Join Bridge Winners
All comments by Doug Bennion
1 2 3 4 5 6 7 8 9 ... 26 27 28 29
You are ignoring the author of this comment. Click to temporarily show the comment.
2K deals. Don’t need a huge number for ‘good-enough’ results when 2 hands are fixed. Just ran another 2K with similar results within ~ 1%.

No clue what were the ‘normal routes to make’, not reported by the engine.
April 15
You are ignoring the author of this comment. Click to temporarily show the comment.
I can tell you what a double dummy engine ‘thinks’.

It thinks 4 makes about 53%, 4 about 56%, but 3NT is the big favorite at 76%.
April 15
You are ignoring the author of this comment. Click to temporarily show the comment.
It depends in part on the level of contract. In game range in a notrump contract your soft values harden: that Qxx might support partner's hard values, and stop the opponents attack suit. Otherwise soft for notrump, hard for suits.
April 15
Doug Bennion edited this comment April 15
You are ignoring the author of this comment. Click to temporarily show the comment.
Just responding to a couple of comments.

I might have included 4333 and 5332, but thought (when setting things up) the bulk of ‘winning’ leads would be from the long suits, so little point to inclusion; the bot would have little ‘choice’, and I thought I wanted to measure the bot’s ‘decisions’. It wasn’t until I started to produce some results that I noticed the bots quite ‘liked’ a lead from ‘xxx’ (and that’s the short holding that would continue to fare well). I was too lazy to rebuild, so I left those hands by the wayside. Including those shapes would change the rankings very little imo.

Tens. Yes tens are nice cards. So is ‘98’. Super-nice is ‘T98”. One would think JTxx is a stronger holding to lead from than Jxxx, but so would QTxx be better than Qxxx, KTxx than Kxxx, and so on. I had to pull up somewhere in my breakdown of suit ‘categories’.

And anyway it seems it isn’t clear just how much better would be a lead from JTxx compared to Jxxx. I just did a 2K head-to-head with those two, and there is virtually no difference. In fact, the spot from Jxxx was very slightly a stronger lead than the spot from JTxx, and the J from JTxx the overall leader by just a teensy bit.

And Jxxx creams, obliterates, crushes KTxx.
April 11
You are ignoring the author of this comment. Click to temporarily show the comment.
I currently play 1 is any balanced hand including 5M332. I read that as ‘Quasi’. For that matter it looks like 6322 and 5422 would also qualify as Quasi.

I currently alert 1. Will Quasis be alertible, or announced, or neither?
April 10
You are ignoring the author of this comment. Click to temporarily show the comment.
We generally use responder's 2nd call to identify the asset behind the 2 call, so responder was bidding by our ‘book’, which might not be a best-seller.
April 10
You are ignoring the author of this comment. Click to temporarily show the comment.
I was curious how often the opps belong in hearts after a Moscito 1 opening, so I (crudely) simmed it. I gave opener 10-14 HCP, 4 hearts, not 4 spades, not 6+ minors (is that about right?). Opponents make 2/3/4 hearts 20/10/4%. For what that's worth.
April 10
You are ignoring the author of this comment. Click to temporarily show the comment.
It was always curious to me that all 18 approved defenses were for opening bids, leaving us to fend for ourselves calls in other positions.

Until transfer responses to 1 came along, that is.
April 9
You are ignoring the author of this comment. Click to temporarily show the comment.
You don't need thousands of deals, usually a few dozen or scores would be enough. Especially when the choice is down to 2 or 3 cards, you can often zero in on the (almost guaranteed) winner after a dozen or so deals. If holding Ax and the A is ‘best card’ 10 times in 10 deals, you could shut the simmer down at that point and rarely be wrong.

The quality of the proxy deals is critical. If the sample deals bear little resemblance to what was portrayed by the auction, you'd obviously need more sims and even that might not result in the play you would like. For example in the OP if the deals did not include some 5-card club suits, the defensive plays might be ‘suspect’.
April 8
You are ignoring the author of this comment. Click to temporarily show the comment.
You get an extra bid which is the disadvantage of transfer openers. I'd play the same approximate methods I'd play against Multi; transfer overcalls. After 1,

X = like heart cuebid = here like T/O of hearts
1 = spades
1 = weaker notrump range, say 13-15
1NT = stronger notrump range, say 16-18
April 7
You are ignoring the author of this comment. Click to temporarily show the comment.
Yes at trick 4 the bot will never play Q from QJx (although if partner has been dealt Kx, it doesn't cost).

To the best of my knowledge, at trick 5 the bot would be looking at the existing current layout, and wouldn't retain a memory of who previously played which card. And even if it did, it wouldn't be ‘thinking’ at trick 5 that playing Q from QJx at trick 4 would have been irrational. However I could be all wrong about that.

About the J finesse. Remaining cards are J98 and one of 98 has been played. Maybe playing the K is 50%? With just two outstanding cards the layout will be J-98 (play the king, no loser) or J9-8 (any play eventually loses one) or 9-J8 (play the ten, no loser) - - J98 (any play eventually loses one).
April 7
You are ignoring the author of this comment. Click to temporarily show the comment.
When simming, odd plays can result from (1) insufficient sample sizes and (2) inappropriate hand construction. I have no idea what the bot sample sizes are, but if they were goosed to say 100, some errors would go away. Also for this particular deal, East bot announces ‘clubs twice rebiddable’, so (maybe) all the simmed hands have 6+ clubs with no provision for hands like the actual hand.

Also it seems odd that declarer did not take the (losing) club finesse to his KTxx when 3 trumps J-x-x were outstanding.

Playing lowest from equals would make the plays more ‘human-like’, for sure.
April 7
You are ignoring the author of this comment. Click to temporarily show the comment.
Opener has 13-14 HCP, 5 spades, no longer suit, 2000 sims.

Responder has Axxx x Kxxx xxxx, and makes game 45.4%

Kxxx x Axxx xxxx makes game 50.9% (K always useful)

xxxx x Axxx Kxxx makes game 52.0%
April 4
You are ignoring the author of this comment. Click to temporarily show the comment.
I see no limitations to 1NT overcalls. You could go crazy there.
March 26
You are ignoring the author of this comment. Click to temporarily show the comment.
Other. Transfers begin with XX = clubs. 2 = diamonds, your cuebid raise. 2 = hearts. 2 = constructive raise. 2 = noise.
March 22
You are ignoring the author of this comment. Click to temporarily show the comment.
However they seem to be allowed over an opponent's double on the Basic+ chart.
March 22
You are ignoring the author of this comment. Click to temporarily show the comment.
Your best fit will be 5-2 or worse about 23%, 5-3 about 43%, 5-4 about 28%, and at least 5-5 about 6%. Bidding looks routine to me at matchpoints, and did before I looked at the sim, except maybe unfavorable.
March 19
You are ignoring the author of this comment. Click to temporarily show the comment.
5440 hands play about 5 points better when a fit has been found, so those minimums will almost always really grow up. Quick and dirty sims show responder will have at least 3 hearts more than half the time, and the balance of the time responder is a huge favorite to have at least a 4-fit with one of the minors. The best way to find one of those fits is to open.
March 18
You are ignoring the author of this comment. Click to temporarily show the comment.
Transfers. X = hearts, 2 = spades, 2 = balanced 12-14, 2NT = balanced 15-17. 2m = natural NF.
March 17
You are ignoring the author of this comment. Click to temporarily show the comment.
Thanks for the article.

I wrote an early-generation bridge program so have some experience with the foibles of double-dummy play. Bear in mind I’m speaking from memory here, not having looked at a line of code in 10+ years. Also I confess to little first-hand knowledge of how the BBO bots perform, rarely playing them. I’m assuming the principles of play are largely similar to how Buff played 10 years ago.

Your first hypothesis is correct. Declarer bot has sifted through a batch of simulated deals and chosen the card which generates the best results over those deals, in this case a trump lead . Bot with Kx under AQJT… deals a bunch of hands based on his perspective and sees that the K and the x lose equally, so are equivalent holdings.

At this point I think the programmer has a choice. He can randomize the play of equals, which apparently is what the BBO bots do, or he can direct the bot to play the lowest of the equivalents, here the trump x, which is what a human would do .

Which leads to your second point. The bots apparently discard randomly from Q5 when the 4 remains out, and sometimes ‘fool’ the opps by playing the Q.

So l think your hypothesis compresses to ‘the bots play randomly from effectively ‘equivalent’ cards’.

If the programmer wanted, he could ‘humanize’ bot play by specifying it always play low from equivalents. Or if he wanted to inject a random element of falsecarding — play a card other than low 5% of the time, or whatever.

There are other ‘cosmetic fixes’ a programmer might undertake. In the example case, bot declarer might have cashing cards on the side, say AKQ in a suit. The batch of simulated hands might not reflect any ‘bad’ things happening to an initial cash of one or two of those honors, then getting around to drawing trump. So a cash would be equivalent to a trump play. That behaviour would look weird to a human — cash side ace, draw a trump, cash side king, draw another trump. I seem to recall writing code to mitigate some of that kind of goofy play sequence.

A point about the bot leads. They do lean conservative. I once built a 100K sample of double dummy leads, compiled and ranked the results. If I wasn’t so lazy, I would publish the results. Suffice to say, as you pointed out, double dummy engines ‘dislike’ leading from honors, often defaulting to xxx and xxxx. . How those bot preferences relate to human play, is problematic.
March 15
1 2 3 4 5 6 7 8 9 ... 26 27 28 29
.

Bottom Home Top