Hey there - I'm a programmer from germany and currently working on a scientific paper on Ai - I need a precise definition of the expression "Ai" - can somebody give me a hint on where to get a good one? thx

Fiesch

Hi,

I also wanted to have a definition of AI and that is why I made one by myself. You can find it on the address:
http://dobrev.com/AI/definition.html

If you are lazy to read all the paper then here is the definition in short:
AI will be such a program which in an arbitrary world will cope not worse than a human.

Here the arbitrary world can be for example one game.

which human? I've met some very smart ones and some very stupid ones, the average seems to lean heavily toward the stupid end of the scale, esp in western countries.

 > > AI will be such a program which in an arbitrary world will cope not
 > > worse than a human.
 >
 > which human? I've met some very smart ones and some very stupid ones, the
 > average seems to lean heavily toward the stupid end of the scale, esp in
 > western countries.

Western countries don't necessarily have a greater proportion of stupid people - they just have infrastructure that allows stupid people to express themselves to a greater audience.
 (whoops - I just expressed myself  8^)

The problem of "which human" is a real one. Even if we try to define an "average human" we need a way to measure humans to decide what the average is. Any system of measurement of a human (on the basis of intelligence/stupidity) must already have a definition of "intelligence". My solution to this problem would be to adjust Dimiter's AI definition to:     "AI will be such a program which in an arbitrary world will cope not worse than 50 percent of  humans."

Virgil.

 > Western countries don't necessarily have a greater proportion of stupid
 > people - they just have infrastructure that allows stupid people to express
 > themselves to a greater audience.
 > (whoops - I just expressed myself  8^)

I agree that the stupid people do get to express themselves more, but I still argue that we have more of them, further more I argue that the more we increase the standard of living and life expectancy and decrease things like war the stupider they are going to get.

 > solution to this problem would be to adjust Dimiter's AI definition to:
 >    "AI will be such a program which in an arbitrary world will cope not
 >    worse than 50 percent of  humans."

the new definition definitely resolves the problem.

but in general this definition leads into some very interesting lines of argument, for example take a tic tac toe program that has a internal graph of all possible moves and can calculate from that which branches provide the best chances of winning and thus is VERY good at tic tac toe, are we happy that within the bounds of the game the program is intelligent?

 > > Western countries don't necessarily have a greater proportion of stupid
 > > people - they just have infrastructure that allows stupid people to express
 > > themselves to a greater audience.
 > > (whoops - I just expressed myself  8^)
 >
 > I agree that the stupid people do get to express themselves more, but I
 > still argue that we have more of them, further more I argue that the more we
 > increase the standard of living and life expectancy and decrease things like
 > war the stupider they are going to get.

 Ok, the natural selection argument works for me.

 > > solution to this problem would be to adjust Dimiter's AI definition to:
 > >    "AI will be such a program which in an arbitrary world will cope not
 > >    worse than 50 percent of  humans."
 >
 > the new definition definitely resolves the problem.
 >
 > but in general this definition leads into some very interesting lines of
 > argument, for example take a tic tac toe program that has a internal graph
 > of all possible moves and can calculate from that which branches provide the
 > best chances of winning and thus is VERY good at tic tac toe, are we happy
 > that within the bounds of the game the program is intelligent?

Dimiter's page does cover the issue of programs designed for a particular world. The word "arbitrary" in the definition prevents us applying the definition to any one single world and drawing any conclusions.
 I do have some other concerns with Dimiter's definition:
 1. It gives no bounds to the variety of worlds and definition of "cope". I would prefer the definition included "...in an arbitrary world, given an arbitrary objective...". Part of being a human and having a distinct identity relies on our ability to choose what our objectives are. Since different humans in the same situation will set themselves different objectives, I consider any test of "intelligence" should exclude factors based on choosing an objective.
 2. It can never be used to create a general finite test of AI (maybe constructing such a test is as hard as actually creating AI). It is good as a thought experiment because it does provide a means to answer questions like the one you raised: "Is my TTT program artificially intelligent?". Using the definition, we only have to identify one other world in which the TTT program would perform worse than most humans to say that it is not artificially intelligent.

Virgil.

 > > solution to this problem would be to adjust Dimiter's AI definition to:
 > >    "AI will be such a program which in an arbitrary world will cope not
 > >    worse than 50 percent of  humans."
 >
 > the new definition definitely resolves the problem.
 >
 > but in general this definition leads into some very interesting lines of
 > argument, for example take a tic tac toe program that has a internal graph
 > of all possible moves and can calculate from that which branches provide the
 > best chances of winning and thus is VERY good at tic tac toe, are we happy
 > that within the bounds of the game the program is intelligent?

Actually, this perhaps deals with another aspect of the definition with which I was not happy. The definition provides a pure yes/no answer to the question of whether a program has AI. It doesn't easily tell you whether one program that fails is closer to passing than another program until the programs get "smart" enough to do better than at least a finite portion of humans (after that you can rate AI based on the proportion of humans that it beats).

From a games perspective, it means that no game produced so far contains any AI.

To assess game AI, I would temper the definition to:
     "AI will be such a program which in a specified world, given an arbitrary
    world state and an arbitrary objective will cope not worse than 50 percent
     of  humans."
We can then answer the question "is my program intelligent within the bounds of this game world?". For simple games like TTT the question is then easy to answer and is testable (using a representative sample of humans).

 > > > AI will be such a program which in an arbitrary world will cope
 > > > not worse than a human.
 > >
 > > which human? I've met some very smart ones and some very stupid
 > > ones, the average seems to lean heavily toward the stupid end of
 > > the scale esp in,western countries.
 >
 > Western countries don't necessarily have a greater proportion of
 > stupid people - they just have infrastructure that allows stupid
 > people to express themselves to a greater audience.
 > (whoops - I just expressed myself  8^)

My opinion about this is that in the west countries the life is easier and when the world is more simple in it cope more people.

 > The problem of "which human" is a real one. Even if we try to define
 > an "average human" we need a way to measure humans to decide what
 > the average is. Any system of measurement of a human (on the basis of
 > intelligence/stupidity) must already have a definition of
 > "intelligence". My solution to this problem would be to adjust
 > Dimiter's AI definition to: "AI will be such a program which in an
 > arbitrary world will cope not worse than 50 percent of  humans."
 >
 > Virgil.
 >

You ask which people I mean. My definition is not for AI but for Intelligence. If you can cope in more worlds than me this means that you are more intelligent. The test which I proposed depends from the person which chooses the test worlds. One AI can be more intelligent or less and can pass one test and fail on another. For example I made a test world and gave it to my students. All of them fail on this test despite that the world was simple (it was Tick-Tack-Toe game). I think that my AI should to cope well in this world but for the moment no one human passed this test.

This test and another more simple one will be added as examples in the next version of our product Strawberry Prolog (it will be released in a week). So you can take this tests and to check how intelligent are you ;-)

 > Hi,
 >
 > I also wanted to have a definition of AI and that is why I made one by
 > myself. You can find it on the address:
 > http://dobrev.com/AI/definition.html
 >
 > If you are lazy to read all the paper then here is the definition in
 > short:
 > AI will be such a program which in an arbitrary world will cope not
 > worse than a human.
 >
 > Here the arbitrary world can be for example one game.

When you use this definition as a test for AI it has a very strong bias towards the negative answer. You only have to identify one arbitrary world in which a human performs better in order to categorize the program as non-AI. Suppose we have a "typical" human (called Fred) as our reference. We then take another "typical" human (called Bruce) as a pretend AI program. Since they are two different humans, there will be worlds in which Fred does consistently better than Bruce, and other worlds where the converse applies. Using the definition, Bruce does not exhibit AI. As far as I can see, any program that does pass the test must be exhibiting super-human intelligence, not merely AI. The definition needs to allow for some failures in some proportion of the set of arbitrary worlds.

 Virgil.

 > When you use this definition as a test for AI it has a very strong bias
 > towards the negative answer. You only have to identify one arbitrary world
 > in which a human performs better in order to categorize the program as
 > non-AI.
 > Suppose we have a "typical" human (called Fred) as our reference. We then
 > take another "typical" human (called Bruce) as a pretend AI program. Since
 > they are two different humans, there will be worlds in which Fred does
 > consistently better than Bruce, and other worlds where the converse applies.
 > Using the definition, Bruce does not exhibit AI.
 > As far as I can see, any program that does pass the test must be exhibiting
 > super-human intelligence, not merely AI.
 > The definition needs to allow for some failures in some proportion of the
 > set of arbitrary worlds.

and here you hit one of the leading problems of AI, you've created a definition under which we are no longer intelligent. the counter also happens often when you create a definition which identifies things we don't believe to be intelligent as being intelligent. personally I believe that it is impossible to form a definition of intelligence which identifies all the things think are not intelligent as such and identifies us as intelligent. some of the things we strongly believe to not be intelligent are in fact intelligent or we arn't!