using game programming to develop animal-equivalent agi

From: Kevin Osborne (kevin.osborne@gmail.com)
Date: Thu Mar 16 2006 - 05:46:28 MST


the list has been a little quiet for a couple of days, so I thought
I'd gag up something I've been mulling over for a while...

I'm thinking it might be possible to develop evolving in-game
lifeforms of trivial general intelligence. and I think that the code
is not only possible today, but that it's already here.

let me explain my basis in contrast to the kurzweil/moravec notion of
replacing existing brain elements with inorganic equivalents to
produce a thinking being - with the contrast being that I see the
moravec approach as being top-down, while I believe there is a
distinct possibility of a bottom-up path-to-intelligence.

two more elements are needed for this position: evolution and reverse
engineering. of course the validity of a moravec-mind is needed as
well, but let's just make than an assumption for this case that can
(and I'm sure has) be discussed at length in other threads.

start with an immersive 3D gameworld with a simulation engine running
newtonian physics and genetic biology. there must be hundreds of
running codebases today that could be merged to provide this.

then, you reverse engineer the properties of an amoeba. you define
it's sequenced genome in logic, and you assign these properties and
abilities to in-game programmed objects, and then you let them loose
in a newtonian phyics based simworld with elements, seasons, sunlight
and resources. Remember that much of existing science, and physics
especially, is just reverse engineering; gravity existed long before
we named it so, as did carbon. We now have a very, very complete
understanding of the sub-cosmic and super-quantum realms; more than
enough to replicate organic existence with incredible realism.

now we know all the answers to how this world should behave, and how
this ameoba should develop, as it's already happened in our past; we
have a perfect history of it all preserved for comparison. if the
world starts to diverge markedly, you can tweak the ratios that
determine random changes via scarcity of resources, overpopulation,
climate change etc that drive evolutionary development. you can
program the basic imperatives of an amoeba; feed, grow, compete,
survive, divide. yes, the simulation is more complex than my basic
examples. but I think rigorous apllication of attention to detail
could overcome any inconsistencies or incompleteness.

if you build on this, again and again, in painstaking detail, you
should eventually get to something substantial. how many data points
are required to map the complete genetic evolutionary development from
an ameoba to a fieldmouse? alot? sure. an infinite amount? no. we have
those records, preserved for posterity in thin layers of earth; we can
trace every step, and then provide a mirror environment to reproduce
it. and if we can get from ameoba-like to mouse-like satisfactorily,
can we then progress our bottom-up approach from mouse-like to, well,
dog-like? monkey-like? man-like? you tell me.

if this is all sounding too outlandish or complex to simulate at the
technical level, there's a another 80's tech staple thats coming out
if it's own winter to help out: Procedural Generation (PG). This is
how the gameworld terrains of 80's titles generated their in-game
enivronments when gig's of animations on a dvd weren't an option. This
procedural approach says you don't have to create a world by hand; you
just set up the rules, press go, and let the world generate itself.
This is also the technique used by demoscene developers; check out
what 64k(!) of procedural generation can do here:
http://awards.scene.org/nominees.php?cat=9

Wil Wright is the gaming legend is currently championing this, and
Spore is the game due out this year which is currently knocking my
socks off.

the first third of this video is a must-watch:
http://video.google.com/videoplay?docid=8372603330420559198&q=spore

and a writeup of the game (and PG) is here:
http://en.wikipedia.org/wiki/Spore_(game)

Wright isn't just using PG for textures or tiles; he's also using it
for biomechanics and evolutionary characteristics like swimming,
walking, hunting, killing, eating and mating. None of the in-game
creatures in Spore are modelled prior to installing the game. every
single one evolves using an array of possible physical development
paths to come to life as completely unique organisms that execute
motile ability in a completely (un)natural way. If you haven't watch
the video linked above, I'd recommend doing so now.

Now the interesting part to all this is that when playing Spore,
you'll be competing against AI-generated evolving orgranisms; and I
think if the simulation is realistic enough, and the properties of the
organisms detailed enough, then we'll be starting to enter the
god-mode lobby of the SA you lot have been fascinated with lately.

It seems more and more that if we use sufficently comprehensive narrow
AI to generate
simulations of our in-historic-light-cone world that we can create
glimpes of real intelligence in our simulations.

There are growing examples of in-game AI generating AGI-like eruptions
of higher-intelligence-like behaviour. One the recent (elder scrolls
iv?) mmorpg games had testers walk into a in-game-evolved
previously-unvisited villiage where none of the townspeople would let
them in. turns out the villiage guards had gotten drunk and gone
hunting and gotten lost, and then bandits had raided the villiage.
each narrow-AI indivdiual guard|bandit|townsperson had combined to
generate an effect more like that which could be envisioned by a
higher intellect.

there's also this older example from 2001:
"This ability to discriminate has already produced some real surprises
for testers and creators alike during development. In one situation,
there was a creature that kept losing a stone-throwing game to another
creature. To get revenge, the first creature heated a rock, planted it
in the pile of stones to be thrown and then fell on the floor laughing
when its rival burned its hands. All this happened without
intervention from the player."
- http://news.bbc.co.uk/2/hi/entertainment/1237848.stm

The next-gen consoles with their Cell processors are supposedly ideal
for the 'Procedural Syntheisis' needed to implement Procedural
Generation - does this mean the next glimmer of basic AGI will come
from the living room instead of the lab?



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT