Help offset server costs by donating. This is totally optional. Any overages will go to library fines or new books.
AI and Diplomacy
Hi all, I just came across this article
and thought that some might be interested. Basically, the Deepmind team have been considering how to move their AI system from two player games (chess, Go etc) at which remarkable success has been achieved, towards multiplayer games involving cooperation as well as competition. So Diplomacy has been chosen, largely because of the complexity of player interactions, the changing role of said interactions as gameplay continues, and the constant need to evaluate alliances with other players.
It's an area that in my own vision of the near future, I imagine to have been solved, so it's interesting to see the progress in current research.
I particularly enjoyed this comment towards the end "Teaching an [AI] agent to utilize other players as collaborative pawns to ensure victory does bring up a series of concerns..."
Comments
This seems very much applicable to our Player of Games discussion!
Interestingly, they're playing a game with no inter-player communication. It's about inferring what would be in the other players' best interests and planning based on that. I wonder what sort of model its building up about the other players, and how that changes throughout the game?
I suppose that's to avoid the need to build some kind of parsing engine to infer what somebody else is offering as a deal? But you'd have thought that as a minimum they could have had a selection from a list of options - "make alliance", "make non-aggression pact", "threaten X unless Y" and so on. That would at least begin the process of imitating negotiating.
Nah, parsing, semantic exraction, and sentiment/honesty analysis are easy tasks now. What makes it more complex is having a whole additional channel of input, and hence needing to create more complex "models" of the others players. And it's another channel of output, as you don't just have to decide on board moves, you also have to decide what to say to whom during the negotiation. Plus, you can infer things from who is talking to whom (and how often and for how long) as well as what they're saying to you.
I think you're right - Deepmind was built (I suspect) using a set of assumptions that the situation on the playing area (and implicitly the gameplay until "Now") was sufficient to construct a strategy for going on. There was no facility for including communications of any kind with the opponent, and the games first chosen explicitly avoid such.
At the same time, it feels like this restriction defeats much of the point of choosing Diplomacy as target - unless they intend to have a second iteration when comms are allowed. You'd think there would be other multiplayer games where you have to implicitly weigh up from the visible board who to attack and who to support, which don't rely so heavily on interpersonal communication.
So governments and corporations will be using all the scraped data about you in order to “win” whatever game they are playing with/against you. Happy near future!