[mikeh]

I strongly distrust software that purports to evaluate the likely outcome of specific leads against specific contracts. Such simulations are biased by the assumptions made by the programmer. What are the constraints on the 1N bid?

What shapes and ranges are used? Merely specifying, for instance, that 1N shows balanced hands with 15-17 comes nowhere close to real life. Some 17 counts are too good for the range. Some 14 counts ought to be included. Some hands with a 5 card major should be included, while others ought to be opened 1 Major. And so on.

And for the 3N bidder: what are the parameters for that? And so on.

That ignores potentially more important factors. How does the programme assess the play? If it is like deep finesse, then it is hardly appropriate. Many contracts that can be made double dummy are doomed single dummy, and vice versa.

I like hand generators for bidding system develoment, but not (at the current state of the art) for play or defence analysis.

[kfgauss]

mikeh, on Sep 15 2005, 12:32 AM, said:

I strongly distrust software that purports to evaluate the likely outcome of specific leads against specific contracts. Such simulations are biased by the assumptions made by the programmer. What are the constraints on the 1N bid?

What shapes and ranges are used? Merely specifying, for instance, that 1N shows balanced hands with 15-17 comes nowhere close to real life. Some 17 counts are too good for the range. Some 14 counts ought to be included. Some hands with a 5 card major should be included, while others ought to be opened 1 Major. And so on.

And for the 3N bidder: what are the parameters for that? And so on.

That ignores potentially more important factors. How does the programme assess the play? If it is like deep finesse, then it is hardly appropriate. Many contracts that can be made double dummy are doomed single dummy, and vice versa.

I like hand generators for bidding system develoment, but not (at the current state of the art) for play or defence analysis.

The definition of the 3N bid is in my first post (top of page 2 of this topic) and is very reasonable, I think. The definition of 1N is simply 15-17 balanced. Perhaps this affects things slightly but I really doubt it's a significant factor here (but who knows, I could be wrong). 1N-3N situations are amongst the easiest to model and I think the data can be quite relevant in such situations.

The analysis is double dummy (this is a much more important point than the above I feel). I'm fully aware of the limitations of this analysis, but at least it's clear what it is. I find it fairly interesting that a spade is a better double dummy lead here (assuming you accept the definitions of the bids).

Maybe this is an unfair context to put the discussion in, but I'd suggest that if you think a heart is a better lead than a spade that you should really think about and try to explain why one of the following is true (for the sake of discussion & figuring out which is better, not for my sake or because you're required to do so, etc):

1) a spade is better than a heart at double dummy, but for these [insert various reasons/factors here] reasons, a heart lead is better than double dummy analysis would suggest when compared to a spade lead.

2) against mom & pop, a spade lead is better than a heart lead, but against experts who don't evaluate their hands rigidly and who bid 1N off-shape sometimes, a heart lead has these [insert reasons] advantages over a spade lead.

I understand that everyone likes their own judgement, but this doesn't seem to be the sort of hand where our experience is really going to give us good data. We'll recall a couple of hands where leads from the 3-card suit are better pretty strongly -- "wasn't that weird" -- and forget the slightly greater number of times that the spade lead worked perhaps. (If you find this "model" of how so many people came to believe a heart lead is better unrealistic, I'm not actually claiming that it's true -- just putting it up as a type of possibility.)

Andy

[mikeh]

I agree with several of your comments, including the danger of relying upon our remembrances of experience.

I am still not persuaded, which may be a reflection of stubbornness on my part However, I will try to addrsss your questions.

There is, for example, an increased tendency for declarer to misread the situation after an ostensibly long suit ♥ lead (I admit that I would prefer to be leading a 2 or a 3, rather than a 6 on this issue). This can result in gains in several ways, including deciding that the outstanding ♥ are 4=4 rather than 5=3, and therefore knocking out a side Ace rather than taking a winning finesse. These marginal gains (if they exist) literally cannot be calculated by your simulation, which plays knowing all of the cards.


Further, if you lead a ♠, partner, when in, may not know to switch to ♥ when it is correct to do so. This is, again, likely to be a low-frequency cost of a ♠ lead but one that will never be taken into account by your program.


Another defect in the computer analysis is that it ignores the real world likelihood that a declarer will misguess a suit. This flaw actually strengthens the case for the ♠ lead, since the computer assumes that declarer will always get a 2-way ♥ hook correct, thus to some degree minimizing the real-world cost of a ♥ lead.

I am trying to be objective, hence my making this point


To me, however, the worst flaw in your approach arises from the effect of the imp and scoring tables combined with the reality that, at the other table, the auction (and both the contract and the lead) may be different. Thus in some real world situations, your partner may be on lead or the opps may be in an entirely different spot. Not everyone plays the NT range your simulation assumes.

Therefore, in the real world, on some of the hands on which the ♠ lead saves points, that savings may be meaningless. If your teammates play, for example, a 13-15 nt range, and the cards are 15 opp 10, your teammates will be in 1N. If they make 150 and your lead lets your opps make 630 rather than 600, you lose the same 10 imps.

But if your lead holds them to 8 tricks, then you have won 6 imps. Your simulation cannot consider this, because it assumes identical auctions.

What if the 3N bidder held a maximum: a hand that was in your teammate's mind worth 4N. And opener had a max and bid 6N.

They go down when the natural ♠ lead leaves declarer misguessing a two way ♥ finesse. They are -100. At your table, your leads result in either -690 or -660. You lose 13 imps either way. How do you account for that in your method?

You don't: not because of any fault on your part, but because of the limitations of the method. Your method calculates a cost of 30 points for the ♥ lead. But that 30 points may be meaningless.

In the real world, the actual gains from setting a usually making contract are much greater than one would calculate if one assumed that the contract would be the same at the other table.

Your analysis assumes that a large number of small gains offsets a small number of big gains. This may or may not be valid, but the analysis depends for its validity upon assumptions as to the contract at the other table. As I hope I have shown, once you try to factor in the reality that the contract may be quite different, without anyone doing anything weird, than your calcualtions may break down.

All of this is a long-winded way of saying that in my view a small ♥ is likely to be very slightly better at imps than anything else, bearing in mind that the contract is likely cold

[Chamaco]

kfgauss, on Sep 14 2005, 05:42 PM, said:

Expected total points (opps non-vulnerable):

Hi Andy,
thanks for your efforts!

All I want to point out is that quantifying cost-benefits in terms of TOTAL POINTS, makes this analysis more suitable for IMPS than pairs.
Needless to say, the difference in analysis IMPS/MP can be substantial.

I did argue that I would lead a diamond at PAIRS, but would lead a major at IMPS: and also the small subset of simulations I ran, suggested (to me) that diamond would be equal or less likely to waste a trick than other leads, but that indeed, it was less likely to defeat the contract.

The result of your simulations are quite compatible with these observations in my view.

It would be interesting to see if Jack can analyze the outcomes in terms of board % at pairs, rather than total points.

Finally, one question: did you include 5332 with a 5 card major in the 1NT opener ?
I think this should be accounted for in the analysis.

Also, it would be fair to include for the 3NT bidder hands with 4333 with a 4 card major.

======================

A side note to people who do not like simulations for chosing leads:
I do agree that it's hard to set criteria that will be followed in the real world by all opps.
However, it is reasonable to use some criteria that ill be hpefully followed by the majority of people, and even if the simulations are not perfect, a large number of runs should compensate for the occasional deviations (e.g. offshape 1NT, upgrade/downgrade, etc).

[kfgauss]

Thanks for the interesting points. The two most important ones seem to be:

1) On a heart lead, declarer may play for hearts 4-4 instead of taking a finesse and go down

2) When the other table is in a different contract, overtricks matter less (or not at all) compared to setting

It's hard to think about these and know how large these effects are, but they're well-taken.

The second suggests that one should try a double dummy analysis of how often the contract makes on different leads and see how they stack up then. I don't have that capability with Jack but will look around for a program that can do this for me (I really dislike the "expected total points" output). I'm a fan of overtricks, so I'm not so convinced that they're worth so much less than the amount the total points analysis values them (note that they're already undervalued compared to imps playing the same contract at total points), but your 6NT example is good -- certainly a heart is worse then.

The first gets at something impossible to get around given double dummy analysis, as you note. Running the sort of test Hrothgar thought I ran (ie having Jack play the hands instead of using double dummy analysis) might help here, but Jack isn't going to be the greatest card player. Some combination of this and the double-dummy analysis (with % making) could be quite good though.

I'm not sure I'd give up on the spade lead yet, but I do see some of the things the heart lead has going for it that aren't captured by the analysis using Jack.

Andy

[kfgauss]

Chamaco, on Sep 15 2005, 06:47 AM, said:

It would be interesting to see if Jack can analyze the outcomes in terms of board % at pairs, rather than total points.

Finally, one question: did you include 5332 with a 5 card major in the 1NT opener ?
I think this should be accounted for in the analysis.

Also, it would be fair to include for the 3NT bidder hands with 4333 with a 4 card major.

I'd love to get a result-by-result printout (ie how many times does ♥6 set 2 tricks? give up 3 overtricks? etc), but I don't have this capability. Jack seems only to play IMPs (or total points really).

I think Jack includes all balanced hands in its 1N openers, but I'm not actually privy to that information. I also can't make it include all 4333's in its 3NT bids.

Best,
Andy

[Chamaco]

kfgauss, on Sep 15 2005, 07:15 AM, said:

I think Jack includes all balanced hands in its 1N openers, but I'm not actually privy to that information. I also can't make it include all 4333's in its 3NT bids.

Best,
Andy

Hi andy,
if Jack can read pbn files, I can send you a pbn files with, say, 500 or 1000 files that satisfy the constraints I mentioned (e.g. including 5332 with 5cM for opener and 4333 hands with 4c major for responder).

This way you could rerun the test including those hand-types.
In the real world, 5332 major 1NT opening and 4333 with responder's hand are frequent enough to shift by some amount the outcomes of the simulations.

If you are interested, send mail at

m_casadei67*REMOVETHIS*@tin*REMOVETHIS*.it

[han]

Hi Andy,

I was wondering if you could run the exact same experiment 3 more times and post the data. It would be nice to see if the numbers you get are the same. It is hard for me to judge whether 1000 hands is a large enough data-set.

[Al_U_Card]

1NT-3NT....always lead a major! Spade at MP and small Heart at IMPS.

[cf_John0]

Assistant leading Heart 6 or 7!

[mike777]

---

Here is the hand.

I lead T of D.

[mikeh]

Proving what?

[mike777]

Small errors in the beautiful and simple inputs of expectations create deformed outputs. Only after extensive surgeries does an "optimal" result appear.