vDiplomacy doesn't have an AI, unfortunately.

There probably wasn't enough data to train it on all the variants, not to mention how hard it would be to get someone to code it.

I'm interested in trying to code some bots to play vDiplomacy (I still haven't heard back from the mods yet). I think I would try to go with a more algorithmic approach than an AI-based approached, but this project is still in the very, very early stages.

What you're looking for is in webDip (https://webdiplomacy.net/), and the AIs there are just dummies.

Algorithmic would be interesting! I found the following article from 30 years ago prototyping one, actually:

https://diplomacyzines.co.uk/home/contents/rulebooks-rules-and-diplomacy-sets/artificial-intelligence-and-diplomacy/

Right now I'm still working on being able to scrape all of the information from the files on here, it will be a while before actually playing even on classic becomes a possibility

Meta Ai has done this.

If we can,we should make a AI on vdip.
However to let AI learn so many variants is not a simple task.

Maybe we can train Chatgpt,or Tongyi AI or some AI which are well-developped to do this?
I think that's cool!

Meta's Cicero is specifically built to play diplomacy, and is intrinsically different from AI like ChatGPT and whatnot, something like ChatGPT does not have the planning algorithms that Cicero has that allows it to actually know what it's doing and react to the situation properly. You can't just throw any AI at any problem and expect it to handle it, they are built for different tasks. The best shot would be that Cicero might be able to handle Diplomacy variants depending on how its coded, but that's hard to say, it has been trained on only one map so far.

Yeah, I've studied up on AI quite a bit, and in my knowledge, Cicero wouldn't be able to do any map-shuffled variants. It might possibly be able to do Classic but with units in different places, like FranceVsAustria. This is pure speculation, but training an AI to play a game is much easier than training it to talk, so I'd guess that its press element would break first. So if any multi-variant AI were to be developed, it would probably need to be gunboat only.

And I'm not sure if there's enough material on Diplomacy online for ChatGPT to outperform humans past the early game, if at all. I guess if you needed a neural network(which is smarter than an algorithm), you could build a Diplomacy engine in Python and train it on that. Otherwise it won't work at all for variants.

That's why I think an algorithmic approach would be better. There's not enough games finished to train a machine learning model on a map-by-map basis. But if you could program an algorithm to take in a graph with SCs and pieces on it, you could get a pretty good result, I think. Maybe? We should all develop an AI and test them against each other. I think I'm almost done with my variant-map-scraping program.

Yeah, that could be pretty cool for testing out the various algorithms. It might even be possible to code it just using human gunboat logic.

uh.
i think we should let AI try gunboat first,then 3-4 people map with diplomacy,then bigger maps.

Well, teaching it to do diplomacy is definitely the problem. Basically, training an AI to play gunboat diplomacy requires good programming knowledge and being able to make a neural network. I cannot say I know enough about what would be required, but it is definitely possible. Teaching it to play in press games, on the other hand, requires the following:

1. Understanding the board state and necessary objectives. Ex: I'm France. Germany is strong, so I need to ally with England.
2. Communicating, in human language, with other bots to try to achieve those objectives. Ex: "Hey England. Germany is getting too strong. Want to ally?"
3. Relating messages received to current board state. Ex: France wants to ally against Germany. I've got a plan to attack France, but Germany will likely profit more than I can.
4. Giving an intelligent response for the scenario. Again, this sounds simple enough, but there are so, so many subtle factors to take into account. Do you trust Germany? Who's the better player? What level of honesty is best? Lying is a strong strategy for a stab, but telling the honest truth will build rapport and may ultimately save you.
5. Do this for any given variant, with little to no textual data about the normal messages sent.
6. Ideally, outperform humans in this. I'm quite confident AI will eventually beat humans in gunboat quite easily. But combining that with the diplomatic element is much harder than it would seem at first glance. A map with fewer players would help, but it doesn't solve the fundamental problems.

Anyway, that's my "expert" opinion on the topic. Relevant xkcd:
https://xkcd.com/1425/