There's a project dedicated to building diplomacy AI here: http://www.daide.org.uk/wiki/Main_Page (although, I must admit, I haven't given it a thorough look over). It appears to be solely related to determining optimal moves outside of the 'talking' part of diplomacy (essentially a gunboat AI).

I see two sticking points though in coding an AI that can play the diplomatic aspect of the game.
First I need a huge set of training data for the algorithm, thus I'd need access to a huge number of games (which both this site and webDip have). What isn't straight forward is that I need to be able to see all the press for all those games (which isn't available to the general public at this point).
Second I'd need to figure out how to cull and parse that press data for relevant press and somehow categorize/classify the wide variety of press into a list of diplomatic "moves" an AI could make and a way to interpret a player's press to the AI.. Obviously this is an immense task that I'll never be able to finish or tackle myself, so I envision an incremental method, the AI would have some very rudimentary diplomatic options and be coded to give a small set of predefined queries (and to accept those same queries) analyze what 'choice' it wishes to make in light of that diplomacy and then to issue (or not) a valid response (and either act in opposition to the agreement or in support of the agreed interaction). Examples would be: 'Let's bounce in <territory X>', 'Can I have <territory X>?', '<Territory X> is yours', 'Let's DMZ <territory x>', 'Support me to <territory X> from <territory Y>', 'Team up to work against <country X>?'

My first step though would be to create a tactical AI, which knows how to play gunboats. This could be used to fill in if players CD so as to minimize the game changing effect that CD's can have. I don't see a huge issue in attempting to code this up.

Anyone else have experience with learning based algorithms that might want to collaborate on this?

Thanks for the link. Being in the middle of the thesis, I haven't really looked into anything yet..

Obviously, if the training set is large, you'll have better results with the diplomatic piece. But you can probably get away with a small sample to start things. I wouldn't worry about culling irrelevant things too much - while it would help, I don't think it'll be totally necessary. You'll want to build something very similar to a SPAM filter and train it to classify communiques (you should be able to train it to classify a message, or parts of a message, as irrelevant, so it can do the culling for you).

Pulling the diplomatic 'moves' out of the text can be tricky, and it does seem like a better interface would be a move suggester or a separate syntax outside of the free-form press box.

I'm pretty new to learning-based algorithms, but I've been thinking about working on a similar project myself. I'd be interested in collaborating on something open source.

https://sites.google.com/site/diplomacyai/ - This AI responds to rudimentary press agreements.

So first step is probably to head on over to DAIDE and see what they are doing so we don't have to relearn everything they've already figured out. Second step would be to see if there is any interest in having such an AI on this or the webDip sites and get approval from Oli or Kestas. Then we need to decide how much we are going to redo ourselves and how much we want to borrow (if open source) from DAIDE.

Right. What languages are you strongest in? I come from a C/C++/C# background, but can pick up just about anything. webDip is in php right?

I have some php/mysql background, and C/C++/Python/VB/(matlab?) are my languages. Really I do everything in matlab and adapt it to C/C++ or python...

I wish you the best of luck!


http://www.diplom.org/Zine/F2003M/Norman/ai_article.html
http://www.perada-magazine.eu/pdf/1761/1761.pdf