Mike,
 
Actually, I'm not sure what the difference is between your proposal and the "game" proposal.
 
In games like SimAnt or Creatures, the game-play basically just consists of exploring a simulated environment full of other organisms....
 
I think your suggestion is basically in the same vein...
 
The risk with these simulated-world approaches is that one is missing the data-richness of the "real" world....  Of course one could then move to an "Ender's Game" style approach -- hopefully not in detail (!!!), but in the sense of having a simulation that draws some of its data from the physical world...
 
-- Ben G
-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Mike Deering
Sent: Thursday, December 12, 2002 10:02 AM
To: [EMAIL PROTECTED]
Subject: Re: Games for AIs (Was: [agi] TLoZ: Link's Awakening.)

This whole approach of successive games is interesting but let me suggest a different route to AI teaching.  Borrow the biological model.  Simulate simplified ecological environments.  Start with a simple organism, perhaps a worm, in a simplified environment with obstacles, rewards, penalties, and other organisms (never alone).  As the AI learns to master the contingencies of the environment gradually evolve the organism and the environment to more complex forms maintaining a contiguous logical path to the final human form.  And if you are lucky and your algorithms are good one day your AI will look up at the simulated sky and scream, "I want to talk to whoever is in charge!  And I want to know what the heck is going on!"
 
Mike Deering.

Reply via email to