MIT scientists teach a computer Civ II

[The_J]

​

Scientists from the Massachusetts Institute of Technology, who have been researching on the field of computational lingustics, have been dealing with Civilization II and actually told the computer how to play it.
They improved the actual winning rate of their own AI vs. the FreeCiv AI to 78% by letting it read the manual, and evaluated it on several instances with 1 on 1 matches.

Anyone who is further interested in this topic can get the actual research paper "Learning to Win by Reading Manuals in a Monte-Carlo Framework" here and a presentation here.
And for the people on CFC there are threads in the Civ5 forum here, in Civ4 here and in Civ2 here.

And the most important thing for us: They got that manual from CFC .


If anyone has links with press coverage of this topic, then feel free to post them .
Thanks to everyone who has opened threads about this topic and posted the news here .

News links:

• Yahoo News: Computer learns to read manual to improve gameplay, world domination next
• TechEye.net: MIT plants Terminator seeds with Civilization II despot computer
• io9.com: Computer teaches itself English so that it can play Civilization
• slashdot: Computer Learns Language By Playing Games
• ScienceDaily: Computer Learns Language by Playing Games
• PCGamer: PC AI sucks at Civilization, reads manual, starts kicking ass
• EscapistMagazine: MIT Teaches Computer to Read, Conquer the Planet

 

[Camikaze]

Here's an article about it.

 

[killmeplease]

what the civilization version do they used?

 

[The_J]

Directly from the abstract of the paper:

We apply our approach to the complex strategy game Civilization II using the official game manual as the text guide.

And i added some of the already posted links .

Edit:

To apply our method to the Civilization II game, we use the game’s open source implementation Freeciv.

 

[LordNwahs6]

Interesting article. Was the built-in AI for Freeciv and Civ 2 programmed differently?

 

[AlpsStranger]

Wait, literally read the manual and "understood"?! (I'll assume they at least mean "read in *.txt" or something)

I'm either missing something critical or I've really underestimated technical progress in this area.

 

[ddd123]

also HOW can you learn how to play civ reading the MANUAL????

not even a human being can do that...

 

[The_J]

Wait, literally read the manual and "understood"?! (I'll assume they at least mean "read in *.txt" or something)

I'm either missing something critical or I've really underestimated technical progress in this area.

Oh, text mining is quite developed, so that's not a supprise. There are sure some quite extreme problems with it, but getting infos out of text can be done.
But you already found the big thing in there: It's not really about the AI playing the game, it's about the AI understanding the manual .

also HOW can you learn how to play civ reading the MANUAL????

not even a human being can do that...

Oh, there are some basic instructions in the Civ2 manual, there's a short tutorial in it.

 

[Petek]

My understanding is that the project consisted of two stages. In the first stage, the computer at first played random moves against the AI. It then went back and analyzed its moves, determined which gave the best results and adjusted its play accordingly. This process was iterated. Eventually the computed achieved a 46% winning percentage. In stage two, the computer then "read" the manual and picked out text, such as "Build your city on a plains or grassland square with a river running through it if possible." It then added this information to its strategy and in doing so increased its winning percentage to 79%. The innovation was in the second stage.

 

[Horizons]

Naturally they would pick Freeciv because it's simplistic.

 

[Minor Annoyance]

Shame no one writes good manuals anymore.

 

[The_J]

Naturally they would pick Freeciv because it's simplistic.

I guess the bigger point is that the source code is freely available and that you so can just plug your AI into it, which you can't do with any commercial game .

Shame no one writes good manuals anymore.

Sad, but (nearly) true.

 

[AlpsStranger]

Naturally they would pick Freeciv because it's simplistic.

Every baby step is important on the road to sex robots playing "The Most Dangerous Game" legally never again beating a computer at chess various important industrial applications.

 

[darkedone02]

Now Give the computer the manual of Portal 2 and teach it how to play it and successfully beaten it

 

[GoodGame]

I guess the bigger point is that the source code is freely available and that you so can just plug your AI into it, which you can't do with any commercial game .

IDK, it sounds like a semi-sensationalistic gimmick to not use the Civ2 AI. But I am impressed for an AI to reduce a text description to logical routine and then follow it, regardless of whether or not it led to victory.

 

[Genocidicbunny]

also HOW can you learn how to play civ reading the MANUAL????

not even a human being can do that...

Because a human being does not think like a computer AI.

What it did was form associations and learned what actions are possible by analyzing the text. The manual also pretty well details all the various actions that could be undertaken. Then the AI can go ahead and apply what it 'learned' to the actual game, whereupon it further refines how it plays and learns when different actions are appropriate.

In the end, it is still just analyzing a situation and picking a task from it's repertoire that is most appropriate to perform. The novel idea is teaching it these actions from a game versus essentially hardcoding them in.

Please note that I am being very simplistic here and may not be entirely accurate. Read the articles or the paper if you want more info.

IDK, it sounds like a semi-sensationalistic gimmick to not use the Civ2 AI. But I am impressed for an AI to reduce a text description to logical routine and then follow it, regardless of whether or not it led to victory.

It's a lot easier to hook into a game where you can directly access the source code. If they were to have the AI run on top of Civ 2 then they'd need to have some way to pass the game state from the game to the AI. In FreeCiv its a much simpler matter.

 

[ddd123]

What it did was form associations and learned what actions are possible by analyzing the text. The manual also pretty well details all the various actions that could be undertaken. Then the AI can go ahead and apply what it 'learned' to the actual game, whereupon it further refines how it plays and learns when different actions are appropriate.

there is no tactic in the manual
this is just plain impossible

if that was true then they could apply it to chess who has a "manual" much more shorter and learn the game

i hardly doubt that could work

manual =//= tactics
knowing the rules doesnt teach you to play

both for human and ofc for ai

 

[Petek]

The computer did not learn how to play the game solely by reading the manual. The computer was first taught by other methods and achieved a winning percentage of 46% against the AI. After reading the manual and applying what it learned, its winning percentage increased to 79%.

 

[ywhtptgtfo]

The computer did not learn how to play the game solely by reading the manual. The computer was first taught by other methods and achieved a winning percentage of 46% against the AI. After reading the manual and applying what it learned, its winning percentage increased to 79%.

Yes, I was trying to poke holes in that paper, but it seems they have done the sufficient controls to justify their claims.

It appears they formulated an objective function to maximize the Civ 2 score. This led to some X degree of improvement. Afterwards, they used the document information to adjust the weights that were trained in the previous steps (or retrain the weights?) which led to another Y degree of improvement.

I'd say this thing probably wouldn't work in many games, because most manuals give crappy hints and some strategy games have pretty decent A.I. However, it's a novel approach that marries natural language processing (NLP) with standard artificial intelligence problems (instead of just using A.I. to solve NLP problems).

I'd give this paper a thumbs up even though I deem their finding to be lucky and conditional.

 

[Superluminal]

Cool experiment. I'd like to play against the MIT AI in Civ5.