Re: BN prior prob

Rich Neapolitan (neo@megsinet.net)
Mon, 07 Jun 1999 07:42:39 -0500

My experience is that if you use an equivalent sample size, the variance in
the relative frequency is unchanged when you propagate down the network
(which makes sense intuitively). However, it increases when you propagate
up and the increase is relative to the number of paths from which
information is coming up. For example, if we are instantiating children,
two such children results in a bigger increase than one, three bigger than
two, and so on. This also makes sense when you think about the proportion
of the equivalent sample that goes into the calculation.

Rich

At 01:02 PM 6/6/99 -0600, Bob Welch wrote:
>Dongsong :
>
>If you start with an imprecise or "uninformative" prior, then updating with
>new evidence will generate a more precise or informed posterior -- how much
>more informed depends on the likelihood's ability to discriminate.
>
>But Bayesian updating does not necessarily result in more precision -- the
>prior could have been very precise and the evidence a surprise given that
>prior -- this results in a less precise posterior.
>
>Eventually (as more and more "surprising" evidence is observed), the
>posterior becomes more precise as it concludes that the original priors were
>incorrect.
>
>Bob
>
>-----Original Message-----
>From: Dongsong Zeng <zengdong@pilot.msu.edu>
>To: uai@CS.ORST.EDU <uai@CS.ORST.EDU>
>Date: Friday, June 04, 1999 12:28 PM
>Subject: BN prior prob
>
>
>>Hi,
>>
>>if my Belief Networks starts with unprecise prior probabilities, can I get
>>precise results? How and Why? Any advices are highly appreciated.
>>
>>Thank you very much.
>>
>>
>> --
>>Dongsong Zeng
>>
>>Email:zengdong@pilot.msu.edu
>
>