Hello,

Last spring, before my hard drive crashed, I started hacking togeather
an workbench for bootstrapping an AI in Squeak. ( www.squeak.org ).

My only notable accomplishment from that effort was to hack togeather a
black and white "laplankian" algorithm (I don't remember the name, I'd
have to search my HD for it..) It wasn't a very good implementation but
it succeded as a proof of concept.

I now have the time and resources to resume the effort.

I have decded that the best approach for a next phase is to repair the
badly-broken X-server for Squeak, use it to run an instance of ZSNES, a
super Nintendo Emulator, at it's minimum resolution, and then map the
controls into my AI.

My hope is that I can find a game that is sufficiently "Learn as you
play" that a purely learning-based AI can succede in demonstrating some
competiency at playing the game.

My tentative selection in that regards is Zelda (3), While there are a
number of places where language skills would be very useful, it mostly
relies on pictograms and graphical gestures to communicate the most
important information.

The specific performance goals of this project are:

A. To demonstrate an AI system that can interract with the game in
scaled real time. and has the capability to -- I intend to down-clock
the game so that the other CPU can have enough time to process the
stimuli at human-equivalent speed. In this regards the limiting factor
is my patience and not the available CPU power.

B. To show that the AI is capable of reacting to its environment, for
example that it can detect incomming attacks and make a recognisable
effort to avoid them.

C. That it is able to navigate in the game world, that it can
demonstrate the ability to find appropriate paths through obsticles...
-- this would be demonstrated if it can navigate through an environment
without resorting to some brute-force algorithm.


These are the minimum success criteria.... The project will be an
astounding success if Link (the charactor's name) can make it out of the
first dungeon... If we leave it alone for a few days and come back to
find him facing off against Gannon (the final "Boss") then we've got
something really interesting on our hands... (I'm not expecting that,
well er, trying not to expect that. ;)

I'm not sure wheather I want an AI raised in this way out in the real
world... It might carelessly kill someone expecting that he would be
carrying faeries who would bring him back to life... =P (who wouldn't!)


Anyway, the research goals for this project are:

1. To develop a working cybernetic loop.
2. To test out my crazy theories about learning algorithms.
3. To discover what types of motovational programming are required to
cause the AI to demonstrate enough curriosity that it will explore its
environment and hence be able to discover what it needs to get around
the game. -- as opposed to some degenerate behavior such as doing
nothing at all or performing repetitive behaviors in areas where there
are few dangers.



To get this to work I need major help in untangling the brokenness of
the current Squeak X-server, (It is quite severe. =( )

I also could use some experts with the X protocol in general and getting
the ugly ugly details worked out. -- that's been a killer in my previous
efforts.