[inquiry]

no doubt justin's point is correct, for what it is worth, bridgebrowser allows you to test for holding such as QJxx, KJTx, etc.. but it ignores 10's (we all know how important ten's are). So the question could be post lead for honor fifth verus honor fouth, or two honors fifth, two honors fourth, or no honor fifth versus two honor fourth, etc.

I have said all along I am not convinced by the data as there are too many other factors. I suspect when the bidding tells you to lead short, you will be often more often than not, for instance. This was more to use the new feature in response to a question jilly sent me about a post free made on his blog.

[FrancesHinden]

I think Justin's point is a very good one.

99% of the bridge playing world lead 4th highest of their longest and strongest against the auction 1NT-3NT. If they come up with a short suit lead instead, it's often for a serious reason.

I don't know what the technical term is, but there's a word for starting with

:when people have led their 4-card suit from 4-5 it gains them MPs

and arriving at

:therefore it is right to lead your 4-card suit from 4-5

[hrothgar]

FrancesHinden, on Oct 11 2006, 10:11 PM, said:

I think Justin's point is a very good one.

99% of the bridge playing world lead 4th highest of their longest and strongest against the auction 1NT-3NT.  If they come up with a short suit lead instead, it's often for a serious reason.

I don't know what the technical term is, but there's a word for starting with

:when people have led their 4-card suit from 4-5 it gains them MPs

and arriving at

:therefore it is right to lead your 4-card suit from 4-5

I just went and Goggled the expression "Logical Fallacy" and waded through several pages of different variants of "All men are mortal, plato is a man, ...)

Apparently, there are all sorts of different formal types of logicial fallacies. From my perspective, the one closest to this example is "Insignificance"

http://onegoodmove.o...llacy/insig.htm

Odds are there's something else from statistics

[cherdano]

Jlall, on Oct 11 2006, 07:29 PM, said:

As usual I have issues with coming to any conclusions from bridgebrowser studies. For instance, I am sure that when I lead a 4 card suit with 4-5 against 1N-3N my average imp score would be better than my average MP (and imp) score when I lead my 5 card suit.

Why is this? Well, I would almost always lead my 5 card suit. When I did lead my 4 card suit, I would have an excellent reason (like having QJT9 and xxxxx). Basically, when I lead a 4 card suit I am doing it for a very strong reason and expect it to be right a majority of the time. When I lead my 5 card suit it is often because I have nothing better to do, and think it's right to lead my long suit. I bet this is true for a lot of players, so basically you have severely biased data.

I think the conclusion of Ben's analysis is very simple: the average bridge players (average among the population included in the BridgeBrowser data) leads the 4-card suit a tad too rarely at MPs (but they do it too often at IMPs).

However, I am not sure this is very useful knowledge. It may be more interesting for a bridge teacher than for an expert, who already has a much better opening lead batting average than the mass of players included in the data here.

An analysis like that in a purely expert field might be interesting, though. It could tell us "when experts disagree on whether to lead from the 5-card or the 4-card suit, those preferring the 5-card suits are right more often than not".

Arend

[Echognome]

Interestingly enough, I feel this would be an ideal problem to put to a DD solver. Simply make the parameters for a 1N - 3N auction, give opener 5422 or 5431 and see what leads are best for various specifications (#tricks, %chance to defeat the contract, etc). Presumably, someone could then run this for say 10,000 or 100,000 deals (depending on how much time they have) and wouldn't it give you a fairly reasonable idea?

[Free]

I also think this needs to be solved DD, not with BB... The data is flawed, but it may be useful for other purposes.

[inquiry]

Free, on Oct 11 2006, 06:03 PM, said:

I also think this needs to be solved DD, not with BB... The data is flawed, but it may be useful for other purposes.

Curious claim. I wonder why double dummy is fine, but bridgebrowser is "flawed". You do know that over thousands of hands, bridgebrowser data and double dummy data agrees. How often have you looked at real bridgebrowser data? Do you have a subscription? Do you have a DVD of data or a couple of CD's? Just wondering on what you make this claim.

The BRBR data is as good or as bad as you can define the hand selection data. If i select opening lead from a five card suit headed by AKQJT, I suspect I will get great results for that lead, but that is not very useful.

[hrothgar]

inquiry, on Oct 12 2006, 03:07 AM, said:

Free, on Oct 11 2006, 06:03 PM, said:

I also think this needs to be solved DD, not with BB...  The data is flawed, but it may be useful for other purposes.

Curious claim. I wonder why double dummy is fine, but bridgebrowser is "flawed".

I think that that Justin's post explained the problem with using Bridge Browser to study this type of issue.

If I use a Double Dummy solver to study this issue, my input will be an unbiased sample of N hands.
If I use BridgeBrowser, the initial sample will be biased.

Garbage In, Garbage Out

[sfbp]

FrancesHinden, on Oct 11 2006, 12:11 PM, said:

99% of the bridge playing world lead 4th highest of their longest and strongest against the auction 1NT-3NT.

tch tch Frances, I can't leave that one untouched, an actual numerical claim with easily testable hypothesis

My measurements show it's in the 65-68% range. I checked the bbo main room, bbo tourneys, okbridge main room, okbridge tourneys, and this number is consistent to within 1 percent.

bbo main room is about 68%
bbo tourneys 65%
okbridge main room is about 68%
okbridge tourneys 66%

The only exception was the world championship data (very small data set) about 63%. Here it may well be that the data is woefully mucked up anyway, as the PBN files were rather disorganised, and frequently the full hand data was not available. So this one should be taken with a grain of salt. I never released this data set, there were far too many oddities and irregularities for it to be useful.

(added about 2 hours later)
I realised over supper that this 1/3 was a bit too high. Of course programmers make errors, I had been expecting somewhere in the 80's but not 99 and not 69, to be perfectly honest.

The problem was in a nutshell, there is more than one long suit in certain hand patterns - 4432 4441 5521 5530 and 6610 to be exact. And I'd saved the "longest suit" years ago without ever using the data, and without ever thinking about this problem.

Anyway it's fixed now, and 14% is a lot more than the 1% you postulated, Frances.

Sorry for any head scratching that may have gone on.
(end of edit)

I invite anyone to try it who has access to BRBR data.

Stephen

[Flame]

Can we comapare only hands where there was actually atleast one pair who lead from 4 cards and one pair who lead from 5 cards ? (if there are enough sample we can change this to more then one to avoid strange leads)
I think this way we might solve this problem and get a trustable result.

[inquiry]

Couple of issues. To flame's question, yes. If you click on the bar for leads from any suit legnth, you can call up all thos hands, create a board index with those board numbers, and then study just those boards using the full power of the bridgebrowser (spot cards, compare the lead of short on those hands to leads from the 5 card suit. Split by hcp, suit quality, etc. That call up a data set, click on one or more bars, and then re-analyze is the best way to use bridgebrowser.

To Frances, what stephen says is that mathematically, the data suggest the lead from a five card suit when also holding a four card suit is less than your 99%. Unfortuanately, the blanket look at a lead suffers from a few problems. Your five card suit maybe a five card suit promised by declarer. Your partner may have overcalled, opened or preempted in one of your short suits, etc. One has to control for all these "obvious reasons" to lead something other than the longest suit. A simple 1N-3N auction with 54 in the majors, see what the opening lead approaches solving the initial problem, but opens the door to quality of the suits in choosing which to lead.

Finally to Richard, I find the "garbage in, garbage out" potentially offensive, and I am sure Stephen pickett will find it offensive. I will assume by "garbage in" you mean if you are not careful in picking the criteria for choosing a fair hand set. IF I searched only for 4 card suits headed by KQJx and five card suit by xxxxx with RHO bidding that suit, I am sure the majority will lead the four card suit, and in the long run that will give the best results. So I will assume that is what you mean, picking the data set and how to probe it requires careful thought.. a point I tried to make in the very post in this thread.

[hrothgar]

>Finally to Richard, I find the "garbage in, garbage out" potentially offensive,
>and I am sure Stephen pickett will find it offensive.

Stephen's a programmer. I expect that he recognizes that this expression has been commonly used by programmers for decades.

>I will assume by "garbage in" you mean if you are not careful in picking the
>criteria for choosing a fair hand set. IF I searched only for 4 card suits headed
>by KQJx and five card suit by xxxxx with RHO bidding that suit, I am sure the
>majority will lead the four card suit, and in the long run that will give the best
>results. So I will assume that is what you mean, picking the data set and how
>to probe it requires careful thought.. a point I tried to make in the very post
>in this thread.

You were the one who rather peevishly asked why BridgeBrowser was a poor choice for this type of study. The precise quote was

>Curious claim. I wonder why double dummy is fine, but bridgebrowser is "flawed".

The answer should be rather obvious...

You always cite how wonderful BridgeBrowser's automation is. However, in this example you are going to be forced to perform manual inspections of all the hands that you're looking at. You haven't significantly advanced the state of the art from the techniques that Paul Marston used 15 years ago when he was studying this same problem.

In contrast, the combination of a random dealer program and double dummy solver permits you to start with an unbiased sample and to completely automate the process. In turn, you can inspect orders of magnitude more hands.

[sfbp]

Yes, I find your attitude offensive. Certainly not to be answered in its current truculent state.

I recall when you were one of the biggest advocates of this idea. I hope others who believe your grand plans for the future of bridge note how you seem to behave when you get tired of *their* idea. Caveat Fred!

[hrothgar]

sfbp, on Oct 12 2006, 04:48 PM, said:

Yes, I find your attitude offensive. Certainly not to be answered in its current truculent state.

I recall when you were one of the biggest advocates of this idea. I hope others who believe your grand plans for the future of bridge note how you seem to behave when you get tired of *their* idea. Caveat Fred!

Stephen, I have always maintained that Bridge Browser is a great product. I still do. There are any number of tasks where BridgeBrowser is incredibly valuable. For example:

I often like to examine my board results playing MOSCITO and calculate my average score and standard deviation for different opening bids. I normally find that the 1♦ and 1♥ openings score are net winners, my strong club openings don't score particulary well, my 2♦ is a big winner but has a high standard deviation and so on. Bridgebrowser is phenomenal for this type of analysis.

In a similar fashion, many people like to understand where their good scores come from. Are they great at bunny-bashing? Alternatively, do they play a slow steady style that picks up a small number of IMPs across the board. Once again, Bridgebrowser is great for this type of calculation.

However, the fact that BridgeBrowser is the best solution for some types of analysis does not mean that it is appropriate for all types of analysis. Suppose that I went and build the best lawnmower that the world has ever seen. This doesn't mean that I should mount it in my living room and use it as a ceiling fan. Furthermore, if I tried to market my lawnmower as a ceiling fan I'd expect to get called on it.

Bridgebase has some enormous limitations. The most significant of which is that your database doesn't contain any information regarding the partner's definition of different opening bids. If you see a 14 HCP NT opening, theres no way to understand if the partnership is playing a 12-14 HCP 1NT opening and this is a good maximum or, alternatively, whether the partnership is playing a 15-17 HCP NT and the player decided to upgrade a "stellar" 14 count. I have consistently maintained that this issue severely restricts the type of analysis that you can do with BridgeBrowser.

Personally, I think that BrBr (or some similar product) has a bright future. I'm not sure whether it is ever going to be a great money maker - I don't think that Bridge Software has produced many millionaires - however, I think that the ideas that you pioneered are going to make a very significant contribution to bridge.

Lets look forward a few years and assume that you're able to interface BrBr with whatever Full Disclosure has mutated into. In an ideal word, that 14HCP 1NT opening would be tagged with a unique identifier. In a similar fashion, all my Frelling style 2♦ openings would be tagged with their own unique identifier.

Once you have this type of feature in place, you increase the power of BrBr enormously. Case in point: Assume that I open 1NT. BrBr could automatically search for all examples of 1NT openings with the same tag, and create a summary statistic describing the range of the opening bid. Maybe there will be a little graph with a probability density function documented expect hand strength. Potentially, the opponents could even specify whether the PDF should be based on High Card Points, Zar points, or whatever.

[Echognome]

I just wanted to add that I'm not slighting bridge browser either. I think Ben's attempts to analyze the data quite interesting. I am only stating that for this problem I find double dummy analysis to be more accurate. It is not that I think double dummy analysis is perfect either. It is just that we know what we are getting with double dummy, whereas with empirical bridge data, we have one major problem; namely selection bias. If you can adjust or correct for that, then I will be behind the numbers all the way.

The main problem comes when the treatment (in this case leading from the 4 card suit in with a 5 card side suit) is correlated with another hidden variable (here I believe it's suit quality OR it's the unknown auction OR it's major versus minor OR it may be MPs vs IMPs). In the case of the 5 card major in 1NT (which I thought was an excellent idea to consider) the hidden factor was that expert players are more likely to open 1NT with a 5 card major than beginners. I appreciate on this example you controlled for IMPs versus MPs. You can also control the auction by considering auctions such as 1NT-3NT. It seems you can also control for suit quality if you spent a little time. Then you just need to control for major or minor, as auctions such as 1N-3N typically mean that responder does not have a 4 card major, and most definitely not a 5 card major.

I think the main point is that if brbr is to be useful for answering more general questions (and I have no doubt it is useful in looking at your own system), then the studies themselves have to be done carefully and controlling for as many factors as you can.

As per Richard's suggestion, I think that being able to filter (or as Richard used 'tag') our data based on what system we are playing would be a big plus. (Enough so that I would purchase the software outright.) But of course that would take some integration with bbo software. Say that we could keep track of how we did within one system. Perhaps there is a way around that by being able to look only at systemic bids, e.g. 1♦ opening with 9-15 hcp and 4+ hearts, but even then we may pick up some hands that have diamonds where we were playing a natural system all along. I don't know if this is possible or not, but it would make it easier if we could just select our system ahead of time and then all hands we played using that system are marked as such. I believe the easiest solution would be when we load an FD file, it sent a tag and the name of the FD file was stored on the database it would be sufficient. I also believe that those that play more artificial systems are more likely to want to use bridge browser as a tool. However, they are also much fewer in number! So, I can understand if you are not so concerned about catering to their needs.

[inquiry]

hrothgar, on Oct 11 2006, 10:31 PM, said:

You always cite how wonderful BridgeBrowser's automation is.  However, in this example you are going to be forced to perform manual inspections of all the hands that you're looking at.  You haven't significantly advanced the state of the art from the techniques that Paul Marston used 15 years ago when he was studying this same problem.

Quite to contrary, you can automate essentially all the steps... if you are willing to frame the question appropriately. Let's take Flame's question as a jumping off point...

Here is a logical sequence of test, that apply very little initial selection criteria to see how things might progress... you can apply a hell of a lot more criteria if you would like. We will start with a simple search where leader against 3NT has any number of hcp but specifically 5431 distribution, and is allowed to lead any card. Let;s see how this is done..

From the contract menu, pick specific 3NT (the only green box in the list of contracts), i also excluded doubled and redoubled contracts, and today, I choose present data from the defenders point of view (see black box in the upper window below). I choose a single distribution for leader (5431), I could have choosen any two, three, four, up to all possible distibutions if I wanted, for instance, I cuold have choosen 5431 and 9112 for leader. I choose leader to have from 0 to 37 hcp, I could have choose 0 to 2 or 4 to 9 etc. See the green rectangles in the middle of the upper window.

I then searched a BBO database and stopped it when about 1/4 finished with the search. Since this was main room, it is not surprizing there was very few matchpoint hands, but I wanted them to illustrate a later point. But the results of the discovered 25044 hands where leader had 5431 on lead against 3NT, the results for the DEFENDERS this time are plotted in the lower window on the left hand side of the image.

But lets look at the top window for a minute. Each row is one hand where someone lead against 3NT. The hand correspondng to the highlighted hand is displayed on the lower right hands side.. I erased a little data that is hiddend by the yellow box in the hand diagram related to players rating. If you look at the one line, you can compare it with the actual hand. All the contracts were 3NT, we forced that, since we searched by defender on lead, the column headed by HP will list that defenders hcp, on the displayed hand, it says he held 7 hcp (green square) and the line points to him, showing he did have 7hcp. The declarers combined hcp is listed in the column headed HPT, it says they had 26, and if you run down the red line to look, they do (15 opposite 11), the last competitive bid other than pass (with respect to declarer;s side) is listed as 2H (colum headed (Cmp), and if you look, one defender did bid 2H (red line connects them). Note Pas(s) only shows up in this column if the "defenders" never bid. The vul (column Vi) is them (since we searched by defender), and the opening lead was the king of hearts (connected from top window to bottom by blue line).



For completeness there is a few more columns, Trt is trumps for the declaring side and DT is defender's turmps. In NT of course, these are zero. DT and DO are distributional points for declaring and defending side, and once again, by convention, these are considered 0 if in NT.

Now 25,000 hands is too many to much around with unless you have already applied all the criteria you were going to apply. Here we applied basically none. There is a lot of things you can do with this data on the screen, for instance, you can sort by any column header (player, lead, total declarer points, opneing leaders hcp (in this case), or competitive bid. For example, you easily pick only hands where the defending side didn't bid by sorting by the Cmp column and using your mouse to highlight and then display just the ones with pass (click first, then shift click the last in the listing).

But lets proceed with Flames question, cna you look at just the hands where people lead from one of the short suits. The answer to that is yes, easily. On the bottom left hand chart (the lead chart), just click on the bar for imp results when leading from a three card suit. When you do that, you will get only the hands that lead from a three card suit (in this case 1731 hands)

When you plot the leads, you find, not surprisingly, that there were 1731 leads, and all were from 3 card suit, Note the matchpoint hands are gone, because I just clicked on the imp bar. This isn't so much useful at first glance but there is a lot of things you can do with this data. For one thing, you can create a board file, so that future inquiries will be restricted to only these deals (thus, any lead from a five card suit or 4 card suit will be from the same hand a lead from a 3 card suit occured). You can push these hands over into bid analysis and open up the full power of investigating by sequence, suit quality, tricks taken, etc, as I do in this example For instance, lets look at the lower left hand window in the diagram above. First, we should see that the total number deals in 1731 there too... (the little box a few inchies to the left above the blue highlighted tab that says bid analysis, Second, if we look at the table in that same window, we see alisting of opening bids, starting with pass, with 1NT the last opening bid readable, but you can also see part of 2C. The number to immediate right of those bids indicate opening bid by the patnership that willl eventually defend on this hand (since we searched origianlly by defender). So you see that the hand was opened by the defenders in 1H and 1S a total of 458 times. By clicking on those two bars, an dyou can run analysis on how well third best lead works after your side has opened in a major. The next graphic gives that information in two formats...



Above we see when we grab just the hands that open 1M and they play 3NT (again our leader has to be 5431), you can see there was the 458 hands (in a number of places you can see that number). In addition, the results show that after opening 1H the average result in imps for leading from a three card suit was -0.38 imps (top row in the red box in the column headed AvIMP), and for a spade lead a more profitable 0.16 imps. Since we plotted a 3 card lead after opening either 1H or 1S, we see that leading a three card, we see the average result was -0.16 (not surprising, multiple 222 by -.38 and 236 by .16, sum the results and divide by the number of hands (458).

Of course a lot of other factors might go into the figure the result. One such factor might be quailty of the suit. Let's take a look at the quality of the 1S suit opened versus the result...To get the following chart, simply click on the 238 next to 1S and choose plot, then from the options (tricks, hcp, contract, suit, etc...) pick suit.



This chart shows the quality fo the 1S openers suit. I then clicked on the bars that included at least two of the top five honors ,,, but had to include Ace or King. There were 132 such hands, and the average result for the defenders when the suit had 2+ honors including the A or K was a pllus 0.41 imps.

So I will go back and restate the obvious. It is up to the person using BRIDGEBROWSER to frame the question being asked appropriately. Worried about justin's concern that an honror lead from a sequence in a short suit might take precedence over low card from longer suit? Then on the contract page, choose a lead where the card can not be higher than an 8 for instance (see the green box on the top window in the middle). Worried that partners bid might affect the lead, only look for leads where partner has not bid. Create several search results and build board files then probe deeper using just those hands. Wondering rather leading a short suit when weak is better than leading your long suit opposite a passing partner? Restrict leaders hcp and force partner to pass, and compare with similar leads when you have more hcp. A blanket which lead is best with no other application to control for significant variables, is of course, "garbage in", but you can fine tune your requrirements of litterally tens or hundreds of million plays... and should be able to get useful data... For instance, I think I demonstrated that people tend to open 1S with 5-1-6-2 when weak and 1D when stronger in antother thread using bridgebrowser. That was obvious, but we could put some number on the frequency.

[inquiry]

BTW. you can do esoteric things as well. For example, for those of you who think the average player on BBO or OKB is too weak, and so will lead randomly, you can do this...

1) go to player tab, enter minimum rating for opponents (say 55 == which is low level expert, high level advanced, or 60 which is expert, or 65 which is world class).
2) do contract search as played by declarer. This will restrict the oopening leader to having a rating of at least that high (54.5 rounds up to 55, etc).
3. Then build a board index using those hands, and then choose 3NT defend, with the rating index again at the same level. This will give you just the hands where declarer and leader have whatever minimum range you want. At the very least you can check to see why such players choose an odd lead like a singleton (partner bid suit most likely cause), or 3 card suit lead.

The screen looks like this...where I set the minimum rating to 55....


BTW, this is the search screen most players use, and they never get past the entering their own name (or fulvio2002 or fred) in the player field. You can provide virtually unlimited seach criteria (like adding minimum lehman like rating to the results found), suit quality, opening bid(s), vul, range of cards lead, hcp ranges, etc.. and you can then take the found data and prune it and subdivide it even further.

[Cascade]

Free, on Oct 12 2006, 11:03 AM, said:

I also think this needs to be solved DD, not with BB... The data is flawed, but it may be useful for other purposes.

There is another solution we can use single dummy results.

Somewhere I have over 100000 hands played by GIB on the auction 1NT 3NT. On each hand I forced the opening leader to lead from each suit - GIB didn't mind playing the hands four times.

I think from memory I restricted the opening leader to hands with no voids and no six-card or longer suits.

I will try and post some of the results later but I have a busy day and I am away tomorrow so it might not be before the end of the weekend.

Single Dummy is much slower than Double Dummy so is not so popular but does not suffer from some of the objections of DD and BB studies.

[barmar]

One problem with trying to use DD analysis for this is that the DD solver cheats when deciding the opening lead, so that will presumably bias the results. It always chooses the "right" lead. This is kind of related to Justin's point about leading from his 4-card suit when there's something about the suit that makes it inherently more attractive than the 5-card suit.

What I think you need to do if you're going to test this hypothesis is force a particular lead, then perform DD analsys starting from trick 2 (or from declarer's play of the dummy on trick 1). For every hand in the collection, first have opening leader lead from his 5-card suit, then start again and have him lead from his 4-card suit, then compare the results.

[Al_U_Card]

If I follow, we have indicated that DD analysis gives about the same result as actual play.....

Is not DD analysis, the "perfect" solution and play is the real-world approximation?