Can Curiosity Be Programmed?

destinyland writes "AI researcher Jurgen Schmidhuber says his main scientific ambition 'is to build an optimal scientist, then retire.' The Cognitive Robotics professor has worked on problems including artificial ants and even robots that are taught how to tie shoelaces using reinforcement learning, but he believes algorithms can be written that allow the programming of curiosity itself. 'Curiosity is the desire to create or discover more non-random, non-arbitrary, regular data that is novel and surprising...' He's already created art using algorithmic information theory, and can describe the simple algorithmic principle that underlies subjective beauty, creativity, and curiosity itself. And he ultimately addresses the possibility that the entire Universe, including everyone in it, is in principle computable by a completely deterministic computer program."

Comments

Re:Physics of computing the universe

[Urza9814]

Basically - there's no way to store more information in a given area than what it already contains. In order to fully simulate the universe, at full (or greater) speed, you would have to know absolutely everything about absolutely every particle and subatomic particle, etc. And that includes the particles that make up the processor itself.

It's like this: Say you have a 300 DPI printer. You print out a full page of text. Now, you want to fit all the information about that page into some sub-region of the page, printed on the same printer. Ok, so you say you can just shrink the text or encode it in binary or something, which is fine - except somehow also fit the information about the shrunken/encoded text in there. As you can see, you enter a recursive nightmare. And as your printer is a fixed resolution, you would quickly reach a point where any attempts to fit more information results in a blurred pixelated mess.

this isn't exactly new speculation

[Trepidity]

A minority of AI researchers have tackled the problem on and off, and even built some small-scale models of curious agents. One of the classic precursors is Doug Lenat's 1977 system Automated Mathematician, which shifted from the idea of using AI to prove theorems, to instead looking for theorems that would be interesting if they were true (it didn't actually prove them; it was an interesting-conjecture generator). Essentially a model of mathematical curiosity.

Some interesting more recent work is a 2001 thesis that modeled curiosity as a social phenomenon in societies of agents, where agents try to find things that are: 1) new enough to interest its fellow agents; yet not 2) so new that they were incomprehensible in its cultural context.

(I'm an AI researcher, though not precisely in this area.)