tysen2k

Partner could have 2 aces like Luis suggested (if that's your style not to open a weak 2). He could also be something like 6-5 in hearts and a minor. Some people prefer not to open these hands and come in later. Slam might be possible, but I'm not going to try for it. I bid 4♥ and put as much pressure on the opponents as possible. From their point of view it might sound like a sacrifice and they might double.

If I have 2N as a puppet to 3♣ then I'll take it, otherwise I pass even if it is a reverse. We have a total misfit and need to get out ASAP.

Sure. It could also set up 4 tricks, 1 trick, or none. My evaluation programs take each of these situations and multiplies it by the probability of it actually happening. So you get a weighted average of the benefit. As you gain more information about your partner's hand, your evaluation will get more and more accurate. If you've learned that your partner has enough outside strength and support for you to safely establish a 7-card suit, then the probability of extra tricks goes up and your evaluation goes up appropriately. Tysen

I am actually working on this problem right now. I should have something to post within a week or so. I'm looking at quantifying the adjustments that should be made to evaluation based on the opponent's bidding. I agree, I've seen no one quantify this before except for these rules of thumb people keep passing around. "Add a point with AQx if RHO bids it," etc. I'm working on the data now but it seems a lot more complicated than that. For example, AQx is worth more than AQxxx since you're going to run into shortness in both partner and LHO in the second case. I have done work on quantifying how evaluation changes when partner mentions a suit (how much does Qxx in trump support go up in value?): Improving Hand Evaluation Part 1 Improving Hand Evaluation Part 2 These articles are a little long and technical, but I'm also working on a "for dummies" version of these findings that can actually be used at the table. I might post that in another week or so. I've been finding time and again that hand evaluation is very dynamic and these articles attempt to quantify this revaluation based on statistics instead of just guessing that "this hand should get better by 2 points because of these honors." Tysen

My only advice if you're trying to invent a response system w/o a negative response is make sure you are able to stop in 1NT when that's the right place to be. I personally use 1♦ negative and a transfer positive system similar but not identical to Meckwell.

This was the full hand: [hv=d=w&v=n&n=skqxxxhktxxdtcqxx&w=sjhqxxdkjxxxcatxx&e=satxxxhajxdaxxckx&s=sxxhxxxdqxxxcjxxx]399|300|Scoring: IMP[/hv] I chose to pass. Down 4 for -800. At the other table, South faced the same situation and chose to make an SOS redouble. After North bids 2♥, E-W decided not to go for the penalty and ended up in 3N making 5. Afterwards I was wondering if maybe Redouble is best. I know that spades are splitting 5-1 or 6-0 with the spades behind my partner. He might not make many tricks at all. Even though rdbl bumps it up to the 2-level, I see 2 added bonuses: trumps might not split as badly opps might overbid us (like at the other table) I'm still not sure what the best action is in general, but I was curious what the concensus here was.

Against opponents who know what they're doing: [hv=d=w&v=n&s=sxxhxxxdqxxxcjxxx]133|100|Scoring: IMP[/hv] LHO PART RHO YOU 1♦ 1♠ Pass Pass Dbl Pass Pass ?

I picked 4♦ too. 1♦ is a big mistake.

Ah, but it does seem to work. If you look at how the different distributions actually behave, you'll see this. Look at this segment of my previous data for hands with voids only: Real Zar 5-3-1 8-4-2 5-4-4-0 1.519 1.200 1.667 1.600 6-4-3-0 1.624 1.600 1.667 1.600 5-5-3-0 1.643 1.400 1.667 1.600 7-3-3-0 1.697 1.800 1.667 1.600 (To refresh your memory this is the number of tricks you are better than a 4333 hand). These are vastly different shapes but they all practically take the same number of tricks. The fact that some have only 5&4 card suits and others 7 card suits doesn't seem to make much of a difference on their trick taking ability. Zar's predictions are all over the map but the ones that count shortness only are right on. Every method that counts points in some way will have certain hand types with the same value. For example Zar assigns the same value to 6331 and 5440 even though they have vastly different trick taking abilities (and 5332 with 4441). Tysen

I really don't want to use fractions of a point either. I can't handle those fractions at the table (although I do know some people who love them). I myself have been using straight HCP+531 and then just making a mental note if I have lots of aces or queens in my hand. It's not as accurate as either BUM+531 or Zar, but it's pretty close to Zar in terms of accuracy. The thing that pushes me to use just HCP instead of the more accurate Zar is the fact that I'm familiar with the ranges of numbers. My system is written in terms of these familiar points and I don't want to have to convert my weak NT into 25-28 or whatever it turns out to be. However, I've been playing around with alternatives in case you are comfortable with Zar's point ranges. Here's what I've found. Zar is very accurate in terms of high card values. The ratio of 3-2-1-0.5 has been known for decades. It's the distribution scheme of Zar that loses the accuracy. So my alternative is to use Zar's 6/4/2/1 for high cards and one of these distribution schemes: 8/4/2 for void/singleton/doubleton or 1*longest + 1*next - 2*shortest Both of these are practically the same in terms of accuracy so take your pick. They are both right between Zar and BUM+531 in terms of accuracy. Note that for the first case you'll have to add a base of 8 points to get back to Zar's scale. In the second case you'll have to add 6 points. Alternatively you could subtract 16 or 12 points from your requirements for game, slam, etc. So if you pick the second option, you'll be using 40 for game, 50 for a slam and 55 for a grand. Hey, nice round numbers... :D Tysen

Okay I've split the data that I posted on my yahoo site by tricks (the tricks the evaluator predicts, not the actual tricks taken). Here are the results: 8 9 10 11 12 13 Overall HCP 1.21 0.43 0.69 0.73 0.59 N/A 0.66 HCP+321 1.07 0.31 0.65 0.68 0.54 0.40 0.59 HCP+531 1.05 0.26 0.62 0.69 0.61 0.93 0.56 Zar 1.05 0.27 0.62 0.70 0.62 0.60 0.56 BUM+321 1.07 0.30 0.63 0.66 0.55 0.50 0.57 BUM+531 1.04 0.24 0.60 0.66 0.60 0.59 0.54 These numbers represent the average number of tricks in error that each evaluator predicted away from those that were actually taken (smaller is better). Note that this is from the sample of hands that always take at least 9 tricks (thus the 1+ error that is always seen when only 8 tricks are predicted). I'll use this dataset since it's the one that has consistantly put Zar in the best light that I've seen. HCP by itself never predicts that it will make a grand (the highest combined on these 13,000 hands is 34 HCP). This is another artifact of limiting the hands by eliminating those hands where NT is the best contract (again to make Zar look better). Also notice that the evaluators that only use 321 for distribution look like they are the most accurate for slams. Don't let this fool you; it only means that if the hands have enough points to predict a slam with only 321 then it's a pretty solid one. It misses a lot of potential slams when it predicts less tricks. Comments are welcome Tysen

Yes it does. Someone (I think Peter Cheung) did a large study of okbridge hands and compared how single-dummy declarers did compared to double-dummy data. He found that overall, the DD result was within 0.1 tricks of the SD result. At high contracts, DD declarer has more of an advantage (can drop those queens). At partscores the DD declarer was at a disadvantage, and at games both DD and SD were very close. So DD data does have its disadvantages, but it's the best we have right now. Just understand its limitations and always use common sense.

It's sort of backwards to look at the hands that make a certain number of tricks and then look at a standard deviation of points (normalized or not). It's better to look at the collection of hands that your evaluator predicts belong at a certain level and then look at the standard deviation of the obtained tricks. So instead of looking at hands that take 10 tricks and then seeing how many hands have >51 Zar, we should be looking at those hands that have 52-56 Zar and looking at the standard deviation of the # of tricks taken. Then repeat that for 26-28 Goren, etc. They both have std deviations in terms of tricks so it's an apples-to-apples comparison. My previous experiments did this (although for all hands, not by level). I could break it out by level and see if one method is more accurate at certain levels. But when you look at overall performance by combining them again, you'll get what I started with. Tysen

Exactly. Which is what I've been trying to say for a long time. Simply looking at hands that make grands and then seeing how many of those hands have at least x points is meaningless. The methodology you describe is exactly what I've done in my experiments. What I've called error is the standard deviation from the expected value.

I often bid games with 25 including distribution. I'm just saying that maybe you should back down on your point requirements and see how that changes things in your study. We're trying to study methods of evaluation not simply aggressiveness.

Wait just a minute here. Am I reading this right where you only have the Goren hands bid games if the total is ">26" (that is 27, 28, etc.)?!? You list Zar as bidding game as >51 so I can only assume this is the case. I don't know anyone that requires 27 points for game. I often bid game on 25, so maybe this needs some looking into. Tysen

I'm not saying that this kind of hand exists. It's simply the average value you get when you add up all the zar distributions on every hand and divide by the number of hands. Question: how did you come up with 11 as being the average value?

That is comparing HCP+531 to Zar. I've never said that HCP+531 is better than Zar. It frequently performs close, but worse than Zar. What we are talking about in this thread is the value assigned to distribution, not high cards. Zar has the advantage of counting aces at a proportionally higher value, which is very key. We're comparing distribution to distribution. BUM-RAP essentially uses the same high card valuations as Zar, so if you compare BUM+531 to Zar you are essentially just looking at how the two systems value distribution since the high card components are identical. In that case, BUM+531 outperforms Zar (the comparison you convineintly left out of your quote). Tysen

Let's continue this discussion in the "useful concept" thread.

Zar, I'm not sure how you can say this. Since when does an "average" have to be a whole number? It's like having a 3.77 GPA and having someone say, "Sorry that's not a valid number so you only have a 3 GPA."

Again I have no idea what this proves. When you're finding contracts that make game you say the criteria is >51. Is that really >51 or is it 52-61 (not the hands that would bid slam). When you do it for small slams are you doing it for >61 or only 62-66? If you're doing it on >51 then all my evaluator has to do is bid a grand every time and I'll score perfect on all your methods. [a few edits made] Tysen

Hi Inquiry, Yes I agree with your reasoning about grands off aces, etc. And I'm not saying that anyone would bid based on points only. However, all the evaluators are treated equally and would all select a grand off an ace. I was curious and you can do this yourself on the published data. Limit the hands to only those that can take 9-11 tricks (11702 of the hands). If you run the analysis on those hands only you get an average error in the number of tricks: Zar off by 0.641 tricks HCP+531 off by 0.640 tricks BUM+531 off by 0.627 tricks I also agree with what you said about misfitting hands. But again all the evaluators are treated the same. You can add in factors to adjust for fit for Zar, but you'll adjust for the others as well and get the same result. Tysen

Depends on how extreme the 2-suiter is. 4441 is better than 5422 but weaker than 5431. The following is the real strength of the different hand patterns in how many more tricks they take compared to a 4333 hand. Tricks Tricks Real 5/3/1 4-3-3-3 0.000 0.000 4-4-3-2 0.296 0.333 5-3-3-2 0.339 0.333 5-4-2-2 0.595 0.667 6-3-2-2 0.660 0.667 4-4-4-1 0.810 1.000 5-4-3-1 0.864 1.000 6-3-3-1 0.918 1.000 7-2-2-2 0.999 1.000 6-4-2-1 1.154 1.333 5-5-2-1 1.183 1.333 7-3-2-1 1.208 1.333 5-4-4-0 1.519 1.667 6-4-3-0 1.624 1.667 5-5-3-0 1.643 1.667 7-3-3-0 1.697 1.667 6-5-1-1 1.703 2.000 7-4-1-1 1.712 2.000 7-4-2-0 1.923 2.000 6-5-2-0 1.964 2.000 The "real" tricks is the real point value that you would want to assign to that pattern. Ideally you'd want to use a separate valuation for each pattern, but that's too much to memorize at the table. The last column shows how many tricks 5-3-1 points predicts using 3 points per trick. Note how close 5-3-1 comes in almost every case (and weight them by frequency). 5-3-1 for shortness comes closer than any other distribution system that doesn't use fractions. Tysen

If you guys actually read the post where I gave the data you would see what those numbers represent. The explanation for what those numbers are is right there, guys! And since you guys are questioning the fact that these are just numbers without supporting data, I've posted a spreadsheet which shows the hands and my generation methods. It's an excel sheet that has a random sample of about 13,000 hands, but the data shows the same thing as I get with millions of hands (that file would be too large). I'm even making this be the sample of hands that eliminates NT hands and always makes 9+ tricks (the sample that makes Zar look the best by comparison). The spreadsheet is in the files section of this yahoo group (I don't have any other way of posting it): http://games.groups.yahoo.com/group/bridgeeval/ Unfortunately you have to join the yahoo group in order to get access to the file, but it's there for all to see. Tysen

Please, please, please stop saying that this methodology shows that x method is 2 times better than y. If my system bids a grand on every hand, it would be shown to be the best method out there by this logic! The comprehensive data that I posted under the "Zar accuracy question" post shows that HCP+531 (no fit) is close to Zar (no fit), but still worse than Zar. However, that's using raw HCP. Zar, you said a few weeks ago that you would run experiments using A=4.5, K=3, Q-1.5, J=0.75, T=0.25 and see how that does. Have you looked at that yet? Again let me reiterate that I'm not trying to push one method over another and I'm not trying to trash Zar. I just want to make sure that experiments and statistics are reported in a scientific and not in a misleading manner. Tysen