[hrothgar]

I've long been skeptical about rating system. I still am. However, this idea may have caught my fancy

While posting on a different thread, I had the germ of an idea. I’m not sure to what extent people would find this sort of functionality interesting or not. If there is sufficient interest, I might consider trying to build something.

I propose the following: Integrate “practice room” type bidding with some kind of cluster analysis algorithm (K-Means Clustering, a Classification Tree, a Neural Network or some such)

The idea is fairly simple. People use the web site to bid hands against one another. (A real live person will be sitting North, South, East, and West). A large number of players will bid each of the boards using a specified standard system like BBO Advanced. Each player will always sit in the same direction.

Once we have a sufficient large sample (say 250 North bidding 250 boards or some such) we can then start to crunch the data and start to group players by their bidding choices. Hypothetically, we might conclude that players, (1, 8, 13, 15, 23, 36, 40, 42, 43, 49, …) all bid in a similar manner to one another. Players (2, 57, and 138) form their own isolated little group… yada, yada, yada.

I’d be interested to determine whether

1. It’s even possible to create consistent clusters
2. Whether these clusters would map onto anything useful (skill level or some such)
3. Whether we could train a neural net to perform this type of analysis.

Imagine a situation where a new player joining BBO was presented with a bidding quiz. The results of this quiz could then be used to recommend an appropriate cohort of potential partners…

For example, you might be told that you are part of the “Gazelle” cohort or the “Llama” cohort or some such and that you might enjoy partnering other llamas….

Come to think of it, the precise same techniques could be applied to card play choices and the like.

Regretfully, I don’t think that Bridge Browsers database has enough consistent comparisons to enable this type of data mining application…

[helene_t]

Not sure why you would expect to find clusters. A cluster would be something like "wj players" or "acol players". If all people are supposed to use the same bidding system, clustering would be quite minimal, I think, although you might be able distinguish those with SA as their native language from those who have learned some rudimental SA but still often default to their, say, Acol background. But even among SA-adopted acolists you would see a continuum of more or less adapted acolists.

Those cultural things nothwithstanding, I would expect one big multi-dimensional cloud with the dimensions corresponding to:
- skill level
- scientific vs fast arrival
- aggressive vs conservative
- traditional vs modern

and a few more dimensions. It would be quite fun to identify those principal components (and/or clusters), I would expect bridgebrowser to have plenty of data for that purpose but I might be wrong, I haven't thought of a particular method for identifying those components. If everyone has been faced with the same, say, 250 decisions then it's easy just to feed the matrix to a principal component procedure. What we do have is a very sparse matrix, not sure how to work with that.

[barmar]

While exploration of the data might reveal some clustering, I'm not sure how this would translate to a rating system. I suppose you could correlate the clusters with some other standard -- e.g. if there's a cluster containing players like Meckwell, Zia, Lauria, Versace, Helgemo, etc. we might consider everyone in the cluster to be world class. But outside this upper echelon, I'm not sure what you would correlate with.

[cherdano]

barmar, on Feb 6 2008, 03:24 PM, said:

While exploration of the data might reveal some clustering, I'm not sure how this would translate to a rating system. I suppose you could correlate the clusters with some other standard -- e.g. if there's a cluster containing players like Meckwell, Zia, Lauria, Versace, Helgemo, etc. we might consider everyone in the cluster to be world class. But outside this upper echelon, I'm not sure what you would correlate with.

I am pretty sure there is no cluster containing Meckwell, Versace, Zia and Helgemo

[han]

Richard are you actually suggesting this as something practical, or is this thread just a theorectical discussion?

[matmat]

hrothgar, on Feb 6 2008, 03:19 PM, said:

For example, you might be told that you are part of the “Gazelle” cohort or the “Llama” cohort or some such and that you might enjoy partnering other llamas….

I'm more of an Okapi myself, though some liken me to a Duckbill Platypus.

what about people from other parts of the world where 2/1 is not necessarily the system of choice? what about all those people whose bidding is OK, but card-play exquisite?

Why would my ranking have to depend on the ability of whoever happens to be sitting opposite me in the ranking/bidding booth?

[helene_t]

A simple idea:
- Collect for each pair of players (not "pair" in the usual bridge sense!) the data points (N,m) where N is the number of situations they have both faced and m is the number of times they made the same decision.
- Fit a beta-binomial model to the data
- Use, for each pair, the fitted beta model as a prior for p in the model m~binom(N,p)
- Compute for each pair the posterior mean of p.
- Now you can either do a hierarchical clustering based on the posterior p-means or you can somehow "translate" the posterior p-means into a covariance matrix and do principal component analysis.

[Codo]

If I understood your model right, its aim is not to judge the players into exellent/advanced/poor etc.

But the goal is to put them in different camps. F.E. all Gazeles rebid 2 NT after 1 M 2 m to show 12-14. All antilopes rebid 2 M with this hand.
Other points are the strength of their weak two and three openings, the use of Lebensohl etc. Errors are another point too, but not the only thing.

So if I decide for 4sf is just invitational, Lebensohl, etc, and my error rate is about x %, I am a rhino.

When I search for partners on bbo, I will find other rhinos quite easy. And I will have more fun with other rhinos then with crocodiles.

Sounds like a nice idea.

[helene_t]

Or maybe on the BBO profiles replace the flags with "favorite animal". Would be less controversial and maybe just as fun.

[hrothgar]

han, on Feb 7 2008, 01:24 AM, said:

Richard are you actually suggesting this as something practical, or is this thread just a theorectical discussion?

From my perspective, this thread is currently sitting in the fuzzy boundary between theory and practice.

I suspect that the main reason that I started think about this is that I am playing product manager for a honking big piece of Statistics software and I suddenly have access to Neural Networks, Clustering algorithms, PCA tools, Partial Least Squares code, and a 1001 different tools that could be used to analyze various bridge related questions.

Its an interesting thought exercise to think about how this might be applied... With this said and done, if we were able to come up with something "good" that looked like it might met some real market need, I'd be happy to try to code something up in MATLAB.

I suspect that my original idea about creating an permanent floating appeals committee would probably be far more useful. However, I fear that the folks who would have the most to gain from such a system - small local events - would probably be the least likely to use it...

For what its worth, I like Helene's suggestion about looking at the branching depth

[Al_U_Card]

Bidding systems are many and varied.

So are defensive methods.

Hand play is...errr...welll...hand play. That is where any comparison could and should be made.

[hrothgar]

Al_U_Card, on Feb 7 2008, 05:18 PM, said:

Bidding systems are many and varied.

So are defensive methods.

Hand play is...errr...welll...hand play. That is where any comparison could and should be made.

I agree with you in principle, however, declarer play and defense if often predicated on the specific auction. It's unclear to me how best to handle this whole issue (plus there is that whole sticky issue of signalling agreements, defensive carding, and the like)

In theory, you might be able to get around some of these issues by providing everyone with the same canned auction and then letting them focus exclusively on declarer play and defense.

In my original posting, I recommended that this would work best if everyone trying to use the same standard system...

[Al_U_Card]

hrothgar, on Feb 7 2008, 09:59 AM, said:

In my original posting, I recommended that this would work best if everyone trying to use the same standard system...

Kind of like some IQ tests, they would be biased towards those more familiar with the methods in question.

Bidding and defensive "signaling" are languages and thus developmentally dependent. Declarer play obviously depends on effectiveness of defense, but against set hands with set defenses (or by GIBs) the better players would rise to the top.

Declarer play would certainly seem to be the place to start, at least....

[gwnn]

helene_t, on Feb 7 2008, 03:41 PM, said:

Or maybe on the BBO profiles replace the flags with "favorite animal". Would be less controversial and maybe just as fun.

"♠A???????? crazy emu!"