The ridiculousness of some double dummy analysis

[thepossum]

Hi all

 

I know DD is central to a huge amount of computational bridge, and also analysis of human bridge performance. Many people also know that it is extremely limited in true analysis of hand potential scores and contracts and should be used with caution. It concerns me that it seems so integral to much research, analysis and development and that seemingly many people are unaware of the limitations.

 

I have been testing out many of the products/tools on the market, some better than others in how they handle these limitations. However it does concern me when major products come up with totally ridiculous par scores and contracts that no two partnerships (of any level) would ever reach. What is the point of a piece of software saying that you missed a totally and unfeasible par score.

 

And it doen't just concern me in Bridge. There are many more "serious" areas of life and work where the kinds of people who do DD analysis are determining critical things in people's lives, how to fly aeroplanes, drive cars, whether people are entitled to government benefits, insurance, trading with our pension funds, predicting health care, involvement in all areas of government policy, healthcare, economics, finance etc. Very dangerous indeed. And please note, I dont necessarily mean the people who build the stuff and know all the limitations; its the less informed people who use hi-tech products/analysis in all areas of our lives who do not understand limitations, who work in organisations where technical limitations are ignored or silenced etc etc

 

PS I'm not trying to donwplay the seriousness and importance of Bridge :)

 

regards P

[LBengtsson]

I work in Algo-Trade (investment bank) and me and my traders all fear what happens if stock market go again into Bear territory.. Human response to falling shares, equities can be rationale. Algorithmic Trades in Bear market is more fluid, instant. Who knows if we let computers finally run ALL our lives. Meltdown???

[helene_t]

DD analysis is objective - we can discuss the meaning of correlations between dd results and other variables without the complications of different ways to obtain sd results.

 

Also, dd software is freely available, multi platform and very fast.

 

More research on how to calibrate dd results to come closer to sd results would be useful.

[barmar]

While DD analysis is not always perfectly representative for any particular hand, I think the expectation is that it's just as likely to err one way or the other. So these errors cancel out, and statistics based on DD results are reasonably valid.

[hrothgar]

It concerns me that it seems so integral to much research, analysis and development and that seemingly many people are unaware of the limitations.

 

 

Can you provide any examples of which you speak?

In my experience, the debates around double dummy methods have been pretty robust and anyone serious is well aware of the examples in which DD analysis is likely to be biased.

 

There are many more "serious" areas of life and work where the kinds of people who do DD analysis are determining critical things in people's lives, how to fly aeroplanes, drive cars, whether people are entitled to government benefits, insurance, trading with our pension funds, predicting health care, involvement in all areas of government policy, healthcare, economics, finance etc.

 

There are certainly concerns about inappropriate use of computer algorithms.

Conflating this with double dummy analysis feels like you have a hammer and are searching for a nail.

[rhm]

It seems to me that the opener has not understood what DD analysis is all about.

 

Let's make it clear: Double dummy analysis is not the same as single dummy analysis.

However, DD analysis is available as software and is fast, while there is still no single dummy analysis as software, which can match good players analysis.

 

DD analysis on printed handouts are only a courtesy and of course need to be taken with common sense.

 

When it comes to double dummy statistics the following has been observed.

 

When a deal is played at many tables in a certain contract the average result is frequently very close to the double dummy result.

This has been checked on millions of deals played online.

 

This does not mean that the deal was played double dummy neither by declarer nor the defense.

It just means that DD analysis is usually a good predictor for the result when the deal is played.

 

This makes DD analysis good for simulations, for example to find out what a hand is worth in certain bidding scenarios.

For example should I bid game with a specific hand when partner say opens 1NT.

You could randomly deal thousands of deals ,where your hand is the one you are interested and which gives partner a 1NT opening and see how often 3NT would make double dummy with the particular hand in question.

 

On any one deal the double dummy result may be wrong but over many deals the result will be quite precise what would have happened on average at the table.

 

Rainer Herrmann

[TylerE]

Complaining about double dummy analysis being "unrealistic" is like complaining that 2+2=4.

 

It is what it is, by definition.

[Trinidad]

If you are not happy with double dummy solvers, it is perfectly fine to feed a problem to a single dummy solver. I have often used Jack to analyze bridge questions.

 

The problem is that Jack is not free software. Another is that it doesn't mimic your actual opponents (but what software would mimic Aunt Agatha?). On the other hand, Jack's declarer play and defense are realistic (much better than e.g. GIB).

 

Rik

[Cyberyeti]

The key is recognising when double dummy analysis doesn't work.

 

Conversations like

 

"We missed 4♥ on board 10"

"But double dummy says it doesn't make"

"Yeah, but that requires an opening lead of small diamond from Axxx instead of the really obvious lead"

 

don't do double dummy analysis much good, but as long as you can ignore them and also accept where larger scale analysis is not as useful as it might be (hands where there are 2 way guesses mainly), then you can get useful data from it.

[johnu]

If you are not happy with double dummy solvers, it is perfectly fine to feed a problem to a single dummy solver. I have often used Jack to analyze bridge questions.

 

Are you sure Jack has a single dummy analyzer? It was my understanding that all the bridge playing programs used a Monte Carlo double dummy analysis to determine play.

[Vampyr]

DD analysis is objective - we can discuss the meaning of correlations between dd results and other variables without the complications of different ways to obtain sd results.

 

Also, dd software is freely available, multi platform and very fast.

 

More research on how to calibrate dd results to come closer to sd results would be useful.

 

Sure, by having the solver take into account the best percentage lines by declarer. But determining the percentage play would require a lot of information about the auction, so an analysis would not apply to every table.

 

Note that when people give each other single dummy problems, or post them on these forums, they give the auction unless it is truly irrelevant. That is not often.

[etha]

Are you sure Jack has a single dummy analyzer? It was my understanding that all the bridge playing programs used a Monte Carlo double dummy analysis to determine play.

 

 

 

 

jack can play the hands. If there is a two way guess he will take the most likely guess. That removes most of the perceived problems from the dbl dummy analysis. it is more of a problem in defense where Jacks biggest weakness is probably not always giving declarer a guess in some positions where it can. it doesn't use pure monte carlo. for example if there was ever a two way guess that you can always get right double dummy using pure monte carlo it wld I think never take the guess knowing it can always get it right next trick. You can see this in some primitive programs that cash aces to delay the guess they know is "working". jack has some in built methods for avoiding this.

[barmar]

There's a related thread on Bridge Winners: Can we trust the statistical analysis in the Bridge?. This specifically discusses using DD analysis to determine "declarer's advantage" -- the advantage declarer has compared to defenders, because they can see all their assets. Someone has actually compared a large number of hands played by an expert player on BBO (Jimmy Cayne) and compared his results to DD. There's a small DA for NT and low-level contracts, and not much for slams (I suspect this is mostly because bidding is pretty accurate at determining whether slam is good, so there's not much need for clever declarer play). This information can be found in the related thread Statistical Analysis of "Declarer's Advantage".

 

Anyway, the conclusion is that since DA is fairly small, DD analysis is statistically useful. While the study only tests DA, it's not unreasonable to extrapolate to other aspects of the game.

[cherdano]

I think the OP has a good case.

I mean, there is a book that claimed you can decide whether to lead high or low from QJxx against NT based on double-dummy analysis. Well, anyone who has thought about this holding for 2 minutes will realise that leading low will gain or break even compared to leading high when playing single dummy due to likely declarer misguesses. (Say declarer has A8 in hand opposite KT2 in dummy - double dummy it doesn't matter what we lead, single dummy leading low will likely gain. Conversely, with Kx in dummy opposite Txx in hand we in theory have to lead high, but in practice we will also succeed by leading low. Etc.) A clear case where adding double dummy analysis not only doesn't add value, but actively makes us dumber.

 

So you'd think that this book would have been laughed out of the room. Instead, it's the most influential book on opening leads in a while...

[cherdano]

Rainer has posted something like the following many times - yet it's wrong, and I think it's worth one more effort to explain why:

When it comes to double dummy statistics the following has been observed.

 

When a deal is played at many tables in a certain contract the average result is frequently very close to the double dummy result.

This has been checked on millions of deals played online.

This is correct, but it does not imply anything of the following:

This does not mean that the deal was played double dummy neither by declarer nor the defense.

It just means that DD analysis is usually a good predictor for the result when the deal is played.

 

This makes DD analysis good for simulations, for example to find out what a hand is worth in certain bidding scenarios.

If I predict every day of the year that there is a 50% chance of rain in Edinburgh, then *on average* my prediction will be very accurate. That does not mean at all that my prediction is useful whatsoever.

 

Similarly, if on average declarer makes as many tricks single dummy in 4H as double dummy, that does not mean the same will be true when he has a 4-3 fit where he needs to guess whether to play for trumps 3-3 or to play on a side suit to break.

[smerriman]

So you'd think that this book would have been laughed out of the room. Instead, it's the most influential book on opening leads in a while...

I don't think it has been influential because everyone thought the double dummy analysis had no limitations (as per OP) and realised they were leading wrongly. All of the comments on it I've seen have been fully aware of the limitations and the fact the specifics are not practical, but it has still been useful from a high level.