Ai

[unmerged(6303)]

Here's an interesting paper I found a while ago:

http://www.cs.biu.ac.il/~sarit/Articles/13.ps

Someone put together a diplomacy AI as part of an academic AI research project. He claims it runs okay, although it sounds like it hasn't been tested very thoroughly.

More stuff here: http://www.starblood.org/daide/news.html

Looks like they have some AIs you can download.

 

[bombshellboi]

I dont think you guys have to worry too much about the AI for this game - Paradox Interactive may have many faults and their games may be a little rough round the edges (Crusader Kings more than the rest) in some areas - but I really dont think AI will be a problem.

AI seems to be the one area in all Paradox games that seems to be consistently decent - again room for improvement (but all games have room for improvement) but otherwise good.
What I would be concerned about is them giving the game a REALLY crap Diplomacy engine - which would defeat the purpose of the game; that being Diplomacy.

No - i fully expect the AI in this game to be just as decent as the AI in Paradoxs' other titles. And if anyone actually says that the Paradox AI is bad then i would refer them to the AI in Koeis' Romance of the 3 Kingdoms 8, the only example of a games series where the AI actually got WORSE in a sequel...

 

[unmerged(39426)]

The difficulty in designing an AI for Diplomacy is not in making one that can handle the moves (very simple) but in making one that can imitate a human player in how it decides when/where/how those moves should be made... to put it simply yet not clearly it is difficult to design an AI that will be able to put the diplomacy in a game of Diplomacy

Actually, the problem is greater than you state. To date, there hasn't been an AI that could get the MOVES right. And the moves are the easiest bit. I've been told its the size of the decision tree. Start with 3 units, each with (say) 5 choices. Thats Spring 01. Then each of those units has 7 choices in Fall. Then, miracle of miracles, you build two new units, and now the computer has 5 units that have between 3 and 8 choices, including convoys, supports, etc. All the moves have to be dependant on what the OTHER ~28 pieces in the game are doing. The ability to make good decisions on this scale has never been successfully programed. Hey, I'm a dancer, not a programmer, I probably have somehow misstated technical details above, but the idea that getting the moves right is easy is mistaken -- an AI that could play in a no-press game would be a wonder and a delight.

Hence my argument for an opening book. Perhaps some randomness within it, or a "personality" modifier where it tends to defend or tends to attack or some such . . . Anyway. I am looking forward to seeing what Paradox can do with this. It aint gonna be easy.

 

[Therlun]

I dont think you guys have to worry too much about the AI for this game - Paradox Interactive may have many faults and their games[...]
AI seems to be the one area in all Paradox games that seems to be consistently decent [...]
No - i fully expect the AI in this game to be just as decent as the AI in Paradoxs' other titles. [...]

i completely disagree.

on a list with all the things that count for paradox the AI was until now no point for them.
the paradox AI is extremely bad, and even with huge advantages it is isanely incompetent.
it does not even know one third of the key concepts of the game it was created for!

it is tied up in a very strict, unflexible script engine which makes its decisions often wrong, illogical or outright stupid.
the paradox AI is not able to win a situation where random button smashing would be a victory.
simply put, the paradox AI does not know what game it plays.

sorry if this seems overly negative and aggressive, but even calling the AI of EU2, Vicky or the HoIs "half decent" is a big exaggeration.

 

[TheArchduke]

I think you are exaggerating a bit, the AI of Eu2, Vicky has to cover with many things and exploiting the AI is not a problem in any game, I have seen so far.

 

[Therlun]

i really dont want to troll, i like several paradox games and their approach to creating games in general, but the AI is IMO one of their biggest weaknesses until now.

there are two things an AI has to do.
act, and react.
the paradox AI can neither.

i dont expect a humanlike computer, but an ai that is not able to handle a key-concept like attrition or shiprange, that cannot build up or keep a closed frontline, that cannot see when it is utterly defeated, simply is weak.
an exploitable AI and an AI that is not able to play its own game is something very different.

 

[unmerged(7276)]

The AI in vicky is the best i think.

If the AI could simply distinguish between being defencive and offensive it would help. That is if hes just defending he doesnt need to try and invade you with 4k armies. Also simple things like building decent sized armies

 

[unmerged(39426)]

Ok, you see my point, I think. If general consensus is that the Paradox AI can't handle other games (I have zero experience here, I haven't played a Pdx game) then there is every reason to pester them to work hard on the Dip AI from the beginning. The fruits of such labor would be substantial -- I can imagine the first company to bring decent AI to markey would be in a position to maximize on the rewards . . .

 

[bombshellboi]

I actually think that the Paradox AI engines have been semi decent; like I said there is lots of room for improvement in many areas of paradoxs' games - if you compare how bad the AI is to say the fact that after nearly a dozen games Paradox remains "stuck" in a very dated looking 2 dimensional engine or as you said an inflexible "event" system then the AI is pretty good.

Also compare the Paradox AI to the AI in some other games ive seen like for example Koeis AI in Romance of the 3 Kingdoms 8 - the most idiotic Ai ive ever seen.
For example the AI in Victoria/Eu2 wont declare war on you for no apparent reason - at least not normally; you have to either have crap relations or they have a casus belli on you. And the AI in Eu2/Vicky certainly wont break their alliance with you and declare war on you just because your army is occupied fighting a mutual enemy.
This happens in RTK 8 - you are allied with a neighbouring Lord to fight a more powerful enemy. I have seen this neighbouring Lord declare war on you because your army happens to be on the other side of ur territory fighting your common enemy. That is a Stupid AI!

Comparing the AI in PE games to that, and to games like Pax Romana *shudders violently at remembering that game* shows that the AI paradox puts out could be a heck of a lot worse.

Also Diplomacy by nature is a relatively simplistic game - in fact every incarnation of a computerised version of the game ive ever owned has been so simplistic ive found it sleep inducing boring.
So I would think that PE would have much more time to devote to improving the AI as they wouldnt have to concentrate on other things like an extremely complex Economic system (Ala Vicky) or a mind boggling marriage/family system (ala CK)...

EDIT: Actually now that I think about it this is where Paradox usually falls down - they start out with a vision of a great game with so many options and avenues to explore. And they put all their time and energy into making one or two unique parts of the game really really good they neglect other areas and it brings the whole game down.

I still maintain that with the exception of CK (whos AI was abysmal) the AI in PE games is semi decent. Not worth worrying too much about where there are so many other horrifyingly bad areas that could better benefit from PE's attention. Not that PE's games are bad - I enjoy them all with the exception of CK (and I dont like that one because I waited an additional 1 and a half to get it in the UK after the rest of the world and the game still didnt work like it was supposed to).

 

[unmerged(33865)]

It would seem that the first test for the AI tactical engine is to be able to win, as any power, against 6 other powers that remain in civil unrest. For this level, the AI would have to be able to move toward supply centers, grab unoccupied ones, attack with support, build reinforcements, and convoy armies (to win efficiently as England). As discussed previously, there are probably too many combinations to fully explore the tree of possibilities, so some way of approximately evaluating the desirability of a position that could be achieved, say by the end of the winter turn, with ones own units, would be necessary.

The second step would be to develop an AI that could win while the other powers support themselves defensively in a logical, randomized fashion, i.e. if Warsaw was threatened but Sevastapol was not, Army Moscow would support Warsaw, but if both were threatened by two units, it would support one randomly.

The third step would be to enhance the AI so it would win against 6 other powers playing using the AI developed above, without making use of the fact that they were known to be using that algorithm.

This sketches out a beginning of an approach called a genetic algorithm, where subsequently developed algorithms are tried out against all prior ones. One doesn't always apply a Darwinian 'survival of the fittest' rule, because situations can arise where AI X could be beaten by AI Y, and AI Z can beat AI Y, but loses to AI X, which gets you nowhere, even without introducing probabilistic outcomes. Therefore, some randomization between various AI algorithms, by power or turn or both, would probably be appropriate.

The good part about this is that once a candidate tactical algorithm has been written, you can just run it through multiple games against others and see how it works. Having an experienced human player review the games to comment on the apparent blunders could possibly help the developer tweak the tactical AI.

The drawback is that this doesn't yet factor in diplomacy, which should substantially change the behavior. One way to simulate this in testing tactical AI's is to start some games with an approximate representation of some permanent alliances, by an alternative country arrangement, i.e. if Germany and France were allied, have all the French supply centers and units be German at the start of the game, including having Paris, Brest, and Marseilles be considered German home supply centers. Thus other powers could be run against such alliances. This kludge could be replaced by actual alliances when that was developed on the diplomatic side of the AI, which I will get around to posting on eventually.

 

[CatKnight]

It would seem that the first test for the AI tactical engine is to be able to win, as any power, against 6 other powers that remain in civil unrest. For this level, the AI would have to be able to move toward supply centers, grab unoccupied ones, attack with support, build reinforcements, and convoy armies (to win efficiently as England).

I think Bryan's putting together some good discussions with his two posts.

An AI for Diplomacy, as has been discussed, would need to handle two things: 1) Tactics (moving, attacking, defending, builds) and 2) Diplomacy. We'll get back to '2' in a moment.

Bryan's test would definitely be a good starting point. If the other positions are inactive, can the AI win?

The next step would be to start adding some of Director's and other peoples' thoughts on personalizing the AI. I'm sorry, but I have to reject the argument that the permutations of moves is too much for a computer AI to handle. Chess is a perfectly good analogy, with each side having 16 pieces.

From what I understand (I could be mistaken) a chess program is strongest in the beginning - because it has opening playbooks with simple if-then variants built in - and the ending when it's become crystal clear what the objective is. Chess AIs sometimes struggle in the middle game and have to start putting together tactical priorities.

The setting of 'playbooks', 'priorities', etc. can be put together by giving the AI a specific personality - a specific winning strategy. For example, France has four possibilities: 1) Try to keep everyone peaceful long enough to secure Iberia, then fold into one of the other strategies - 2) Attack Italy... slower, but if Austria's not doing anything it has potential - 3/4) Attack England or Germany, preferably with the other's help. There are other possibilities, but this is more than enough to feed the AI. At the beginning of the game the AI 'picks' one - and has a certain set of moves to prosecute.

The middle game, with switching alliances, could still be dicey but at least then an AI shouldn't implode in the beginning unless they're simply hit by too many opponents.

The end game with little diplomacy shouldn't matter to the AI. By now it should be well aware that either 1) it has the advantage, find a vulnerable SC and exploit... or 2) someone else has the advantage, and it's attack now or never. The latter part would be a simple SC count, not dissimilar from various games' BB ratings.

Bryan's argument makes a lot of sense to me - teaching the AI to evaluate its potential winter positions, then how to find against a randomly defensive supporting AI, then how to win against THAT. Once the AI is able to do that, human players watching the games can help determine purely tactical blunders that need to be fixed.

2) Now....Diplomacy (interaction) is going to be a problem...there are two possibilities:

A) Simply don't. All SP games are 'no press'/'gunboat.' There are human PBEM games that run like this. If the tactical AI is decent, then this should be doable.

B) All the AI *really* needs to know is 'who's my friend? who's my enemy?' (We'll ignore deception and non-aggression pacts/DMZ pacts for the moment.) If the AI has a (loyal) AI friend, then the computer treats them as one power. That would help it handle supports, convoys and the like. Enemies will get a taste of the tactical AI above.

The real problem I see is AI friend with human. We don't want to give the human limited control of the AI pieces, that's too vulnerable to exploit.... so we have to worry about coordinating attacks, including supports and convoys.

Diploming with the AI....is doable. I would recommend Paradox try to find Trust or Betray: Legacy of Siboot, c 1987 by Chris Crawford. It involves you and six AIs with distinct personalities. You're trying to get information on other contestants from them, and they want information on what you know. What you learn could determine how you do in the night's contests. Each AI has ratings of fear, trust and love for the other players, determining how likely they are to trust you with information, or betray what they know about you to others. Something similar could help determine 'relations' between the six AIs and player, which would trigger a warlike response or desire to work together as above. This will definitely require a lot of work and discussion.

 

[Powerslave]

I have faith in Paradox and I think they are going to make a superb game with a good AI. However if Paradox game Diplomacy at some level atleast plays like the boardgame Diplomacy I think that they should make the AI heavy moddable so that we can create our ow AIs. Paradox have plenty of intelligent players that I think could improve on the AIs delievered with the game.

Well atleast it would be fun to have tournaments to see which one could create the best AI.

 

[unmerged(39426)]

I think Bryan's putting together some good discussions with his two posts.

An AI for Diplomacy, as has been discussed, would need to handle two things: 1) Tactics (moving, attacking, defending, builds) and 2) Diplomacy. We'll get back to '2' in a moment.

Bryan's test would definitely be a good starting point. If the other positions are inactive, can the AI win?

The next step would be to start adding some of Director's and other peoples' thoughts on personalizing the AI. I'm sorry, but I have to reject the argument that the permutations of moves is too much for a computer AI to handle. Chess is a perfectly good analogy, with each side having 16 pieces.

Ah, sadly, this analogy is false. Yes, there are 32 pieces in Chess and 34 in Dip, but this is where the analogy falls apart -- in chess, each side has 16 pieces, but only one piece moves at a time. In Dip, of course, every piece moves at the same time, increasing the complexity of the board exponentially.

FYI, this isn't a new problem, nor one that has gone unaddressed before. There have been several attempts to write a Dip AI over the years. When I lived in NYC a fellow actually had one coded that, if you inputted the other six players moves first, could come up with a respectable set of moves. He was tweaking around with "alliance Matrixes", as well, along the lines you discuss above -- who is and isn't your friend. The AI was not horrible -- until it got to 1903. When it completely fell apart.

That isn't the only iteration of the Dip AI. There was, and maybe still is, a newsgroup somewhere devoted to the project, and I think there has even been a gunboat game or three on the judges where the various AI's play each other, and one engine SEANAIL, by the fellow from NYC, even played in a gunboat game against humans, as Turkey, eliminated in 1904, IIRC.

All of these efforts have been by programers tackling the problem in their free time -- but you know programers, for some of them free time is just twice as much time as they put in at work. And, this is a problem that these people have been working on (some of them) for over ten years. Ten. Years. Ok, with no budget and no deadline -- but ten years is a long time to be hacking at code . . .

So why am I such a negative Nellie? I really DO want to see something good here. I just want to put the word in Pdox' ear that producing a capable AI for Diplomacy is monumentally more complex than it at first appears. The chess analagy is a popular and exceedingly dangerous one. If you have access to some chess code, and have the ability, try to get it to make a move for three pieces, black and white, at the same time, not knowing what the other side is going to do. Watch as the processors heat up to the temperature of the surface of the sun! Go out for a sandwich as the engine grinds to a halt under the levels of complexity!

An opening book, some decent interaction scripts, a randomizer for choices within branches of a tree -- all of these are doable, and will help tremendously. The hated Hasbro AI badly fails the "win the game against civil disorder" test. An AI that can pass the "win the game against supports and holds" test would be a huge improvement. I have high hopes!

Chris

 

[unmerged(2005)]

Well, my hopes are a wee bit higher than the AI being able to solo if all other six powers are in civil disorder! If a Chess programme can beat Kasparov & Co, I mean it should be possible to design a tactical AI able to beat me in a nopress game.

And why should a diplomatic AI be designed to be a smart bargainer? In 1989(?) CivI was able to bargain. Okay, not brilliantly, but that was the computer stone age... Why shouldn't an AI be competent in negotiating demilitarized zones, reciprocal supports, non-aggression pacts, move combinations, transfer of SCs, mixes thereof etcetc., while emulating human behaviour (vengefulness, carebearism (okay, not sure I want that in a PC game as that's pretty contrary to the spirit of Diplomacy, chainsaw dipping, terrorist styles (you know what I mean ), etcetc.).

The hope here is of course that unlike other PC games, the AI won't be an afterthought, but the main preoccupation of the development team. After all, implementing the rules should be doable in no time. Many have already done that.

Really, we can't compare this effort to Paradox's EU type games. Those AIs suffered from the tons and tons of figures, decisions etc. Such complexity makes it very tough to come up with a half decent AI. That said, a Diplomacy AI is also a daunting task - just in a different way.

 

[unmerged(48389)]

I've been interested in diplomacy for about a year but never managed to play a game.

Please correct any of this if I'm wrong, but I do have an opinion on Diplomacy AI.

An exploitable AI would simply kill diplomacy. And this is a serious problem. Computers, by their very nature, are designed with consistancy in mind. 2+2=4 today, 2+2=4 tommorrow, and 2+2=4 next thursday. An AI France with a good Relationship with Germany might ALWAYS choose one opening, and whereas a sour Relationship with Germany or a Good Relationship with Italy chooses another.

This is not a laughing matter. The AI would remain an efficient player until humans can figure out what its plans are by its moves, at which case you are reading its cards and it is not reading yours. Even the best tactician can only cut even against perfect knowledge, and not even that sometimes.

An exploited AI might result in the ultimate proof of failure...never winning a game. The AI will make or break the single player release of this game, but the strength of the AI won't be known unless at least a dozen games are played against it.

I would suggest that the AI not even pursue the rule oriented (X, if Y) line of thinking with Diplomacy. There has simply got to be something highly cool in the works for this--I'm sure Paradox knows all of this. And, the challenging part is that the AI can't cheat, or this will also ruin the game.

Some people have suggested randomness as a AI tool. This is better than nothing at all, but randomness is not the cure I'd expect it to be.

The first thing you learn about "randomness" is that there is a random chance involved. The next thing is the chance of it happening. The clincher is knowing the AIs next two moves in advance, as I can do in various computer games, HOI included.