Dr. Matthias Heger wrote:
Performance not an unimportant question. I assume that AGI has necessarily
has costs which grow exponentially with the number of states and actions so
that AGI will always be interesting only for toy domains.
My assumption is that human intelligence is not truly general intelligence
and therefore cannot hold as a proof of existence that
AGI is possible. Perhaps we see more intelligence than there really is.
Perhaps the human intelligence is to some extend overestimated and an
illusion as the free will.
Why? In truly general domains every experience of an agent only can be used
for the single certain state and action when the experience was made. Every
time when your algorithm makes generalizations from known state-action pairs
to unknown state-action pairs then this is in fact usage of knowledge about
the underlying state-action space or it is just guessing and only a matter
of luck.
So truly general AGI algorithms must visit every state-action pair at least
once to learn what to do in what state.
Even in small real world domains the state spaces are so big that it would
take longer than the age of the universe to go through all states.
For this reason true AGI is impossible and human intelligence must be narrow
to a certain degree.
I would assert a few things that appear to contradict your assumptions
(and a few that suppport them).
1) AGIs will reach conclusions that are not guaranteed to be correct.
This allows somewhat lossy compression of the input data.
2) AGIs can exist, but will operate in modes. In AGI mode they will be
very expensive and slow. And still be error prone.
3) Humans do have an AGI mode. Probably more than one of them. But
it's so expensive to use and so slow that they strive diligently to
avoid using it, preferring to rely on simple situation-based models (and
discarding most of the input data while doing so).
4) When humans are operating in AGI mode, they are not considering or
using ANY real-time data (except to hold and replay notes). The process
is too slow.
The two AGI modes that I believe people use are 1) mathematics and 2)
experiment. Note that both operate in restricted domains, but within
those domains they *are* general. (E.g., mathematics cannot generate
it's own axioms, postulates, and rules of inference, but given them it
is general.) Because of the restricted domains, many problems can't
even be addressed by either of them, so I suspect the presence of other
AGI modes. Possibly even slower and more expensive to use.
I suppose that one could quibble that since the modes I have identified
are restricted to particular domains, that they aren't *general*
intelligence modes, but as far as I can tell ALL modes of human thought
only operate within restricted domains.
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com
