suokko

Me and Derric would like to play again.

How about always rebiding 2♦ with balanced minimum? That would leave responder all bids to show his hand type and strength making auction a lot simpler.

I open 1♦ because spade contract is very likely to require 4 card support to be a good one. We lose easily trump control if partner has only 3 card support. Rebid problem after 1♠ response? Nope. My system has 4♠ showing exactly this type of hand ;) But without that agreement I would still force to game opposite 1♠ response.

Those hand types are so rare that they don't dominate scores. But far more important error is that 5♣ won't be doubled ever. But I updated the original simulation with requested parameter changes and increased production limit to thousand deals. I also added statistics for north holdings with these limits.

I felt like I at least have to double. But biding over passing is also clear winner based on my quick simulation. Too bad scoring function doesn't support doubled contracts yet so I only simulated undoubled both. But undoubled both 5♣ and 4♠ makes biding winning choice by 5.3 IMPs on average. Doubled contracts would require some C coding and more complex rules when 5♣ will be doubled. I won't be doing that at least today but if there is enough interest I might do the doubled simulation some other day. Rules for preempt and T/O double were my quick toughs how to limit the hand types but if you see flaws on those rules I can repeat the simulation with improved hand limits. EDIT: I updated the simulation limits based on feedback and added statistics for north holdings. But still 5♣ not doubled because that would take quite a lot more coding. EDIT2: I added doubled contracts to simulations. I also improved ew hand definitions that slightly increases total tricks and makes biding win a bit more IMPs.

Because 3-3-4-3 is bad shape for T/O double and I choose not to double with it. My partner doesn't have any better shape so we end up defending even tough holding majority of hcp. I'm happy to score +100 in a part-score hand. Of course MP is completely different game where +100 isn't enough.

Me and Derric would like to keep practicing. Maybe we will be ready for our f2f league starting in January :)

Nothing special realy. A lot of small details. Some errors probably were missed by ruthless kibs because they didn't cost anything. We found something to improve from over 20 boards. No surprise that kibs were telling we are bad players. We were bad compared to our opponents but that is just relative to our opponents. Most of ruthless kibs wouldn't be favorites to win if playing against our line up on Saturday. I prefer to play against world champions and lose. That is the best way to learn play better. I like to keep our agreements publicly shared. They are still far from including all details but they are complete enough to give general idea how the system is supposed to work. https://googledrive.com/host/0B4izKEmfNQIBN2Y4Zl96N0ZvMUU/Juho_Pauli_2015.pdf Google decided that my notes have something inappropriate. It is blocked for sharing until manual review. But original odt file is still shared.

Not much about system. We collected minus mostly from worse judgement at biding and some errors in hard defensive situations. Our opponents showed how little errors they can make. Of course it helps biding judgement when one gets more familiar with the new system but that shouldn't be our (me&Derric) excuse for mistakes.

Probably: 1) ♥2 2) ♠5 3) I'm not on lead to 1NT

Fairly easy. It depends how easy it is to calculate the value that you want. But there is tertiary selector operator for if-else statement and one can use check length of suits, HCP, controls, specific suit quality with simple functions. It's not turing complete programing language but powerful enough for most cases. Bonus is ability to ask number of tricks for declarer. But that is poorly optimized feature making that kind of simulation fairly slow. The average speed is only about 10 contracts per second when the libdds can be a lot faster with the newer api.

The implementation stars with a ordered pack. Next it reorders it with single shuffle iterating all cards. From that pack order it generates the hand presentation for analyse step. For next deal it takes the shuffled pack from previous deal and does same suhffling step the first board. This makes the pack order weakly part of rng state. That may make rng plus pack state period a lot longer than what rng period would be alone. Of course there is chance that period start to repeat with some low multiplier to rng period but it is fairly unlikely.

I use the direct single board upload if only checking a few selected boards. But I often like to check every board to see if there was any card play tricks that I missed. The interface isn't very friendly for keeping the tournament hands browser open while playing cards in table. That makes me close it often leading to quite many clicks to upload next board. For the 28 board team match I prefered to prepare a bit and save boards. That makes switching boards easy with single click on redeal.

It would make it a lot easier to open a teaching table to analyze recently played boards if it was possible to select recent tournament or team match as deal source. Ability to at least have own recent boards would be helpful. If possible I would also like to be able to select any recent tournament for coaching purposes. Currently I have to remember to save the boards to my own hands folder. If I forget to do I need to manually save them one by one.

https://github.com/suokko/dealer/blob/master/dealer.c#L651 I think that older version before my changes is used to generate BBO deals. At least all information points toward that. It has fairly good prng called from shuffle code. It uses 16*52 bits of prng stream to generate a board. That code is not aimed to generate high quality deals but good quality deals for statistical analyze of different bridge situations. That code requires around 1000 CPU cycles to generate a board with some tweaking it could probably take about 500 cycles per board. The shuffle implementation is a bit different to original shuffle algorithm. It should give correct probabilities for each card (as far as I understand probabilities) if we had true random numbers. I agree that true randomness is very hard.

Please apply here if you'd like to play vs JEC/Garozzo on Saturday. We (suokko & Derric) want to try it again.

I never even try to make perfect simulation that takes all unlikely parameters into account. To me opponents passing in this situation is balanced enough effect to both lengths. It would probably have more effect to hcp frequency for south making the lowest values less likely. Of course that reduces slightly spade blockers effect but it isn't going to reverse the effect.

That is agreement that I always have. I only drop stayman from the NT system bids. I have seen too many penalty doubles with solid suit and no real defense that can beat 1NT but not our other games.

I have agreement that biding promises more spades than hearts. We transfer to hearts with a weak hand and 5 hearts and 4 spades. Only mildly invitational with 5 hearts and 4 spades bids stayman. So this question is more about agreement details than actual random parameters. But simulation shows: $ cat simu.descr predeal north SAQJ,HJ93,DKQT8,CJ64 condition shape(south, 54xx + 45xx + 55xx) && hcp(south) < 9 action frequency "heart length" (hearts(south),4,5), frequency "spade length" (spades(south),4,5), $ ./dealer < simu.descr Frequency heart length: 4 140195 5 164470 Frequency spade length: 4 124103 5 180562 Generated 10000000 hands Produced 304665 hands Initial random seed 1419266048 Time needed 3.185 sec So more small cards in spades makes spade length more likely because of hcp limits.

Squeeze was superior play for over trick instead of the finesse. Because it works with either 4♥+4♠ in same hand or 4♠+♣K in same hand. That is better odds than simple finesse where 4♥ at south makes finesse less than 50%. But seven is very hard to bid without fit because it requires that combined very good diamond holding to be a good contract.

We lost 47-121. Some mistakes and they played very well.

I can do a written analyze at 20 boards per hour speed. If that was my day job it would take only 3 months to analyze 9000 bridge hands. But actual arguments for practical randomness in IMP vs MP games. Normal winner score in MP is around 57-65% in competitions where skill level difference aren't big. If a tournament is about 40 boards then single board top-bottom difference is 2.5% in end results. But in practical play my opponents cannot score more than 2/3 of top with a good play. That translates to about 1.6% at stage each board which is 1/9 of score difference between winner and average. But even a good play can be often countered with a less important decision that still score a few points back to side that didn't have the more important decision to make. At similar IMP competition winner score is +1.5-2.5 IMP per board per comparison. But a single vulnerable game swing is 10-14 IMPs per comparison. That translates to about 0.25-0.35 IMPs per board and those decision are pretty common with only one side having decisions that affect the make or not make situation. So even a normal game swing has a higher magnitude in IMPs than extreme maximum MP games offer. But if we factor in some extreme IMP swings with small slam their effect can be even twice the effect that games have. That makes IMP scoring more volatile because smaller number of important decision provide relatively higher magnitude changes to the results.

http://en.wikipedia.org/wiki/Fisher%E2%80%93Yates_shuffle That is by far easiest and fastest dealing approach. I have tried implementing the 96bit entropy dealer but it was slower because even table look up ncr calculation is slower than the simple shuffle implementation. Also that shuffle repeated from previous generate hand instead of always from original ordered pack improves statical quality of generated hands.

If dealing is done like I suspect there is minor bias from the process. Starting from ordered pack causes a bias always but there is statical test for the dealer program that proofs it generates random hands when dealing a very large number of boards in a single run. That is easy also reason why it works because random shuffle is applied to randomly ordered pack from previous board so even a poor prng can make statically good shuffles.

MP and IMPs emphasize very different skill sets. Of course those same skills are required in both forms. If a player has some areas that clearly weaker it can make results quite different in MPs and IMPs. Also cross IMPs is fun to play but a good biding decision by opponents can easily dominate the score in a specific board. In MPs good card player can often produce relatively large improvement to the score with a good defence. In my experience it is clear that in MPs it is a lot easier to play stable good score (55+%) than in cross IMPs. Teams is completely different IMP scoring where good teams have a decisive advantage even in short matches.