[thorvald]

---

Trying this on the 2012-version North bids 5♥ over 4♣. Bidding 7♥, simply just seems wrong

This was from a challenge, and I gained 3 IMP by accepting I would go down

[smerriman]

thorvald, on 2022-September-02, 12:08, said:

Trying this on the 2012-version North bids 5♥ over 4♣.

Not too relevant here, since it also plays 4♣ as a help suit slam try rather than a cue bid.

But it's the exactly the same as discussed in previous threads; the GIB database will never, ever, bid a grand slam unless holding the appropriate combined number of points. North can tell that's impossible in this situation, so every single simulated bid will lead to an identical result of 6♥.

(And indeed, over the help suit slam try, 5♥ ties 4N and 6♥, it's only picking 5♥ based on an arbitrary ordering. It didn't consider 7♥ then because it wasn't allowed to).

If enough hands where you can make grand turn up, then it will leap to 7♥, since that will work better on average. It is not capable of realising that if you go slow, you can simulate to find 7♥ if it's best later.

[thorvald]

smerriman, on 2022-September-02, 14:58, said:

Not too relevant here, since it also plays 4♣ as a help suit slam try rather than a cue bid.

But it's the exactly the same as discussed in previous threads; the GIB database will never, ever, bid a grand slam unless holding the appropriate combined number of points. North can tell that's impossible in this situation, so every single simulated bid will lead to an identical result of 6♥.

(And indeed, over the help suit slam try, 5♥ ties 4N and 6♥, it's only picking 5♥ based on an arbitrary ordering. It didn't consider 7♥ then because it wasn't allowed to).

If enough hands where you can make grand turn up, then it will leap to 7♥, since that will work better on average. It is not capable of realising that if you go slow, you can simulate to find 7♥ if it's best later.

When simulating and calculating that on average 7♥ will give net positive larger than 6♥, it probably lacks the the percentage used to bid grands, by pure odds it is 2-1, but in reality most want better odds - and it is probably not calculating the odds for getting doubled.

[smerriman]

thorvald, on 2022-September-02, 15:45, said:

When simulating and calculating that on average 7♥ will give net positive larger than 6♥, it probably lacks the the percentage used to bid grands, by pure odds it is 2-1, but in reality most want better odds - and it is probably not calculating the odds for getting doubled.

No, all of that part works perfectly fine. It doesn't need to use human methods for working out what percentage it needs to make to be profitable; all that is computed from the simulated hands and theoretically should work better than the human methods.

(But of course, as usual, a low number of simulations is not always 100% accurate).

[thorvald]

smerriman, on 2022-September-02, 15:50, said:

No, all of that part works perfectly fine. It doesn't need to use human methods for working out what percentage it needs to make to be profitable; all that is computed from the simulated hands and theoretically should work better than the human methods.

(But of course, as usual, a low number of simulations is not always 100% accurate).

The simulation does not take into factor that the other table might stop in Game. I just had such a board tonight, where I calculated the odds in favor of a grand. It went down and my opponent played 4♥

---

7♥ looks to be with odds.

I also tried to let the bot bid my hand, and noticed, that after the response 6C, there was no simulation and the bot just bid 6H

[smerriman]

thorvald, on 2022-September-02, 17:51, said:

The simulation does not take into factor that the other table might stop in Game.

Right. But there is no chance a robot could ever figure out what humans at other tables would do; trying to add that in would more likely cause disasters than solve anything. Assuming the same contract at the other table is very reasonable given the robots' consistency.

[thorvald]

Adding a rule that a grand needs 75% chance of winning before considering it would be enough

[smerriman]

Hardcoding a completely arbitrary figure like that in doesn't really make any sense at all to me, and wouldn't fit into the entire concept of the way it bids, so I can't agree with you there.

[thorvald]

smerriman, on 2022-September-03, 04:18, said:

Hardcoding a completely arbitrary figure like that in doesn't really make any sense at all to me, and wouldn't fit into the entire concept of the way it bids, so I can't agree with you there.

Using Milton Points, distribution points etc are al arbitrary values :-)

Learning about risk - reward in gaming is very interesting

If I offer you a bet on toss of a coin giving you odds 6:5 on tails you would (and should) take it, but a factor might be how much is at stake.

There has been a lot of studies in that area - what is your limit on these odds (a car, a house or something even more valuable) ?

But as we both know the chances of getting any changes in GIB looks to be non-existing

[thepossum]

How many hands does it simulate to have any real chance of bidding 7 (1 or 2 perhaps)
I appreciate there are thousands of tables being played simultaneously. Trying to do some sums on CPUs and threads and processing time and how much is actually done on servers rather than on client machines?
One would think the requirements to simulate all likely contracts once beforehand should be enough - with a decent number of sims - not every time
Maybe BBO use a small cluster of Pentiums

As a small test I reduced my number of simulated hands to around 10 and was starting to get a few >60% chances and only used a few percent of a second of an 8-thread processor - bot sure exactly how much resource it used. It takes a lot of hands to get 10 though

Does it have any rules at all (common sense) or is everything just DD sims

Surely South could be thinking along which slam to go for but not North

[smerriman]

thorvald, on 2022-September-03, 04:33, said:

Using Milton Points, distribution points etc are al arbitrary values :-)

Precisely, that's the point - simulations are all about trying to avoid using these arbitrary values / rules of thumb that humans have come up with as much as possible.

There was a classic example on BridgeWinners a while back when, defending 6NT at IMPs, a robot defender won a trick, had the setting trick in hand, but failed to cash it, despite the fact it was clear that doing so would allow the contract to make on some layouts (and make it did).

Humans would always cash the winner, because your only aim is beat the contract.

The robot saw that cashing the winner would guarantee down 1, and its return would sometimes allow the slam to make.. but it would also take the contract down 2 or more tricks a significant proportion of the time; enough to average out to a better IMP score, and was actually theoretically better.

Introducing hardcoded values would result in so many more bugs - the most obvious would be how are you going to define sacrificing anyway, or a two-way shot for sacrificing when it might make? Or situations when it's completely clearcut both tables will be in slam, and all you're doing is programming in a losing strategy.

Agree there is definitely a lot of interesting aspects of game theory here, but while there are lots of things I'd change about GIB, but I wouldn't want to go near this one. Especially when the hand in question is easily resolvable via other means.

[thorvald]

smerriman, on 2022-September-05, 00:54, said:

Precisely, that's the point - simulations are all about trying to avoid using these arbitrary values / rules of thumb that humans have come up with as much as possible.

There was a classic example on BridgeWinners a while back when, defending 6NT at IMPs, a robot defender won a trick, had the setting trick in hand, but failed to cash it, despite the fact it was clear that doing so would allow the contract to make on some layouts (and make it did).

Humans would always cash the winner, because your only aim is beat the contract.

The robot saw that cashing the winner would guarantee down 1, and its return would sometimes allow the slam to make.. but it would also take the contract down 2 or more tricks a significant proportion of the time; enough to average out to a better IMP score, and was actually theoretically better.

Introducing hardcoded values would result in so many more bugs - the most obvious would be how are you going to define sacrificing anyway, or a two-way shot for sacrificing when it might make? Or situations when it's completely clearcut both tables will be in slam, and all you're doing is programming in a losing strategy.

Agree there is definitely a lot of interesting aspects of game theory here, but while there are lots of things I'd change about GIB, but I wouldn't want to go near this one. Especially when the hand in question is easily resolvable via other means.

Yes, I remember that board, and Bots have one advantage over humans. A bad board will never affect the next boards :-)

And yes, game theory is very interesting.

But I don't think you can just base the play purely on statistics.

As an example:

You are playing against a vulnerable 4Ma and make a simulation (with 20 boards), where 1 card will defeat the contract in one simulation, but giving away an overtrick in 15 other.

Statistically you will lose in the long run playing that card, but when you look in the greater perspective, it is rare that the Imp will cost the match, but 13 Imps might well be decisive :-)

This is in fact the other side of playing safe, where we pay a small insurance to make sure the contract will win against a bad layout.

And I agree, that it is not where I would start if I ever got the chance to improve GIB