Re: Today's evil Bayesian math problem

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Thu Sep 11 2003 - 16:37:20 MDT

  • Next message: Chuck Kuecker: "RE: Excel - was RE: How do you calm down the hot-heads?"

    Hal Finney wrote:
    > Eliezer writes:
    >
    >>In fact, for any j != k, p(Wj|Wk) == 4/9, provided that the only other
    >>information given is as described above.
    >
    > That's pretty surprising. In fact, it seems that an even more surprising
    > fact is true: the probability of a white chip on every round j is always
    > 3/8, the same as it is on round 1. And the probability of white chips
    > on both of any two different rounds j and k is always 1/6. From this,
    > Eliezer's formula follows, as p(Wj|Wk) = p(Wj&Wk)/p(Wk) = 1/6 / 3/8 = 4/9.
    >
    > I can prove at least the first statement...
    >
    > I imagine a more complex proof could show that the probability of white
    > on both of any two rounds is constant, and presumably this could be
    > generalized to any n rounds. But I don't see how to prove all this
    > in a simple way.

    Evil Hint #3:

    The classical version of the problem has a bowl where we sample without
    replacement - take the chip, and throw it away. Suppose there are M red
    chips and N chips. If you don't know what you got on the first round, the
    probability of getting a red chip on the second round is still M/N. The
    probability of getting a red chip on the second round, given that you got
    a red chip on the fourth round, is just (M-1)/(N-1), since there's one
    less red chip to go around. We can think of sprinkling the chips over
    holes, instead of drawing them in successive rounds, and it doesn't matter
    if you swap around the labels on the holes.

    The evil version of the bowl, instead of running out of the chips already
    sprinkled, has more and more of them, like a P2P file-sharing network.
    But that doesn't change anything important.

    -- 
    Eliezer S. Yudkowsky                          http://singinst.org/
    Research Fellow, Singularity Institute for Artificial Intelligence
    


    This archive was generated by hypermail 2.1.5 : Thu Sep 11 2003 - 16:49:36 MDT