Slashdot Mirror
The State of AI In Games
CowboyRobot writes "Carnegie Mellon researcher Alexander Nareyek has an article at ACM Queue describing how the role of graphics as the leading technology in gaming is being replaced by advancements in artificial intelligence. The author targets some game AI: 'Early milestone demonstrations for the publisher, press presentations, and other hype-generating events do not promote inclusion of a globally/consistently good AI, but instead focus on one or two 'absolutely unexpected but genius outcomes of revolutionary new and complex AI procedures' (did you spot the ironic tone?) that provide the necessary 'wows'.' He concludes by suggesting: 'we need to see more effort to make AI functionality available for the designers/artists.'"
The biggest advance in artificial intelligence so far came when the reaper bot started saying things like.
"I destroy you"
"Cry to your mom. b!cyh"
etc. I mean that was most humanlike characteristic ever to be properly emulated.
was reading about the AI that was used for LoTR battle scenes. this would be a helpful tool to get a mm online game going. the startup of one of those huge massively multiplayer games makes it difficult for companys to get a foothold. the ai of the mm game worlds inhabitants has been very poor imo. much worse than a non mm game. I have yet to find AI in any game that did not just make me laugh at some point in time. halo was not half bad but still way off sometimes.
this is not a Sig.
David Wong's main point was that the video game industry thrives on novelty, and that said novelty doesn't exist anymore now that we have fairly realistic real-time 3D capabilities. He forgot about the fact that graphics aren't the only novelty that gets gamers' attention. AI is arguably the weakest obvious element of gaming today, so near-future advancements in it would probably be as impressive as advancements in graphics were decades ago.
Rob
Having read the article, I think that the author seems to forget that though we know have pretty much reached the summit of the quality of graphics, AI is not the only direction that future games will have to be aiming for; I believe physics engines are taking a pogressingly important role in video games, and this is an area which I think developers will concentrate on, as it is showing more and more importance for gamers.
The problem is that now, AI is as "basic" as physics engines are. If you look at the best selling games, they probably have bad AI (XIII comes to mind, that game's AI is hideous) -- and an average physics engine. If you ask me, developers will have to find the balance between developing both, because these elements are becoming big enjoyment factors in player experience, as the players realise that good AI makes for a more realistic experience, and allows the player to progress in several ways instead of having to follow the same, linear route (a feeling which decreases replayability a lot). As for the physics engines, it's just something else which more and more players see as important; just imagine the possibility of great puzzles with a great physics engine.
A game with both great AI and a realistic physics engine ([cough]Half-Life 2[/cough]) just needs some good level design for these to come through, and that is how you will create the ultimate player experience.
The author of the article says the AI is more than just how the NPCs react, but nowadays I think the emphasis has to be put on this type of AI anyway, as it is still kind of apalling in most games. If a game has good AI, it can allow more freedom of choice to the player as the NPCs will know how to react accordingly to the player's actions and decisions. If NPCs are ready to act in "real time" to any situation, it is less a case of the player having to adapt to a game, following a linear route and ending up against a horde of enemies and some prescripted event, but more a case of the game "adapting" to what the player chooses to do, hence giving the player greater freedom and a better experience.
As for the evolution of physics engines, one has to see how important this is becoming; there are so much more progression possibilities for a game with a good physics engine. This allows the player to interact with his/her environment in a much richer way, and use it much more as he/she would like to have it exploited. Coupled with a good AI, this can make for an amazingly realistic game and a much better playing experience, and in my opinion, these are the departments which developers have to aim expanding on.
"Really, I'm not out to destroy Microsoft. That will just be a completely unintentional side effect" -- Linus Torval
Did the parent read the article? I don't understand why you'd be puzzled. The author specifically states that game AI is different from research AI and almost all games use the same FSM and decision trees. Only games such as Creatures or Black and White use "advanced" research AI techniques like genetic algorithms or neural networks. FSMs and decisions tree operate at a basic level:
If in this state with these conditions move to state Q
otherwise move to state R
So, half-life and deux-ex II both use the same basic FSMs and decision trees, and "advancement" doesn't come simply from increasing CPU power or more RAM (those things obviously help though) but also the programmers/designers ability to conceive of possible scenarios that could occur. So, if you want an enemy to flank you and kill you then someone has to think ahead of time "Well gee, what state would this bot need to be in before he would go ahead and try to find a way to get behind the player to flank him?" Then, the programmer has to find a way to get the bot to flank you (i.e. find a path to get behind you without you seeing) and then kill you. That's not exactly a trivial task, especially with so many possible scenarios involved.
David Wong does say that graphics advancements have leveled off in recent years, but he's not making this point at the expense of ignoring advancements in game AI or physics engines. His point is that pitching the same old rehashes of rigid game genres with some incremental technological advancement to give it novelty value is leading the industry nowhere. Besides, whereas graphics are (by definition) a fundamental element of video games, AI (especially the pathfinding sort mentioned in this article) and physics play only marginal roles in just a few existing popular genres, and are by no means present in all video games. How is better AI going to matter in multiplayer games where there's more than enough REAL AI (human players) available? Or how is a new physics engine going to significantly alter the way we play, say, strategy games that don't depend on realistic physics for their essential gameplay? Looking to AI or physics as the technology that will solve the video game industry's many problems isn't the answer.
Memetic AI for Neverwinter Nights. This stuff is pretty innovative, at least in theory, and I hope they eventually get to the 1.0 phase with more than just the groundwork done. The approach it takes might be an interesting, even intuitive, way for building the NPC/creature behavior from ground up.
There are games where the AI is very good and the computer plays better than the average or even above average player. Chess for instance.
But would it make your game sell better? I bet most gamers won't like that when the harsh reality of a uber bot keeps grinding their noses into the dust.
Given enough effort a very good programmer can easily make a very good bot that beats everyone in Quake 3 or something similar. In many modern games (e.g. FPS shooters) most gamers will NOT win against a well designed bot. A great quake bot can camp near perfectly, know exactly where you will respawn, monopolize items extremely well e.g. it'll always know when the megahealth is going to respawn if it took it, and if the opponent took it, it can still make a very very good guess too - based on it's estimates of opponent health/armour and damage so far). It'll have perfect hearing - based on sound loudness and direction, and it'll have near perfect aim (the accuracy limit depends on ping). If such a bot has a railgun and you somehow sneak up on it undetected, shoot and don't kill it, you'd most likely get shot in the next frame. You're unlikely to kill it with the first shot, coz it would be hogging the mega and armour - but since you always get killed in one shot, you never get to hog the mega or armour and so you always get killed in one rail shot. Your only hope is a map where it can't hog both, but then it means you can't hog both either, which means you just lose less badly, or a map where aim doesn't count at all (doh).
You can talk about Go, chess and all that, but most popular computer games don't require much brains, coz most gamers don't want to use their brains that much, or don't have much brains in the first place.
In short: most popular computer games aren't really brain games, so AIs aren't at a big disadvantage.
Thus it won't surprise me that someone with brains can construct an AI to beat them in games designed for such people.
But what's the point? If you really wanted to play an AI with semi-decent tactics just play against one of those cheaters with an aimbot. Some of the better cheaters might have slightly better tactics and intelligence than the average AI bot (most cheaters aren't that smart tho in human terms). If you're a good enough player with a good enough connection (low latency) you can usually beat em, and then they whine about as imaginatively as a bot.
I disagree with some of the author's observations about the potential for improved AI in the gaming industry.
1) The author neglects the fact that multiplayer gaming capabilities in games have reduced the need for developers to include well-designed single-player AI. For example, in the game Battlefield 1942 most people will play multiplayer over single-player since the single-player bots don't have enough AI to work in teams or operate vehicles effectively.
Since a majority of the people won't play singleplayer, there is no incentive for the developer to improve the single-player AI.
2) Lack of thorough analysis of AI techniques used in recent games. He covers some techniques from academia, but he should have interviewed more developers and which AI techniques they employed, and the trade-offs involved.
3) A few errors in the article, like referring to a Counterstrike Bot as being from Microsoft.
Later this month Valve will be rolling out an official CounterStrike bot in their "new" game CS:Condition Zero (which has a controversial development history). Any CS bots before this one were developed in the mod community.
I'm sure the author could have interviewed the Condition Zero programmers and ask them what AI techniques they had to utilize for this new bot.
A key point I do agree with is that there is not a common AI standard that game companies are using, compared to the video and audio standards.
No matter how fast the CPUs become, gaming AI will not improve in general until developers can easily generate core AI behavior with little effort.
There's different types of AI depending on the genre.
There's strategy games. Just because it's a game with very little chance (i.e chess) doesn't mean that writing the AI is easy. Though the Civ3 AI is really good, I don't consider it by any means infallible. Perhaps the best example for AI in strategy is Magic: The Gathering. Years ago, they came out with a PC version with AI. The problem was that the AI could never understand the strategies behind different deck and thus couldn't play them effectively. I think designing a good MTG AI would be much tougher than a good GO AI.
There's also reaction-based games (FPS, 3rd person, etc). In this case, the goal is not for the AI to be better than the player, but to emulate the behavior of a human player. UT had much better AI than Quake, simply because the bots weren't omniscient.
Lastly, there's RTS AI, which is a mixture between the two mentioned above. Like the MTG problem, RTS AI generally uses the same generic strategy regardless of opponent, but has lightning reflexes.
Depending on the game genre, different approaches to AI are necessary. In strategy games, simply winning is a good objective. In reactionary games, since computer reaction time is better than human reaction time, creating a good AI player is almost trivial, but creating a realistic and fun opponent is a challenge.
-- Political fascism requires a Fuhrer.
If AI will become the next generations "killer app" (for lack of a better term) will we see things like AI cards? I know that most AI is processed by the CPU, but if the AI is too advanced might a separate add on card work well? Also, will we see some sort of basic AI programming environment. Something like DirectAI or OpenAI? I hope they start working on AI more than graphics.
Morrowind was a great game, but it wouldn't work as a MMORPG. However, with adanced AI it would be like entering the world of Mournhold.
As far as current games pushing AI forward; it seems like Fable from Molyneaux is the leader (we hope).
This encouraged other NPC's to find alternate routes to the player, if available, without having to put in a bunch of hints and such. They'd just "naturally" try to flank the player if they could.
Sorry, can't remember the name of the book. "AI Game Programming" was in the title, though.
PHEM - party like it's 1997-2003!