@draug: "Interesting idea. A "learning" AI that takes the previous results and uses them to determine what is best. Wouldn't that make it a weak player at first, but one of the best learners on the site?"

That's what this means: "...built something like a genetic algorithm and ran it through a few dozen evolutions"

Remember... each game is different... moves early can screw you late.... so 'learning' is such a tricky thing.
A basic genetic algorithm works like this:
Put in a bunch of random moves. Build several random move 'players'. Play against one another. After losing X number of times, do one of the following things:
- Add a new random line
- Update part of a current line
- Remove a line
Put him back into the fray.

After one wins Y times in a row, make more 'players' based on his rules. Most of the other 'players' will die off. This is an 'evolutionary jump'.

Basically, run as a batch process to play the game as quickly and efficiently as possible, to try and run through several hundred evolutionary jumps.... possibly thousands. Then you'll have a player that is seemly randomly playing, but does really well...

Adapting this to diplomacy is extremely difficult. You can't just 'plug in a bunch of random commands', because not every move is possible in every game. You have to adapt to where your units are and such, and that really fouls up with the command/rule list. It's an interesting problem to tackle... perhaps use a neural network to read in the other players input... adjust moves (maybe number units?) to only valid ones in some programatic language? Who knows... should be fun to code, though. :)

@Abge - heya! Yeah, busy as hell at work. Missing being around here, though...

I hear ya. A month into a new job myself. Not much time for webDip (except today when I'm stuck in the lab installing OSes).

@orathaic Being omniscient would offset the AI's inherent disadvantage of having it's algorithm on github. Opponents could determine a series of moves that would put the AI in an unrecoverably bad position. It could be an interesting challenge, particularly since the human players' moves would not be known to the other humans.

In a pinch, the human players could all cooperate, coordinate moves, and play the AI like a puppet. But are any of the humans lying?

@tendmote: the code being on github doesn't mean you can predict what it will do in a given situation. It isn't too hard to write unpredictability into code, look at code which generates model of the 'unpredictable' weather system.

I can know what the code is, publicly, but without running it i can't possibly know what the result will be.

And the 'initial conditions' which you would run here would be based on how the AI 'feels', in weather initial conditions are what it is like today; Any good AI would 'learn' these feelings in the way fortknox describes above.

Yes, your cheating AI could be an interesting challenge, but it wouldn't be diplomacy.

@orathaic I'd run the code and know the outcome :)

And I agree, it wouldn't be diplomacy, but it would be something *even geekier*... at least one AI player in the game behaving deterministically with every other players' moves as input... it would be diplomacy with a non-human intelligence thrown in.

The more I think about it the more I like my idea for a new variant.

But running the code still wouldn't tell you the outcome. There is always a random factor involved and the seed can easily be pulled from something uncontrollable in the system so you'd never know exactly which of several outcomes would be chosen. Plus, *you* wouldn't know all the other players' turns. The AI wouldn't know it until the turn ran, so you wouldn't know it until the results came back. And if you did manage to cheat and altered your orders, that would alter what the AI did as a result of your new orders.

Wouldn't you design it so that the AI could have a variable aggression level? When the game starts, the aggression level could be set randomly. Then when deciding between a conservative move and an aggressive move, another random roll of the dice based on the aggression level would be used. How are you going to predict that?

@Draugnar Yes you could introduce a random element, but the variant I'm imagining wouldn't have it. It would be a deterministic algorithm, where the inputs are the other players' moves and the state of the game. And the AI would be *good*, i.e. it would play to win for itself. Now, the human players cooperating would always have enough information to know what the AI would do if they communicated honestly and ran the algorithm. So depending on how good the AI was it would be a fair fight... but the human players would want to win for themselves too, so might *not* be honest with each other and try to get the upper hand.

Details would matter but... it would be weird. And eerie. The AI would be a potentially stoppable psychic-zombie in a war of all against all.

@rolling dice, That was one of the great things about diplomacy, the lack of random chance... no dice rolling required, moves are completely deterministic... and somehow you want to re-introduce it. I beg you, go play risk! *looks up to heaven* :p

@Alderian And since the AI would behave deterministically, it would still be in the spirit of the Diplomacy game, i.e. chance plays no role. An AI that used anything random in it's input would spoil the "chance plays no role" aspect.

And an omniscient AI that used the other players moves as inputs would also preserve the human players' honesty and forthrightness as primary factors in the game. Total honesty and full disclosure would make the AI an easy mark, you'd know what it would do. The AI is in fact under human control, indirectly, and in a way that threatens to destroy all the other players and win for itself.

It's weird...

@orathaic no dice! a deterministic AI with knowable inputs!

@tendmote: You've already admitted that playing against an omniscient AI wouldn't be diplomacy, and my comment was directed at rolling dice not at you...

My point is against the pseudo-random and in favour of deterministic AIs which are unpredictable, by definition. I'm thinking some kind of neural network weighting tactical moves within strategies, and some weight to 'try something new' for sampling strategies/tactics which it hasn't got any information on...

@orathaic I don't mean to be pedantic, but how is what you're suggesting unpredictable? If it's deterministic, and the inputs are known, then it's predictable. I can run the code with the state of the game and see the result on my laptop.

The weird thing I'm suggesting is *given that it's predictable*, make the players' moves part of the input to the algorithm, so they can vie for influence over the AI.

Yeah, the weird thing is, you're not suggesting playing against an AI at all...

And i'd have to refer you back to my earlier comment on building Intelligence, using neural networks, with varying weights (which form parts of the 'input' to the algorithm, but are in essence part of the program's current state - though it can iterate through possible moves, ie think about them, and use the resulting guesses about what will happen to alter it's own weights)

I'm assuming you don't know what weights the program currently has, because it is unfair or cheating to go into it's mind and read it's thoughts, just as much as telling it what you're going to do gives it an unfair advantage

@orathaic If an AI was built into webdiplomacy, the algorithm would be known because the software is AGPL licensed ( http://webdiplomacy.net/AGPL.txt ) which (iamnotalawyer) means that the code would probably end up published on github with the rest of the code. *You* might not read the AI's "mind", but everyone else would. It would be public knowledge. I surely would run the algorithm against every possible input to see which outcomes were beneficial for me. Anyone who could, would.

Now, if you simply take into account that the AI's algorithm is known (and it could still be implemented as an "AI", whatever definition of that you choose), it's more fun to have the players moves be an input to it's algorithm, because then you can work to influence it's behavior, instead of just flat-out knowing what it is beforehand because it's only based on known, past information.

Here is the kind of AI I envision and it would be used as an assistant as well.

Looks ta historical record to determine best "raw" move.
Looks at recent supports and aggressive or defensive behavior to determine friendliness level and evaluates that against friendliness level of the "raw" move, possibly choosing a different move that better equates to the friendliness level.
The player could adjust the friendliness level up or down one based upon communication and gut feeling when used in "assistant" mode.

@Tendmote, the code is public, but FOR ExAMPLE the messages which players send are private, any AI would need to learn (draug's example above just examines historical games, but my example would also try predicting moves from the current position) - you can go ahead and run it against all possible inputs, (assuming you have the computational power to run hundreds of trillions of simulations) but at that poont how do you which answer is the ONE which the AI will pick?

You don't know the internal state of the AI, which is based on it's experience and learning (whether that is Historical games or actual games played by the AI)

The hard bit is getting it to parse press, though i've seen sggestions for a schene for handling press, with limited verbs, like SUGGEST, AGREE, SHARE, etc...

Anyway, back to the original point of this thread; because of this thread, i wrote a basic websocket ping server, which is cool; really new technology, with server-side js and capable of broadcasting of events to clients.

Really interesting, but i'm not sure how easy it would be to incorporate it into webdip (i didn't see a good php/apache websocket implementation)

I can imagine a server where the database is treated like a client, and the server broadcasts messages to all connected clients. Newly connected clients can download from the db... Hmm

I agre about the deterministic bit, ans keep the learning algos general and simple to start... & from a long list, like the title: lolol
"Behavior in a Shared Resource Game with Cooperative, Greedy, and Vigilante Players" http://arxiv.org/pdf/1306.3127.pdf

I agree and