Artificial intelligence in 1v1 diplomacy

[Squigs44]

Someone just code this up already

[teacon7]

I would love to see Albert coded into webdip. We'd all lose, but still. I wonder what it would take to translate the AI to this code framework ?

...it's probably really hard, and that's why no one has done it.

[Restitution]

Someone just code this up already

The intersection of people who can build NN's and diplomacy players is probably pretty small.

[bozotheclown]

It would be interesting to see what AlphaZero could do with an imperfect information game like 1v1 Diplomacy. However, the estimate for the AlphaGo Zero hardware is $25 million, and AlphaZero likely costs more, so someone has to talk DeepMind into trying Diplomacy.

[captainmeme]

It would be interesting to see what AlphaZero could do with an imperfect information game like 1v1 Diplomacy. However, the estimate for the AlphaGo Zero hardware is $25 million, and AlphaZero likely costs more, so someone has to talk DeepMind into trying Diplomacy.

the CEO of DeepMind actually attended WDC 2017 Oxford, it's not impossible that this could happen

[Claesar]

It would be interesting to see what AlphaZero could do with an imperfect information game like 1v1 Diplomacy. However, the estimate for the AlphaGo Zero hardware is $25 million, and AlphaZero likely costs more, so someone has to talk DeepMind into trying Diplomacy.

the CEO of DeepMind actually attended WDC 2017 Oxford, it's not impossible that this could happen

Did anyone explain 1v1 to them?

[bozotheclown]

It would be interesting to see what AlphaZero could do with an imperfect information game like 1v1 Diplomacy. However, the estimate for the AlphaGo Zero hardware is $25 million, and AlphaZero likely costs more, so someone has to talk DeepMind into trying Diplomacy.

the CEO of DeepMind actually attended WDC 2017 Oxford, it's not impossible that this could happen

Does anyone have some connections?

[BobMcBob]

What about an AI for all variants of Diplomacy? I don't think it would be too much harder and you could have it all on the one AI.

[Puddle]

@BobMcBob, I would think that the task of teaching the AI how to Gunboat would be quite a bit simpler than teaching it how to play full press. Setting aside the need to pass the Turing Test to really be able to play full press, even just doing something like what Albert does wouldn't be quite a step up from just learning to gunboat.

[RoganJosh]

I think 1v1 diplomacy is to simple - AI would probably solve it completely. That would be pretty boring.

[Squigs44]

I think 1v1 diplomacy is to simple - AI would probably solve it completely. That would be pretty boring.

I'm not sure there is a complete solution. Too much guesswork involved.

The hard thing about diplomacy is its turn structure. In chess, or go, one decision is made at a time. White makes 1 move, then black makes one move, etc. In diplomacy, two players are simultaneously making up to 17 decisions that all interact with one another. So although there are fewer turns in diplomacy, each turn requires multiple complex interlocking decisions. Also, the fact that you are taking your turn at the same time as your opponent means you are working with incomplete information at every step.

[Restitution]

I think 1v1 diplomacy is to simple - AI would probably solve it completely. That would be pretty boring.

I'm not sure there is a complete solution. Too much guesswork involved.

The hard thing about diplomacy is its turn structure. In chess, or go, one decision is made at a time. White makes 1 move, then black makes one move, etc. In diplomacy, two players are simultaneously making up to 17 decisions that all interact with one another. So although there are fewer turns in diplomacy, each turn requires multiple complex interlocking decisions. Also, the fact that you are taking your turn at the same time as your opponent means you are working with incomplete information at every step.

You can solve a probabilistic game. Nash equilibriums.

[Squigs44]

I think 1v1 diplomacy is to simple - AI would probably solve it completely. That would be pretty boring.

I'm not sure there is a complete solution. Too much guesswork involved.

The hard thing about diplomacy is its turn structure. In chess, or go, one decision is made at a time. White makes 1 move, then black makes one move, etc. In diplomacy, two players are simultaneously making up to 17 decisions that all interact with one another. So although there are fewer turns in diplomacy, each turn requires multiple complex interlocking decisions. Also, the fact that you are taking your turn at the same time as your opponent means you are working with incomplete information at every step.

You can solve a probabilistic game. Nash equilibriums.

You might be able to solve for the best move-set at any scenario with your given information using probability. But you can't solve for a "complete" solution in the sense that you can't guarantee that an AI will win every game it plays.

[Restitution]

I'm not sure there is a complete solution. Too much guesswork involved.

The hard thing about diplomacy is its turn structure. In chess, or go, one decision is made at a time. White makes 1 move, then black makes one move, etc. In diplomacy, two players are simultaneously making up to 17 decisions that all interact with one another. So although there are fewer turns in diplomacy, each turn requires multiple complex interlocking decisions. Also, the fact that you are taking your turn at the same time as your opponent means you are working with incomplete information at every step.

You can solve a probabilistic game. Nash equilibriums.

You might be able to solve for the best move-set at any scenario with your given information using probability. But you can't solve for a "complete" solution in the sense that you can't guarantee that an AI will win every game it plays.

No, but that's not what solving a game means. Solving the game means that there exists an algorithm which can play perfectly.

[RoganJosh]

You can solve a probabilistic game. Nash equilibriums.

Nash equilibrium doesn't exist at every stage in this game.

You might be able to solve for the best move-set at any scenario with your given information using probability. But you can't solve for a "complete" solution in the sense that you can't guarantee that an AI will win every game it plays.

Solving means to describe the optimal strategy. In this game, there is usually no "best move-set." Solving, in this case, means to describe the distribution of frequencies by which you should play each possible move-set. Then you randomize...

Also, no AI in any game is guaranteed win, just let AI play against AI...

[Restitution]

You can solve a probabilistic game. Nash equilibriums.

Nash equilibrium doesn't exist at every stage in this game.

?

If diplomacy is solvable then there necessarily exists a Nash equilibrium for any decision

[Squigs44]

Also, no AI in any game is guaranteed win, just let AI play against AI...

Depends on the game. In a game of Nim, depending on who goes first and the starting heaps, you could build an AI that would win every single time.

[RoganJosh]

If diplomacy is solvable then there necessarily exists a Nash equilibrium for any decision

What can I say. In 50/50 guess for the win there is no Nash equilibrium. I think we don't use the same definition of "solve."

In Nim, depending on who goes first and the starting heaps, you could build an AI that would win every single time.

Even if AI plays against AI? Both would win?

[Restitution]

What can I say. In 50/50 guess for the win there is no Nash equilibrium. I think we don't use the same definition of "solve."

The nash equilibrium is to do option A 50% of the time and option B 50% of the time, anything else would be exploitable. That is exactly what a nash equilibrium means... I do not mean to be rude but I think you might be thinking of a different concept, or have not been taught the concept correctly.

Like, the nash equilibrium on rock paper scissors is to guess each option 33% of the time, which means it's a solved game. There is a GTO solution.

If 1v1 Diplomacy is solvable, there is an optimal (meaning non-exploitable) strategy. Which necessarily means that there exists some Nash equilibrium for every decision.

Even if AI plays against AI? Both would win?

Pure sophistry

[RoganJosh]

The nash equilibrium is to do option A 50% of the time and option B 50% of the time, anything else would be exploitable. That is exactly what a nash equilibrium means...

Sorry, but that is not what a Nash equilibrium is. A Nash equilibrium is a situation where both players made a choice, and neither player would want to change their mind even if they were told what the other player chose.

Soo - at least it turns out we had the same definition of what "solve" means! :)