Diplomatic Go

(Most of the below post is probably pretty obvious to anyone who has played diplomatic games before or has any knowledge of game theory. I don’t belong to either category so I wrote it for myself and for others like me :slightly_smiling_face:)

I made this comment in the middle of the game:

This was a bit naive of me, as has become clear to me while playing and discussing the game, so let me amend my position.

Here’s how I might have reasoned earlier (not a real quote, just approximating my thought process):

Players without any clear goals, who are still able to affect the game, is in my opinion a bad thing. This is because I believe having ways to predict and reason about the opponents is very important, both for making the game enjoyable and strategically deep. Thus the game rules should specify a utility function that is as fine-grained as possible. Ideally, every player should try to “play rationally” to their best of their ability, under the specified utility function.

What wasn’t clear to me then was that, even if it were possible to force everyone to “play rationally”, this wouldn’t necessarily be a good thing. Consider this example:

We play go with 3 players, but the players take turns submitting moves. The goal of each player is to maximize their own score on the board. We further dictate that you should play under the assumption that both your opponents are only trying to minimize your score, without caring about their own.

(If it seems weird that each player is playing under some assumptions that look false, just think of it as each player optimizing for their best possible worst case scenario, given no information about the others.)

(Also ignore for now how impossible it would be to play 1v2 normally. Assume we are already in an endgame with some immortal groups on the board or something. The details don’t matter for our thought experiment.)

The result is a game where we can actually define “perfect play” just like we do in normal go, since we’ve basically reduced it to a perfect information 2-player game for each individual player, “me vs everyone else”.

That may sound pretty good, especially if you like being able to “read” properly like in normal go and so on. The bad news is that we’ve completely removed the diplomacy from the game.

For a less extreme example, consider the game we just played. Would it be a good thing if we somehow could require each player to always play for a win if possible, or else a draw if possible? No, not really. Strong diplomatic play may be to make a threat to play “irrationally” later on, to force cooperation from another player. Note that you may not actually have to make any “irrational” plays, only make a credible threat that you will. But if the rules forbid you from following through on such a threat, then it can’t be credible and won’t work.

I think these examples demonstrate that in some way there is an inverse relationship between the two desirable properties of being able to “read” (or more generally predict the other players), and being able to do diplomacy.

So does this mean that a more fine-grained utility function is bad for the diplomacy? No, it just means that we have to completely embrace the fact that the given utility function is merely a suggestion, and we can never fault a player for playing “irationally”. Still, the utility function given in the rules plays an important role. It tells you what to play for, in the absence of other considerations. It gives you a goal from the absolute beginning of the game (before you have made any friends or enemies), and with a fine-grained utility function you will almost always have something to play for all the way to the end.

If you want to ignore the given utility function, you can. But it’s there when you want it.

All this to say that I’m still very much of the belief that a more fine-grained utility function is better for this game. It should be clearly specified in the rules, but it should also be clear that it is different from the objective rules of the game, since you can play however you like.

I believe one can say a bit more about what constitutes “good” play, and when it’s appropriate for a player to seemingly go against the given utility function, in the context of repeated games. For my last proposal a few posts above, where a prize pool of $ (# players) * (# intersections) is distributed, I like to think of this as each player paying $ (# intersections) to participate in the game. Then we can imagine a long string of such games, and the goal is to make money in the long run.

But I don’t think we have to write “imagine this hypothetical scenario, and play accordingly” in the rules. That might be helpful for training an AI to play strongly, but I think most humans will quite naturally converge towards the same opinion of what constitutes strong play, just based on our psychology and the human context of participating in a board game.

5 Likes