barmar's take on things:

Quote

What's written in the notes is my best understanding of how the GIB code works. There are some comments in the code that explain what it's supposedly doing, and that's what I put into the web page.

I think there's some random variation to keep it from being too predictable, but I'm not sure. I specifically remember seeing some code that randomizes the card when splitting honors, because of restricted choice.

And uday:

Quote

Signalling is tricky. GIB signals only in very, very, very limited situations, and even there, it signals only when the the "correct" card ( which is granted a bonus ) beats the other candidates.

I downloaded a set of about 20000 hands from Arena challenges on the MyHands page to do some analysis. Since both sides of the challenge are included, this means a lot are doubled up (though that actually lead to some interesting results).

Let's solely take the case where a non-trump Ace is lead to a robot partner in 3rd position who is capable of signalling.

Case 1: Partner has King and at least two cards no higher than 8.

Expected result: GIB should play high.

Actual result: GIB played high on 305 of the 310 occasions.

Of the other 5:

3 were due to freak hands such as: holding KJ987652 over dummy's QT, when GIB played the 2.

However, this isn't an explanation, as there were 19 cases of similar freak hands (>=7 cards), such as holding KQT97652 over dummy's J83, when GIB correctly signalled with a higher card.

1 involved dummy having a singleton, when GIB played low from K653 over dummy's T.

However, this isn't an explanation, as there were 50 other cases where dummy had a singleton and GIB signalled correctly, such as this one holding exactly the same hand but over a singleton 2 in dummy:

The other involved playing low from KQ53 over 8642 against a small slam:

However, it correctly signalled in 10 other slams, eg holding KQ52 over 93.

Case 2: Dummy has King, partner has Queen and at least two cards no higher than 8.

Expected result: GIB should play high.

Actual result: GIB played high on 108 of the 111 occasions.

Of the other 3:

One was with Q74 over dummy's KJ8 against a small slam

However, it correctly signalled against the other 10 small slams, so that doesn't appear to be relevant, such as this one.

One was with QJ986 over dummy's KT43 and one with Q986 over dummy's KT; in neither case could it hurt to signal as far as I can tell.

Now of course, the fact that GIB 'normally' plays high with a K or Q over a K only helps us if we know what GIB does in other circumstances. First, does GIB signal with a Q when dummy doesn't have the K?

Case 3: Dummy does not have King, partner has the Queen and at least two cards no higher than 8. (I'm excluding cases where GIB drops the Q.)

Expected result: GIB should play low.

Actual result: GIB played low on 170 out of 184 cases.

In 14 cases, GIB played high for no apparent reason. A couple of examples:

Q964 over 85

QJ853 over T74

but signalled low in other similar circumstances, eg:

Q864 over 73

QJ832 over T9

What about if GIB has a doubleton?

Case 4: Partner has a doubleton Tx or worse, just to rule out any chance of unblocking being important.

Expected result: according to the above rules, GIB should play low; though I would be happy for GIB to play high from doubletons in suit contracts.

Actual result: GIB played low on 97 of 209 hands.

Wow. It seems GIB plays at random from doubletons. Here's an example of playing high from 63 over Q94, yet low from 63 over K98.

What about a trebleton?

Case 5: Partner has a trebleton, Txx or worse.

Expected result: GIB should play low.

Actual result: GIB played low on 234 of 263 hands.

It seems to play low most of the time, but still plays high a fair chunk of the time for no apparent reason. Here's an interesting example where it plays low when dummy plays 9 from T93, but high when dummy plays the 3 due to a slight variant in the bidding. Perhaps it sees its card are equals and thinks signalling doesn't matter anymore? (This doesn't cover all examples though).

More?

Case 6: Partner has 4 or more cards, Txxx or worse.

Expected result: GIB should play low.

Actual result: GIB played low on 185 of 223 hands.

Still a general tendency to play low, with occasional high cards for fun.

Let me check one more thing

Case 7: Partner has a singleton.

Expected result: GIB should play high, and low, simultaneously.

Actual result. GIB played high, and low, on 105 of 105 hands.

Consistency at last!

So in conclusion, if you lead an Ace, GIB will play high if:

- it has the King, except for about 2% of the time when it forgets it has the King
- it has the Queen over dummy's king, except for about 3% of the time when it forgets it has the Queen
- it has the Queen not over dummy's king about 8% of the time when it decided its Queen was really a King
- it has a doubleton, and a coin flip comes up heads
- it has something else, and simply feels like confusing you about 14% of the time.
- or maybe it just didn't have a low card to play.

[gwnn]

Awesome, thanks for this. Case 7 especially was very instructive.

[gwnn]

Any chance you could do some analysis on its opening leads?

[Phil]

Also, I am convinced of the following:

1. GIB plays UD suit preference when following suit. When I have done this myself, it tends to find the correct shift.

2. GIB will always shift from 2nd highest in middle game.

[smerriman]

gwnn, on 2017-April-24, 03:52, said:

Any chance you could do some analysis on its opening leads?

I could so - what sort of info would you be interested in? Unlike GIB's signals which are clearly buggy, I tend to trust its lead algorithm is accurate, even though assuming partner plays double dummy isn't always a great assumption. But I guess one thing that comes to mind is, say, the probability GIB has an honor when leading small against 3NT, or something like that. (AKA, why all finesses on the opening lead fail.)

[smerriman]

Phil, on 2017-April-24, 08:01, said:

Also, I am convinced of the following:

1. GIB plays UD suit preference when following suit. When I have done this myself, it tends to find the correct shift.

2. GIB will always shift from 2nd highest in middle game.

Interesting - I thought #1 that was basically ruled out as impossible (not quite sure what you mean by 2). Can you provide a couple of demo hands to make sure I'm understanding what you're suggesting - I can see if the hands I have prove otherwise.

[gwnn]

smerriman, on 2017-April-26, 22:52, said:

I could so - what sort of info would you be interested in? Unlike GIB's signals which are clearly buggy, I tend to trust its lead algorithm is accurate, even though assuming partner plays double dummy isn't always a great assumption. But I guess one thing that comes to mind is, say, the probability GIB has an honor when leading small against 3NT, or something like that. (AKA, why all finesses on the opening lead fail.)

edit: I deleted the previous post as I was just super ignorant about this!

Are you saying that this is correct?

Quote

If it leads an honor that's part of a sequence, it uses standard honor leads (K from AKx, A from AK doubleton). If it leads from a long suit, it leads 4th best (but see above: it doesn't always lead its long suit). When leading from 3 small, it leads top of nothing against NT, low against a suit contract.

If so, cool! I'm just not sure it is?

This post has been edited by gwnn: 2017-April-27, 02:54

[smerriman]

Alrighty, let's do some lead analysis!

As well as looking at when GIB 'broke the rules' as to which card to lead, I used double dummy analysis to see how well this card compared to playing the 'normal' card of the same suit.

This isn't perfect - for example, at IMPs, GIB may have simulated that a killing lead may defeat the contract enough times to make it worth it, even if that wasn't very often. Or, if not for its 'special' lead, it may have chosen to lead from another suit. And of course, simulations are telling it which suit to lead in the first place. However, I would still expect that if GIB is simulating correctly, the statistics should find GIB coming out on top when it decides to break the rules.

For now I've just analysed suit contracts (will look at no trump independently later), and excluded cases of GIB leading a trump, as that's pretty special and I didn't have as much data there. (It generally seemed to follow the same rules, other than of course the rule about not underleading aces, but still with plenty of exceptions mixed in).

The Rules

In order of priority:

1) High from doubleton
2) K from AK+
3) A from A high
4) Top of sequence of 2 from KQ+ down to T9+ (it treats 10 as an honour here)
5) Top of interior sequence of 2 (JT or T9)
6) Low from trebleton
7) Fourth high

All of these rules generally appear to apply when leading trumps as well, except as expected it skips the rule for leading A from A high. Since of course I don't have anywhere near as much data for leading trumps as non-trumps, I'll exclude trump leads from the analysis below (but similar exceptions pop up).

Now, let's see how well it performed.

1) High from doubleton

GIB had a touching doubleton on 603 occasions and always led high.

However, when the doubleton wasn't touching, in 79 out of 2379 hands (3.3%) when leading from non-touching doubletons, it lead low.

This including occasionally underleading an Ace (of the 199 occasions it lead from A8 or worse, 7 times it underled the Ace), with absolutely no logical reason to do so on the ones I checked:

low from A7, when Ace would defeat the contract

low from A8 - luckily didn't cost on this occasion, but surely would a lot of the time given this bidding/distribution

low from A2, again I can easily see it costing a trick especially given this bidding, but never gaining.

It occasionally lead low from other honours, and perhaps somewhere amongst the list there is an example or two where there is a reason for doing so. However, good luck coming up with a simulation-based excuse for leading low from 64, 74, 83, 73, and 84.

How did these leads work out for GIB from a double dummy perspective?

In 72 of the 79 occasions, the 'normal' card achieved the same double dummy score as the card GIB played. On 7 occasions, GIB's card was *worse*. On 0 occasions was GIB's card better.

Score: GIB 0 - Rules 7

2) K from AK+

GIB played the K on 1061/1061 occasions. Whew. Though, given the other results, I kind of find this surprising.

Score: GIB 0 - Rules 0

3) A from A high

Excluding doubletons and holding AK, of the other 1143 times GIB lead from a suit with an Ace, it underled the Ace on 238 occasions (21%).

Many of these are ludicrous, such as this one.

GIB seems to have a penchant for leading low from weak aces.

When leading from Axx, GIB underled the Ace on 70/182 occasions (38.5%). Using double dummy results compared to the Ace, it didn't work out well - on 52 occasions it didn't matter; it gained on 2 hands, but lost on 16.

When leading from Axxx, GIB underled the Ace on 52/190 occasions (27.4%). Dummy double results were a bit closer, but still in favour of the Ace: +6, =35, -11.

It achieved similar results for some other hands, but still at a loss.

Score: GIB 19 - Rules 48.

4) Top of sequence of 2

GIB led from a suit headed by KQ+ down to T9+ on 2480 occasions. On 38 of these (1.5%) it played low.

Most of these fit no pattern, though GIB did lead low from JTxx on 8/220 occasions.

As usual, some of the rest simply made no sense, such as this one .

Score: GIB 2 - Rules 7.

5) Top of interior sequence of 2

When the normal lead would have been top of an interior sequence, GIB decided to go against the trend on 12/213 occasions (5.6%).

On 4 occasions it played an unsupported king from a suit headed KJT - double dummy comparisons showed this didn't cost on any of the occasions, though it didn't gain either.

The other 8 involved it playing small - costing it a trick on one occasion.

Score: GIB 0 - Rules 1.

6) Low from trebleton

On 1936 occasions, GIB had a trebleton where the normal card was to play low. GIB ignored this on 44 occasions (2.3%). On these occasions it decided to lead an unsupported honour (or occasionally a 9 from J9x or 9xx). On none of these did GIB gain a trick.

Score: GIB 0 - Rules 3.

7) Fourth high

On 2270 occasions, GIB was meant to play fourth high. It played something else on 59 occasions (2.6%) - never gaining a trick.

Can a simulation really tell GIB that it's much better off leading 5 from QT875 in a side suit?

Score: GIB 0 - Rules 5.

Overall summary

Over 12164 hands, GIB 'broke the rules' on 470 occasions.

Half of these were due to it deciding to underlead an ace. Compared with leading the ace, this gained a trick 19 times, but cost a trick 48 times.

The other half came from an assortment of other situations. Compared with the 'normal' card from the same suit, only twice did this gain a trick - while it cost a trick on 23 occasions.

Final score: GIB 21 - Rules 71.

[gwnn]

So cool! Thanks a lot. wrt rule 6, you mean Txx or higher, right? It leads high from 9xx or lower I think (vs NT)?

[smerriman]

The above was just for suit contracts where it definitely plays low from xxx (as per the system notes). I'm still trying to figure the exact 'normal' rules for NT - especially with regards to honor leads involving AK/AKQ - but while the system notes agree with you, preliminary analysis suggests they are plain wrong, GIB leads low from xxx against NT as well, which kind of ruins everything.

[gwnn]

OK I misread your previous post, sorry. This is so useful. I'll wire you 10% of my future masterpoints in gratitude (don't expect too much though, it's still *me*).