This would be the paper, everyone:
http://www.vetta.org/documents/IDSIA-12-06-1.pdf
Shane - first you smack down the Goedel machine, and now AIXI! Is it
genuinely
useless in practice, do you think? Hutter says one of his current
research priorities
is to shrink it down into something that can run on existing machines...
Of course it's genuinely useless in practice!
What AIXI does is to continually search through the space of all
possible programs, to find the one that in hindsight (based on
probabilistic inference with an Occam prior) would have best helped it
achieve its goals -- and then enact that program.
This is not something you can do on a realistic scale.
Sure, you can view any learning system as a specifically biased way of
"searching through the space of all possible programs", but this insight
gets you about .1% of the way toward designing a thinking machine...
To make an analogy: if you have enough time, you can generate a literary
masterpiece by randomly typing characters until a masterpiece happens to
appear. However, "scaling down" this strategy to reasonably small
amounts of time doesn't work very well. In fact the strategy has nearly
nothing to do with appropriate strategies for generating literary
masterpieces given realistic amounts of time. And the parallels one can
draw (yes, random variation plays a role in creative insight, etc.)
really don't get you very far toward knowing how to realistically
produce a literary masterpiece.
-- Ben G
-----
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/?list_id=11983
