Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Parsely]

*looks at title*

Something something SHODAN.

[Sergarr]

Those game mechanics sound broken, and should be fixed.  The AI isn't broken for not exploiting the mechanics.  I'm slightly disappointed the AI doesn't detect those optimal strategies dynamically and use them, but only because I have a high opinion of GalCiv AI from the OS/2 days.

Finding optimal strategies is a difficult task for an A.I. to do. Mainly because there's no perfect criteria for optimality of the strategy. In a well-designed game, there cannot be an obviously optimal strategy, and non-obvious optimal strategies are usually a byproduct of exploiting game mechanics, which modern A.I.-s aren't trained to do. You can, but it'll lead to AI-s using bugs and exploits, and that's a kind of thing you want to avoid.

[E. Albright]

I think that this is the definition of good AI. 

[...]

I have not listened to this FEAR dev interview thing, so maybe I am misunderstanding you. 

I'd agree with you that what they describe is very AI-y in a technical sense even if it's "smoke and mirrors" by (rosy-eyed, naive) popular understandings of AI. The dev's paper on it was an interesting read: http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf

[Oneir]

I think that this is the definition of good AI. 

[...]

I have not listened to this FEAR dev interview thing, so maybe I am misunderstanding you. 

I'd agree with you that what they describe is very AI-y in a technical sense even if it's "smoke and mirrors" by (rosy-eyed, naive) popular understandings of AI. The dev's paper on it was an interesting read: http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf

You had a non-printing character that was messing with the link. I nommed it.

[Rolan7]

Speaking of AIs which play optimally with no regard for intended gameplay, any NES game wins due to the Playfun algorithm!
Sadly it only sees a few seconds ahead, but it's neat to see it find glitches:
https://www.youtube.com/watch?v=xOCurBYI_gY&feature=player_detailpage#t=900

(I skipped the long initial discussion and its development process on Mario 1, but they're good too).

[Retropunch]

I had this conversation with a good friend of mine who is a programmer at a well known computer game company specialising in FPS games.

He said that the problem wasn't making AI good, it was that when it becomes too good the player just thinks it's cheating. They experimented with 'amazing' AI and had to purposefully let the player win. This isn't about the computer having amazing aiming - lets say you've got 6 bad guys coming up on you (a standard action game scenario), if they all took flanking positions, shot at you from cover, and oppressively moved forward - you'd really, really struggle even if they were being a bit stormtroopery

To counteract this, you have to make the AI basically stupid or the player really strong - however if the player is really strong, they found that s/he just didn't realise what the AI was doing (or that it was impressive) as they could just dive into them and blast them to bits. Even if they really paid attention, the fact that they weren't really suffering made the AI seem pathetic.

I do think that 4x AI has no excuse though - it just seems lazy programming a lot of the time.

 

[Shadowlord]

Finding optimal strategies is a difficult task for an A.I. to do. Mainly because there's no perfect criteria for optimality of the strategy. In a well-designed game, there cannot be an obviously optimal strategy, and non-obvious optimal strategies are usually a byproduct of exploiting game mechanics, which modern A.I.-s aren't trained to do. You can, but it'll lead to AI-s using bugs and exploits, and that's a kind of thing you want to avoid.

On the other hand, if your AI could find bugs and exploits for you, that sounds like the kind of thing that would save time that would otherwise be spent trying to convince human players to find them and actually tell you about them rather than exploit them (Unless, of course, you want to leave them in).

[BurnedToast]

Those game mechanics sound broken, and should be fixed.  The AI isn't broken for not exploiting the mechanics.  I'm slightly disappointed the AI doesn't detect those optimal strategies dynamically and use them, but only because I have a high opinion of GalCiv AI from the OS/2 days.

You're right about the mechanics, but the stardock devs are either incompetent or too arrogant to admit they are wrong.

The "all X" strategy was the optimal strategy since the beginning, and countless patches + 3 expansions later nothing changed (they did nerf it very slightly by making it slightly less efficient, but not enough to matter), and the AI was never changed to take advantage of it.

They actually did change combat around a little, it used to be (iirc) the attacker always fired first... which made "all weapons" ships even more powerful. All they'd have to do is make it random who wins a tie (or just let both ships die) which seems like a trivial change... but nope.

As for it being an exploit, I disagree. You're following all the rules of the game to their logical conclusion. It would be an exploit if, for example, level 2 factories were supposed to make 10 hammers and they made 1000 due to a bug... so you purposely built a ton of level 2 factories and never upgraded them. But I guess, what's an exploit and what is not is a very grey area.

I had this conversation with a good friend of mine who is a programmer at a well known computer game company specialising in FPS games.

He said that the problem wasn't making AI good, it was that when it becomes too good the player just thinks it's cheating. They experimented with 'amazing' AI and had to purposefully let the player win. This isn't about the computer having amazing aiming - lets say you've got 6 bad guys coming up on you (a standard action game scenario), if they all took flanking positions, shot at you from cover, and oppressively moved forward - you'd really, really struggle even if they were being a bit stormtroopery

To counteract this, you have to make the AI basically stupid or the player really strong - however if the player is really strong, they found that s/he just didn't realise what the AI was doing (or that it was impressive) as they could just dive into them and blast them to bits. Even if they really paid attention, the fact that they weren't really suffering made the AI seem pathetic.

I do think that 4x AI has no excuse though - it just seems lazy programming a lot of the time.

Your friend is absolutely 100% correct for FPS games, or any other game that requires heavy "twitch" skills. The computer can react 100 times before you even realize there's a situation to react to, and it never makes mistakes or misses, and has no problem coordinating dozens or hundreds of units all doing different tasks. That easily offsets any disadvantage it might have for being unable to innovate or adapt to the situation. FPS games either let the player "cheat" (AI dies in 1 - 2 bullets, player soaks up 100's of bullets with regenerating health) or the AI does literally just let him win sometimes.

However, strategy games are a completely different thing. Computers are very bad at that sort of thing, and humans are very good at it. Giving the human a chance to use his big, impressive, creative (but slow) brain dooms the AI, unless they make the game so horribly complicated the human can't (or won't) properly manage things (and most people wouldn't enjoy a game like that). Even the current state of AI which is, honestly, usually pretty pathetic without cheats is not easy to make and to create an AI that's actually "good" at a strategy game without cheating would pretty much require you to create a strong AI.

[Shadowlord]

I remember Wardell's constant boasting about the AI, too.

[pisskop]

o.o

[Rift]

The AI in paradox's grand strategy games is actually pretty impressive.
In particular the CK2 one, which manages to somehow run cohesively well despite insane amounts of ai-actors is really technically impressive.
Whats more, they always work on their ai in their titles as long as they are still working on the title itself.
While some of it is a bit -illusion tricky- [like various event-driven stuff], its executed very well so that unless you are outright looking for it, or play a insane amount the world and its actors feel very alive. The Ai's execute [mostly] fairly sound strategies militarily[and politically] speaking, rarely do insanely stupid things, while still maintaining a degree of unpredictability.
I would say love/hate relationships between actors and non-war-related politics are severely lacking[pretty much just grouping up to assassinate you, or i'm going to rebel unless you do what i say and here's these friends who all agree with me.], hopefully the next dlc[which is on relationships] will help that a  bit.

[snelg]

Love the stalker ai. Agree that the ai in fear was good as well, they somehow managed to make the enemies really entertaining and interesting even without a large variety. Also had a ton of fun with the creature ai in black and white.

As much as I like paradox's games I'm not sure what to think of their ai. The one in crusader kings is entertaining but also pretty poor att the actual game and kind of predictable. It is impressive though how they managed to have so many characters at the same time.
EU4 is pretty good I think (though I haven't played the game as much as the others). Not sure if it's just because the player really got taken down a notch with the Mana Point/Monarch Point addition (one country WILL have less than four) or if the ai is actually good. I do like the changes to relations from EU3 and that they're better at grouping up on someone than before. Which is something I think a lot of their games could benefit from.
Victoria 2 had a very passive ai which did okay-ish but would miss things like colonization of Africa if the player prioritized it and took it all.
The ai in Hearts of Iron 3 was absolutely horrendous at launch. For instance the French ai would miss to garrison the German or Italian border and the like. It was salvaged somehow and got decent in later expansions but the player can still walk all over it even on the harder difficulties.

[Sergarr]

The AI in EU4 is not actually good, they just sorta bolted in a few winning strats for war, like gathering ALL THE TROOPS into one DOOM stack.

Also unlimited range and no attrition on naval units. Yes, the AI in EU4 cheats. Badly.

[Leonon]

Here's another one that's a custom AI playing and exploiting a game.

http://lbrandy.com/blog/2010/11/using-genetic-algorithms-to-find-starcraft-2-build-orders/

[Aichuk]

I wish more people would pay attention to A.I instead of Graphics. I mean, Graphics are awesome, but A.I is the future and Fear's A.I holds up better than its graphics..