I found myself wondering Jeff Orkin's muse about "the future of AI in an increasingly multiplayer-dominated world." He asks whether the AI in the virtual world will be limited to roles for which no player would want to engage. This is a religious topic in the games community, easy to miss. But then along the way comes a report of a "massacre" of Chinese players by Korean players in Lineage and old questions surface: better to trust the world to AI (and game design) or the players?
Jeff's post "AI, Inner-Life, and the Multiplayer World" extends the discussion from an earlier Gamasutra article by Bruce Blumberg ("Anticipatory AI and Compelling Characters" ). Both are notable reads. A theme that emerges in both discussions is the distinction between an intelligent act and its signal. In other words, behaving intelligently is not always sufficient to communicate that intelligence - the need to communicate and divine intent is critical. As Blumberg wrote:
The absence of anticipatory behaviors that predict significant changes in motivational state (and hence behavior) is a consistent weakness of many AI models of behavior. As a result, changes from one motivational state to the next often appear startling and ultimately inexplicable. One source of the problem is that by modeling motivational contexts as finite states, such a system only knows how to display that state when the system is in that particular state.
Analogy with the lessons from Kismet ([1.]) and those involved in social robotics might be possible. The ability to express intent and social cues via gaze, for example, can go a very long way in communicating something more than just what it does. By contrast, is Deep Blue intelligent or just smart?
In a way, Dmitri's work ( Cultivation Hypothesis) along with Ted's theory of players aping the virtual world AI (ref) could suggest a pitfall of "AI effects" in virtual worlds. A contrasting view is to leave it up to the players to do more, and leave the AI to running NPC rabbits and shopkeepers.
But if people-quality intentions are important, consider their pitfalls...
Recently billsdue ran a post (Financial Times source) of how South Korean players in Lineage are "are ganging up to obliterate the Chinese, whom they view as greedy and rude." Stories like this are hard to parse from hype (the other media effect), but they do point to a deeper conundrum:
How much content should a player be allowed to control?
The sad news is: There is no AI.
What is currently termed AI is a patched set of techniques akin to very limited scripting.
An AI that would simulate a human is decades away. (at least)
The same goes for robots that could be mistaken for human beings.
So, unless you need to simulate very simple behaviors, you better find a way to interest humans in the different roles your game mechanics requires.
Posted by: Darth Pixel | Mar 12, 2006 at 13:03
Ack! FT report sounds pretty awful, actually. Lineage is outside my direct experience, but I guess it has mostly either Chinese or Korean players. I have mostly played games that have players from a lot of different countries, so the idea of everybody ganging up to kill, say, Turkish players would be absurd.
Everyone knows that anyone could be hiding behind a particular avatar, but they still pre-judge them based on their appearance. Stereotyping must be even easier in a game environment of "us" and "them."
Posted by: CherryBomb | Mar 12, 2006 at 17:36
Many forms of competition (warfare and poker for extreme examples) don't have any communication at all, but you must assume your opponents are intelligent.
Games I think are in a tough situation. The computer can't do anything that the average player can't explain. How smart is the average player? Computer chess now exceeds most humans -- the computer can make moves that the average player can't understand. If it weren't so easy to identify illegal moves, I'd bet human players would complain the computer cheats. Games have to boost a player's ego by being smart enough to be a challenge, but not so smart that the player has a hard time winning.
I think it's funny that now that we understand chess, we say a chess playing machine is not intelligent. Humans will be saying machines aren't intelligent right up to the point when the machine smacks us over the head and disagrees.
Posted by: Ken Fox | Mar 12, 2006 at 18:43
The Internet is moving towards user-generated content; why not Internet gaming?
Posted by: Mike | Mar 12, 2006 at 22:33
Nate, I've been trying to figure out the connection between the Lineage issue you raise and AI. The former seems to be an extreme and perhaps methodical response to the farming/RMT/ninjalooting constellation of issues we've debated here many times. Apparently some Lineage players think the debate is over (and in a game as PvP-heavy as that, is it really a surprise?).
As far as AI goes, I think the articles you link to raise some good issues -- e.g., the presentation/representation issue for example, often comes down to social cueing. AI thus far has been great for pathfinding and such, and not so good on the social/emotional side of things. As this changes, I believe we'll see new kinds of gameplay emerge and new dynamics in online worlds. I don't think AI in online worlds will be just about better strafing or navigating, nor just about making better rabbits and vendors. What it will become is something I can't really talk about right now. :)
Posted by: Mike Sellers | Mar 13, 2006 at 07:59
I think we're running into the same old problem here of the difference between "play" and "game."
It reminds me of when I was in college studying to get a degree in writing. One of my best professors used to quip, "The problem is, becausse everybody can write... everybody thinks they can write."
Not all *play* is a game, and not all *gaming* is play. Some play has no rules (and, therefore, would be most hard to model), and some games are "played" with about as much fun as people approach their day jobs. It's an important distinction, because game designers need to dillineate between those aspects of the game that need to be "rule based" from a "we need rules in order for things to make sense," and those aspects that can be more free-form; i.e., fun without rules. One code of good GMing from my past in live RPGs, for example, was "don't forbid anything without good cause." If you start having "game rules" without taking into account that they will have an impact on "fun," you'll have disappointed players.
When you say "there is no 'artificial intelligence' from a *gaming* perspective, Darth, I have to disagree. How much AI do you need to play, for example, BlackJack? Or Cribbage? Or, as has been pointed out, chess? One of the "grandest" strategy games of all time? If the computer can mimic the behavior of an intelligent human player... that's artifical intelligence, within the context of that game.
When we get into the social aspects of gaming -- which can occur in any game, but is explicit in live RPGs -- modeling "intelligent" behavior becomes, of course, much more complex. Fighting with swords is just the beginning. Shopkeeping was brought up as one of the acts that many folks might not want to roleplay. I'll bet that in a "no NPC world" you'd actually get plenty of people who'd like to play the innkeeper/trader. It takes all types in the real world, the same, I've seen, is often true in virtual spaces.
So the real question becomes, how to make it the *most fun* for the most players, right? If the AI is good enough to do that... rock on. If players can take over roles that used to be relegated to AI... more power.
If gangs of one real-world type of player are smiting other groups... how is that an AI issue? I'm confused...
Posted by: Andy Havens | Mar 13, 2006 at 18:36
"how is that an AI issue?">
Person or virtual machine, who do you trust?
The reported incident felt a good catalyst - but I had hoped to avoid making too explicit a link with PvP (Player-versus-Player, as a form of player created content), knowing how those discussions can go.
However, let me put it this way. If the most direct expression of player created content is PvP combat, then what does it say about the creators of such content if it can lead to such outcomes such as is reported?
Twist the perspective, would we be all better off if it were PvE (Player versus Environment) and just everyone *thought* it was PvP?
I think it also interesting to wonder more subtle variations (vs. the direct one): what was wrong with the environment (its minions and AI) so that it may lead to such a fester? Was there some adaptation that were possible in the environment?
disclosure: I'm a fan of a number of PvP games, not picking on 'em, just wondering a deeper imbalance...
Posted by: Nate | Mar 13, 2006 at 19:00
"Or, as has been pointed out, chess? One of the "grandest" strategy games of all time? If the computer can mimic the behavior of an intelligent human player... that's artifical intelligence, within the context of that game."
Just to quibble I'd argue that, from what I understand, chess programs support Darth Pixel's argument rather than refute it. Computers are "good" at chess because they can crunch millions and upon millions of moves while mechanically churning through game algorithms. Humans are good at chess not because they can crunch millions of moves but because they are capable of insight and creativity. Humans going up computers in a game of chess are a test of one methodology versus another.
Incidentally, the ability to chew over millions of calculations in a second will only take you so far. Consider that no one has figured out how to produce a decent AI opponent for the game of go.
Posted by: | Mar 14, 2006 at 01:56
Go is a very tough game for computers, but to say that computers aren't decent because they can't play "strong amateur" level yet seems short sighted. I wonder what percentage of humans can beat a computer? Are the humans who can't unintelligent?
It strikes me that most on-line games have poor AI not because that's the current state of the art, but because games choose to have poor AI.
Nobody wants to walk into a shop and get his ass handed to him by a shrewd NPC shop keeper. Nobody wants to fight a smart Onyxia that always breathes flame on the maximum number of people.
Posted by: Ken Fox | Mar 14, 2006 at 09:34
Nate wrote: would we be all better off if it were PvE (Player versus Environment) and just everyone *thought* it was PvP?
In a March 11, 2006 PvP Debate at MMORPG.com between Garrett Fuller and Frank Mignone, Fuller started by saying: "I love beating other players knowing that they are angry on the other side of their PC."
If Fuller's motivation for engaging in PvP play is typical (and I suspect that it is), just how far would NPC AI have to go before it would be an acceptable substitute for human opponents?
Even if we could reach that point, how would it satisfy those players who want to take out their RL grievances on other players in the virtual world? Or should virtual worlds not be expected to satisfy such desires?
Are expectations for NPC AI higher for PvP gameplay than for other forms of player interaction with NPCs? If so, why? Should NPCs be believable emotional actors?
Bearing in mind the Eliza effect, how far should that go? Would it really be wise to allow NPCs with superior AI to present themselves as human players? Is there any serious danger of widespread "you're not who I thought you were" feelings that one's trust has been betrayed?
Should products with "nearly-human" AI be required to bear big warning labels? ("CAUTION: Some players in this game are not human") Or is it OK for developers to play with gamers, to tease them into wondering whether any character they encounter could be a bot?
"Turing Test: The MMORPG"?
--Bart
Posted by: Bart Stewart | Mar 14, 2006 at 12:03
"The Internet is moving towards user-generated content; why not Internet gaming?"
Very good point indeed, - why not? - is The Web a big reading hall or is it caleidoscope of images, amateur programming, fractals and dollhouse-like play?
AI twist in this:
"I think it's funny that now that we understand chess, we say a chess playing machine is not intelligent. Humans will be saying machines aren't intelligent right up to the point when the machine smacks us over the head and disagrees.
EXACTLY, - as technology evolves, we understand less and less how stuff works and become increasingly passive users of all this tech... and YET we are so snobbish when it comes to assessing "smartness" of machines
Play can be very passive and simple activity which does not require much of intellect of humans ;) Let us admit, average player does not do neither much number crunching nor much problem solving in the modern MMORPGs... So, modelling a machine (AI) at such run-about-poke-around level is not much of a challenge, is it?
& very soon players will be (are) manipulated by rather simple algorythms - just like we watch art or listen to the music...
then AI would be really as real and simple as a set of logic-oriented scripts tied into machine (fast) number crunching, algorythmic 3d generation, force-feedback peripherals manupulation - & no one needs faking emotions really to test if this is advanced enough...
Posted by: Alex | Mar 14, 2006 at 12:51
Normally, you don’t have interesting and engaging pvp games, you have interesting and engaging pvp moments. When those moments happen to arrange themselves (rarely) into some pleasurable and/or (so-called) dramatic form, then we call those beaded moments the greatest game ever played or the fight of the century. But your Saturday afternoon football/baseball/basketball whatever match is much more likely to be mundane than memorable, and pvp is much more often chaotic, gank-filled, frustrating, and unfun than it is terrifically wooty. Nevetheless, despite all the unfun, it’s that occasional focus of exhilaration, those relatively rare moments of intense pleasure, that lead us back to spin the wheel.
So, if humans are really naturally good at creating (usually through little conscious effort on their behalf) these popup moments of exhilaration, but really quite bad (usually through very serious, conscious, and unfortunately overly narrative/social efforts) at extending these moments into fights of the century, then what to do about it? Some would say lay the narratives down, asphalt in the plot lines, pass out the guild badges. But that really doesn‘t seem to work. Extending the pvp experience from moment to moment seems to me, necessarily, a bottom-up kind of thing. Good players can do it, sometimes -- with a good fight becoming analogous to a good dance. But the moons and whims have to be perfectly aligned.
So that’s the role of the pvp AI: not to be a good player alone, but to be good player that extends the moment.
Howard Cosell revolutionized sportscasting simply by creating a plot line -- by overlaying a context on top of a game that gave the players and events in that game some broader set of values and meanings. But that was most definitely a top-down (and usually obviously artifical) thing. In more participatory games, the AI needs to be Howard Cosell from the bottom-up. AI needs to take revenge. AI needs to taunt, to be nasty. But, most of all, AI needs to have a context, a set of values and meanings beyond the immediacy of the game.
Yul Brynner in Westworld. Rutger Hauer in Blade Runner.
Pvp AI shouldn’t be about story or plot or narrative or overlay. AI pvp should be about the same thing pvp players are about: themselves.
Posted by: dmyers | Mar 14, 2006 at 13:09
One of my pet peeves arises above: fun. What makes anyone think that MMOs should be "fun" or that "fun" is a desirable outcome?
Satisfaction, achievement, simply passing time - there are many other reasons to log on to your preferred MMO than a search for "fun". I see them as the virtual equivalent of "keepy-uppy" or batting a tennis ball against a wall.
The huge attraction of what passes currently for "AI" is its predictability. PVE Players do not appreciate AI that even approaches the unpredictability of a human player. PvP players strive for positions of superiority, where the unpredictability of opponents is as muted as possible.
MMOs could easily employ much more sophisticated and unpredictable AI than they choose to use. The brake on doing so is an appreciation of the commercial disadvantage this would give their games over games with more predictable, manageable, beatable AI.
There is no popular art or entertaintenment that I am aware of in which the more difficult, challenging examples have proved to be consistently more profitable, or to have achieved wider public acceptance, than simpler, more obvious forms. When it comes down to it, most of us want to log and win. We want dumber, slower human opponents or AI. Maybe we would ideally like them to be only very, very slightly dumber and slower than us, but we certainly don't want them to be even very slightly smarter and faster.
After all, where would the "fun" be in that?
Posted by: Bhagpuss | Mar 14, 2006 at 19:13
"Go is a very tough game for computers, but to say that computers aren't decent because they can't play "strong amateur" level yet seems short sighted. I wonder what percentage of humans can beat a computer? Are the humans who can't unintelligent?"
The point I'm trying to make is that humans and computers take very different routes to get from A to B. Again, a computer plays a decent game of chess by crunching millions of numbers and mechanistically churning through rote algorithms. That's not how human beings play chess.
Why does this approach work well? It's because chess is a fundamentally mathematical game that is well suited this kind of mathematical analysis. The problem with go is not that it is "unmathematical". The problem with go is that a decent AI requires crunching billions of numbers rather than mere millions and processing on that scale is beyond the capabilities of modern processors at the moment.
So there's one kind of limitation right there on the current approach to AI. But there are other types of limitations as well. How well does the computational/algorithmic approach work when you try to apply it to something like an NPC in an RPG? You wrote:
"Nobody wants to walk into a shop and get his ass handed to him by a shrewd NPC shop keeper."
How do you code that? A friend of mine was recently victimized by a real life shrewd shop keeper. Try sitting in on a real life transaction where both sides go back and forth, appealing to the other party's interests, using flattery and greed, feigning disinterest, etc. etc. etc. I can easily whip up some simulation code that says "If player's last offer (n) was x, issue counter offer of x/2 + y where y is the difference between offer(n-1) and the current offer(n) yada yada yada" but that is a very poor substitute for the real thing.
Posted by: lewy | Mar 14, 2006 at 21:24
"the real thing" lol Nobody knows. Computer intelligence may end up working differently from human. Turing was clever to design a black box test.
Shop keepers can act a lot smarter than they do. Furby is smarter and more enjoyable to interact with than shop keepers. Technology isn't the problem. Game designers build vending machines shaped like humans. Why? Because players want vending machines. The whole charade of selling drops and buying repairs from humanoid vending machines is so stupid that it must exist because players want it to work this way.
Posted by: Ken Fox | Mar 15, 2006 at 00:37
Because players want vending machines.
Considering that AI in MMOGs hasn't gone much beyond pathfinding and human-shaped vending machines, what evidence do we have that this is what players want? We've never given them an alternative, so we have no way of knowing that this is actually what they prefer. Single player games with interesting and engaging AI have sold well, The Sims being the prime example of this. Until we try putting better AI into MMOGs, we won't know how players will respond to it.
Posted by: Samantha LeCraft | Mar 15, 2006 at 01:16
"Shop keepers can act a lot smarter than they do. Furby is smarter and more enjoyable to interact with than shop keepers. Technology isn't the problem. Game designers build vending machines shaped like humans. Why? Because players want vending machines. The whole charade of selling drops and buying repairs from humanoid vending machines is so stupid that it must exist because players want it to work this way."
The point is that the term "AI" in games really isn't accurate. It's being abused. There is a vast gulf between what the PhD's at MIT are talking about when they refer to "AI" and what exists in 99% of the games out there. Game "AI" is all about faking it--it's a series of algorithms and routines that are there to give the illusion that you're playing against another player. In other words, game AI is about duplicating the evidance of a thinking process, not duplicating the process itself. That, I think, is what Darth Pixel was trying to point out and what I was agreeing with.
Getting that shop keeper to appreciate your jokes is a job for those PhD's at MIT, not for game designers. That's a very, very thorny problem indeed. After they figure it out it will trickle down into games.
Posted by: lewy | Mar 16, 2006 at 18:24
Getting that shop keeper to appreciate your jokes is a job for those PhD's at MIT, not for game designers.
I disagree. Well, mostly. You latched onto humor, which seems to be among the thorniest of AI (better, artificial psychology) issues -- part neurology, part cognition, part culture. But games have pushed the boundaries in graphics, physics, and user interfaces among other areas. Why should we wait for someone else to solve problems related to AI? Who better than game designers to make NPCs more engaging, immersive, and socially plausible?
Most of the examples above have been about direct confrontation, zero-sum situations, something that classical game AI has focused on forever and which hardly leads to more convincing, immersive AI. Consider instead how it might change and expand gameplay if NPCs had emotions, remembered you, and had relationships with you and with each other.
This brings up all sorts of unexplored possibilities, both positive and negative. Suppose for example you're interacting in a virtual world and re-create the Milgram experiment (not that I'd recommend this for a game!) with socially and emotionally responsive AI-driven NPCs, but with a human in the role of the participant. If no one gets hurt, but AIs neverthless show all the signs of pain and distress, is this ethical or not? If AIs can begin to pass a sort of emotional Turing test, where do the ethical boundaries lie?
'O brave new world that hath such creatures in it...'
Posted by: Mike Sellers | Mar 16, 2006 at 20:24
"I disagree. Well, mostly. You latched onto humor, which seems to be among the thorniest of AI (better, artificial psychology) issues -- part neurology, part cognition, part culture. But games have pushed the boundaries in graphics, physics, and user interfaces among other areas. Why should we wait for someone else to solve problems related to AI? Who better than game designers to make NPCs more engaging, immersive, and socially plausible?"
I think the biggest barrier to the game industry pioneering AI research are the time and resource requirements. Inevitably it seems to me that AI researchers who are trying to duplicate consciousness have to tackle the issue of what consciousness is in the first place. Trying to build a machine that understands language, for instance, leads to the question of how human beings understand language. And that is not a question that's well understood by anybody as far as I can tell, whether it's anthropologists, neurologists, computer scientists, etc. If the shortest path to producing artificial intelligence is to duplicate human intelligence then the barrier becomes the enormous amounts of basic research that needs to be done in the medical/neurological realm, research which is really just getting off the ground. I really can't see private companies (game designers) expending the resources necessary, given that those resources could be huge and the payback is uncertain.
Posted by: lewy | Mar 17, 2006 at 04:04
The fact that game AI is all about faking it, does not lessen what I see as the point of the article. What I was trying to convey was that by recognizing the central role of anticipation in our perception of human and animal behavior we perhaps can add machinery that provides those cues. In some cases it may be as simple as adding a "prepare to transition from this state to the next state" call in an FSM that has the effect of providing a cue to the player as to what is to happen next. In some cases, it may be easy to do, in other cases hard. But until we recognize the central role of anticipation in our perception and understanding of behavior we are not going to do the stuff that is like falling off a log. Indeed, my intuition is doing the easy stuff will get you 80% there.
Posted by: bruce | Mar 17, 2006 at 08:07
I think you're largely right, Bruce, that there are a bunch of fairly easy-to-implement aspects of apparent intelligence (such as visible anticipatory micro-behaviors) that lead us away from, or hopefully over, the Uncanny Valley. I think there's also a whole lot of deep and unexplored territory as lewy says, though I don't think this is necessarily the province of academics over game developers. Nor is it necessary to have a complete understanding of consciousness or language generation to do this. We learn by doing.
I don't know that game AI is all about faking it though, or that it will always be that way. Is game physics all about faking it? It was once. Now I'm not so sure; there the line between simulation and simulacra is blurred beyond division.
I think we're nearing this with artificial characters. Not that we're yet approaching "more human than human" (Tyrell, apparently paraphrasing Baudrillard), but I think we're quickly going to be beyond the easily dismissed "faking it" territory.
On a different note, I wonder whether anticipation is all that central to perception and understanding. It's an important component, but just one of several (e.g., attention, recognition, recall, emotion, cognition, decision) that goes into behavioral creation -- unless you just fake it with a FSM. :)
Posted by: Mike Sellers | Mar 17, 2006 at 11:09
The following points aren't in any way novel; they've been known for years:
* simple (hidden) rules can generate complex behaviors [Conway's Life, Braitenberg's Vehicles]
* even minimal appearance of behavioral complexity can be perceived as evidence of intentionality [Weizenbaum's Eliza]
If producing plausibly intelligent NPC behavior is this simple, why do NPCs seem so dumb at everything but combat (and not terribly bright at that)?
I can only imagine a few possibilities:
1. I'm wrong, and plausible behavior isn't simple to achieve. Some hard technical breakthrough is required to make interactions with NPCs plausible.
2. Implementing plausibly intelligent NPCs actually is conceptually simple; developers just don't think there's enough value in smart NPCs to justify the time/cost.
3. Implementing smart NPCs is non-trivial but doable, but developers don't think it should be done because they don't believe that players really want smart NPCs. (Seems especially true if all players are assumed to be zero-sum, competitive, combat-oriented gamers.)
So does one of these three positions explain the state of NPC AI in MMOGs today? Or is there some other explanation?
--Bart
Posted by: Bart Stewart | Mar 17, 2006 at 12:47
I have to choose just one? :)
My take is that some plausible behavior (Eliza+) is pretty easy to achieve, but dig down just a little ways and you hit some hard problems. OTOH, a lot of those hard problems are still manageable with the right architecture, just difficult. At the same time, most game developers approach this area with an overall lack of psychological understanding or much intuition as to what might be important to maximize believability. So their early efforts don't pay off and they become convinced that more work isn't justified and that players (i.e., people like them) don't want this anyway. Instead we get beautifully rendered graphics of people that end up coming off as really creepy. If AI garnered as much focused interest as does graphics, consider the changes that we might see to parallel going from 64x64 sprites to normal-mapped real-time animated full 3D scenes.
In looking for counter-examples to the "AI isn't worth it" refrain I keep going back to The Sims (which is significant given the game was released more than five years ago). In some ways it's brain-dead AI (and having worked with the software, I think I can say this without disrespect). OTOH, it's way beyond facile systems like Eliza. In terms of "what players want" it has still sold more than any other PC game. And I still get the feeling that many game developers think this was somehow a fluke or faked or something. It's cognitive dissonance: they just can't accept that even minorly "smart" domestic NPCs would be attractive as a key part of gameplay. It's much easier to think about frame rates and shaders, or optimized A* and flanking moves.
This is starting to sound like sour grapes on my part, which isn't what I mean at all. I have great hopes for better, more engaging and believable AI in games. There are a number of interesting research centers for this in and around academia too; I don't believe that shuts out game developers though. No group has a better handle on how to explore some new area and present it in a way that is meaningful to people. Hopefully we'll start to see the fruits of this in believable AI in the not too distant future.
Posted by: Mike Sellers | Mar 17, 2006 at 13:36
Mike Sellers wrote:
In terms of "what players want" it has still sold more than any other PC game.
WoW will soon surpass the Sims as the top revenue generating game, and it has AI that makes the Sims' look intelligent. I'm not sure that using individual games as examples of "what players want" is very useful long-term.
(I'm not disagreeing or agreeing with your point on AI, incidentally.)
--matt
Posted by: Matt Mihaly | Mar 17, 2006 at 13:49
WoW will soon surpass the Sims as the top revenue generating game
But not in terms of units sold. WoW would have to about quadruple its current total sales for that to happen. This is still strongly indicative of "what people will pay for" and the potential payoff of more believable AI.
... and it has AI that makes the Sims' look intelligent.
I'm baffled. What are you thinking of? I enjoy WoW, but not for its deep AI.
Posted by: Mike Sellers | Mar 17, 2006 at 14:01
Sure, units sold, but that's like saying the games industry is bigger than the movie industry while ignoring DVD sales. No doubt some people will pay for improved AI.
I'm baffled. What are you thinking of? I enjoy WoW, but not for its deep AI.
Yeah...that's why WoW's AI makes the Sims look intelligent. WoW's AI is rock-dumb.
--matt
Posted by: Matt Mihaly | Mar 17, 2006 at 15:00
(Oh. D'oh. Must read before posting. Requires intelligence -- the non-artificial kind.)
Posted by: Mike Sellers | Mar 17, 2006 at 16:43
Some random thoughts about AI:
For the sake of this discussion, split AI into two groups: behavior and speech.
Behavior - In a combat-oriented world, more intelligent behavior AI means monsters being better at combat. Most players don't want this; they want to wade through 10,000 mindless orcs and slay the lot. They don't want the 10,000 mind-ful orcs to set up a trap and kill their character.
Of course, behavior also include emotions, gestures, etc. Barring "pet" NPCs, these behaviors tend to be tied into speech...
Speech - Getting speech AI to work has a few problems...
1) To "talk" to the AI, players would have to type in semi-valid sentences. Players are lazy, and prefer clicking on a menu, which mostly negates the need for speech AI since menus can only have a few options. If players could actually talk using speech recognition, things might be better. Is SR good enough? Maybe, but probably not.
2) The speech then needs to be parsed using NLP, which is doable, but yet another technology. Cheezy NLP works too.
3) The speech needs to be understood, which is trickier, although hacks exist.
4) A conversation model must exist, which is even trickier than the minimum NLP/understanding work necessary. Simple conversation models are possible, especially if players are going to be talking about buying stuff or asking about the weather.
5) NPCs need to generate sentences to "speak". This mostly a matter of sprintf ("I don't want to talk about %s now." or "%s is a thief.") A really powerful/flexible sentence-construction system would be better, but it's tricky too. (Just look at the problems that machine translation systems have.)
6) Players don't like to read text. Because of the "%s" issue above, only text-to-speech will work. Text-to-speech doesn't sound terribly great yet, and certainly isn't able to convey any emotional depth.
Combine all the above difficulties with the fact that the world is populated by intelligent players that make even the best AI look stupid, and most big-budget games would rather invest in more eye candy and larger worlds. Eye candy and large worlds are paying off at the moment, so why risk working on AI?
Posted by: Mike Rozak | Mar 17, 2006 at 22:33
Well Mike I disagree with your premeses and reasoning, but I can't really point out specifically how right now. Or maybe you're right. We'll see eventually.
As to your last question though: Eye candy and large worlds are paying off at the moment, so why risk working on AI?
Going up against the strengths of existing worlds like WoW seems like lunacy to me, especially when players have demonstrated a willingness to pay for what these worlds aren't providing (e.g. better, meaningful AI). Finding differentiators for new worlds is an important aspect of any product strategy.
Posted by: Mike Sellers | Mar 18, 2006 at 00:36
I agree with differentiation, and I'm working on AI myself as a differentiator, although probably not the same facets that you're emphasizing. I am particularly interested in conversations, but the first version(s) will inevitably be unimpressive.
However, I (cynically) think that most games won't touch AI with a 10' pole. (They may claim to have a new-fangled AI feature, but it won't be anything significant.)
Oblivion may change a few developer's minds if it works. However, I still haven't seen a MMORPG with NPCs even as good as Morrowind's NPCs, and that came out 3-4 years ago, plenty of time for at least one MMORPG to head that way. (If any are out there, I'd like to know. Seed seems to be into AI, but I haven't checked it out.)
To get even more cynical: Although I haven't played MUD I & II much, I suspect Richard Bartle's creations had more impressive AI than any/most contemporary MMORPGs.
Posted by: Mike Rozak | Mar 18, 2006 at 01:09
Mike Rozak>I suspect Richard Bartle's creations had more impressive AI than any/most contemporary MMORPGs.
The default behaviour for mobiles in MUD2 was basically an expert system. I had to dumb it down though, because it kept beating the players.
Some of the human mobiles (eg. the old man) have plans, which they will execute properly and will adjust on the fly if things change (ie. they're reactive). They don't use a full-blown planning system, though, in that they don't compose arbitrary sequences of operators to achieve their goals; they use more like canned plans. There's a route-planning system, but it uses way points so is fairly easy to exploit if you know where the way points are (although it does some optimisation, so it's possible the way point may be shaved off).
I expect we'll see more intelligent mobiles eventually, especially given the number of times I've had to point people at my Skotos articles when they tell me they have a brand new way of creating quests using AI.
Richard
Posted by: Richard Bartle | Mar 18, 2006 at 06:48
Matt> WoW will soon surpass the Sims as the top revenue generating game, and it has AI that makes the Sims' look intelligent. I'm not sure that using individual games as examples of "what players want" is very useful long-term. <
The two games are also using AI in very different contexts. In the Sims, you are in God Mode. No matter how smart the Sims get, you will still be vastly more powerful than them. I suspect if anything, a smarter Sim AI would make players feel more powerful and more significant. On the other hand, in a combat game, a smarter AI makes you feel more powerful if you consistently beat it. Once it consistently beats you, you are likely to walk away. For WoW to improve their AI, much of the research should go into teaching the Orcs when and how to lose gracefully.
Posted by: Hellinar | Mar 18, 2006 at 09:21
You know what I find most strange about this conversation are the related ideas that (1) developers don't learn much from other developers, and (2) developers don't know what consumers want.
Does the game industry work by throwing mud at the wall to see what sticks?
Man, if that's true, I feel really sorry for any of you in academia trying to influence game development. I had no idea.
Posted by: Ken Fox | Mar 18, 2006 at 10:04
This is a very interesting dialogue. MMOs are something of a special case when it comes to their use of AI because, as has been mentioned, their designs (historically) have had a set of requirements that in many ways precludes the use of more advanced AI techniques (as Mike Rozak correctly states, "Most players don't want this; they want to wade through 10,000 mindless orcs and slay the lot").
Whether it's possible to design an entirely different kind of MMO that makes much better use of AI is something of an open question, and one I won't try to answer.
However, I wanted to drop in to clear up a couple of misconceptions:
> [Darth Pixel]
> The sad news is: There is no AI.
> What is currently termed AI is a patched set of
> techniques akin to very limited scripting.
I'm sorry you hold AI in such low esteem, Darth. Unfortunately, this is not at all true -- or at least, you've been playing the wrong games. AI in the game industry may be an entirely different thing from AI in the academic world, but there's a surprising amount of sophisticated AI that's increasingly being applied in a huge number of games across almost all genres -- Jeff Orkin's work on Monolith's 'F.E.A.R.' being one of my favorite examples.
It's hardly something you can refer to as "scripting" any more, even if many games do use scripting in limited contexts in addition to AI.
> [Ken Fox]
> You know what I find most strange about this
> conversation are the related ideas that (1)
> developers don't learn much from other developers,
> and (2) developers don't know what consumers want.
> Does the game industry work by throwing mud at
> the wall to see what sticks?
There may be some truth in (2), but I'd like to state for the record that (1) is completely inaccurate. Developers learn from one another constantly, sharing ideas through constant dialogue, mentoring, and cross-pollination of ideas, and via forums and media such as the Game Developer's Conference, the AIIDE conference (www.aiide.org), Game Developer magazine, www.gamasutra.com, www.thechaosengine.com, and the various Charles River Media books (particularly the Game Programming Gems and AI Game Programming Wisdom series).
> [Mike Rozak]
> Combine all the above difficulties with the fact
> that the world is populated by intelligent players
> that make even the best AI look stupid, and most
> big-budget games would rather invest in more eye
> candy and larger worlds. Eye candy and large worlds
> are paying off at the moment, so why risk working
> on AI?
It seems to me that eye candy and large worlds aren't actually paying off for anyone except Blizzard at the moment.
Regarding your assertion that intelligent players obviate the need for AI (and the assertion that there's significant human intelligence in MMOs contradicts at least 90% of my experiences in World of WarCraft, but I'll play along!) ...
I'm reminded of a quote by Mark Twain: "We can't all be heroes because somebody has to sit on the sidewalk and clap as they march by." Human players are great at doing the things they want to do, but a world also needs characters who will do things that human players aren't willing to do.
World of WarCraft seems to do fine with its vendor NPCs and scripted mobs, but I don't doubt for a second that there are plenty of other kinds of viable MMO designs that would revolve around a level of AI that's much more on par with where the rest of the game industry is heading.
Posted by: Paul Tozour | Mar 18, 2006 at 20:52
Random comments:
- Developers not learning from one another - As Per Paul Tazour, they do. Sometimes I think they learn too well, copying previous successes too much.
- Developers not knowing what the customers want - I know that I want to play a game with good AI. I know that Mike Sellers wants to play a game with good AI, but not necessarily the same AI I'm thinking about. I can ask existing players if they want good AI, but I won't get a reliable anwer because (a) half the players will say "Who needs AI? WoW rulez!", and (b) most players are not informed/imaginitive enough about what AI is and what it's capable of to know if they'd like to play it. They'll know if they like it when they see it though. Therefore, I know I will have at least 2 players, maybe more, but there's a lot of uncertainty in the market size. More thoughts on http://www.mxac.com.au/drt/LawOfNewInventions.htm.
Games going for eye candy and scale - I think that AI could be useful too, but given the top-rated upcoming games in MMORPG.com, how many are emphasizing AI? These are how I read each of the top game's marketing messages:
#1 - Star Trek Online - Too early to tell. They only have mocked up screen shots.
#2 - Vanguard - More EverQuest-1 than EverQuest. Like EQ-1, but bigger and with better graphics.
#3 - Pirates - PvP ship combat, different setting, but I don't recall them touting AI too much. They do mention personal NPCs, so that could me more work on AI.
#4 - Chronicles of Spellborn - While they don't say this, I read, "Like WoW, but better."
#5 - Age of Conan - Eye candy, battle formations, more adult (violent/selacious?) content.
#6 - Warhammer online - War everyplace. Still too early to tell.
#7 - Hero's journey - Customize your character. Lots of quests in instances.
#8 - LOTRO - It's in middle earth and has great eye candy.
...
#17 Seed - AI is an important feature.
And then all the way down the list, no mention of AI as a top-5 feature (as far as I know).
Mike Seller's AI-intensive MMORPG, whatever it is, doesn't seem to be listed yet. I haven't bothered to list mine because I'm a very long ways away, and I'll be lucky to get a rating higher than the bottom 10. :-)
Posted by: Mike Rozak | Mar 18, 2006 at 22:42
As much as I like AI, it's not surprising that it's not something people gush about when thinking about their next MMOG. AI is the kind of thing that should be transparent to all but the AI geeks.
As for our game, we aren't saying much about it yet. You'll know when we do. :)
I have some thoughts on general features for post-WoW games but I'm saving them for their own posting.
Posted by: Mike Sellers | Mar 18, 2006 at 23:42
A few comments:
Human played characters may be able to start having some of the more 'real' reactions in the next 10 years. We are already developing fairly sophisticated AI routines, that could use Web cams to interpret people's reactions/feelings/etc. What if this became another 'controller' along with a mouse/keyboard/joystick. And if the reactions of a player could be transposed onto characters?
Some obvious issues, is that many people play half asleep, with blurry eyes, while drinking soda. They may not show an emotion that matches what their character would feel. But its worth pondering.
===
I have pondered the 'no-AI' mmorpg for the last few years since I first started to play Ultima Online and Dark Age of Camelot, and a lot now that I play World of Warcraft.
I do not feel that this will ever occur. While Everquest allowed us to setup our characters in a bazaar and sell items while we were off doing other life activities, I can't imagine this happening too much. Systems such as auction houses work better. And I do not know if I want a completely player driven economy/marketplace. Personally, I'm not confident that many game companies could fulfill that task, without the expensive, lengthy, and difficulty of learning economic theory, and making a system that could suit that (particularly since we still do not have a fully interactive environment).
A MMORPG without NPCs, could become incredibly boring for many players. Would there be enough that would fulfill mundain activities? I doubt it. Would systems of hierachy/power be abused? Without doubt. Would it be incredibly interesting to start an mmorpg with nothing but an environment and charaters 'dumped' into a starting zone and see what happens? Yes. How they developed housing, an economy, a political system, how they'd enforce rules, etc. But this is a topic for another thread, and a game far down the lines (one that I'm semi-working on).
Lastly, I just want to say that I'm really interested to see where development of NPC interaction goes. When they can percieve and react to their environment, perhaps then aggro range will become more intuitive (if I do this... then surely they will attack me, because A) they see me, B) they dislike me, C) they know I'm a threat). Instead of it being a specific threat radius. Further, if speech can be implemented, how NPCs talk to me could be far more dynamic.
As NPCs become more 'alive', even if we as characters do not, single player and MMORPG games will start to become more immersive. So long as this immersion is rewarding, gaming will advance.
Posted by: Scott Borre | Mar 20, 2006 at 11:30
Paul Tozour> Human players are great at doing the things they want to do, but a world also needs characters who will do things that human players aren't willing to do.
Just out of curiosity, what are some examples of things that human players aren't thought to be willing to do? (In the real world, at least, there doesn't seem to be anything that people won't do....)
It might be interesting to see if there's general agreement on some tasks or roles that are useful in a game world but that no real person could be motivated to perform. I also wonder if such tasks/roles would be the best candidates for improved AI.
--Bart
Posted by: Bart Stewart | Mar 20, 2006 at 13:21
Just saw this report at Gamasutra that the folks behind Havok are releasing related tools for character behavior development.
--Bart
Posted by: Bart Stewart | Mar 22, 2006 at 13:30
My point was the following: Combinatronics is not AI.
Computers are very good at executing very simple tasks incredibly faster than human beings can.
Yet, as of today, most AI techniques used in games include little adaptation. (what happens when you do something for which no response was anticipated)
Even neural networks are a poor approximation of what human beings are capable of.
Posted by: Darth Pixel | Mar 22, 2006 at 20:16
Thank me for coments !!
Posted by: CATROONS | Oct 05, 2006 at 07:19