I played my first game of Diplomacy in 1969, and have played only sporadically since then, some FTF, some via judge. I have played two PC versions of the game and found neither worthwhile. I will assume for this post that the standard boardgame rules will apply to at least one way of playing Paradox' version of the game, and thus will comment about how to make the AI work. I am not a programmer, but took some classes way back when (including some that required punchcards), and have written an Othello (Reversi) game with an AI opponent for my own amusement, and some others. Likewise, I am familiar with game theory, having read a bit and having had one course graduate business course that spent some time on it, but am not a mathematician or economist. I have monitored people (generally with a programming background) that try to write and run financial trading systems, which is a form of AI. Thats enough of the disclaimers.
When I have written or thought about how AI could work in the past, I try to describe what I would do in playing the game, and distill those down to behavior rules, what is usually called an 'expert system'. That is what I would suggest to Paradox, that they download masses of diplomacy games played through on-line judges, analyze those, with the help of experienced players, and use that to help build up their expert system database. After that they would have to build in a couple of necessary twists.
There are two parts to playing Diplomacy, which I will label 'tactics' and 'diplomacy'. If you are running a power already up to 17 supply centers, facing two separated opponents, then diplomacy probably won't matter, if you can find a set of tactics that will force a win (gaining that last supply center) or will have a good probability of a win, as there are numerous situations where it is probabilistic, i.e. an opponent can defensively support either one of two supply centers, and you can attack either but not both with two, so if either or both players randomize, then there would be a 50% chance of capturing a supply center. If the situation elsewhere is static, and this situation persists, i.e. neither the defender nor attacker can bring reinforcements, then the attacker would eventually win by attacking the unsupported supply center. Therefore, the AI should first analyze the board to see if there is a forced or probablistic win tactically, and if so, not worry about diplomacy much that turn. This is a large, complex problem. If there are 36 units on the board each having an average of five possible moves, then there are 14,551,915,228,366,900,000,000,000 possible sets of moves to consider after one spring or fall move. While many of these possible moves can be ruled out immediately, as they involve defensively supporting a unit you are moving, the number of plausible move combinations remains much higher than in chess, so the ability to follow the tree of tactical possibilities all the way out to a result could well be computationally impractical if a brute force method was applied, only paring down implausible, i.e. self-contradictory, moves. That being said, I would hope/expect that Paradox could come up with a tactical AI that would fare reasonably well in a no-press game, if not then human players would jave a tactical advantage. Some other games compensate for this by giving AI run players production or combat advantages, but that is not possible inthe standard rules of Diplomacy (other than making the one human player start as Italy). The other way to compensate for tactical AI weakness would be to have computer run powers be automatically more suspicious of human run players and be more apt to ally with other computer players. In one game "No Greater Glory", a US Civil War game, the AI was given what the player moves were, which would spoil Diplomacy, so that choice is not advisable, although the a slight chance of a 'spy' learning something occasionally could be tolerated.
I will break this off now and post later regarding some ideas about the diplomacy part of the AI. Hope this helped.
