The AI Kiss

Artificial Intelligence (AI) in MMOGs is a hot topic on Terra Nova this week.  Let us introduce a parallel discussion.   Damion Schubert ("You Don't Want Realistic AI") and Jamie Fristrom ("Manifesto Thingy") banter a fresh insight on an observation made many times (e.g.  lately in these comments): namely that MMOG AI is for the most part simplified because it needs to be a player solvable puzzle.

Damion writes:

In most games, the best way to think about A.I. is to think about it as a puzzle the player needs to solve. It’s roughly analogous to the cards that start face up in solitaire - it’s the visible circumstance that you have to modify your gameplay patterns to deal with. Everquest’s model has persisted because it provides an interesting problem that can be solved either by a solo player or a group.

EQ’s AI is also somewhat deterministic, which is to say that if you know how the AI works, you’re not going to be surprised a whole lot. You certainly could have more randomness thrown in and/or hide more behind the virtual DM’s screen, but having a game where the AI is fairly transparent makes it easier for groups to congeal and plan their strategy. The more the AI requires players to evolve to random circumstance, the less faith players will have in casual pickup groups.

Damion points out that this doesn't mean that AI realism has to be avoided in MMOGs (though it is now), but that it does mean that AI realism will puzzle different gameplay patterns.  In other words realism will necessitate a  different kind of game than we may be used to.

In Richard's earlier comment,  he cites "Who Needs Emotions? The brain meets the robot" - a  book whose conclusion was overviewed with this bullet: "Beware the Passionate Robot."   If the synonym for Passion were Realism, emotional or otherwise, then too beware, for a kiss may not be a gameplay pattern in your world...   yet.

Comments

[Damion Schubert]

1.

To stress: it's not just that it needs to be a player solvable puzzle. It needs to be a player solvable puzzle that works well with groups, and tolerates the off-chance that you've somehow joined up with an idiot.

[Edward Castronova]

2.

There's an element of political theory showing that while it may be possible to design systems that are incentive compatible for individuals, it's much harder to design them as incentive-compatible for groups.

Example: Maybe you can design an auction system such that no one person finds it in her interest to rig it. But it's much harder to do if you assume people can get together in groups.

The theorists typically respond to this by trying to figure out how to keep groups from forming. Groups are costly for their members to maintain, so the idea is to raise these costs. Keep corporations from colluding on prices, for example: you know it as anti-trust law. But discouraging grouping is antithetical to the/one point of a mog.

In games, I always think with sympathy about mog designers, who always have to imagine "What if a 100-person guild decides to go after this and screw it up?"

Here Damion's going at this another way: he's designing fun puzzles for groups, of 2-100 people. By extension of the above, that's got to be much harder than designing fun puzzles for individuals. In fact, I am having trouble getting my head around the idea at this hour.

Well, it's interesting. Never thought of the theory/game interface that way before.

[magicback]

3.

One key factor in design is the need to maximize content and by this you have to consider groups (otherwise why not just play single-player games).

Thus games and activities start to look the same because developers appear to have come to similar conclusions as to what would maximize enjoyment while reducing cost.

Using the casino analogy: some people like to play slots, some like to play the tables, while others like private poker games. Casino may try something like Bacarrat, but hey their happy with blackjack tables.

Thus, I think it is less about complexity or depth, but more about people wanting to have some form of consistency and essentially know what to expect out of their enjoyment dollar.

In this way I think people like a set of few basic classes or sufficiently dynamic yet highly predictable "puzzles". This enable individuals to know their places within the group and reasonably expect what they can and should do.

"We need 3 nukers, 4 tanks!"
"We're short 2 tanks, so let's go find some pickup-group players, and idiots are ok as long as they can soak damage!"

When we play, we want a safe and fun place to play. We really don't want to deal with unpredicable and perhaps unreliable people, so we go solo. Or we really don't want to play online alone, so we go groups.

Thus, we have NPCs with "Quest here" signs, have NPCs with their "threat here" signs. Likely the next enhancement will be quests with "manufacturer suggested retail number of x class, y class, and z class players" displayed for easy reference.

And this leads to research on the perspective that people use RMT to shape their enjoyment curve to their liking as the developer is unable to offer the specific enjoyment curve they like.

[Raph]

4.

The observation that we can make AI too smart for players is a VERY old one. Consider how Robotron would play if all the robots made a beeline for the player and the humans. The game would be over in seconds.

Sorry to sound like Richard with his "read the MUD-Dev archives" thing. ;)

[magicback]

5.

Yes, but's it's interesting to observe how history repeats in variations and to hear whether there is any new perspective.

For example, WoW increased relative to other MMo UI customization leading to gameplay augmentation, which affects how new AI will be programmed, which leads to new gameplay dynamics based on the new AI programing, which leads to....

Now, Guildwars to select-a-skill-set paradigm may become fashionable :)

[Mike Sellers]

6.

Very interesting topic. Both Damion and Jamie have made insightful comments.

"AI" is a huge topic though. Thus far the discussion seems to have focused on what we all know, which is combat-oriented AI. The consensus seems to be that making mobs slightly more dynamic or tactical would be great, so long as they don't smear the player at every encounter. This makes sense; it's a special case of general game balance. And Damion rightly extends this to providing robust but not overpowering gameplay in the face of imperfect player grouping.

But I think that "better AI" doesn't just mean better Quake-bots, or monsters that are just a better combat-puzzle. IMO, that's a very small (if currently important) slice of the "better AI" pie. In fact, I'll go so far as to say that AI has the potential to open up new kinds of gameplay -- for both individuals and groups -- that will change our concept of MMOG play, extending it beyond the ruts we've worn so deeply for the past twenty years.

That is, the addition of AI that is interactive, emotional, and socially plausible (beyond combat situations) opens up new kinds of gameplay, especially when combined with MMOGs, that don't fit into the formalized (almost fossilized) gameplay categories in our current games.

I haven't said much about what we're working on, but much of it focuses on this area of "better AI" as it applies to creating better, broader gameplay -- outside of the well-trodden paths of combat AI and pathfinding. Damion's right that anything in this area has to contend with the grouping of players in MMOGs. Our belief and experience is that socially interactive and plausible AI (including emotional AI) takes both solo and group MMOG play in entirely new directions.

Do non-combat uses of AI hold interest for people here?

[Matt Mihaly]

7.

Definitely holds interest for me, Mike.

Richard wrote a good series of articles (the location of which I can't remember) on auto-generating interesting and relevant quests using NPCs that seemed like a good start in that direction. I recall him admitting he's never implemented it, but it really set off a firestorm of speculation in my mind for awhile after reading them. Maybe Richard could post a link?

--matt

[Edward Castronova]

8.

add my voice to those looking for more from AI than a combat puzzle.

but speaking of combat, how do movies handle the Great Foil character? doesn't it always have some key weakness that the hero exploits? the problem in a mog context is that the key weakness, if static, becomes common knowledge. "you hit the dragon on his left toe." yawn. maybe a dynamic weakness model. Also: in films, sometimes the enemy or a key ally changes sides at the last minute, because the hero is just so darned good. that might work. are there other film industry approaches?

anyways - would love to see richard's dynamic quests.

i also think the affective element of AI is going to be enhanced, as several mentioned in the previous AI post.

[Theo]

9.

As a player, I enjoy a less predictable experience, and I find it sorely lacking in the MMOs I've played. In City of Heroes, for example, Archvillains are challenging not because they surprise the players by using their abilities in creative ways, but because they're given phenomenal boosts to their hit points and regen rate, along with a powerful offense. What archvillain battles amount to are six to eight guys button mashing against one mob for five to fifteen minutes. AV battles are really one of the least exciting aspects of the game for some players.

PvP in CoH, on the other hand, allows you to fight opponents with wildly different powers and varying degrees of competency. With City of Villains on the horizon, it's possible that they might hit pretty close to the mark with player-created conflicts. Traditional PvE AI for the masses, player created content for the more...discriminating(?)...palette.

As a side note, nothing ruins the immersion of a game for me more than running past hordes of bears, spiders, wolves and orcs who are just wandering around the woods alone, waiting to be killed by me.

Just observations from a player. :)

[Thabor]

10.

The observation that we can make AI too smart for players is a VERY old one. Consider how Robotron would play if all the robots made a beeline for the player and the humans. The game would be over in seconds.

I dislike like way that the term AI has been watered down to include basic behavior. I wouldn't even categorize such a swarming behavior as AI, much less "smarter".

From my point of view AI should imply perception, prediction, and decision making from a pool of multiple possible states.

The beeline pursuit behavior is one of my pet gripes, and probably one of the most exploited NPC behaviors in any game once obstacles start being factored in.

A real "smart" pursuit AI would path towards a predicted location instead of a current one. It might try going a different way around an obstacle instead of always trailing in a player's wake. A step up from that might try to predict if it is really a good idea to charge around a blind corner without any friends..

Player's asking for smarter AI are placing the emphasis on the intelligent, rather than the artificial portion. The current flight / pursuit behavior models are worn pretty thin to an experienced audience.

AI pretty generally includes any pattern that includes the semblence of intelligence now, but the bar is raised by an educated audience.

Even in many single player tactical combat games current AI implementations aren't capable of presenting a challenge without a singificant power or numbers advantage.

[Damion Schubert]

11.

Consider how Robotron would play if all the robots made a beeline for the player and the humans.

Um, most of the robots do, in fact, make a beeline for the player (albeit a slow one). This predictability is in fact what makes the game playable - you against 100 robots becomes a lot harder when all the robots figure out how to dodge.

[Flatfingers]

12.

I'd also like to see more non-combat AI.

Actually, I'd like to see more AI, period, and at three levels:
* individual (NPCs & critters)
* group (flocking, ecosystems, emergent behavior)
* system ("the game" has goals)

The problem is that players are semi-schizophrenic: They want predictability so that they have a good chance of being able to compete, but then they complain that the game is boring.

How do you satisfy both of those desires? How do you let NPCs and groups and the system do what they "want" but still provide enough predictability to let players develop functional gameplay models?

Would a game whose AI is so advanced that it constantly "changes the rules" be successful?

--Flatfingers

[greglas]

13.

The comments on Damion & Jamie's blogs are well worth checking out too.

The only point I'd make is that if you want realistic AI, you've got it and it's called the PvP MMOG -- which I take is Theo's point about CoH. The reason the majority doesn't like PvP in MMOGs is that PvP = PK pWning of n00bs, or worse yet, group PK pWning -- so everyone can't win. See the comments about realistic AI forcing people to play easier games.

I'm with Damion that you can have a great game with extremely "dumb" AI -- see e.g. Tetris, or the EQ epic mob raid. Pix. The "play" in that kind of raid, though, is funny -- it's the play of organizing the raid, coordinating the various classes, dealing with team inexperience and mistakes, etc. In other words, you might say it is really is a kind of subtle *cooperative* form of PvP. ;-)

Just a thought, but an alternative might be to have a game involving only one extremely smart opponent that one engages and fails to fully defeat, rather than require 1000 dead rats for another ding.

[Damion Schubert]

14.

Would a game whose AI is so advanced that it constantly "changes the rules" be successful?

For a game to be fun, the first and foremost rule is that it has to, on some level, be 'predictable'. I'm not saying the same thing has to happen every time, but the player has to know what to expect enough that he can take his own skills and apply them to the new problem. (This is, one will note, my thumbnail interpretation of Raph's Theory of Fun).

You can shake things up, mind you. Magic: the Gathering frequently adds new cards to the mix to keep things interesting (and sell more boosters). But you know that some things are going to be the same - white cards will be defensive, red has lots of ways to do direct damage, and mana cost tends to escalate with card power. There are patterns you can detect, and if you observe them, you feel like you can respond to them. When you do so, you feel really smart.

When a game is TOO random, though, it feels capricious and mean, and players won't tolerate that for long. Players don't mind losing, but they want to feel like the fight was fair. So your AIs CAN change the rules, but those rules on some level have to be something that a player can observe and respond to.

[Michael Chui]

15.

Richard wrote a good series of articles (the location of which I can't remember) on auto-generating interesting and relevant quests using NPCs that seemed like a good start in that direction. I recall him admitting he's never implemented it, but it really set off a firestorm of speculation in my mind for awhile after reading them. Maybe Richard could post a link?

Mayhap...
Notes from the Dawn of Time at Skotos?

[Michael Chui]

16.

You can shake things up, mind you.
When a game is TOO random, though, it feels capricious and mean, and players won't tolerate that for long.

I thought I'd point out that Raph points that out, too. (Though I loaned my copy to my econ teacher and he hasn't given it back. Fair, since I have two of his papers, and haven't given them back to him. *hums*)

It goes to question overall goal of the game, which is theoretically a maximization of Fun for the player(s). An AI that imitates human intelligence perfectly is fun only for a player who has fun interacting with human intelligence. Most players who show up enjoy interacting with stuff that behaves stupidly enough that they can take'em out, but intelligent enough that it's not a cakewalk. They need a shot of "Yeah! I did it!" at the end of the fight.

Making it realistic is the game the engineer plays: it's fun to create something realistic, because that's at the edge of his skills, and accounting for his players doesn't measure up. The engineer gets the Fun injection, but doesn't realize the game he plays creating the game isn't the same game the player plays, so he doesn't understand why it's not fun for the player.

[Samantha LeCraft]

17.

Most players who show up enjoy interacting with stuff that behaves stupidly enough that they can take'em out

If we're still talking about combat AI, sure. But what about interacting with AI characters in non-combat situations? Could AI that has nothing to do with combat or killing lead us to new forms of gameplay -- gameplay other than "kill monster, get loot"?

[Theo]

18.

>>The only point I'd make is that if you want realistic AI, you've got it and it's called the PvP MMOG -- which I take is Theo's point about CoH.<<

-greglas

Pretty much, though to be clear I think a good PvP system is just a starting point. Players creating content for one another, then resolving the encounters with a combination of PvP and PvE would seem to be the real deal.

[Paolo]

19.

"Do non-combat uses of AI hold interest for people here?"

Yes, Mike Sellers, it does to me. I make my case about it here.

Here's a snippet:

AI isn’t synonymous with killing efficiency. Artificial Intelligence is about behaviour and intellect, period. The A to Z of good AI is from an NPC who strays to take a smoke, pee, snack, or browse a magazine instead of something else to displaying fear, shock, anger, etc. when faced with your actions.

[magicback]

20.

In regards to non-combat AI,

I would be happy if someone added an Eliza-like oracle that can hold a decent conversation, but mostly entertain by freely giving out words of wisdom, messages of protent, wit to spare, etc.

The question of the week once the oracle is added is "is that a GM or an AI." Now that would be fun.

On, and to spread this content over a longer period...you can only query the oracle once a week. Might even get a buff out of it.

Now the question is do you want an NPC bartender or a PC playing a bartender? Guess it depends on the game/world.

Frank

[Theo]

21.

I'd rather have a PC bartender, which begs the question "who wants to play a bartender?"

[Matt Mihaly]

22.

Michael Chui wrote:

Mayhap...
Notes from the Dawn of Time at Skotos?

Yes. Start with article #23 and go upwards.

--matt

[Thabor]

23.

The only point I'd make is that if you want realistic AI, you've got it and it's called the PvP MMOG

Actually it isn't. The closest thing you get to a good challenge in most MMOGs is the duel. Regular PvP generally has too few constraints, and real players exploit holes in the system.

An AI isn't capable of the sort of meta-gaming and exploiting players are. It is only provided with the facilities the designer codes in. Although, for lazy designers that usually does included perceptual knowledge that wouldn't normally be available to a properly simulated entity.

Its probably actually more of a challenge to implement the ignorance portion of an AI than the intelligence (knowledge) portion.

[Michael Chui]

24.

If we're still talking about combat AI, sure. But what about interacting with AI characters in non-combat situations? Could AI that has nothing to do with combat or killing lead us to new forms of gameplay -- gameplay other than "kill monster, get loot"?

If you restrict yourself to "game" games, then all NPCs are either decoration or obstacles. And in obstacles, I separate them into "allies" and "enemies", where allies are those you have to obtain something from and enemies are those you have to defeat in some way.

Only if you take out the directional component implicit in the tone of the game can you have NPCs that are truly humanlike. (Notice I don't say intelligent, since that wouldn't be humanlike at all.)

The reason is fairly basic. If you have a goal, then all things are either relevant or they're not. If they're relevant, then they exist as a stepping stone. This is just as true in real life as it is in virtual life. If you're trying to become President of the United States, the opinion of a Celtic hermit who's lost his way in Madagascar is little more than an amusement. You've just got a couple million allies (constituents) and a large smattering of enemies (rival political parties and candidates).

[Mike Sellers]

25.

Well that's a commonly held view, for certain. It's also emblematic of the classic young male 'core' player's view: task-oriented, linear, and analytic. Nothing wrong with that, but it's a small slice of the gameplay pie. When you open up gameplay to more social (not just chatting), inclusive, and synoptic forms, the game doesn't cease to be a game... but a lot of things do change. 'Kill monster get loot' is no longer the sum total of the game.

MMOGs, IMO, are ideally suited to such play, being as they are inherently social -- even though we've done everything we can to eliminate that side of thing as part of the gameplay itself.

Now, making AI that enables such broader gameplay... that's the challenge.

[Flatfingers]

26.

Mike Sellers> Now, making AI that enables such broader [social, etc.] gameplay... that's the challenge.

Exactly. On a practical level, you're talking about something that's sufficiently hard to do that doing it would chew up a lot of your development time and money. You could wind up with great AI... but no game.

I think can see a couple of paths from here to better game AI. One (short-term) is that we keep nibbling around the problem to discover what simple heuristic tricks work. Conway's Game of Life, flocking sims, Eliza, Braitenberg's Vehicles, and the Looking Glass/Ion Storm "smart guard" are all examples of how relatively simple rule can create the appearance of semi-intelligent complexity. This approach might get us closer to social AI, but it probably won't be satisfactory for a long time (if ever).

The longer-term (and likely more successful) approach will probably be someone creating a general-purpose AI that can be given context-specific knowledge and then just "plugged in" to an application. Instead of trying to design AI into an app from the inside (while trying to write the rest of the app, too), we'll wind up with a standard interface for AI -- you just pick the one you want, install it, teach it things that are specific to your app, and you're up and running.

The best current example I'm aware of along these lines is Doug Lenat's Cyc project, but there are probably others in the works by now.

Until these "AI-in-a-box" tools and standards start to come online, we'll have to make do with simple heuristics. Focusing on how to get maximum useful believability with minimum rules will probably probably be a more effective route than trying to hard-code social intelligence into NPCs.

--Flatfingers

[Thabor]

27.

I'm not sure what exactly would constitutes a "social" AI. Discussion of AI usually tends to revolve around "combat" AI, because the general usage of AI is decision / control centric.

Also due to the perceptual limitations of those systems I'm not sure how an AI could maintain any social relevance. The most socially advanced AI would probably be the least interesting to me. IE: It wouldn't want to talk to me.

[Samantha LeCraft]

28.

From yesterday's New York Time's article on last week's Artificial Intelligence and Interactive Digital Entertainment conference (article found at http://www.nytimes.com/2005/06/07/arts/07arti.html):

But one of the broadest and most powerful approaches to artificial intelligence may be one that does not focus on determining specific behaviors. (''Does the computer general know that it should use tanks and artillery together?'')

Rather, it is a move to structure programs so that they absorb available information and then generate their own strategies to achieve sometimes-contradictory goals (''protect the hostages'' versus ''kill the enemy,'' for instance).

New sorts of advanced AI are already in developed. The US government is currently funding research and development of AI with believable emotions. It's only a matter of time before this sort of AI shows up in games.

[Mike Sellers]

29.

Thabor said, Discussion of AI usually tends to revolve around "combat" AI, because the general usage of AI is decision / control centric.

I'd say that most discussions of AI revolve around combat because it's easiest, most quantitative, and what many AI designers (especially in games) know and care about.

OTOH, there are significant examples of non-combat AI in games and game-like applications, starting with The Sims, and most recently with "Facade" and with others yet to come.

How will these be different from combat-specific AI? Well, how are graphical games different from text-only ones, or how are online games different from single-player ones? Adding in a new element of this magnitude changes everything.