TI Second Order Randomization 

AV ftp thorhild.cs.ualberta.ca:pubTR91-04.ps.Z 

AV mail britta@cs.ualberta.ca  

OR ALBRT 

LT TR 91-04 

AU Scott D. Goodwin 

MN February 

YR 1991 

AB ABS We examine the problem of reference class selection within
the context of Bacchus and Halpern's combined probability logic.
We identify some problems with the inheritance mechanism proposed
by Bacchus and introduce a new approach based on the idea of
second order randomization.  Under this scheme, the independence
properties of particular predicates are based on the independence
properties of randomized predicates.  This a natural extension to
first order randomization, where the degree of belief in a
proposition about a particular individual is based on the
statistical properties of randomized individuals.  The flavour is
much like that of maximum entropy approaches but we avoid
problems such as syntax sensitivity. This approach solves
problems involving conflicting sources of statistical knowledge
that present difficulties for man;y other approaches.