So, who's interested in making a program to play Diplomacy?

We could call in Draugbot. In addition to being a mediocre player, we could program it to occasionally DDoS the site. It would be just like old times.

That's a very interesting article, abge -- thank you!

CSteinhardt linked me to a Diplomacy program last year. It could play gunboat well, not so much with actual diplomacy. It actually can figure out stalemate lines and does well at the end game, and its openings aren't bad.
He and I tested several games, each playing one power while the AI plays the rest, and it actually took effort to determine who the bots were. Also, it wasn't easy to win.
I can email him and ask what the program was.

I'd be very interested in a link to that! I've been thinking about writing something that plays gunboat for a while.

This might also be of interest to some:

http://dl.acm.org/citation.cfm?id=2609605

It's a paper about looking in a database of past Chess games for positions which are similar - but not identical. The technique is aware of the way the pieces move, so it tries to produce positions that experienced players would agree are similar. It would be interesting to come up with something similar for Diplomacy..

Here's another good article about go AI:
http://www.wired.com/2014/05/the-world-of-computer-go/

thx

The graphic in the Spectrum article (posted by abge) is quite irritating. When making an artistic pattern with go stones on a board, it doesn't have to be a legal board position (i.e., captured groups are ok), but at least put the stones on the intersections instead of centering them in the squares.

I think they were trying to make it look like a brain?

Yeah, I see that it is a brain. Go art doesn't have to be a logical, or even legal, board position.

I just don't like that they didn't place the stones on the intersections.

b

I've thought a lot about what it would take to teach a computer how to play gunboat. Machine Learning was by far my favorite course in college. I may just get around to actually giving it a shot sooner or later. PPSC would be far easier to accomplish successfully; WTA would be significantly more difficult with the added logic to prioritize stopping solos. I'm also not too sure how you would program the AI to determine when to put the draw up. The logic to make any individual move wouldn't actually be too difficult, you would just likely have to make some specific logical instructions on an individual nation basis as well as making general rules overall. A big key is to add in some level of variability such that the AI doesn't become predictable.

I'm not sure if PPSC would necessarily be that much easier than WTA. PPSC gunboat isn't just an attack in all directions, try to grab as many SCs as possible type of game. You still have to try to make friends and work with people to be successful. A tough part of designing a gunboat AI would be to understand and leverage the psychological aspects of the game.

I think risk tolerance would also have to be a design parameter toward figuring out when to switch from playing for a solo/reducing the draw vs solidifying to prevent a solo.