Join Bridge Winners
All comments by Richard Willey
You are ignoring the author of this comment. Click to temporarily show the comment.
Looks like there are two separate discussions going on

1. How should we run seeding (in general)
2. How should the seeding system account for missing data

#1 Would appear to be a special case of “How do we develop an accurate rating system for bridge”…

(Does anyone disagree that if we had an accurate rating system, we could use this for seeding purposes?)
March 22, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
From my perspective, it makes sense to treat this as a “missing data” problem.

The field of Statistics has a bunch of different methods to handle situations where specific observations are missing from a data set. “Mean replacement” or “median replacement” are one (well known) technique.

Different techniques get used depending on why the data is “missing”. For example, there is one set of methods that one might employ if the data is “MCAR” (missing completely at random). In this example, my suspicion is that the data is not MCAR. The set of players where we are missing seeding information share certain characteristics (for example, they are new players or they are players who have not yet played in a top level event in a US national)

From my own perspective, it might make sense to use the strength of one's team mates as a predictor of the strength of the new pair. I would expect that a team that is looking to augment will invest more time / effort in investigating the (potential) skill of a new pair than anyone else. As such, there should be a reasonable good match between the average skill level of the existing team members and the folks that are being added.
March 22, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
The purpose of the letter that Jacobs is citing is to warn perspective Uber drivers that their insurance won't necessarily cover them, not to warn the passengers…
March 18, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
> Oh, convention/agreement disruption exists. It is a
> violation of the CoC .. we are required to know and
> be able to disclose our methods.

I don't think that anyone disagrees that players should be able know and disclose their methods.

What distinguishes convention disruption is the notion that some sets of methods should be treated differently than others. If your conventions are on the approved list and you screw up, its “just bridge”. If you make the mistake of playing methods that Bobby Wolff disapproves of, he will try to invent a new form of punishment…
March 18, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Given how secure the ACBL's hand generator was (perhaps is), you should be thankful that you were using shuffle and play…

Speaking of which, does anyone know whether the migration to Dib Deal actually happened?
March 17, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
“What bothers me most is that everybody, in particular
the professional players, accepted this kind of stuff
for years. Why?”

What bothers me is that so much of the current work is fixating on identifying past wrongs rather than improving the systems that are currently in use…
March 16, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
@Jurgen

The best protection against the type of issues that you are raising is an appropriate separation between the (training and test sets) and the final validation set.

On a philosophical basis I don’t tend to trust anyone’s motives. Therefore, I prefer to develop systems and processes that will work well regardless of what particular set of users might be involved. Looking at the existing process, my main complaint is that the complete dataset was publically available at the same time that the hypothesis was being developed. In a perfect world, I hope that it will be possible to identify an additional corpus of hands for a final validation round.
March 16, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
The simplest claim would be one specific card

Personally, I'd wait and see what kind of results you're getting with the validation set before deciding precisely what to use…
March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
FWIW, I strongly agree with Fluffy on this one.

Keep things simple. If at all possible, you want to develop a single, simple claim that provides the greatest degree of predictive power.

1. B+Z will go down just as hard if they are found to have one set of signals as 10
2. The more competing claims that you introduce, the more room there is for the defense to pick holes in your presentation.

Simple is better…
March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Here is an interesting inverse test.

How often does a player signal an “impossible” card?

For example, one that is visible in dummy…
March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
If you want something as simple as possible, I'd simply compare the results of your analysis to the CDF of the binomial distribution.

For any given card, there is a 25% chance that it is in the “correct” hand.

Start by checking how likely it is that that you have X successes out of Y trials.

If you want to get a bit more complicated, it should be possible to construct a 2X2 contingency table and use a Chi Square test for independence.

<Actually, given that B+Z don't always signal with a given card, the Chi Square might run into trouble>
March 15, 2017
Richard Willey edited this comment March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
@Nic

You're putting the cart in front of the horse. If / when you are able to propose an actual hypothesis, I might be able to suggest a way to test it.

Please note my use of the word “might”.

I know a number of ways to compare whether a pair of univariate distributions are the same. For example, visual inspection, and Anderson Darling test, a Kolmogorov Smirnoff test, various resampling techniques. However, if your hypothesis is more complicated then we're in the world of comparing multivariate distributions. I'm not aware of any well accepted multivariate version of Anderson Darling or the like. Ultimately, I think that any kind of testing methodology would probably need to be based on resampling techniques. Boot strap you test set and validation sets to generate multivariate confidence bounds and see whether there was much overlap. (Greg Lawler might be able to suggest something more formal)

I am somewhat surprised at the claim that all of the signal are independent of one another.

For example, do you often see cases where B/Z make the signal for clubs followed by a signal for
spade? Or makes the signal for hearts then spades then diamonds?
March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
@martin

I very much agree with your statement

1. The more complicated the model that is being proposed, the more likely that you are overfitting the data

2. The more complicated the model that is being proposed, the easier it becomes for the defense to discredit any one piece of the model

In part, this is why I am being so insistent that Nic provide a precise description of the hypothesis before moving on to the next part of the data set.

I have long held a different philosophy regarding this type of analysis than many other folks. Specifically, I don't believe that it is necessary to “breaking the code” is should be a necessary component of this process. Rather, it should be sufficient to demonstrate that the pair in question is behaving in a manner that is consistent with transmitting a signal.

With all this in mind, if I were trying to craft a hypothesis based on this data, it would look something like the following (Please note, I don't know if the following holds true, however, I consider it to be any interesting line of inquiry)

B+Z's signally method is based on the following type of method

1. A signal that shows a suit that is followed by
2. One or more signals that show honors or other information about the hands

What we expect to see is a pattern in which you have “bursts” of signaling activity that start with one of four unique signals followed by one or more additional signals. Moreover, one should not see any of the “unique” signals in sequence.

As such, I would very much focus the analysis on trying to demonstrate whether the signalling activity naturally falls into discrete “blocks” and the ordering of the information inside these blocks.
March 15, 2017
Richard Willey edited this comment March 15, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
If you can't provide a good written description of your hypothesis, you're going to have an awfully hard time convincing people that your tests have any validity.
March 14, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
> I'll let you generate the statistical model for the next step which is…

> What are the minimum number of matches between the signals
> and hands that you expect for each of the 10 signals if we
> examine the next match played by BZ in EBTC 2014.

I can start to take a look at this after I get done with work. However, before doing so, I'd like to get some data in a more useful format.

1. A formal description of each of the signals.

(I believe that you are alleging that there is a signal for each of the 4 suits as well as the cards A-9, but I want to make sure)

2. A CSV file with the following columns

A. Board number
B. Hypothesis 1: T/F
C. Hypothesis 2: T/F
D. Hypothesis 3: T/F

K. Hypothesis 10: T/F
L. Has card A of Spades
M. Has card K of Spades
N. Has card Q of Spades

AH. Has card 9 of Clubs

3. Do you believe that any of the hypotheses are independent of one another? Are there any that seem mutually exclusive. For example, I would assume that B+Z don't ever signal two suits at the same time…

4. How many boards will you be testing?

5. Do B+Z ever fail to signal? By which I mean, how often do we see hands where they have the key cards but don't bother to generate a signal.
March 14, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Nic, one weakness of the process tat is being used is that you are actively involved in leading the development of the hypothesis while having access to test / training / validation sets. You should either be doing one or the other.

Prior to releasing the test set, I would like to see a precise description of the hypothesis being tested.
March 14, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
FWIW, I think that Richard Bley is (obliquely) making a very good good suggestion.

Assume for the moment that B+Z are not cheating. Rather, they are simply very very prone to fidgeting. All other things being equal, we would expect that they types and likelihood of fidgeting would be consistent over time.

Regardless of whether they are declaring or defending, dummy or declarer, we'd expect to see similar behavior.

In contrast, if the fidgeting is design to convey information then we would expect that the types of messages to change depending on what role they are in.

If I were trying to make a case, I'd start by focusing on this rather than trying to completely break B+Z's code…
March 13, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Congratulations Justin!
March 13, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Given that B+Z were not convicted the first time around, the record is hardly perfect.
March 10, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
I am most certainly denigrating what I believe to be a flawed process that will not lead to any kind of conviction.

Doing something for the sake of doing something is a really good way to waste time and energy and can actively set back your cause.

I don't see anything wrong with pointing out that folks are on a fool's errand.
March 9, 2017
.

Bottom Home Top