BBO Discussion Forums: State of the Art4 - BBO Discussion Forums

Jump to content

  • 2 Pages +
  • 1
  • 2
  • You cannot start a new topic
  • This topic is locked

State of the Art4 The state of play

#21 User is offline   Free 

  • mmm Duvel
  • PipPipPipPipPipPipPipPipPipPip
  • Group: Advanced Members
  • Posts: 10,728
  • Joined: 2003-July-30
  • Gender:Male
  • Location:Belgium
  • Interests:Duvel, Whisky

Posted 2012-April-18, 10:01

View PostCthulhu D, on 2012-April-17, 21:42, said:

Curious here: why doesn't it make sense to use pre-programmed optimal play to a suit stuff (like suit-play), if simulation doesn't generated a clear best line. So have a pre-calculated table of all suit combinations and when it's declaring it looks up the one it has and makes that play.

As far as I know current programs already use some form of suitplay analysis in certain situations. However, this is limited to deals where only 1 suit can be played in several ways, and entries aren't any problem.

Expanding this to 4 suits wouldn't work btw. It doesn't take tempo or control into account, combinations of suits need a certain amount of entries,... It's just too simplistic.
"It may be rude to leave to go to the bathroom, but it's downright stupid to sit there and piss yourself" - blackshoe
0

#22 User is offline   Cthulhu D 

  • PipPipPipPipPipPip
  • Group: Advanced Members
  • Posts: 1,169
  • Joined: 2011-November-21
  • Gender:Not Telling
  • Location:Australia
  • Interests:Overbidding

Posted 2012-April-18, 19:52

Not sure I get this. For starters, there is already a fall back rule. It is 'from any DD equivalent holding, play a random card.'

So the objections to having fall back rules at all are illogical. There is currently a fallback. It is just not very good.

View Postbarmar, on 2012-April-18, 08:23, said:

Whenever I think about programming in tables like this, I think back to Mike Lawrence's book on card combinations. It's organized into chapters with several hands where declarer has to figure out how to play the same card combination, but has different inferences from the way the opponents have bid and/or played, or you're in a different contract and have different constraints. Tables of card combinations usually make simplifying assumptions, like you have adequate entries to both hands, there's no danger hand, etc. When you put them into the context of an entire hand, things aren't nearly so simple.

So while there may be times when a table like this can be used, figuring out when you're in such a situation is difficult.


Surely in these cases (as for Free's) though simulation is going to produce different values! If there is a danger hand, the simulation will find this! It's only when the simulations tell you there is no difference in what card you play from from your suit holding that you should resort to manual rules.

In that case, it doesn't even matter if your rules are very good, or very detailed.

Quote

It doesn't take tempo or control into account, combinations of suits need a certain amount of entries,... It's just too simplistic.


If the simulation isn't finding this either then you have other problems. Remember, we're only talking about situations in which DD, there is no difference what card you play. The reason is to avoid the below:

Quote

GIB doesn't try to "get into the opponent's head", so it doesn't know that declarer doesn't know the location of the Q. As a result, we often see boneheaded defensive plays -- declarer leads a ♣ towards dummy and West puts in the Q, taking away the guess completely. West assumes South is playing double dummy, so will always guess right, and it's not giving up anything by playing the Q.


I am not sure why objections like 'if there is a danger hand' hold true - surely, surely if there is a danger hand your simulations will determine that and then the cards will not be DD equivalent.

It's just to prevent cases like Barmar is stating above, where all the cards are DD equivalent. If you KNOW the cards are DD equivalent, there is little complexity - entry problems, danger hands etc would have been revealed in the simulation. You just need a totally generic fallback position. It is unlikely to cost if your fallback position is wrong either - remember the cards are DD equivalent...

Then you can play a basic rule. Note that 'second hand plays a random small card from any holding with a single H' is sufficient to solve the problem above, and handles all permutations in which the two, three, five or 12 cards are DD equivalent.

If I am mistaken, please tell me why, but I am not sure why a rule such as 'from any DD equivalent holding headed by 1 honor, play a random small card in second seat' is inferior to the current rule of 'play any random card.'

Heck, I'm 99% sure that you could just change the fallback rule from 'random card' to:

* From any DD equivalent holding with no honour or sequence, play a random card
* From any DD equivalent holding headed by an single honour, play a random small card
* From any DD equivalent holding headed by an honour sequence, play <random card from the sequence>
* From any DD equivalent holding headed by more than one non touching honor, do... something. Maybe cover an honour with an honour otherwise play a random small card.

I just pulled these out of my ass.

Then if you can show me any scenario in which both these are true:

A) Monte Carlo simulations would show that the cards are DD equivalent
B) Applying these four rules costs over simply playing a random card.

I will give you a gold star and commit internet seppuku. ;)
0

#23 User is offline   kgr 

  • PipPipPipPipPipPipPip
  • Group: Advanced Members
  • Posts: 3,423
  • Joined: 2003-April-11

Posted 2012-April-19, 10:41

See http://www.bridgebas...__1#entry631300
I think that this is another disadvantage of DD analysis. It will delay decisions (it has to take anyway) to later tricks and take extra risks of ruffs in the mean time.
If it can finesse in trumps, but will never do, and go down if trumps are 4-1 and it will delay playing trumps from the top. Because the risk of a ruff is smaller then trumps 4-1. Then it will later play the trumps from the top anyway, having risked a ruff in the mean time.
0

#24 User is offline   Scarabin 

  • PipPipPipPip
  • Group: Full Members
  • Posts: 382
  • Joined: 2010-December-30
  • Gender:Male
  • Interests:All types of games especially bridge & war games.
    old bidding systems & computer simulation programming.

Posted 2012-September-30, 02:17

At my level of competence, it seems logical to believe:

most robots follow monte carlo rather than pragmatic reasoning because it's quicker to program and gives better results on average

monte carlo could be improved quite simply by superimposing rules on drawing trumps, playing low cards when not affecting result, and possibly on combining chances (although theoretically monte carlo should cover this?)

when pragmatic reasoning programming becomes sufficiently developed (probably requires someone like Fred who is expert at both bridge analysis and programming) it should out perform monte carlo (which has limited potential for improvement without changing into true double dummy analysis, which is equivalent to cheating, or into pragmatic reasoning).

OK, tell me I am wrong but this is where I stand.

Edit: Looking back I find I have restated my original thesis, so let me try to add something new:

Barmar has pointed out the limitations of single dummy analysis like suit play but does not a human expert plan by examining the combinations in each individual suit, taking into account any inferences as to the distribution of the missing cards and then combines individual suit plays to give the optimal result for the complete deal.

I do not see any insuperable barrier to reproducing this in a computer program. Whether we rewrite suitplay or use a table of probabilities is a point of detail?
0

#25 User is offline   FM75 

  • PipPipPipPip
  • Group: Full Members
  • Posts: 496
  • Joined: 2009-December-12

Posted 2012-October-03, 17:28

View PostScarabin, on 2012-September-30, 02:17, said:

At my level of competence, it seems logical to believe:

most robots follow monte carlo rather than pragmatic reasoning because it's quicker to program and gives better results on average

..snip..

I do not see any insuperable barrier to reproducing this in a computer program. Whether we rewrite suitplay or use a table of probabilities is a point of detail?


It seems clear that you simply do not understand the reference work that was provided in your earlier thread.

It is nice to know that you do not see an insuperable barrier...

When can we expect to see your program in competition?

Perhaps you should read up on some game theory. Its foundations go back to the 40's.

What is your human rule for this situation?
After an opening bid of 2NT, the player in the balancing seat always doubles.
How should the "declaring" pair handle this in their bidding? Will the balancing player (team) modify its behavior? How? When?

(I am not trying to hijack this thread. I only pose a game theoretic straw man to suggest that. in fact, games of imperfect information tend to be perforce stochastic.)
0

#26 User is offline   Scarabin 

  • PipPipPipPip
  • Group: Full Members
  • Posts: 382
  • Joined: 2010-December-30
  • Gender:Male
  • Interests:All types of games especially bridge & war games.
    old bidding systems & computer simulation programming.

Posted 2012-October-03, 22:05

View PostFM75, on 2012-October-03, 17:28, said:

It seems clear that you simply do not understand the reference work that was provided in your earlier thread.

It is nice to know that you do not see an insuperable barrier...

When can we expect to see your program in competition?

Perhaps you should read up on some game theory. Its foundations go back to the 40's.

What is your human rule for this situation?
After an opening bid of 2NT, the player in the balancing seat always doubles.
How should the "declaring" pair handle this in their bidding? Will the balancing player (team) modify its behavior? How? When?

(I am not trying to hijack this thread. I only pose a game theoretic straw man to suggest that. in fact, games of imperfect information tend to be perforce stochastic.)


Thanks for your interest. As regerds your first point it's very possible I have not fully understood the papers I read. I'd find it easier to judge if you will do me the courtesy of specifying which paper, and even better which precise part, I failed to apprehend?

I'll pass on your second point which seems to be weak sarcasm and unnecessary.

The answer to your third point is, unfortunately, probably never. I am a painfully slow programmer and I check each step meticulously, mainly because I hate having to debug large batches of code. I do produce spin-offs from time to time and these have value for me. Perhaps I may publish some of these. My aim is not to produce a world champion program but rather one which would give a good player a good game and teach a beginner to play good bridge.

I have read up on game theory, also statistics and probability. That's how I made my living. Unfortunately my roots go back to the thirties.

I do not have an answer to your next question, any program I write will only deal with bidding systems with which I am familiar. To compensate it will cover these as completely as possible.

Perhaps it's another gap in my understanding but I am not sure your last statement actually means anything. If you mean games with imperfect information must be programmed with monte carlo simulations this would meet serious opposition from people whose opinions are highly valued, as well as by me.

Slainte,

Ian
0

  • 2 Pages +
  • 1
  • 2
  • You cannot start a new topic
  • This topic is locked

1 User(s) are reading this topic
0 members, 1 guests, 0 anonymous users