Very interesting except they only really trained the AI for "gunboat Diplomacy" where you are not allowed to communicate with your fellow players.
Any serious Diplomacy player knows that the moves on the board, coming once a day or once a week, are far less important than the interpersonal relationships and written communication that actually creates the game. Convincingly beating humans at Diplomacy would also involve passing the Turing test.
I've never played Diplomacy, only read about it. But, I feel like you don't have to convince your opponents that you're human to win. The players could know that their opponent is a bot and still attempt to strategize with it, game the A.I. to get it onto their side, etc.
I suppose everyone might gang up on the A.I. immediately, but I don't think it's necessarily a forgone conclusion that they do.
Three billon human lives ended on August 29th, 1997. The Survivors of the nuclear fire called the war Judgement Day. They lived only to face a new nightmare: the war against the machines.
The meta is very important in Diplomacy. If the AI always wins when it gets down to 1 human vs 1 AI, then humans will conspire to prevent that situation from arising. If the human-AI team always wins, humans will conspire to prevent other humans from working with the AI. In the end, I think the reason "an AI" can't be a perfect Diplomacy player is because no one entity can be a perfect player. Once it's known that that entity is "too good" at situation X, the other players will prevent it from getting into situation X.
I have played (and won even!) Diplomacy and I agree about what makes the game fun.
But it is interesting that they argue these AI are able to infer people's alliances and reliability based on their actual in-game behavior by modeling how they would act in that scenario and training against that.
That's pretty neat, there must be a lot more information encoded in who you choose to support than I'd have thought.
It makes me wonder whether the AI would do better or worse than the human attempting to manipulate it. I have this feeling that we will be shocked to discover that the human's inevitable betrayal was inferred based on behavioral changes that fit the betrayal scenario the best.
It's not especially hard to imagine even with communication in a fairly reliable way (no misunderstandings from NLP) that an AI might easily learn to guess when a human was acting in a way that didn't... quite... match the behavior of an actual ally.
> where you are not allowed to communicate with your fellow players
Technically there is no textual communication. You could still communicate with players via signalling e.g. Choosing support commands for that player in places that are physically unreachable to you. Agents may learn to do that if players interpret them in the same way
I’d love to see an AI learn to add all the right historical rhetorical flair to their emails for the country it is currently playing. :) (Humans can’t help but sign off as “The Czar” when playing Russia in 1901, for example.)
Surely there must be bodies of dip communications and write-ups available from forums like Apolyton, if nothing else.
Most of the popular strategies for playing Diplomacy include telling people that the player is going to do one thing and then deliberately doing a different thing. Those strategies mean that essentially the player has to lie in order to be successful.
This means that essentially what we're doing with these kinds of AI experiments is headed into the: "How do I teach the computer how to lie" territory. And that seems bad for a whole host of ethical reasons.
19 comments
[ 3.6 ms ] story [ 4.5 ms ] threadAny serious Diplomacy player knows that the moves on the board, coming once a day or once a week, are far less important than the interpersonal relationships and written communication that actually creates the game. Convincingly beating humans at Diplomacy would also involve passing the Turing test.
I suppose everyone might gang up on the A.I. immediately, but I don't think it's necessarily a forgone conclusion that they do.
https://youtu.be/_Mg7qKstnPk
Meeeeeep, ding, meep, ding, meep ding meep ding ding ding.
The war against the machines.
Wuu- wub wub wub, wub wub wub wub. Wuu- wub wub wub, wub wub wub wub.
https://youtu.be/yePgFSls1x4
https://open.spotify.com/track/7zcVFsEFDUawXS6ZPnqQHF
https://music.apple.com/no/album/against-the-machines-origin...
What about keeping a puppet ally alive to disrupt a draw?
(And why did I think the parent comment was about 4-way draws?)
But it is interesting that they argue these AI are able to infer people's alliances and reliability based on their actual in-game behavior by modeling how they would act in that scenario and training against that.
That's pretty neat, there must be a lot more information encoded in who you choose to support than I'd have thought.
It makes me wonder whether the AI would do better or worse than the human attempting to manipulate it. I have this feeling that we will be shocked to discover that the human's inevitable betrayal was inferred based on behavioral changes that fit the betrayal scenario the best.
It's not especially hard to imagine even with communication in a fairly reliable way (no misunderstandings from NLP) that an AI might easily learn to guess when a human was acting in a way that didn't... quite... match the behavior of an actual ally.
But it would sure help =)
Technically there is no textual communication. You could still communicate with players via signalling e.g. Choosing support commands for that player in places that are physically unreachable to you. Agents may learn to do that if players interpret them in the same way
Surely there must be bodies of dip communications and write-ups available from forums like Apolyton, if nothing else.
No. In the normal version, players negotiate. That is the point of the game.
This means that essentially what we're doing with these kinds of AI experiments is headed into the: "How do I teach the computer how to lie" territory. And that seems bad for a whole host of ethical reasons.
> In the full version, players can negotiate. DeepMind tackled the simpler No-Press Diplomacy, devoid of explicit communication.
So it won’t be joining you on games night.