[Art]

...Then learns how to play it even better once it get the chance to read the game manual!!

Cool article at:
http://www.ripten.com/2011/07/13/mit-ai-computer-teaches-itself-english-to-play-civilization-v-skynet-is-pleased/

[Freddy]

Interesting though a bit lacking in detail.  So they just showed the computer the manual and it read it.  How for example did it understand, did they have to teach it English first ?  I'm guessing it's a bit more hand fed than they let on.  Just not enough to go on here.  Thanks though 

[Art]

Freddy, the article indicated: MIT decided to develop an AI that was essentially thrown into a game of Civilization with virtually no instructions on the game’s tasks or on the language that the game was written in. Creepily enough, after some trial-and-error, the AI began winning the games 46% of the time. Feeling that 46% wasn’t good enough, the MIT guys decided to give the AI the manual to the game.

...an AI that can think and rationalize its way through a videogame. The AI then began winning the game 79% of the time.

I think it said quite a lot about the AI they developed.

[Data]

Yes interesting, the Ai managed to learn and then make good use of what it learned, doesn’t sound much but to my mind it is.

For sure a step in the right direction.   

[Freddy]

Yeah I am with you Art, but just how much did they help it out.  You just don't write some code and have it suddenly understand English - Chatbot developers have been trying this for years, it's the holy grail.  If it was that remarkable how come it hasn't made the news headlines ?  If it's that advanced then this is not something trivial - all I want is more information...

[victorshulist]

I don't think this lives up to the hype either.

From : http://web.mit.edu/newsoffice/2011/language-from-games-0712.html


So initially, its behavior is almost totally random. But as it takes various actions, different words appear on screen, and it can look for instances of those words in the instruction set. It can also search the surrounding text for associated words, and develop hypotheses about what actions those words correspond to. Hypotheses that consistently lead to good results are given greater credence, while those that consistently lead to bad results are discarded.

 
Ok so a classification type system.     This series of actions , with this degree of probability , could relate to this action word.

Semi interesting ... but they make it sound like it understands the totality of a phrase or sentence as a whole.  Which I very very much doubt...  being an article from July 2011. ..I think if it was as they portray, it would beat a Turing Test.

Good article and nice approach though.  This work will feed into a larger Natural Language Understanding system, but probably won't be the central piece.

When we can ask it questions about the manual it read, then we can say it has understanding.  Anyway my two cents

[Art]

@Freddy,
One problem is that this event is NOT noteworthy as an event like say Kim Kardashian visiting Paris or London or some interesting place. It is not something that garners the attention and acclaim of the general public. People are nosy when it comes to other people and those events...not what a machine does or can do.
Watson is all but forgotten in spite of that accomplishment. People are only entertained for the moment, then they quickly move on.

@ Victor,
I think their experiment speaks greatly in line with their original goal and it performed its task better that expected. I don't recall their project even being geared toward competing in a Turing Test nor is their success the same as a Turing Test. Perhaps with proper coding, it could.

You see, you both are typical examples of how complacent we, as a society have become. We are USED to having machines around us in practically every aspect of our lives, computers, smart devices, robots, etc. So a computer can play chess and GO and run households and adjust inventories, etc., it is not as big a deal as it once was. It is no longer MAGIC! (to paraphrase Arthur C. Clark)  Aside from some abject trickery by lesser known individuals I would have to side with the success of the researchers at MIT, a quite prestigious institution, not given to hype or inflating their own ego.

Sometimes we simply have to take things for face value...for what they are. In the same framework, I'd certainly like to know more about this program and its potential applications. Just my thoughts....

[Freddy]

Yes you are right, this kind of thing is not media worthy most of the time - but Rollo Carpenter made the news here in the UK on a couple of channels when he won the Loebner prize so I do hold some hope.

I wouldn't say I was being complacent about this - I was in fact challenging it and not accepting it at face value...  I'm not sure where you got that from, perhaps I phrased things incorrectly.

Peace 

[victorshulist]

  I think the confusion here, as usual, stems from different people's ideas of what the word 'understanding' means, the conceptions we each have, and the expectations we have of a machine that claims to have it.

[Art]

Yes guys, please forgive my wrong choice of wording as I should have said that you have become accustomed to or used to various forms of AI. Complacent was simply a wrong choice for me. English being my second language. MY first one was as an infant with goos and gahs!

On the contrary you both are always looking for ways to improve and seek refinements in the field of AI and interaction with your projects and your peers.

Again...my bad...sorry!!
 

[Freddy]

No problem at all Art, I think we all knew what angles we were each coming at and were respecting our individual views.  It's all good

[victorshulist]

I should have said that you have become accustomed to or used to various forms of AI.

Ok, that I very much agree with.

No problem at all Art, I think we all knew what angles we were each coming at and were respecting our individual views.  It's all good

I second that.

[Bragi]

I think that a combination of n-grams, random forests, logistic regression and a whole lot of computing power should be able to do the trick.
It sounds a lot like an advanced kaggle competition: all about machine learning.