[uday]

Being bored, and trying to postpone walking the dog in 34-degree weather, I looked at the set of deals in our live database


There are about 137 K deals in there. Mostly very recent deals, a handful of manually uploaded goulash deals, no doubt, but almost exclusively hands dealt by our dealer, including many 'best hand' deals


When South had the AQ of spades ( 6774 times ), west had the king 33% of the time and north 33% ( rounding to integers )

When West had the AQ of diamonds ( 4998 times ), north had the K 32% of the time, S had the king 36% of the time

When North had the AQ of hearts ( 5054 times), west had the K 31% of the time, south 37%


I'm fairly confident it is only bec of best hand that S has the K more often than the others but I don't really see a problem here. If I had to, I could find all the original poster's hands from our archives and do the same thing on those but that would be a bit of a chore.

Anyway - it was an interesting exercise, not that I know if i demonstrated anything

[tx10s]

1eyedjack, on 2015-January-22, 13:50, said:

I am missing something but I don't understand why (excluding total columns) there are 4 columns when I would expect 2.

Every finesse that is working for one side is losing for the other. Say you (human) are South, North has AQ and East has K. This is a working finesse for the robot, failing for human.

So I would expect this to count +1 in both the "player lose" and "robot win" columns. And I would expect this double entry to balance for each example, so that the total for the player lose column should never depart from the same total for the robot win column. But in your spreadsheet these totals do differ. What's that all about?

The last two columns cover the total. The first set of 4 columns just break down the finesses by hands played by the player and hands played by the robots, thus the last two columns are sums as you describe.

[tx10s]

hrothgar, on 2015-January-22, 19:36, said:

Here's the rub...

People are very good at finding patterns in data. If you spend long enough staring at a random set of hands, odds are your going to find something weird about it.

On this set of hands, you're upset because the finesses don't work
On that set of hands, you're annoyed because E/W gets more HCPs than North South
On some other set of hands, the clubs never break well

It is for this reason that you start by specifying a hypothesis, and then test your hypothesis using a completely different data set.

This notion is fundamental to any kind of serious analysis...

I have a hard time taking your comments seriously. I simply looked at various randomly selected tournaments to determine how the finesses were splitting. They were looked at after the fact without recording which actual finesses were taken. I may be only a Bronze Life Master, but I can guarantee you that I can identify finesses and which direction they favor. The fact that the split so consistently averaged 40% over so many sets of tournament is a very strong indication of built in bias. The fact that the split "miraculously" changed to 50% after my posting leaves a strong indication that someone found the bias and "fixed" it. I suspect it was an input variable to the dealing program, but obviously I could never prove it.

[barmar]

tx10s, on 2015-January-23, 13:59, said:

The fact that the split "miraculously" changed to 50% after my posting leaves a strong indication that someone found the bias and "fixed" it. I suspect it was an input variable to the dealing program, but obviously I could never prove it.

There have been no recent changes to the dealing code. The last change was when we added the best-hand feature a few years ago.

What you're probably seeing is "clumping". Random values are not always evenly distributed in any particular sample, there are often bursts of one type or another. What's important is that you can never predict them.

http://en.wikipedia....oisson_clumping

This is a common misconception that many people have about random data. If you look at a picture of lots of dots, and it looks uniformly fuzzy, it's probably NOT random. A random picture would have several concentrations of lightness and darkness.

[hrothgar]

tx10s, on 2015-January-23, 13:59, said:

I have a hard time taking your comments seriously. I simply looked at various randomly selected tournaments to determine how the finesses were splitting. They were looked at after the fact without recording which actual finesses were taken. I may be only a Bronze Life Master, but I can guarantee you that I can identify finesses and which direction they favor. The fact that the split so consistently averaged 40% over so many sets of tournament is a very strong indication of built in bias. The fact that the split "miraculously" changed to 50% after my posting leaves a strong indication that someone found the bias and "fixed" it. I suspect it was an input variable to the dealing program, but obviously I could never prove it.

I'll see your "bronze life master" and raise you an MIT graduate degree...

The only thing that your posting are doing is demonstrating that you don't even understand what issues are being discussed, let alone being able to make any meaningful contribution. The insinuation that the dealing code was "fixed" based on your expose would be hysterical, if it weren't so sad...

[barmar]

hrothgar, on 2015-January-23, 16:48, said:

I'll see your "bronze life master" and raise you an MIT graduate degree...

Not to mention the years you spent developing statistical software at The Mathworks.

[uday]

Richard's credentials are impeccable but ... kumbaya isn't just a river in Egypt, as they don't say.

I've seen nuttier bugs than this. I'm convinced there is no bug here, but perhaps I need to demonstrate this with hard data from the archives.

When we released our html-based rewrite of our 'Just play bridge' game, someone complained that we kept redealing the same hands ( and that other than that, the game was fun). Took us a while to realize he was serious, but turns out that under certain cache settings, some players were getting the same hands redealt a few times.

uday

[hrothgar]

tx10s, on 2015-January-23, 13:59, said:

The fact that the split so consistently averaged 40% over so many sets of tournament is a very strong indication of built in bias. The fact that the split "miraculously" changed to 50% after my posting leaves a strong indication that someone found the bias and "fixed" it. I suspect it was an input variable to the dealing program, but obviously I could never prove it.

Here's a less conspiracy minded explanation regarding what happened

1. You noticed a pattern in a given set of hands
2. The pattern was real, but it happened by chance
3. Accordingly, when you examined a larger set of hands the pattern did not repeat itself

If you look back across the thread, you'll note that all my posts emphasize the need to test your hypothesis using a fresh set of hands.

[barmar]

To make my point about clustering more concrete, suppose you threw a coin 1000 times, and the result is like this:

TTTTTTT H T H TTTTT HH T H TTT H T HH T H T H TT HH T HHHH TTTT H TT HH TTTT H T H TT H TTTTT HHHH TT HHHHH T HHHH TT HHH TTTT H T H TT H TT H T HH
H TT H T H T HH TT H T H TTTT H T H TTTT H T HH TT HHH T H TT HH T HH T HHHH TT HH T H TTTT H T H T H TT H TTT H TTT HHHHHH T H TTTTT HHH T H T H T HH T H TT
HHHHHHH TTT HHHH T HHH TTTT HHH T HH T HHH T HH TT HH T H TTT H TTT H T HH TT H T H T H TT H T H T HHHH T H T H T H TT HH TT H T HHHHH T H T H TTT H TTT H
H TTTTT HH TTTT H TT H T HHH TTT HHH T HHHHH TT HHH TT HH T H T HH T HHHH T HHH TTTTTTT H T HHH TTT HH T H TTTTTTT H TTT HHH TTTTTTTT HH T H
H T H T HHH TT H TTT HHHH TT H T HHHH T HHH T HH TTTTT H TTT H T H TTT H T HHH TTT H TTTTTTT H T H TTT HHH TTT H TT H T H T HHH T HH TTT H TT HHHH TT
T HHHH T HHH TTT HH TT H TT HH T H T H T HH TT H T HH T H T H T H TTT H T H T HH T HHHHH T HHHHHH T HHH TT HHHHH T H T HH T H TTT HH TT HH T H TT HHHH TTT
HH T HH T HH T H TT HH T HH TTTT HH T H TTT H T H TT HH T HHHH TTTT H T H TTT HHH TT HHHH TT H T H TT H T HH TT HH T HH TTT HH TT HHHH T HHH TTT HH TTT
HH T H T H T H TTT HH TT HH T H TTT HHHHHHHHHH TTT H TTT HHHHHH TT H T HHH TTTTTT H TT HH TTTTT H T H T H T HH TTT H T HHH TT H TT H TTTT H TT HHH
TTTT HHH TTTTTT HH T HHHH TT HH T H T H T H TTT HH T HH T HH TT H TTT HHH TT H TTT HHH TTT HHH T H T H TTT H TT H TTTTT HH T HH T H T H T H T HHHH TT H
TT H T H T HHHH TTT H TT H TTTTTT HHH T H T H T HH T H TTTTTT H TT HHH T H T H T H T H TTTTT H T HH TTT H TTTT H T H TT HH TT H T H TTTT H TTTT H T H TT HH

If you asked most people if this looked random, they'd probably say no, because they see lots of long runs of the same face: two runs of Tails near the beginning, two runs of Heads in the middle of row 8. But this is actually the expected nature of random sequences. If there weren't many sequences like that, it would have to have memory so that it could avoid generating the same results that have been produced recently. If each event is independent there is no memory, so nothing prevents clustering like this.

This incorrect intuition is the source of the"gambler's fallacy": if it's been a while since you threw Heads, then Heads are "due" so you should bet on it.

[tx10s]

barmar, on 2015-January-23, 17:46, said:

To make my point about clustering more concrete, suppose you threw a coin 1000 times, and the result is like this:

TTTTTTT H T H TTTTT HH T H TTT H T HH T H T H TT HH T HHHH TTTT H TT HH TTTT H T H TT H TTTTT HHHH TT HHHHH T HHHH TT HHH TTTT H T H TT H TT H T HH
H TT H T H T HH TT H T H TTTT H T H TTTT H T HH TT HHH T H TT HH T HH T HHHH TT HH T H TTTT H T H T H TT H TTT H TTT HHHHHH T H TTTTT HHH T H T H T HH T H TT
HHHHHHH TTT HHHH T HHH TTTT HHH T HH T HHH T HH TT HH T H TTT H TTT H T HH TT H T H T H TT H T H T HHHH T H T H T H TT HH TT H T HHHHH T H T H TTT H TTT H
H TTTTT HH TTTT H TT H T HHH TTT HHH T HHHHH TT HHH TT HH T H T HH T HHHH T HHH TTTTTTT H T HHH TTT HH T H TTTTTTT H TTT HHH TTTTTTTT HH T H
H T H T HHH TT H TTT HHHH TT H T HHHH T HHH T HH TTTTT H TTT H T H TTT H T HHH TTT H TTTTTTT H T H TTT HHH TTT H TT H T H T HHH T HH TTT H TT HHHH TT
T HHHH T HHH TTT HH TT H TT HH T H T H T HH TT H T HH T H T H T H TTT H T H T HH T HHHHH T HHHHHH T HHH TT HHHHH T H T HH T H TTT HH TT HH T H TT HHHH TTT
HH T HH T HH T H TT HH T HH TTTT HH T H TTT H T H TT HH T HHHH TTTT H T H TTT HHH TT HHHH TT H T H TT H T HH TT HH T HH TTT HH TT HHHH T HHH TTT HH TTT
HH T H T H T H TTT HH TT HH T H TTT HHHHHHHHHH TTT H TTT HHHHHH TT H T HHH TTTTTT H TT HH TTTTT H T H T H T HH TTT H T HHH TT H TT H TTTT H TT HHH
TTTT HHH TTTTTT HH T HHHH TT HH T H T H T H TTT HH T HH T HH TT H TTT HHH TT H TTT HHH TTT HHH T H T H TTT H TT H TTTTT HH T HH T H T H T H T HHHH TT H
TT H T H T HHHH TTT H TT H TTTTTT HHH T H T H T HH T H TTTTTT H TT HHH T H T H T H T H TTTTT H T HH TTT H TTTT H T H TT HH TT H T H TTTT H TTTT H T H TT HH

If you asked most people if this looked random, they'd probably say no, because they see lots of long runs of the same face: two runs of Tails near the beginning, two runs of Heads in the middle of row 8. But this is actually the expected nature of random sequences. If there weren't many sequences like that, it would have to have memory so that it could avoid generating the same results that have been produced recently. If each event is independent there is no memory, so nothing prevents clustering like this.

This incorrect intuition is the source of the"gambler's fallacy": if it's been a while since you threw Heads, then Heads are "due" so you should bet on it.

Well Mr. MIT (I graduated from Cal Tech myself), you apparently need a basic lesson in probability. As you know, with one finesse, (or coin flip), there are two possible outcomes, Onside (I will call Y, and Offside, I will call N). With two finesses, the combinations are NN, NY, YN, and YY, thus 25% will have no finesses, 50% will have 1 finesse and 25% will have 2 finesses. I am sure you understand that this can be expanded to any number of finesses, and a basic Excel spreadsheet can calculate the number of combinations for an number of finesses. I found that 20 tournaments had an average of about 280 finesses. 280 finesses have a total of 1.94267E+84 possible combinations for the 281 total outcomes (0 to 280 Y) These average 280 finesses had an average of 112 finesses in favor of the player. There are 9.5056E+80 combinations for 112 or less finesses. Thus, the chance of getting 112 or fewer finesses on side out of 280 total finesses is about 1 in 2000. My data showed the same 1 in 2000+/- outcome (within a fairly narrow margin)all 9 separate times. For your example, the chances of getting 400 or fewer heads in 1000 coin flips is about 1 billion to one. In other words, the greater the number of coin flips or finesses, the less likely the final count will be very far outside (as a percentage of the total) from the middle outcome, as in a bell shaped outcome distribution. When I did the analysis, I was not sure what the outcome would be, I just had a feeling that the hand distributions were not random, and unlike others, I actually accumulated the data to see what the distributions actually were. Your problem is you are looking at small cluster distributions rather than total outcomes, which is what I did.
If you would like to tell me what is wrong with my calculations, I am willing to listen. If you think I am lying or did not obtain the data correctly, I posted two sets of 20 tournaments with my calculated outcomes. I do not know if you are associated with BBO, but someone from there can double check my calculations if they want to.
BTW, I am not assuming some conspiracy that the finesse distribution just happened to change after my post. I had complained to BBO support service about the distributions and they said that their programmers would look at the issue, so I am assuming someone looked and "corrected" the situation. Since BBO has long insisted that the hands were totally random, I would not expect them to ever admit that they made the change.

[Bbradley62]

tx10s, on 2015-January-24, 20:23, said:

barmar, on 2015-January-23, 17:46, said:

To make my point about clustering more concrete...

Well Mr. MIT (I graduated from Cal Tech myself), you apparently need a basic lesson in probability...

Hrothgar is Mr. MIT, not Barmar. I guess they don't teach Reading Comprehension at Cal Tech.

[Antrax]

I think some introductions might be in order. Barmar is one of the "programmers who will look at things". uday is another. They both work for BBO and they'd know if the code changed since your complained about it, as they're the ones who'd have changed it.
Hrothgar is one of two posters I know of in BBF which line of work was doing the sort of analysis you're trying to do here.

To address your point, as someone who has considerably less background than people here, it seems they're saying that if you look at all the hands BBO has dealt to best hand tounaments, and consider splitting them into subsets based on the user who declared. That's like looking at clumps in (hopefully) random data, and what they're saying with the coins and all that is that you shouldn't expect each clump to be uniform. In other words, at some point in time for some user, more finesses would succeed or fail than would seem likely a-priori. You apparently were that user before, but there's no reason for you to expect it to continue, as you see now.

[1eyedjack]

I have no credentials. I did a bit of maths including stats up to age about 20, and have lived longer than that again without any practice.

It seems to me that if your sampling method is sound, and your sample size is adequate, then it should be possible to draw a conclusion, subject to a level of confidence, regarding whether the underlying data is random, by limiting your observations to that sample without requiring further observations to be made.

No doubt the making of further observations, which are consistent with your earlier conclusions, would add confidence to your conclusion, as would a larger sample size in the initial exercise. But if the initial sample size is sufficient, I would expect the "clumping" effect to dissipate.

Is that wrong?

[Antrax]

Think of it this way: dozens of people have looked at the clump they define, and everything was okay, so they never posted here. Thus, seeing one guy complain and drawing conclusions is like throwing a million-sided die and then being impressed by a one in the million chance occurring.

[hrothgar]

Bbradley62, on 2015-January-24, 22:47, said:

Hrothgar is Mr. MIT, not Barmar. I guess they don't teach Reading Comprehension at Cal Tech.

Actually, Barry also went to MIT (I originally met him at the MIT bridge club)

Can't swing a dead cat in the New England Tech industry without hitting an MIT grad...

[hrothgar]

tx10s, on 2015-January-24, 20:23, said:

Well Mr. MIT (I graduated from Cal Tech myself), you apparently need a basic lesson in probability.

I'm not arguing about your knowledge of probability, but rather your sampling methodology.

You constructed one data set that shows a result.
You are unable to replicate this same result with other, more recent data sets (and claim that the dealing algorithm has changed)
Other people have also constructed surveys based on your claim and have produced different results.
Color me unimpressed.

I haven't carefully re-read your postings, however, as I recall you never even provided the raw data that you were working from.
(You provided summary spreads that you said demonstrate that the deal's are weighted)

A few years back, there was a program being discussed on this site called Bridge Browser which maintained a pretty extensive database of hand records.
As I recall, Inquiry has a copy.

It should be possible to mine this database and perform a pretty detailed statistical analysis of historical records...
(This is not to be construed as an offer to do said work, especially since I am still unburying myself from a month in Vietnam and Myanmar)

[ArtK78]

hrothgar, on 2015-January-25, 06:02, said:

Actually, Barry also went to MIT (I originally met him at the MIT bridge club)

Can't swing a dead cat in the New England Tech industry without hitting an MIT grad...

Is swinging a dead cat a popular pastime in New England?

[hrothgar]

ArtK78, on 2015-January-25, 06:11, said:

Is swinging a dead cat a popular pastime in New England?

Its a popular expression...

I recall a very silly discussion in which a bunch of us tried to figure out how wide an area swinging a dead cat would cover.
We eventually decided the the key variables were

Angular velocity
Time since death
Temperature

[Bbradley62]

hrothgar, on 2015-January-25, 06:35, said:

Its a popular expression...

I recall a very silly discussion in which a bunch of us tried to figure out how wide an area swinging a dead cat would cover.
We eventually decided the the key variables were

Angular velocity
Time since death
Temperature

No mention of size of the cat or arm length of the person doing the swinging?

[tx10s]

Bbradley62, on 2015-January-24, 22:47, said:

Hrothgar is Mr. MIT, not Barmar. I guess they don't teach Reading Comprehension at Cal Tech.

You are correct, I did miss that. Your nasty comment is noted but totally irrelevant.