Re: using game programming to develop animal-equivalent agi

From: Phillip Huggan (
Date: Thu Mar 16 2006 - 13:46:46 MST

What interests me about AGI is that it can become a recursive physics learner; potentially spitting out engineering blueprints that can solve many of our problems. Are there any baby-AGI projects, AI projects or game environments that have the potential to model a recursive Philosophy learner? I'm not refering to game-theoric economics AIs that allocate scarce resources to maximize portfolios gains, or anything like that. I guess the key is to for the AI to understand the "neurology" of a game-agent's goal systems. And for the AGI to implement solutions that appeal to rearranging the game environment or modifying agent sensory stimulus, rather than tinkering with the "hypothalamus" of the agent's goal systems. I'm not ready to invest the time in creating such an environment, but if there is already a game out there that can be tweaked a little, I'll contribute.

Ben Goertzel <> wrote:

> 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.

If you want to experiment with this in practice, I suggest you may
want to join the AGISim project, a spin-off of the Novamente AGI
project which is involved with creating an open-source 3D environment
specifically for AGI development, based on the open-source
CrystalSpace game engine. The Sourceforge page for AGISim is here:

The one to email if you're interested in becoming involved is Ari
Heljakka, "heljakka at iki dot fi"

For example, one task that has not yet been done is to integrate the
ODE physics engine into AGISim (it already words with CrystalSpace so
this is not a fundamentally hard task, we just haven't gotten around
to it yet).

However, this kind of physics engine is much too simplistic to
support any kind of realistic physics-based genetics. In fact, we
don't yet know how to make a physics engine that supports
physics-based genetics, because no one yet understands the physics
underlying protein-folding very well.

Nevertheless, I think one could probably make an interesting
artificial-genetics-based artificial life framework within AGISim,
which would be nifty because then the adaptive, evolving artificial
organisms could interact with humanoids controlled by Novamente ;-)

Regarding your more general point, I doubt very much that AGI is going
to emerge spontaneously from AI bots inside computer games. This is
for a number of reasons, including

* the first AGI is probably going to push the limits of computing
capability, requiring a network of machines with a lot of RAM, whereas
AI's running in games need to work on the kind of hardware setup a
typical gamer can afford

* a baby AGI is going to be very stupid and boring to play with, which
is not a great gaming experience. Note that Creatures, while cool,
was never a very popular game; and Black and White 2 has much less
focus on teaching the AI's than Black and White 1 did. I see no
evidence that teaching AI babies is actually considered a fun gaming
experience by any substantial portion of the gaming market.

I do think that gaming-style simulation software like CrystalSpace can
be very useful for AGI and Alife research, but it must be used in
somewhat different ways than is appropriate for commercial gaming.

Of course, once you have created a reasonably intelligent AGI, you may
find ways to deploy it within gaming environments for fun and profit.
For instance, an AGI-powered NPC would be really cool in an massive
multiplayer online role-playing game like World of Warcraft or Final
Fantasy 11. But the AGI would have to already be pretty clever to be
marketable in that context. And of course one may find even more
interesting things to do with a clever AGI...

 Yahoo! Mail
 Use Photomail to share photos without annoying attachments.

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