From: Ben Goertzel (email@example.com)
Date: Mon Oct 06 2003 - 07:28:30 MDT
Yes, the AI design you suggest would be prohibitively computationally
complex, -- in a world, intractable, now or in a hundred years.
Think about the inclusion-exclusion theorem from set theory, which is a
necessary part of elementary probability. To compute conditional
probabilities involving combinations of n items, it requires estimation of
the probabilities of 2^n combinations. To avoid this, one makes heuristic
For another way to look at it, look at Marcus Hutter's recent work on the
AIXItl artificial intelligence system, which uses ideas from statistical
decision theory (inclusive of Bayes' Theorem). Provably intelligent, but
computationally totally intractable.
As someone else said in their reply to you: more than coding tricks and
compression are needed, clever approximative heuristics are needed, and the
system for managing, balancing, tuning and adapting these clever
approximative heuristics is called a mind.
My own Novamente AI design relies heavily on probability theory, but
deployed in a certain way, within a particular overarching framework ... and
we're well aware of the limitations of this approach, which essentially
arise from the inability to fully expand the above-mentioned
inclusion-exclusion formula in most contexts, and the ensuing need to make
(implicit or explicit) heuristic independence assumptions.
-- Ben Goertzel
From: firstname.lastname@example.org [mailto:email@example.com]On Behalf Of Metaqualia
Sent: Monday, October 06, 2003 12:36 AM
Subject: Feasibility: 100% Bayesian systems
Would the set of all bayesian probability data necessary to replicate a
100IQ human be prohibitively large and therefore impossible to store on a
modern supercomputer? (exclude the visual cortex, just think common sense
reorganized in a bayesian fashion)
Would the lookup time for a moderately complex thought be too long in such
system? (imagine that all probabilities are stored on a hard disk).
Programming tricks, compression, anything goes to reduce the size - but
after nothing more can be done and it's just a matter of rough storage space
and speed, are 100% bayesian human equivalent AIs theoretically possible or
impossible to implement with present off the shelves technology?
Is there any use for alternatives to bayes a cognition paradigms?
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT