I previously posted here claiming that the human mind (and therefore
an ideal AGI) entertains uncomputable models, counter to the
AIXI/Solomonoff model. There was little enthusiasm about this idea. :)
Anyway, I hope I'm not being too annoying if I try to argue the point
once again. This paper also argues the point:

http://www.osl.iu.edu/~kyross/pub/new-godelian.pdf

The paper includes a study of the uncomputable "busy beaver" function
up to x=6. The authors claim that their success at computing busy
beaver strongly suggests that humans can hypercompute.

I believe the authors take this to imply that AGI cannot succeed on
current hardware; I am not suggesting this. Instead, I offer a fairly
concrete way to make deductions using a restricted class of
uncomputable models, as an illustration of the idea (and as weak
evidence that the general case can be embodied on computers).

The method is essentially nonmonotonic logic. Computable predicates
can be represented in any normal way (1st-order logic, lambda
calculus, a standard programming language...). Computably enumerable
predicates (such as the halting problem) are represented by a default
assumption of "false", plus the computable method of enumerating true
cases. To reason about such a predicate, the system allocates however
much time it can spare to trying to prove a case true; if at the end
of that time it has not found a proof by the enumeration method, it
considers it false. (Of course it could come back later and try
harder, too.) Co-enumerable predicates are similarly assumed true
until a counterexample is found.

Similar methodology can extend the class of uncomputables we can
handle somewhat farther. Consider the predicate "all turing machines
of class N halt", where N is a computably enumerable class. Neither
the true cases nor the false cases of this predicate are computably
enumerable. Nonetheless, we can characterize the predicate by assuming
it is true until a counterexample is "found": a turing machine that
doesn't seem to halt when run as long as we can afford to run it. If
our best efforts (within time constraints) fail to find such a
machine, then we stick with the default assumption "true". (A
simplistic nonmonotonic logic can't quite handle this: at any stage of
the search, we would have many turing machines still at their default
status of "nonhalting", which would make the predicate seem
always-false; we need to only admit assumptions that  have been
"hardened" by trying to disprove them for some amount of time.)

This may sound "so easy that a Turing machine could do it". And it is,
for set cutoffs. But the point is that an AGI that only considered
computable models, such as AIXI, would never converge to the correct
model in a world that contained anything uncomputable, whereas it
seems a human could. (AIXI would find turing machines that were
ever-closer to the right model, but could never see that there was a
simple pattern behind these ever-larger machines.)

I hope that this makes my claim sound less extreme. Uncomputable
models of the world are not really so hard to reason about, if we're
comfortable with a logic that makes probably-true conclusions rather
than definitely-true.


-------------------------------------------
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=106510220-47b225
Powered by Listbox: http://www.listbox.com

Reply via email to