> I believe that to be adequate, the code language must incorporate
> something loosely analogous to probabilistic logic (however
> implemented) and something analogous to higher-order functions
> (however implemented). I.e. it must be sensibly viewable as a
> probabilistic logic based functional programming language -- even if
> at first glance it doesn't look anything like that.
Why must it be ... ?
Why probability theory?
Because of the numerous math results (starting with Cox's) arguing
that probability theory is the only sensible way to represent
uncertain knowledge ... and the fact that representation of
uncertainty is critical to any realistically useful KR ...
Why higher-order functions?
Because the only tractable ways we know of to represent highly complex
patterns and procedures are
-- using nested quantified variables
-- using higher-order functions
The former option leads to nasty complexities of a type that are very
unlikely to exist in the brain, and that don't lend themselves well to
adaptive learning. OTOH the viability of the brain using higher-order
functions (neural nets that take coded versions of other neural nets
as inputs) is much clearer, and the use of probabilistic reasoning
with higher-order functions leads to less arbitrariness than the use
of probaiblistic reasoning with quantified variables.
(As it happens NM uses variables as well as higher-order functions,
but I don't consider this necessary...)
Novamente explicitly implements probabilistic relationships and
higher-order functions in its KR ... but this is not necessary...
One could also (for example) use a neural network architecture, in which
-- probabilistic relationships are expressed approximately via the
conductances of bundles of synapses btw neuron assemblies
-- higher-order functions are expressed via neural assemblies whose
inputs are coded versions of other neural assemblies (with the coding
process carried out perhaps partly or wholly in the hippocampus and
routinely done within "working memory")
I give this example to show that the use of probabilities and
higher-order functions may be implicit rather than explicit.... But
when creating an AI for von Neumann machines it seems to me that the
explicit approach is probably better (hence the NM approach)
-- Ben
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]
