Wednesday, November 23, 2022

Diplomacy in the News

It's relatively rare to see a news story in the mainstream media that touches upon one of my interests. Consequently, I was surprised this morning to read about the claim that an "AI agent" named Cicero had "achiev[ed] human-level performance in the strategic board game Diplomacy."  Needless to say, this story caught my attention, as I was once a very avid player of Diplomacy and still retain a great fondness for it, though I haven't actually played it in many years. 

I'm naturally skeptical of these kinds of claims. I likewise lack the specific technical knowledge necessary to evaluate their veracity. Nevertheless, this is quite fascinating to me, since, if correct, it would represent a significant step in the evolution of computing. Mind you, much like chess, the degree to which being good at Diplomacy has any correlation to intelligence is separate question. In high school, my friends and I liked to flatter ourselves because we enjoyed playing "cerebral" games like this. I wonder if this story might be something similar, with the bravado of AI researchers standing in for that of fifteen year-old boys. 


  1. What it means that in the context of playing Diplomacy that is learned to recognize patterns of behavior that players of the games exhibit. Particularly when the goal of the player is to win the game. It's success means that patterns exist however nuanced and complex that can be recognized and learned by the underlying AI algorithms/models.

    In layman terms what the latest round of AI software cracked is pattern recognition. Suitably formatted data can be analyzed using algorithm like GPT-3 and AlphaGo to recognize underlying pattern.

    But to be clear it is pattern recognition devoid of thinking. Take for example image recognition. Suitably trained it uncanny recognize when a cat is in the picture, but it has no idea of what a cat is. It is missing the extra step human do after it glances at a picture or look at something and see a cat there in.

    I made the comment about roleplaying because what research did with AI and Diplomacy was to train the models to WIN THE GAME as defined in the diplomacy rules. It was looking and learning the patterns needed to achieve that singular goal. If the goal of the diplomacy was to do a bit of roleplaying along with playing the game. Then the AI would not understand that and still play to WIN THE GAME per the rules.

    And it would flounder in situations like when a group of causal gamers decided to gang up because one member of the group get too serious about the game and spoil the roleplaying fun they are having.

    Hope this clears up what was and wasn't accomplished.

    As a side note what is will likely mean in the near future for RPGs that our random table generation and procedural content generation (think Traveller) will be kicked up several notches. If there any sort of repetitive pattern that can be discerned whether it is a list of monsters, spells, or a universe worth of star system. One of these models can be trained to act as an assistant to help an author using that content.

    But just like random tables are notorious for producing nonsensical results when used in mass (Again think Traveller especially Traveller 4e Milieu Zero sector data). The lack of context means you will get stuff like a perfectly generated human face but where the earrings don't really don't match at all.

    1. I mean...people are, to a terrifying extent, just pattern-matching their way through life.

      I don't disagree that this isn't true cognition, but God knows a good chunk of what I do in a day isn't either.

    2. I agree it not sentience and yes often we pattern match a great deal many thing as human beings.

      A the end of the day these are tools that will allow us to do more with the time we have. As tools they are neither good or bad but their use can be good or bad.

  2. Isn't duping and betraying your opponents a huge part of Diplomacy? I'm having trouble believing an AI can even participate in that, let alone do it.

    1. That's exactly what the AI had to do and refine to win.

    2. If there is a pattern to the conversations (deception, planning, etc.) the models they can use can sniff it out. But keep in mind what it will look like to you especially using the on-line version of Diplomacy is that the players doesn't engage much in idle chat and very focused on the game.

      They explain it here.

      To build a controllable dialogue model, we started with a 2.7 billion parameter BART-like language model pre-trained on text from the internet and fine tuned on over 40,000 human games on We developed techniques to automatically annotate messages in the training data with corresponding planned moves in the game, so that at inference time we can control dialogue generation to discuss specific desired actions for the agent and its conversation partners. For example, if our agent is playing as France, conditioning the dialogue model on a plan involving England supporting France into Burgundy might yield a message to England like, “Hi England! Are you willing to support me into Burgundy this turn?” Controlling generation in this manner allows Cicero to ground its conversations in a set of plans that it develops and revises over time to better negotiate. This helps the agent coordinate with and persuade other players more effectively.

  3. Ah, Diplomacy. A beautiful and elegant game, at which I am utterly and completely hopeless.

  4. Diplomacy is bluffing game, not much more sophisticated than poker. And Hungary is a shithole of a country, I say that with the full awareness that I was born and raised here.

  5. The only thing I can say about Diplomacy is that, in college, my friend group played it once, and almost stopped being a friend group because of it. If the AIs want to play it, more power to them. Maybe it will keep them from trying to take over the world in real life.

  6. There's previous successful research on "no-press Diplomacy", where the AIs learned human-expert-level tactics and strategy but there was no way to communicate. (Disclaimer: I was distantly involved in some of this.)

    As I understand it, the interesting thing about the new Diplomacy AI is that it has open press - it's generating whatever it wants to say. Expert level Diplomacy players I know rate it highly at tactics, strategy, and communication quality, as far as we've seen in what's been published.

    The possibly important limitation of this Diplomacy AI is that it's not playing completely unrestricted Diplomacy, it's playing the "Blitz" variant - so there isn't enough time for every player to talk to every other player every turn, let alone have lengthy planning sessions. The conversation is mostly tactical, so yes, nobody's expecting idle chat or unfocused messages, playing to the AI's strength.