Robin Hansen writes:
>I argue that not all possible priors are
>rational -- rational priors should be obtained from "uberpriors" via
>conditioning.
I don't see any arguments for that conclusion. You show that
rational Bayesians should agree if and only if that's true. The assumption
is outrageously false; it requires that prior differ only due to valid
reasoning based on the evidence of the beliefs themselves. Human
uninformed beliefs are caused by developmental, genetic, and memetic
processes that have essentially nothing to do with the truth of such
things in the real world. The reason my mother and I differ on the
existence of God has nothing to do with the existence of God.
In addition, it's an odd "norm", at least for humans. We have the priors
we have; to change them requires violating Bayesianism.
>I do *not* assume that everyone has the same uberprior,
>actually, but I do describe plausible constraints on such uberpriors that
>imply common priors.
You assume everybody agree on world_state_function/belief_state and
on probability(belief_state). It's trivial to derive a
belief-state-independent
world state function (i.e. a prior) just by averaging
world_state_function/belief_state
by probability(belief_state). So you do assume everybody has the same
uberprior.
I would actually split your above assumption into two parts; one, that
everybody
has the same uberprior, and two, that everybody agrees on the evidential
status of beliefs. The two assumptions are quite independent, and putting
them this way makes it easier to discern how most people would address
the issue. Rational people understand that uninformed beliefs are minimally
informative about the world and so effectively agree on the evidential status
of beliefs; hence they disagree on the uberprior. phrased
as "what would you expect the world to be like if you didn't exist". My
mother
thinks God would exist even if she didn't; I think if I weren't here God still
wouldn't either. Ditto for differences on this group on the appropriate value
of Cisco stock, plausibility of friendly AI, etc.
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 10:00:04 MDT