[fuzzyquack]

kgr, on 2024-January-09, 11:11, said:

This 2S seems not optimal to me.
http://tinyurl.com/yrf6btz7
Pass if no DBL seems better, and RDBL if DBL.

2♣ seems to be not optimal either.

[lorserker]

fuzzyquack, on 2024-January-09, 13:03, said:

2♣ seems to be not optimal either.

2C is possible.
2S from North seems really bad.

[thepossum]

Can I ask the difference between GiB - the intelligent one - and Ben - the artificially intelligent one?

[fuzzyquack]

Why bot plays the hand differently after the same defense yielding a near top to a few:
http://tinyurl.com/2x8tx2ct
http://tinyurl.com/yqe8yclv

[smerriman]

fuzzyquack, on 2024-January-10, 00:53, said:

Why bot plays the hand differently after the same defense yielding a near top to a few:
http://tinyurl.com/2x8tx2ct
http://tinyurl.com/yqe8yclv

South differed at trick 3.

smerriman, on 2024-January-09, 01:49, said:

Also, I'll post the hand after the tourney is complete (you'll probably see it yourself), but on the latest tourney, Ben responded to Blackwood to show an even number of keycards, and a void.

Unfortunately it didn't have either of those things

---

I guess Ben doesn't have the same concepts of captaining and when you're allowed to differ from book bids that GIB does in its database?

[fuzzyquack]

thepossum, on 2024-January-09, 17:02, said:

Can I ask the difference between GiB - the intelligent one - and Ben - the artificially intelligent one?

I believe Ben can be trained while GIB uses a fixed simulation algorithm. Therefore, you may expect the same ridiculous plays from GIB at any time while the Ben's ridiculousness will evolve.

[lorserker]

thepossum, on 2024-January-09, 17:02, said:

Can I ask the difference between GiB - the intelligent one - and Ben - the artificially intelligent one?

Most important is that they are both very intelligent

For differences in playing ability and style, there is something written here

For explaining differences in how the bots work technically, I will assume that you already know how GIB works, and will highlight where Ben is different.

Bidding
Ben doesn't have any bidding rules programmed, all bids are given by neural networks which are trained on data (trained to bid like Gib but we have also trained it to bid like Argine and to bid like people in other versions). GIB has all bidding programmed and there are over 10 thousand bidding rules.
If you are in a well-defined situation that is covered by a convention, then it's better to do what GIB does. If you are in a situation that isn't covered by system notes, then it's better to do what Ben does.
Basic GIB always makes the bid it is programmed to make (aka book bid) and never simulates. If there is no book bid, then it passes. Advanced GIB does simulate sometimes when it has options, but the simulation is expensive because it's based on double-dummy analysis.
Ben always simulates, and it simulates on a lot of samples, because simulation is cheap for Ben as it doesn't use double-dummy analysis to do it.

Sampling
When sampling hidden hands GIB asks "is this hand possible in the given auction - yes/no".
Ben asks "given the auction, how likely is this sample?". This makes sampling simpler in Ben.

Opening Lead
The neural network proposes some possible leads, then we simulate. As opposed to GIB, the simulation is not double dummy based, so we can do more of it and faster.

Card play
The neural network suggests possible cards to play in a situation (together with the scores showing preference).
Then we do a quick search to decide which candidate card is best. For this we don't use any double dummy either.
The neural network for the card play was trained based on human play.

[fuzzyquack]

[quote name='smerriman' timestamp='1704872250' post='1060537']
South differed at trick 3.
Does it make the defense different?

[helene_t]

lorserker, on 2024-January-10, 05:10, said:

Sampling
When sampling hidden hands GIB asks "is this hand possible in the given auction - yes/no".
Ben asks "given the auction, how likely is this sample?". This makes sampling simpler in Ben.

This sounds as if Ben understands restricted choice while gib does not. Is that correct?

[sandlzrd]

i find that ben is basically a novice it makes plays that older bots do not do i think the millions of hands you have compiled to train ben were from a weaker field than i normally like to play i'm not interested in ben at this point however i lobe bbo

[wvbg]

In an 8-board session 1)robot switched suits at trick 2 vs 3NT allowing contract to make (similar to GIB); 2)played high from touching honors when winning a trick after partner's lead; 3)overbid to 3 poor slam contracts. Not close to usable at this point.

[P_Marlowe]

Hi,

if you train the bidding, how do you inform the human, what a specific bid means?

Does Ben give some rough shape / range information?

Or does Ben determine, that a specific contention is now active and from that follow the responses?

With kind regards
Marlowe

PS I find it great, and I do think it is sensible to try to replace the GIB with something that is
actively worked on.

[plucian]

I am stuck on board 6 after the biding ended. Maybe I made a bid that broke the Ai. Also, I think the Ai should be trained on a lot of defence boards and less on biding.

[lorserker]

plucian, on 2024-January-11, 14:19, said:

I am stuck on board 6 after the biding ended. Maybe I made a bid that broke the Ai. Also, I think the Ai should be trained on a lot of defence boards and less on biding.

really sorry about this.
i was trying to fix this bug today and i thought i did, but apparently i didn't.
new try tomorrow.

[lorserker]

wvbg, on 2024-January-11, 12:30, said:

In an 8-board session 1)robot switched suits at trick 2 vs 3NT allowing contract to make (similar to GIB); 2)played high from touching honors when winning a trick after partner's lead; 3)overbid to 3 poor slam contracts. Not close to usable at this point.

thank you for your feedback.
if you shared the links to the hands where it messed up, it would help me with debugging.

[lorserker]

helene_t, on 2024-January-10, 21:03, said:

This sounds as if Ben understands restricted choice while gib does not. Is that correct?

i haven't thought about it this way, but yes, perhaps it is.
i'm just not confident enough to say that it definitely understands restricted choice and then on the next hand it turns out that it doesn't

[pilowsky]

lorserker, on 2024-January-11, 14:55, said:

i haven't thought about it this way, but yes, perhaps it is.
i'm just not confident enough to say that it definitely understands restricted choice and then on the next hand it turns out that it doesn't

So it randomly uses restricted choice 50% of the time?

[kcostell]

This was from yesterday's Daylong:

http://tinyurl.com/mtmtpfhs

With 11 HCP and AJT96 in Spades, Ben first passed my 2S overcall of 2H then balanced with 3S (3+ spades, 3-8 total points) after opponents competed to 3S. Both seem like pretty gross underbids.

In general, it's felt like the BBO bots (not just Ben) tend to undervalue long trump support

[smerriman]

Yikes. And East opening a weak two then raising his own preempt.. this one was clearly trained on very weak human pairs :/

[undoubling]

IMO the term "AI" is HALF right------------it IS Artificial BUT it is NOT "Intelligent"!! Intelligence IMPLIES the ability to 1) MAKE mistakes and 2) LEARN from them!!! AND AI does NEITHER!! It is constrained by the CODING!!! And the CODING doesn't CHANGE!!!
(At least NOT on it's own, or at it's OWN volition).!!