Programming w/ Artificial Intelligence

[complacent]

I was wondering if anyone knew of practical alternatives to AI in computer applications other than evolutionary programming, or if evolutionary programming with a nerual network is state of the art?
I don't want to know about a list of scripted responses that give games a feel of AI. I am speaking specifically about programs that actually change their responses if given information about a current state, a desired state, and some output.

Also, could such evolutionary programming be useful in game AI? Certainly it wouldn't work with most console games, since there is no storage for an evolving neural network. But, could Half-Life 3 use a neural network to better "guess" at your next combat move?

I wrote a mancala program in C++ that uses a neural network with initially random weights, and used an evolution method where a subset of the current generation was kept based upon number of wins against opponents, and their neural net weights were varied using a specific formula. I was impressed when my own program, which I could initially beat, consistently handed me my own ass after only 50 generations of evolution (which took approximately 4 days to get to.)

 

[AbsolutDealage]

20q

If you havn't seen it before, its a stunning case of real world neural network AI programming. Check it out.

 

[hoppa]

hah i just kicked its @ss with "camera tripod." 30 guesses and it wasn't even close! i guess when something comes in so many forms its hard to define. i answer "sometimes" to nearly every question

 

[Jeffyboy]

usually they teach AI with prolog or LISP or some language they think is AI related... lots of major attempts at developing large scale intelligence usually comes in to the limelight for a moment then disappears. It's VERY difficult since a computer really has no senses... no visual & audio input that is interpreted real-time. The only understanding a computer has of the world is what you tell it.

Intelligence would be the ability to classify objects and create new classifications. Understand language and context. Sort priorities according to where you are in society... it's hard to conceive... I think good AI won't be seen in our lifetime.
Jeff

 

[Cattlegod]

it gave up on "penis"

 

[everman]

I tried for espresso, and it came up with vomit...needs a little work

 

[Vee]

I fugured I was going to give it an easy one with a Jet airliner. Not so easy after all. I won. It took like ~25 questions to get it.
After that, it also volunteered the following "information". Quite interesting.

****
Uncommon Knowledge about a jet
Can you play games with it? I say Probably.
Is it a mammal? I say Probably.
Does it have a trunk? I say Probably.
Does it have claws? I say Yes.
Can it live out of water? I say Doubtful.
Can liquids pass through it? I say Yes.
Is it human powered? I say Yes.
Can it be worn by a person? I say Yes.
Does it live above ground? I say No.
Can it be placed on your head? I say Probably.
Does it have a hole in it? I say Yes.
Do you need batteries to use it? I say Yes.
Would you like to be one? I say Probably.
Does it eat grass? I say Yes.
Is it awake at night? I say Probably.
Can you use it 24 hours a day? I say No.
****

So far, - not too bright.

 

[Train]

holy crap this one floored me, how the hell did it figure this out after only 11 questions?

I am guessing that it is an aluminum can?
Yes , No , Close
11. Is it used in a game? No.
10. Is it outside? No.
9. Would you use it daily? Sometimes.
8. Does it roll? Yes.
7. Would you use it in the dark? Irrelevant.
6. Is it usually colorful? Yes.
5. Is it black? No.
4. Can you play with it? Yes.
3. Does it fit in your wallet? No.
2. Can you use it with your friends? Yes.
1. It is classified as Mineral.

 

[Looney]

It guessed it correctly when i gave it yogurt.

 

[jagec]

The problem with AI is that it takes an insane amount of processing power. It just wouldn't work in a game...unless you wanted to give up on your pretty graphics.

The way computers work is they take solid facts, feed them through predetermined pathways, and get the result. This makes them lightening-fast. AI needs to work on fuzzy facts and preconcieved ideas acquired through experience, and constantly question whether they're using the right pathways or ideas to make a decision. Best to focus on our strengths: Use computers for computing, use humans for the fuzzy decisions.

 

[EightySix Four]

My attempt at penis

I am guessing that it is a muscle(body part)?
Yes , No , Close
12. Can you use it at school? No.
11. Is it awake at night? Yes.
10. Is it usually visible? No.
9. Can you make money by selling it? Yes.
8. Can it cheer you up? Yes.
7. Is it used in a sport? No.
6. Is it shiny? No.
5. Do you clean it regularly? Yes.
4. Does it move? Yes.
3. Does it go inside other things? Yes.
2. Would you find it on a farm? Yes.
1. It is classified as Other.

not bad... some are debatable

 

[Jeffyboy]

Problem with AI is that no one knows how to program AI. No one knows how the brain works either..lol.
Our attempts at AI include backward/forward chaining and goofy neural nets... it's all pretty primitive and made to sound impressive until you see the code that runs it. Give it a few more hundred years and I think we'll start to get closer ;-)

As for "fuzzy facts"... there's nothing really fuzzy in a computer... it's still 1's and 0's with probabilities associated to pathways.

Anyone have an algorigthm for parsing natural english language? It's freaking impossible almost. It's something that practically has to be learned... and i think we don't really know how to make computers learn things outside of what they are programmed to learn.

Jeff

 

[dguy6789]

You wont see 100% self learning computer until they come out with quantom computers in another 10-15 years.

 

[Train]

Originally posted by: Jeffyboy
..
As for "fuzzy facts"... there's nothing really fuzzy in a computer... it's still 1's and 0's with probabilities associated to pathways...

The brain breaks down to just nuerons firing, a lot like 1's and 0's if you ask me, I think its closer than you think

 

[Jeffyboy]

well... being an organic sensory and memory organ... it seems to be sensitive to hormone changes, aging, senses the passage of time and many other things... and I don't believe the brain is digital with nerves firing in different patterns and intensities and feeding information to the brain. The brain grows and wires itself when it's young and brain cells die as you get older. Amazing when our brain cells die we aren't thrown in to a trash bin with all the other AMD or Intel junk ;-) Even if quantum computing happens in the next 100 years... it'll seriously be lacking programmability. In the brain... neurons are firing constantly and usually for no reason and usually nothing you may be aware of... bringing to your attention memories you aren't even trying to retrieve, and even memories you thought were 'deleted'. The brain is immensely complex to say the least.

When quantum computers come out... it'll still be similair to transistor technology but on an atomic scale. On or off... while individual neurons have oogles of pathways... not just a base, emiiter, and cathode. Still... it'll have to be programmed with logic that people understand and I believe that developing software to mimic human intelligence even on a quantum computer goes beyound our capacity within the next 100 years or longer.

Jeff

 

[KentState]

I think the biggest challenge is too decide what is the definition of intellegence. For example, the choices that humans make are a mixer of poor and intellegent choices. Sometimes those choices cause other choices that seem to be nothing more than dumb luck. I would say that computer already make the best decisions, but cannot react to the poor or unexcpeted choices that we make.

 

[Anubis08]

We don't understand our own brains. How can we mimic what we do not understand.

 

[Anubis08]

Even better question. Would we want to?

 

[pakigang]

Just a question, are there any tutorials on programming AI. It just can't be if then else thing. Just curious.

 

[Jeffyboy]

if you want examples... you could get a book called the "Little lisper"... which is about programming in LISP. (an AI language) You might find it a purdy cool language.

Jeff

 

[harrkev]

Originally posted by: pakigang
Just a question, are there any tutorials on programming AI. It just can't be if then else thing. Just curious.

Please define "AI." If you are talking about symbolic processing, this is difficult.

On the other hand, there are simple schemes like genetic algorithms (PBIL is a good one) and optimization schemes (like Q-learning) that are good for simple tasks (including robots). These are good for having a robot learn how to avoid walls and such.

I can give you some links if you are interested.

**EDIT**
Things like this can be done in C, Pascal, Matlab, BASIC, etc.

 

[harrkev]

Double-post. Please ignore

 

[Jeffyboy]

**EDIT**
Things like this can be done in C, Pascal, Matlab, BASIC, etc.

Yeah, you can use any language... but in university or trade school they will generally teach AI concepts with certain languages... LISP, Prolog, Smalltalk and prolly a few others. These languages provide powerful tools for AI. LISP (LISt Processing language) is a symbolic language... and worth taking a look at just so that you can say you did ;-)
Local hospitals here used to use a scheduling system originally developed in LISP. It would find replacement staff based on all the union rules and who-ever is available with the appropriate skills. It worked great but was eventually redeveloped in to C. So, anything is possible with work ;-)


Jeff

 

[harrkev]

Jeffboy. This all depends on the type of AI being taught. LISP and Prolog are used for "top down" AI. You can use anything that you want for "bottom up."

Let me give you an example. Prolog is useful for deducing logical things. But if you have a robot who goes "I think therefore I am", what good is it if the machine blindly runs out the door and gets hit by a car. This is the "top down" approach. You model the higher reasoning skills, but forget the basics such as maneuvering in your environment.

The "bottom up" approach is like looking at a dog. A dog cannt play chess or reason. But it can still fetch your slippers and newspaper. Bottom-up approaches are more interested in ways to acquire and process sensory information, and achieve learning based on this simple sensory input, rather than process abstract logical constructs. I once programmed a simple robot to learn how to maneuver in its environment using something called "Q-learning." Pretty cool stuff.

**EDIT**
Some people have done "top down" approaches to navigation, too. One example is a system from over ten years ago where the robot would take a picture of its environment. The computers would churn for 20 minutes, and the robot would move two feet. Processing power has advanced a lot in that time, but roaches have a lot less power than my Athlon 64, and they seem to do alright.

 

[DrPizza]

Originally posted by: Jeffyboy
if you want examples... you could get a book called the "Little lisper"... which is about programming in LISP. (an AI language) You might find it a purdy cool language.

Jeff

I learned LISP and xLisp in the 80's... I hated the language.

Anyway, I was going to add that I was reading in a book yesterday that was attempting to explain how computers will never duplicate human intelligence. It was some sort of proof that the human mind doesn't act in an algorithmic way and it showed evidence that there are things we can figure out that aren't based on an algorthmic system (or something like that) - I was just skimming it yesterday.. Once I read the whole book, perhaps I'll understand that chapter a little better.