AI Discussion Thread

[unmerged(9145)]

Compare Gal Civ to MOO3, Civ IV, or any other turn-base game and maybe you have a case. But when you compare it to a continuous play game with about 50 x the number of AI opponents, about 1000 more provinces, and an overall level of game complexity that is roughly on par, you do not make anything remotely close to a meaningful case.

We are years away from users having the kind of computer systems available to be able to run a Stardock AI type of model on EU3's scale in real time.

Computers don't give a rats ass one way or another about 'turn-based' games. To a computer ALL games are turn-based; certain instruction sets are executed, in order, eating a number of cpu cycles. When all instructions are executed the process starts over again.

Whether or not the player thinks he's playing a 'turn-based' game or a 'real-time' game is irrelevent. To the computer they're the SAME THING.

This was explained once before on this forum, by another programmer. Perhaps we should have a sticky explaining the concept to the non-programmers, who keep making the mistake of thinking that the external human representation of the game somehow matters to the computer.

'm just saying that you can't compare them because they are not the same type of game so the requirements on the AI and player's system are a lot different.)

And you would be wrong. The "type" of game is a non-sequiter here; all that matters is the complexity of the actions that the AI has to consider in each execution cycle. For this reason, it's legit to compare games like EU and GalCiv - except that GalCiv is more demanding because the range of possible AI actions is far, far greater.

The only applicable point you made here is that the number of opponents is larger. However, since most of these opponents are minors with limited options (you aren't going to run through useless cycles considering actions they can't possibly engage in) this isn't nearly the processor load it looks like.

In any event, I don't see the point in apologia for Paradox AI's. It's quite reasonable for any gamer to expect a decent, challenging AI in EU 3 after so many attempts in previous games. Unless you already know that the AI sucks as bad as it did in EU 2, or the HOI series? That would be some useful information to have for those of us who're planning to purchase the game.

Max

 

[unmerged(42223)]

I have question.

Has anyone ever wondered how when for first time you load up HOI2 (maybe same some of the other games too) after having the computer off all night it takes a significally longer time, compared with the second time, even if that 'second time' is many an hour later? Same thing happens if you mod many files.

My thinking is that this kind of game system 'loads' into ram the whole picture of the scenario so to speak and then runs it. Which would mean, the more files, the more it taxes the system requirements and more complicated it becomes possibly cascading more related quandaries.

Also alliances and removal of countries via annexation immediately speeds up the game. Further considering the Civ series for instance does not have more than 8 separate AI players. Where HOI2 can have well over a hundred.

Now skipping vanilla Max, I modded 54 fully functional countries, which is 53 AI's and 1 human. These play on a rich resource map and have dynamic economies about a powerful as Germany on average. Within six months there are 75 DoW. At scenario start there are 3000 divisions split between sea, air and land. All is a whirl at land, air and sea from one part of the map to other. It is solid, I run it at the fastest speed, well over 1000 divisions are eliminated per year and are replaced. All builds are excellerated.

So I know the Paradox engine works. AI Air is furious against Air, Sea and land, multiple large amphibious landings occur every game involving hundreds of units. So what is the difference betwixt a Battle Mod and one emulating history?

I think your slamming something that could work very well, once it is better understood. My concern is that they change too much without that demonstrated understanding.

When I first cast eyes underneath the hood of HOI2 I was simply amazed at all the detail and calls available to modders, which could only have come from a company with long history of being 'responsive' to its community. And wondered out loud to others that this puppy could be made to zing. Especially after all the other crap I had been working with years prior.

Well, I am even more sure, yet continue to be dismayed by the appearing lack of clear application of the game engine potential. And do not want see a favorable trend dimmished.

 

[Zeitgeist]

Computers don't give a rats ass one way or another about 'turn-based' games. To a computer ALL games are turn-based; certain instruction sets are executed, in order, eating a number of cpu cycles. When all instructions are executed the process starts over again.

Whether or not the player thinks he's playing a 'turn-based' game or a 'real-time' game is irrelevent. To the computer they're the SAME THING.

This was explained once before on this forum, by another programmer. Perhaps we should have a sticky explaining the concept to the non-programmers, who keep making the mistake of thinking that the external human representation of the game somehow matters to the computer.

And you would be wrong. The "type" of game is a non-sequiter here; all that matters is the complexity of the actions that the AI has to consider in each execution cycle. For this reason, it's legit to compare games like EU and GalCiv - except that GalCiv is more demanding because the range of possible AI actions is far, far greater.

The only applicable point you made here is that the number of opponents is larger. However, since most of these opponents are minors with limited options (you aren't going to run through useless cycles considering actions they can't possibly engage in) this isn't nearly the processor load it looks like.

In any event, I don't see the point in apologia for Paradox AI's. It's quite reasonable for any gamer to expect a decent, challenging AI in EU 3 after so many attempts in previous games. Unless you already know that the AI sucks as bad as it did in EU 2, or the HOI series? That would be some useful information to have for those of us who're planning to purchase the game.

Max

I am a programmer, with experience in planning and scheduling AI systems.

It's true that a EU2 can be thought of as a turn based game from an AI perspective. However, your reasoning is still flawed; there is not an infinite amount of time available per turn. Granted, even in a turn based game, there also isn't an infinite amount of time, but players are much more tolerant of a wait between turns in such a game. A wait of 10 seconds would not be uncommon. Now let's consider a wait of 10 seconds in a game like EU2 per turn. A four hundred year game at 10 seconds a day would take 400 * 365 * 10 = 1460000 seconds to complete. This is 405 hours, far more than most gamers would be willing to devote to a single game. Even if that gargantuan amount were available, there are a lot more demands being placed on those precious 10 seconds of time in terms of the sheer number of countries that all need to make short, medium, and long term plans, not to mention all the pretty little things wandering around the map, sounds playing, and other game upkeep functions that must be run (e.g. taxes, random events, revolt risk). Granted not all countries require the same amount of clock cycles-- countries at war for example would require more than countries at peace, for example. But if you expect to keep them all active you can't starve any country of clock cycles for too long at all, as it must constantly be assessing the situation and revising or creating new plans.

Basically it's not about apologia. It's about unreasonable expectations. Of course we all want to see AI that would knock us around and make the game challenging in single player, but the chances of that happening are slim. I just don't think the AI will be able to keep up with a human player given the resource limitations that we have. Even the GalCiv AIs get some boosts, albeit at higher levels and the Intelligent AI is no pushover. Regardless of the machine's perspective, RTS and Turn Based games are two very different genres and the distinction is very important. GalCiv 2 is a good game, but the fact is that it is a lot slower paced than the EU series which allows it to focus more on the AI. It also places severe limits on the number of races, with only a few major races with fully fledged AIs, whereas in the EU series, any nation is capable of emerging as a potent power, and therefore cannot be ignored.

With all of that said, it would be interesting for Paradox to create an AI that takes fuller advantage of slower speeds. I don't know if EU2 does anything like that currently; at any rate, I have never caught the AI behaving smarter at Below Normal than say Very Fast. I realize it would make something of a disjointed game with speed switching, but would still be interesting, at least to me.

 

[unmerged(47626)]

Sounds to be like what you guys need is a university-level history department with access to the Physics departments supercomputer.

 

[unmerged(42223)]

Yep, we've been up, down and back around and haven't concluded a bloody thing...

I'm so sorry...

 

[unmerged(6777)]

...a bunch of stuff that ignores the fact that the CPU has a finite number of cycles per second to execute commands, and also completely ignores that fact that in "turn-based" games (the ones that a moron like myself calls turn-based, that is, becuase the game sits there doing nothing until the player clicks the "end turn" button) the player has a psychological expectation of a delay between the end of his last turn and the beginning of his next turn, whereas "real time" games (games that an idiot such as myself who is perfectly aware that the code does till treat each day in EU1/2/3 as a turn but calls them real time because the turns all roll into one another without requiring the user to comfirm the end of the turn and must also process all AI activities in synch with the user's actions so they are no longer sequentially processed) where the user expects time to flow in a fluid way with little or no jerkiness or interuption; and would thus be a little put out by lag or a maximum game speed of 2-5 seconds per day...

I was going to reply but Zeitgeist did in a far more polite way than I wold have.

 

[Sinbuster]

How to explain my confusion :wacko:

As someone who can't make much sense of the AI files, could someone explain to me what the depth of the AI is when it comes to conflict?

Is it just a matter of: hmm, troops on this border, I'll build some to oppose them if they decide to attack. And, oh, no troops protecting that province I'll look for assets to attack there instead.

or is it a far more complex system?

I mean, the EU and HOI AI is essentially reacting to conditions in front of it as oppossed to creating even more favourable conditions. This is an assumption.

Ideally, I would like a an AI with as much depth as, say, the IBM program that beat Kasparov at chess... the name escapes me... Deep Blue maybe. It was designed, as all good chess players are, not only to plot moves in advance- in its case many moves - but to also consider alternative moves should the opponent not follow the bait.

I know that's a mother of a computer but man I wish the EUIII AI could be built in a similar manner. If the human is doing something, the AI could go through a whole list of things we might me up to, then draw us in to an unfavourable scenario. Imagine an AI that notices a build up of troops on its borders and intentionally leaves a province open, or poorly defended, so as to determine the time, place, and terrain of the conflict. Mindlessly tossing troops at mountain provinces because they are worth more war score is, well... mindless.

Just curious, if this game builds off the AI of all the other titles it'll be excellent, but one can always dream.

 

[Registered]

Ideally, I would like a an AI with as much depth as, say, the IBM program that beat Kasparov at chess... the name escapes me... Deep Blue maybe. It was designed, as all good chess players are, not only to plot moves in advance- in its case many moves - but to also consider alternative moves should the opponent not follow the bait.

Two problems with that. Deep blue is AFAIK the name of ths uper computer that beat Kasparov, something that is somewhat above even our collective budgets. Secondly, EUIII is a great deal more complex then a game of chess.In chess you have a relatively small playing field with a limited amount of moves. EUIII, same as Go, has many possible moves. This makes thinking far ahead impossible for an AI, it has to analyse all possible moves and all possible reactions to them (as far as i understand) doing that on the scale of EUII is impossible even with a very powerfull computer.

 

[Seli]

Apart from the problems nicely presented by Zeitgeist and MrT there is a second important difference between most PI games and the 4x games like the civ series and GalCiv. In the latter case people are content when they get good opposition from one or two of the AI controlled factions. For the PI games people not only want good opposition, but also historical and sensible behaviour.
Not any random country has to do well, but the countries we all know prospered.
And I feel the exploration phase in CIV and GalCiv makes it easier to construct a strong AI.

 

[Sinbuster]

No doubt Deep Blue is beyond all our budgets but is the conflict really that complex?

The provinces are still just squares to be controlled, and each country has only a set number of squares.

I oversimplify of course, but it draws me back to my original question: what's going on inside the AI's head?

Registered, what is Go? Is that a card game or something?

 

[unmerged(8915)]

I've done a bit of reading on AI since this thread started and its an amazing topic that i barely noticed before (unless i was playing a game and complaining)!

anyway, the biggest difference from an AI perspective between a turn based and real time is that the conditions change as the program is conducting its evaluation along the decision tree in real time, while in turn based, the AI can travel further down various nodes in the decision tree and have the luxury of the conditions not changing as its making a decision. added to that is that paradox games have dozens of ai's running at the same time and i think that paradox's efforts are simply remarkable!

if the only computer that can beat the best humans at a simple game like chess costs millions of dollars, years of development and has its entire gigantic capacity geared towards chess, what of us playing eu against our little pcs

go is a wonderful game and although superficially similar to checkers/chess, it is far simpler, yet more difficult to master. hence the fact that GO players can still beat GO computer programs.

 

[unmerged(42223)]

anyway, the biggest difference from an AI perspective between a turn based and real time is that the conditions change as the program is conducting its evaluation along the decision tree in real time, while in turn based, the AI can travel further down various nodes in the decision tree and have the luxury of the conditions not changing as its making a decision. added to that is that paradox games have dozens of ai's running at the same time and i think that paradox's efforts are simply remarkable!

In addition consider that there is also a 'flux' of what the conditions are even going to be. A battle can last moments to days with weather changing and troops added by player or another country AI.

In designing Risk Battle Mod all for the sake of being a war game exclusively I took advantage of some design tweaks that have improved AI proformance (I believe). Namely, there is a 60 day time lag from DoW to ground offensive (this does not apply to the human). As well the cycle of unit production is average 45 days. So, the AI's are not pressed to yank off troops already engaged (if so elsewhere) and they have ample time to redeploy and use the new units coming on line. This assures the AI's never are 'unprepared' for yet another war. Each AI often is engaged in two to three wars at a time for years running or death all without a problem and often with a great performance.

 

[Registered]

Registered, what is Go? Is that a card game or something?

No, nothing as complex as that. It is a boardgame. The rules of it are quite simple. Mastering the game is difficult. Hence why i like it a lot. I used it as an example because no AI has yet been created that can defeat a good human player. I am not a good player (i don't play enough) but even the it takes my computer half a minute of thinking before it makes it's move. Half a minute is long for a computer.

More information on the game is here.

 

[Spricar]

No, nothing as complex as that. It is a boardgame. The rules of it are quite simple. Mastering the game is difficult. Hence why i like it a lot. I used it as an example because no AI has yet been created that can defeat a good human player. I am not a good player (i don't play enough) but even the it takes my computer half a minute of thinking before it makes it's move. Half a minute is long for a computer.

More information on the game is here.

Go is a great game. Still, I believe the main reason why AI can't beat human player in Go is the fact that far less resources were invested in developement of programmes, libraries, procedures for Go than for chess, simply because chess is more popular game and not because the logic and thinking behind Go can't be mastered by AI and/or because it is fundamentally different than that behind of chess or any other board game.

 

[unmerged(42223)]

Without wanting to rant or point fingers, I would like to make a goodly suggestion following a some general observations. Some I am restating, but then so are 'misunderstandings' being restated.

If one AI country can do a thing, then all can. It not then an issue of hard-coding or 'dumb AI' it is an issue of scenario parameters. Much crap falls into this category.

Many exposed parameters that affect AI functioning in game are impacting or 'inhibiting' of one another. I see this in mod section occuring all the time. There are many I do not use for this reason alone. And it seems that 'reasoning about it' is less than useless. It takes much experimentation to track it down and prove it. Much, much so-call lousy AI behavior falls into this category.

Some things are a hard-coded over-site or under-site. However this category is less than it seems is generally believed. But it looks full and much ends being blamed here.

Suggestion:

Post gold after release, I humbly recommend a Paradox sanctioned AI beta test, where some of the AI people at Paradox are available. And some thick skinned moderator or two supervise the rest of modders who care to participate in it.