http://io9.com/5820624/computer-teaches-itself-english-so-that-it-can-play-civilization
Found this on reddit. Does this mean that the AI is gonna rock next patch?
Found this on reddit. Does this mean that the AI is gonna rock next patch?
Computer teaches itself to play civilization V
- Thread starter Karl_Rove
- Start date
- Joined
- Mar 31, 2008
- Messages
- 14,928
I think the guys who made that AI could make a lot of money selling the AI to video game developers. An AI that learns by actually reading the instruction manual and by playing the game would make for an interesting option. Of course, I wonder if the AI learns to use exploits like the player does.
. Of course, I wonder if the AI learns to use exploits like the player does.
I'm sure it does. If I read the article correctly, the computer tries something new and then evaluates how well it did based on that decision. It would undoubtedly run into an exploit sooner or later.
I wonder how much the win percentage would change if the manual were up-to-date
spider1
Prince
- Joined
- Mar 29, 2006
- Messages
- 569
Or, sell that AI as a seperatly packaged Civ DLC of some sort that is always a computer AI, not to be played by the humans, and you could watch it learn the game, and learn your every move and how you play the game. After a while you'd have to quit putting it in as a competing CIV in your games because it would kick your ass every game.
- Joined
- Mar 31, 2008
- Messages
- 14,928
Well, if they want AIs that play like players, this AI would make a great secondary option. Look! It already did all the learning itself! Because it learns, it can adapt to mods better than the actual AI. If Fireaxis wanted to keep it so diplomacy is not gamey, they could try to make the AI only adapt for combat and game mechanics. No more frontline archers!
Or, sell that AI as a seperatly packaged Civ DLC of some sort that is always a computer AI, not to be played by the humans, and you could watch it learn the game, and learn your every move and how you play the game. After a while you'd have to quit putting it in as a competing CIV in your games because it would kick your ass every game.
New DLC, skynet. UU Terminators UA time travel (So they can drop Terminators at any era)
KungCheops
Warlord
It was Civ 2 (technically freeciv) that was being played here. The full paper can be found at http://people.csail.mit.edu/regina/my_papers/civ11.pdf.
I think you guys bashing the CIV AI are being a little harsh here. Their simulation runs on a core i7 and take almost a minute to execute each turn- that would be 8 minutes a turn for 8 AIs and that's for civ 2 on a modern machine!
The really impressive thing isn't that it can beat an old AI, but that it managed to learn how to do it by reading the manual!
Edit: Having read the paper Here's a bit more info:
1) They play 1 vs 1 on a small map.
2) The Algorithm works by trying out a large number of moves with the aim of picking the one that will improve the player's score the most. In this particular technique you sometimes chose a non optimal move, with the idea that it will lead to a more optimal move later on.
3) They claim great success in that they often manage a win inside 100 turns, whereas the CIV 2 AI never does . I suppose this is valid, but that makes the assumption that the CIV 2 AI has the aim of winning the game as quickly as possible.
4) They specifrically don't say what difficulty level they were playing on.....
I think you guys bashing the CIV AI are being a little harsh here. Their simulation runs on a core i7 and take almost a minute to execute each turn- that would be 8 minutes a turn for 8 AIs and that's for civ 2 on a modern machine!
The really impressive thing isn't that it can beat an old AI, but that it managed to learn how to do it by reading the manual!
Edit: Having read the paper Here's a bit more info:
1) They play 1 vs 1 on a small map.
2) The Algorithm works by trying out a large number of moves with the aim of picking the one that will improve the player's score the most. In this particular technique you sometimes chose a non optimal move, with the idea that it will lead to a more optimal move later on.
3) They claim great success in that they often manage a win inside 100 turns, whereas the CIV 2 AI never does . I suppose this is valid, but that makes the assumption that the CIV 2 AI has the aim of winning the game as quickly as possible.
4) They specifrically don't say what difficulty level they were playing on.....
Misterboy
Modern Major General
- Joined
- Oct 12, 2010
- Messages
- 244
It was NOT Civilization V. It was Civ 2... actually FreeCiv.
And the point of the paper is that they used nothing to teach it *language* except feed it the manual! So it's a testament of the Civ manual and the AI's learning capabilities. The Civilization connection, while interesting, is secondary.
And the point of the paper is that they used nothing to teach it *language* except feed it the manual! So it's a testament of the Civ manual and the AI's learning capabilities. The Civilization connection, while interesting, is secondary.
f35acepilot
Needs to post more
The real question we should be asking is whether the AI preferred conquest victory over the other options - in which case, we're in trouble.
It only wins 76% of the time? What a noob.
Be glad, because soon, AI trained chimps are gonna overrun your home town!
When did they add a conquest victory? Conquest is completely different than domination.
jdog5000
Revolutionary
FreeCiv (Civ2 open source clone) has a conquest victory ... plus, this study was playing 1v1 on a small map, so conquest is the easiest way to win.
Polycrates
Emperor
- Joined
- Dec 15, 2006
- Messages
- 1,288
It's not about Civ, but THIS article in a similar vein about Starcraft is really interesting, and both incredible and scary at the same time. Some of it is just stuff a human couldn't do because of the speed of micromanagement in a real-time game, but a lot of the decision-making and threat-mapping stuff is quite applicable to a Civ-type game as well, I think.
