home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!europa.asd.contel.com!emory!wupost!zaphod.mps.ohio-state.edu!rpi!bu.edu!transfer.stratus.com!sw.stratus.com!tnh
- From: tnh@sw.stratus.com (Tim Hill)
- Newsgroups: rec.games.bridge
- Subject: Re: Which inference is better, WAS - "finesse or play for the drop"
- Message-ID: <1ej3p8INN6f@transfer.stratus.com>
- Date: 20 Nov 92 16:29:28 GMT
- References: <1992Nov17.175850.18797@cbnewsi.cb.att.com> <davidm.722210891@questor>
- Organization: Stratus Computer, Inc.
- Lines: 61
- NNTP-Posting-Host: bigbootay.sw.stratus.com
-
- In article <davidm.722210891@questor>, davidm@questor.Rational.COM (David Moore) writes:
- > A British Bridge magazine did an analysis of hand dealt hands about 20 years
- > ago which supported this; because there is some "memory" in a hand-shuffled
- > deck, Q's lie under K's more than 50% of the time. What they actually showed
- > ib the magazine was that hands tend to be flatter with hand-shuffled decks.
-
- Note that this applies to decks shuffled, say, three or four times. Hand deals
- after at least seven shuffles are, for all practical purposes, random. (Sandy
- Kutin pointed to a paper that discusses this mathematically: Bayer & Diaconis,
- "Trailing the Dovetail Shuffle to its Lair," _The Annals of Applied Probability_,
- 1992, 2, no. 2, pp. 294-313)
-
- (When I was a lad, my Mother often admonished me not to "shuffle the spots off
- the cards" on the theory that overshuffling produced flatter deals. Wrong.)
-
- There have been some questions here and on the OKBridge mailing list about the
- randomness of computer deals. Programs that CORRECTLY implement the simple
- algorithms in Knuth [_The Art of Computer Programming, Volume 2, Seminumerical
- Algorithms_, 1969, Addison-Wesley, Reading, Mass.] produce deals that are, for
- all practical purposes, random. OKBridge apparently has it right. I hope the
- ACBL has it right, but I don't know.
-
- However, there are a couple pitfalls* in implementing those algorithms. As
- somebody noted, the dealing program recently posted here hits one of those
- pitfalls. I'm sure that, among the thousands of programs that deal cards,
- a substantial number don't have it right.
-
-
- Tim
-
-
- *The pitfalls:
-
- 1) The least significant bits produced by a linear-congruential pseudo-random
- number generator aren't very random. If the output from your generator
- is an integer in the range [0,2^N) and you want an integer in the range
- [L,U], the best technique is to divide by 2^N (producing a fractional
- result, which can be represented as floating or fixed point), multiply
- by U-L+1, truncate to an integer, and add L. This works very well provided
- U-L << 2^N.
-
- 2) The simplest way for a computer to deal is to pick a card pseudo-randomly
- from the unshuffled deck, then pick another card pseudo-randomly from the
- remaining 51, etc.
-
- If you prefer to "shuffle" the deck in place before dealing, Knuth gives
- an algorithm that is equivalent to the above: Pick a card pseudo-
- randomly from the whole deck (including the bottom card) and swap it with
- the bottom card, then pick a card pseudo-randomly from the top 51 cards
- and swap it with the next-to-bottom card, etc.
-
- Note that only the top 53-N cards should be candidates for the Nth swap.
- This gives each card exactly one chance to end up in each slot.
-
- The pitfall is including the bottom N-1 cards among the candidates for the
- Nth swap. This gives cards originally placed in low slots extra chances to
- be yanked out of those slots. (If this isn't clear, work out the
- probabilities for a three-card deck. You'll discover that three
- permutations are twice as likely as the other three, and that the
- probabilities of a given card ending up in a given slot are 4:3:2 instead
- of 1:1:1.)
-