Artificial Intellligence Questions

[Prometheus98876]

Those who have looked through the Aesthetics section recenlty might have noticed that I am working on a manuscript centered around a highly developed AI engine developed far into the future.<br><br>I have started researching into the science of AI (I reccommend the book Intelligent Machines, Myths and Realities, bu Smith, Clarence as a good book for novices and more advanced readers) so that I can represent it in manner that is reasonably familiar with those whom know about it, and so that it is reasonably scientifically valid, given current AI theory.<br><br>Who wants to give me some of the definitions of AI they might have heard and thought were particularily valid. Or who has some AI theory you really think I should research into, or any advice at all...<br><br>In particular, can you think of any simple tests that we might give a artifical intelligence system that demonstrates advanced human-like intelligence?For instance, we might give it the following as part of the test:<br><br>Suppose we define a Matrix, A in a standard way:<br>A mxn Matrix A, is a rectangular array of numbers of mn real numbers arranged in m horizontal rows and n vertical columns.<br>And that a_ij is the entry in the ith row and the jth column.<br>And we define Matrix Addition in the correct way:<br>If A and B are both mxn and C=A+B, then c_ij = a_ij + b_ij<br>Now, supposing we give the robot the following problem:<br>Z = % + B <br>Ok, % is a stupid symbol for a Matrix, but a human might be excepted to figure that % is a Matrix.<br>Suppose we give all three Matrices, with some entries being variables such as x,y,z instead of variables.<br>If it can give us the final answer, in the rectungular form of C, correctly, given only the information here, could we take this a sign of AI?I am sorry if this was originally posted in a terrible format, my HTML skills are abit off this morning...

Edited July 20, 2005 by Prometheus98876

[softwareNerd]

Suggest you search the forum for threads with 'artificial intelligence' in the title. Here is one (the longest) thread, but there are others as well.

[Prometheus98876]

Suggest you search the forum for threads with 'artificial intelligence' in the title. Here is one (the longest) thread, but there are others as well.

Ok, I have thought of that, and am currenlty looking through it, hopefull both approaches will net some feedback..

OK...I have searched and have found very little of much use in the Forums. A more generic search of the Epist. section might uncover some gold though in regards to artificial intelligence.

Edited July 20, 2005 by Prometheus98876

[Hal]

In particular, can you think of any simple tests that we might give a artifical intelligence system that demonstrates advanced human-like intelligence?

For instance, we might give it the following as part of the test:<br><br>

Suppose we define a Matrix, A in a standard way:<br>

Problems like this are relatively easy for AI to solve, and most can be attacked using heuristics searches (the one you give would be a simple case of constraint satistification). AI has enjoyed high levels of sucess when attacking purely mathematical/logical problems like these, as well as problems occuring in restricted domains as 'translating' calculus word problems into mathematical symbols, or coping in microworlds (eg Winograd's Blocks program).

The real problems arise when you try to design an AI which can work in a real world environemnt, where there are thousands upon thousands of factors to consider, and an incredibly large amount of information to handle. This kind of knowledge is what makes up human 'common sense', and generally provides the background knowledge which allows us to function in day to day life. We are presented with huge amounts of data but are normally able to decide what information is directly relevant to our current interests. Duplicating this ability in a computer has proved almost impossible so far, and this problem, known as the "Common Sense Knowledge Problem" (or the Frame Problem, which is related) is probably that biggest single issue in AI today.

Let's take a concrete example to illustrate what I mean - imagine I wanted to design a gambling program which would help me select a horse which is likely to win a race. The program is given a huge list of data, and has to use this to predict a winner. But think of all the different things which an expert human gambler might consider relevant to his assessment - If a horse is sick, it will be slower. If a jockey's mother has died recently, then he might perform badly due to grief. If the jockey lost his previous race, he might be demoralised and do badly in this one. And so on. This all seems obvious to us, but how would you manage to design a program which 'knew' that a jockey's mother dying would have a bad effect on him? Well, you'd have to give it basic knowledge of human emotions and relations. Then you'd have to give it some way of assessing the effects the mother dying would have on the jockey - did he get on with his mother? Did she abandon him as a child? Again, this kind of thing is just common sense for us, but you'd have to somehow program it all into the computer - if the jockey was a 38 year old man who had lived at home all his life then his mother dying would probably devastate him. But if she had threw him out the house at 18 and he hadnt spoke to her in years, he probably wouldnt care. How would the computer know all this?

The same applies to the fact that the jockey lost his last race. To draw conclusions from this youd have to understand our 'folk psychology' and have a way to assess personalities - is the jockey the kind of person who will perform badly after losing a race? Or will his failure motivate him to try even harder? Again, an expert human gambler will consider these factors as obviously relevant, whereas the computer will struggle.

Even worse, is all the factors which arent relevant. Is it relevant that the race is in Manchester rather than London? Probably not. Does it matter if the jockey has blue eyes? Unlikely. But there are potentially infinite things that are irrelevant, and the computer needs some way of distinguishing between these and the factors that matter. As you give the computer more and more data to consider, your execution speed becomes slower and slower since there is no effective way of sifting out only what is relevant. This is commonly known as combinatoric explosion. Suddenly the task of building a program which can rival a human gambler seems almost impossible.

This kind of thing is even more apparent when you look at how badly computer programs for games like bridge or poker do. The amount of human background knowlege and psychology that goes into being a world class poker player is astonishing, and there's no obvious to rival this purely by statistics and number crunching.

I would consider any of the following to be significant breakthroughs in AI because they arent restricted to artificial domains: machine translating, robots which can function in any human enviornment, passing the turing test.

This article by Daniel Dennett gives a good outline of the frame problem, this is slightly more technical, and this article by Hubert Dreyfus is excellent, but more concerned with philosophical problems and tracing the history of modern (at the time) AI back to its rationalist roots.

[Prometheus98876]

OK, thanks for all that. From what you say here, I am reminded that an AI system should be able to predict factors that are not initially specified, ie predict changes in conditions or allow for unexpected variables, and how it should be able to adapt beyond the initially specified parameters.

Edited July 20, 2005 by Prometheus98876

[Hal]

OK, thanks for all that.  From what you say here, I am reminded that an AI system should be able to predict factors that are not initially specified, ie predict changes in conditions or allow for unexpected variables, and how it should be able to adapt beyond the initially specified parameters.

Yeah, I would say that some kind of learning capacity would be necessary before I would be prepared to call a machine intelligent. In my opinion the best way to get a computer to pass the Turing Test would be to build a robot that is capable of living in a human envioronment (whether a real society or an online virtual world) and learning a language in the same way a child would from interacting with others. While this would obviously be insanely difficult, it seems more realistic than somehow programming knowledge of language into a computer from scratch. It's this "ability to live and adapt in a complex enviornment" that I would want to class as intelligence, even though I've no idea how to define it rigorously.

Edited July 21, 2005 by Hal

[Prometheus98876]

Yeah, I would say that some kind of learning capacity would be necessary before I would be prepared to call a machine intelligent. In my opinion the best way to get a computer to pass the Turing Test would be to build a robot that is capable of living in a human envioronment (whether a real society or an online virtual world) and learning a language in the same way a child would from interacting with others. While this would obviously be insanely difficult, it seems more realistic than somehow programming knowledge of language into a computer from scratch. It's this "ability to live and adapt in a complex enviornment" that I would want to class as intelligence, even though I've no idea how to define it rigorously.

Yeah, I would say so too. Ultimately, the robot portrayed in my novel would have a form of the Turing test applied to it if it was to be created in reality, as it is intended to function in a similar way to a human. Your definition is very good, although I would define it a little more rigourously, perhaps as something like this:

"The ability via means of rational anaylsis, to function in a way that resembles human behaviour, by adapting to changes in multiple complex environments, using information that is beyond its original data input"

[DavidV]

You might already be aware of this, but the Turing test is not a test for intelligence per se. It provides some evidence of human evidence, but such evidence is not conclusive, and limits intelligence to that which comes from human experience. A real test of intelligence should test for instances of rational thought – concept formation in particular.

[Prometheus98876]

You might already be aware of this, but the Turing test is not a test for intelligence per se. It provides some evidence of human evidence, but such evidence is not conclusive, and limits intelligence to that which comes from human experience.  A real test of intelligence should test for instances of rational thought – concept formation in particular.

Yes, that is the case, but it is a part of a comprehensive series of tests that would be should carried out to test such things, at least according to our current AI knowledge.

The events of the novel in which the AI robot I am placing in a novel dealing with such things, ie - life itself is the ultimate test of intelligence...

[Hal]

You might already be aware of this, but the Turing test is not a test for intelligence per se. It provides some evidence of human evidence, but such evidence is not conclusive, and limits intelligence to that which comes from human experience.  A real test of intelligence should test for instances of rational thought – concept formation in particular.

Depending on what is _precisely_ meant by concept formation, and 'rational thought' I think this would be a lot easier than the Turing Test. If I produce a neural network which can learn to identify trees, would you say it had formed the concept tree? And if not, what else would you require it to be able to do?

If a computer could pass the Turing Test, I would have no hesitation in attributing to it both rational thought, and the abililty to reason in concepts. How else would it be possible to speak a natural language?

Edited August 7, 2005 by Hal

[Prometheus98876]

Can anyone reccommend any good books on the subject that dont require me to have done/be studying a year three AI paper? I have found a few books at the library at my uni that are OK, but I like to ask these things so I can as many options to consider as possible.

[skap35]

Can anyone reccommend any good books on the subject that dont require me to have done/be studying a year three AI paper?  I have found a few books at the library at my uni that are OK, but I like to ask these things so I can as many options to consider as possible.

I don't know if this is the sort of information you are looking for, but this site has a lot of articles related to AI in videogame development. It was a great help when I used to mess around with videogame programming.

Edited September 13, 2005 by skap35

[Prometheus98876]

I don't know if this is the sort of information you are looking for, but this site has a lot of articles related to AI in videogame development.  It was a great help when I used to mess around with videogame programming.

Thanks, I have yet to go through this site in much detail yet, however it looks like it might have some information/insights into challenges inherent to our current AI understanding/concepts. It should certainly be interesting to say the least, even though I do not intend to go into game design.

[softwareNerd]

Patrick Henry! Believe it or not, that's the author's name.

I could be mixing up authors, because it has been over 20 years since I read it, but I'm fairly sure his AI book was quite "accessible": not just algorithms and formulae.

[Prometheus98876]

Patrick Henry! Believe it or not, that's the author's name.

I could be mixing up authors, because it has been over 20 years since I read it, but I'm fairly sure his AI book was quite "accessible": not just algorithms and formulae.

As much as I might be starting to get to grips with the algorithms and formulae involved in this sort of thing, they are not quite what I am interested in right now. Lets see if I can get a copy of this thing, from a quick look online I have found many people whom seem to quite like it. The author is Patrick Henry Winston, just to clear up an apparent confusion .

It might be to my advantage to try and find his books on Lisp, as I want to learn this langauge too (partially so I can tinker with the Emacs source code, alot of which seems to be written in Lisp.).

Edited September 13, 2005 by Prometheus98876

[Felix]

The true challenge in AI today is not solving mathematical algorithms. We already have chess computers able to compete with world champions.

Yet we have no robot to 'really' understand language.

The Turing Test is crap. If the computer can fool you, then it's intelligent?

A friend of mine has a wav-file on his PC. If you call someone and play that soundfile so the other person hears the soundfile over the phone, that person believes that there really is another person at the other side of the line shouting and ranting and acting pissed because somebody named Hans didn't give him his money back and that you should please get Hans on the line because it's clear he's there with you and you are just hiding him. This file is really intelligently made, but is it intelligent? I've seen it fool about 5 people, including myself, so it somehow passes the turing test. (I know the test involves another human being and is a bit more strict.)

What we need is a robot who actually can achieve this common sense that was adressed by Dennett.

For this you have to work on how humans organize their sensory input, how the human brain has evolved. How different parts of the brains work and interact. How the signal processing works, how data is stored and how it is processed. If human-like intelligence is what you're after, you won't get it with a cheesy math program.

You need to rebuild a human being.

Edited September 13, 2005 by Felix

[skap35]

For this you have to work on how humans organize their sensory input, how the human brain has evolved. How different parts of the brains work and interact. How the signal processing works, how data is stored and how it is processed. If human-like intelligence is what you're after, you won't get it with a cheesy math program.

You need to rebuild a human being.

That is true. When I was still working with AI related projects, I would read tons of neurology books to try and understand the brain. Then make as best of an attempt I could at emulating that behavior in a computer program. The only problem I ran into is that 1. the brain is so complex that once you get into very much detail, the learning curve goes up. 2. much of the necessary information about how the brain works (in order to emulate it with a computer) is not yet known.

I would suggest not only reading about the human brain, but learn about simpler creatures like mice or insects. It's certainly not human level intelligence, but many of the concepts are the same, and much easier to understand.

I was working on a real-time strategy game long ago, and for the resource-gathering heuristic, I researched the behavior of termites and ants. That is also another area to investigate. If that is an area you're looking into, look up Soul of the White Ant.

[Felix]

I was working on a real-time strategy game long ago, and for the resource-gathering heuristic, I researched the behavior of termites and ants.

Hey, I did that for my bachelor thesis for finding out a new algorithm for optimization.

Thanks for the recommendation. I will add it to my list.