Yes, I guess that's in principle a test of AGI  (& that's why it's worth 
discussing this subject at length).

I don't think in practice it will be anything but an extremely far distant AGI 
*robotic* project - although I haven't thought this through yet.

There are two major problems here.

First, most videogames are played by humans and depend on a human embodied 
player.

Ok, so you forget about those, and concentrate on games that can play by 
themselves without human intervention - like chess, draughts, etc.

Your problem then is presumably - off the top of my head - to define a 
generalizable concept of "move" and "piece", that will enable your 
computational would-be AGI to absorb entirely new games,with new, diverse kinds 
of  moves and pieces.  Again, they should be any games that are comparable to 
chess -  presumably board games with pieces that move?

I doubt that those generalizable concepts are possible.  My initial hypothesis 
is that too many different kinds of moves are possible.

But by all means take this further.

P.S. I do think that you could have sub-AGI programs that creatively explore 
many different lines of movement towards goals and present them to humans for 
final judgment (although I don't think they apply here to games).



From: Ben Goertzel 
Sent: Saturday, June 09, 2012 9:49 PM
To: AGI 
Subject: Re: [agi] The 2 Tests of AGI - generalizability & creativity


Mike T,

A fairly decent purely computational AGI test (suggested by Sam Adams from IBM) 
would be the Video Game Test

--- Learn to play a series of randomly chosen human videogames, based only on 
interaction with the games themselves, and the goal of maximizing score (or 
whatever the goal of each game is)

Of course, this only works if the AGI designer has looked at, say, 1% or fewer 
of existing videogames when building his system.   It doesn't work if the AGI 
designer has hired 1000000 programmers to write game-playing AIs separately for 
each game, and then wired them together

This is not a perfect test, but success on it would be rather compelling...

We briefly discussed this along with the Woz coffee test and others in our 
article on AGI in the recent issue of AI Magazine...

... ben g



On Sat, Jun 9, 2012 at 3:04 PM, Mike Tintner <[email protected]> wrote:

  Sergio,

  The Woz test as I indicated to Bob is indeed extremely complicated. I used it 
only because it's already out there - and is therefore helpful as a *loose* 
guide/image.

  The other isn't really Ben's - it's the basic fetch test a dog faces -  he 
must (and will) fetch a ball thrown by his master in more or less any field -

  this basically means he must (and will) negotiate more or less any unfamiliar 
terrain (within loose limits) -

  he can create and negotiate a course across terrains of grassy clumps,  rocky 
ground, sandy beach,  furniture and furnishings in a building et al - all of 
which will spring surprises

  also, of course, the ball could end up hidden from view in different ways and 
situations

  there's no way the dog could be specifically preprogrammed for every new 
terrain and hidden ball...(nor, by extension, is there any complex "set"  that 
can infer the features of every new terrain)

  if your robot can simply negotiate new .terrain after new terrain somewhat 
like a dog (or all other life forms) and not even fetch a ball - it's AGI

  If we were talking a relatively simple practical starting-point, I would 
suggest aiming for a robot that could negotiate just a few metres of endlessly 
diverse terrains (wh. is more or less what roboticists are attempting now, 
although I'll bet they all still cheat)..

  P.S. I don't think a purely computational AGI project is possible. Once you 
think in depth about the goals of generalizability and creativity, you will 
realise they depend on being implemented by a body with an extensive 
range/spectrum of different lines of movement and observation.  The body is the 
foundation of generality and creativity - it affords the capacity to always try 
out new lines of movement and looking, and handle objects and negotiate 
terrains in new ways.

  By all means try to outline a project that contradicts me. It will be 
interesting regardless.



  From: Sergio Pissanetzky 
  Sent: Saturday, June 09, 2012 7:27 PM
  To: AGI 
  Subject: RE: [agi] The 2 Tests of AGI - generalizability & creativity


  Mike, 



  I like the concept of the Woz test. However, the test itself has three 
problems. It is unfair to those who do not build robots, and it requires the 
ability to recognize images, which is in itself a major test. The third 
problem, it requires considerable computer power, besides generalizability and 
creativity. It would be unfair to those who may have a good idea but lack the 
necessary power, such as me. Do you think it can be rephrased so as to 
eliminate these limitations?



  Can you please explain what is Ben's fetch test?



  Sergio



  From: Mike Tintner [mailto:[email protected]] 
  Sent: Saturday, June 09, 2012 5:37 AM
  To: AGI
  Subject: [agi] Re: The 2 Tests of AGI - generalizability & creativity



  P.S. The Woz Test {"go and make a cup of coffee in this new kitchen")  is a 
test of creativity - of being able to design a course of action without 
specific programming.



  But (correct me) it isn't defined as a test of creativity - and should be.



  Note: there is a great deal of underlying unanimity here  - in the Woz Test, 
Ben's fetch test and similar - but the basic principles involved 
(generalizability and creativity) haven't been clearly spelled out.

        AGI | Archives | Modify Your Subscription
       
       



        AGI | Archives  | Modify Your Subscription   

        AGI | Archives  | Modify Your Subscription  




-- 
Ben Goertzel, PhD
http://goertzel.org

"My humanity is a constant self-overcoming" -- Friedrich Nietzsche


      AGI | Archives  | Modify Your Subscription   



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to