Sponsored By

AI and MMOs - The Controversy

There have been numerous posts on Gamasutra lately discussing various aspects of MMOs and how better AI can help. Coincidentally, this is the topic of my lecture at GDC Austin next week. As a sort of "coming attraction" I weigh in on these issues.

Dave Mark, Blogger

September 6, 2009

20 Min Read

There have been a significant surge of blog posts (and ensuing conversations) on Gamasutra recently dealing with MMOs. These discussions cover a number of topics but there are two consistent threads that run through them - player socialization and the paucity of adequate AI. I find this oddly fortuitous in that I am delivering a lecture at GDC Austin next week on the nexus of these two seemingly unrelated issues.

To paste from the GDC Austin site:


Cover Me! Promoting MMO Player Interaction through Advanced AI
Session Description 
Historically, PvE AI in MMOs has been a straight-forward affair. While this leads to predictability, it also leads to monotony. In online, team-based PvP games, however, much of the attraction is the dynamic nature of the engagement that necessitates that players read, communicate, and react appropriately to changing, even unexpected actions of their enemies. By leveraging more advanced techniques that are becoming common in FPS, RPG, and RTS games, the AI in MMOs can be designed to provide some of the more attractive and engaging elements of PvP games. This, in turn, can lead to more involved team play, greater replayability, and an increased sense of community in the game.
This lecture shows examples of some of the aspects of PvP games that are attractive to players, the AI techniques that can be used to replicate them, and the effect that inclusion of these aspects can have in an MMO environment. The attendee will leave with a variety of concepts that can be included in their own MMO designs.


For reference, the blog posts I refer to above are specifically:

Additionally, I would like to point to an article by Gregory Weir from back in June:

While Edward Hunter's original post about MMO classes wasn't directly related to the topic, the comments took a significant turn when Jonathon Walsh said rather innocuously:

"The problem with PvE focused games is they are inherently working against their greatest strength. The draw of an MMO is playing alongside thousands of other players. PvE works directly against this by splitting up players into small groups and giving them a similar experience to what they could get by hosting up an RPG using a matchmaking service. The other problem with PvE content is that it will always be 'solved' at some point. Very well done randomly generated content and AI could change this a bit but that's definitely unfeasible at this point in time."

While I agree with his point that the draw of an MMO is playing alongside many other people, I disagree that PvE works directly against that. An instanced server does do so as he mentions. This is really no different from a very large game of Left 4 Dead. However, that is not the fault of PvE. Large world games such as back in the days of UO gave you the feeling of playing with or against many people but also the environmental monsters (AI).

Additionally, while I agree with his statement that randomly generated AI content could go a long way towards breaking away from the "solvable" AI encounters that we are used to now, I disagree with his position that this is "definitely unfeasible". That is not the case. There are plenty of techniques that can be used to create unpredictable, yet reasonable-looking actions in AI actors. I actually dedicate a significant portion of my book, Behavioral Mathematics for Game AI to these techniques and will touch (briefly) on some of them in my lecture.

Still, his points were well taken. The discussion continued in ernest, however, when Owain abArawn dropped this bomb:

"Christopher, one reason for the lack of choices in WoW, as you describe, is the cop out that developers all use of substituting hit points for AI. Raid levels aren't necessarily hard because the mobs are skilled or use effective tactics as much as they are insanely overpowered or nearly invulnerable. How much easier is it to give a mob 100,000 hit points than it is to develop an effective, adaptive, flexible AI. This is one reason why developers find themselves on a perpetual treadmill of providing updates, expansions, and additional content. Their current content is static and predictable. Perhaps a better solution would be to implement a dynamic unpredictable spawning system with mobs that have better AI. That way, your content refreshes itself, and players can never be quite sure what they will encounter, where they will encounter it, or what it will do when it is encountered." 

This is very true in many cases. In fact, this state of MMO AI actually harkens back to the early days of single-player AI - give the boss a lot of hitpoints and he is harder to kill than his minions. He is no smarter, however. You simply must repeat the same process that many more times in succession in order to defeat him. And that certainly doesn't make for engaging gameplay. While the spawning system has a little to do with it (and I cover this in my Austin lecture), the actual encounters are what needs to see the most significant change. Regardless, having the sense of the unknown does make for a significant change in how players must approach the game - as a group.

After Christopher Wragg argued briefly that he didn't believe changing AI would help the title topic (that of new classes), Owain responded again in support of the AI-based solution. He asserted that the whole reason that we are stuck in the tank/healer/dps class model is that the AI across many games has not progressed.

"I have used variations on that same formula in UO, DaOC, ShadowBane, Star Wars Galaxies, WoW, Vanguard, Age of Conan, Warhammer Online, and most recently, DarkFall, so I think I am pretty secure in asserting that all those games, over a span of a decade, have all used pretty much the same dogshit industry standard AI model. Raids are only different in the number of friends you need to drag into the party to compensate for the insane number of hit points/damage capabilities of your typical boss. The AI is usually just as stoopid.

Now consider how much success I'd have with those same tactics against a group of other players in a free for all PvP environment. I'd never get the first shot off before being charged, flanked, mobbed and pwn'ed."

This is exactly my argument in my lecture - compare the excitement and fluidity of a game such as Team Fortress 2 to the rote memorization and execution of many of the games he mentioned above. The point is, many of the tactics we use in MMOs against "stoopid" AI don't fly in PvP games.

Christopher Wragg responded:

"Now you miss the point, the lack of *choices* available to players has little to do with AI. If anything simple AI, raises the number of choices available to players (working on the concept of, if you are not being forced to react, you are instead free to act at will). For instance, given more complicated AI in a game like WoW, wouldn't raise the choices available to players, if anything it reduces the number of "correct" actions that can be taken. As such the classes would still have the same restrictions placed on them and more complicated AI would simply mean everyone dies."

This is where I really disagree. In fact, I believe that the exact opposite is true. Christopher claims that the lack of AI increases the number of choices available to players. I can't imagine a scenario in any sort of game where this would be true. I suppose I could see where Chris was coming from in this respect. Against simple AI you can do many different things. Theoretically, you could do an audition for an animated version of "So You Think You Can Dance" - but that won't bring you any closer to your goal of winning the battle. Usually the action set of things that you "must" do to win the battle is limited to a rather narrow "golden path". This is the phenomenon that gives rise to sites like TankSpot. Watch the video, play your part, win the battle.

By including good AI, however, you are significantly increasing the number of possible actions you may have to do in order to win the battle.It is a commonly known fact that the more potential responses that an opponent may take to your actions increases the branching factor significantly. Additionally, the more actions that an opponent can take against you (i.e. through good AI), the more potential reactions you must be prepared to exhibit.

I submit that PvP could be used as an example of the pinnacle of AI - the best we can hope for. However, in a PvP game (whether it be in an MMO or a game such as TF2), the gameplay is far more wide and varied because of the combinatorial explosion that results from many independent people making a wide spectrum of choices.

I also take issue with his overly simplistic (and almost completely binary) statement that "more complicated AI would simply mean everyone dies." This is actually a rather silly notion. One could have said the same thing about giving a boss monster 10,000 hit points. "How can we win against someone that is 100x more powerful than us?!?" That's not at all the case... it just takes longer. In the case of better AI, it is also not the case. The difference being that you may have to actually do more than one type of attack in order to counter the more clever AI.

He continued...

"If the enemy is smart enough to bypass the warrior, then what point having him, if the enemy is clever enough to kill the mage, the mage will die. The classes in WoW are not flexible enough to handle such a scenario. My point, is that without even any increase in AI or change to the enemy abilities, skill choice and use could be made far greater. This raises the number of choices a set of players could make before and during battle. As I was saying before, classes shouldn't be built with such harsh roles built into them. The players should be presented with a scenario, and then given the ability to work out their own way of surmounting it. Clever use of AI should merely alter the nature of the challenge.


If you change AI you MUST alter class as well to balance your game again."

Again (and I hate to pick on Chris, but he is illustrating an unfortunate meme in the industry - by players and developers both), this is a horribly narrow-sighted statement.

Using TF2 again as an example... each class has its own strengths and weaknesses. One obvious example is the use of the engineer's sentry. If we were to say "spies can go through our ranks and destroy sentries so what's the point of having them?" it would be viewed not as a statement of obsolescence of the sentry, but rather a call to arms to other players to make sure that spies don't go through their ranks. In fact, many times, pyros are called upon to patrol the less-active areas of the current battle specifically to look for spies. That is what pyros do well.

The same can be said for other classes. Many classes get mowed down by those same sentries... should we all give up on the battle? No... the preferred method for dispatching said sentries is for a medic and demo to go in with a deployed uber-charge specifically to take down the hardware and allow the others through. It is a big game of rock-paper-scissors. If, for some reason, the other team doesn't have sentries but is overloaded with pyros, my team will counter with heavies instead of demos. Heavies (especially with medics in tow) stop pyros in their tracks.

The dynamics of the team balance - the necessity for the different classes - is specifically caused by the fact that we do NOT know what we are going to be facing. If the opponents were to be completely static and predictable (analogous to poor AI), I could tell you exactly which classes we should have on our team and what we should attack in what order. There is one best way of doing it. The component of unpredictability and the surprise that can result is exactly what causes us to have to make decisions on the fly, communicate what we see, formulate plans, assign the proper person to the role, and execute this plan as a coordinated team - and this entire re-planning process repeats often - even multiple times a minute.

In the end, his final claim that altering the AI forces the designer to alter the classes to re-balance the game is not necessarily true. As I illustrate above, it just changes how we must use those classes. In my Austin lecture, I use the exact example that he mentions... that of an enemy flanking to go after the "squishies" in the back. All this means is that the fighters in front can't simply pull up a lawn chair and be content with their mindless button-mashing. They must be ready, if necessary, to rush back to the aid of their pals - or better yet, alter their tactics originally to be aware of the potential for flanking and prevent it to begin with. See how the increased AI can lead not only to interesting decisions but also more interaction between players?

Quick question to Owain abArawn regarding the following comment:

"Consider the following real world scenario. I am faced with two opponents. The opponent that is closer to me is armed with a knife. Behind him is a person holding a gun. Yes, the person holding a gun is a significant threat, but if I ignore the person right in front of me holding the knife and try to push past him, the person holding the gun never need fire because I will be killed by the knife wielder. That is a failure in intelligence on my part, and in an MMO, it is a failure in AI."

Seriously, dude... did you read my book already? That is almost straight out of chapter 14!

Edward Hunters's second post, Bring Real Socialization Back Into MMOs Or Else!, is more directly on-topic to my lecture in Austin. I particularly was fond of his observation that "... raid groups and 'Guilds' are not social situations in the MMO's today.  There is no difference, typically, in these social interactions and those you would find in your typical IM conversation or Facebook thread." The reason for this is largely because players do not have to communicate much game-related information to each other in MMOs these days. Again, contrast that to online PvP shooters such as TF2 where communication of relevant information is crucial to survival as an individual and success as a team.

Edward's main focus, however, was not on combat but rather on the "RP" aspect of MMORPG. We are agreed that oft-times MMOs are simply virtual chatrooms. This isn't as much my purview, however.

My new best friend, Owain abArawn, commented at one point,

"My comments apply to PvP-centric games, since the AI in PvE MMOs is typically strongly lacking."

I wish this wasn't mutually exclusive. As I mentioned above - and which is the focus of my lecture - I believe that if we move the AI in the MMO space to be more in-line with what we encounter in PvP, we can do a lot towards generating that community that Owain goes on to talk about in his comment.

In An Alternative to Aggro, Bart Stewart gently rails against the industry standard of "aggro" and goes on to suggest some relatively viable alternatives. While I won't comment on his suggestions here, I agree with his premise that the current concept of aggro is a horribly misshapen vestigial relic from the days when games couldn't afford to do expensive processing to determine something as simple as who/what/where/when to attack (not in that order, of course). The result is that all the enemies in MMO games are effectively "on strings". We tell them when to attack (when we are ready), who to attack (the tank), we know exactly how they will attack, etc. All sense of autonomy of the enemy is completely lost. We go back to the idea that MMO encounters are a "solvable" puzzle much like the little sliding number puzzles. All we need to do is push the little tiles around in the right order until we get an arrangement we like. There is no reaction necessary on our part... only control.

With loads of processing power available to us, many (but not all) single-player games have far more complicated algorithms in place that give a far more realistic look and feel to the initial and ongoing interaction with the player. (I give considerable treatment to this in my book and will give some quick examples of alternatives to standard aggro in my lecture.) The odd thing is that, for the most part, we haven't lifted these algorithms out of the single-player world into the MMO world. Really, the algorithms to deal with multiple units aren't that much more complicated than they are in single-player games. What's more, they can be entirely data driven, flexible, and adaptable to changing situations. Once again, if our goal is to have a more engaging interaction with the AI - not unlike what we see and feel in PvP games, we need to imbue our creations with more person-like thought processes. And, as we know, people are a little unpredictable. Reasonable, yet neither random nor rigid.

I was appalled at some of the comments that were defending the notion of simplistic aggro, however. (Johannes Smidelöv noticed this as well.) We have established memes like tank/healer/dps so deeply in our collective MMO-playing psyche that we use them as a defense against what caused those strategies in the first place. That is, current aggro models begat the idea of tank/healer/dps, and yet we defend the aggro system by saying "what will the healer do if he has to heal someone other than the tank?!?" It's almost Stockholm Syndrome-esque that we are so worshipful of the very captor of gameplay creativity that we are so willing to use our own mind-numbing boredom to defend it.

There were flashes of hope, however. Andre Gagne summed it up nicely when he mused,

"With a different type of AI the combat itself would fundamentally change and be, potentially, more interesting."

Amen, my brother.

Farther down, Maxime Doucet echoed the quick example that was suggested by Owain abArawn above...

"The AI should at least be able to compare both possibilities: should I attempt to take down the tank who's standing in front of me, or should I try bypassing him and aim for the healer, who's 20 yards behind him?"

Again, this is an example straight out of my both my book and my lecture. There are plenty of ways to do this that make the character seem reasonable yet varied. (Incidentally, one of the main problems with aggro is that it exposes the fact that the AI knows nothing of the concept of futility... e.g. "I will never kill the tank as long as the healer is buffing him up.") Maxime concluded with the observation,

"This would make the whole PvE encounters much less predictable and more adventurous, and way interesting for an experienced player as well, as the players would encounter mixed groups of different kinds of mobs with widely different AI behaviors and triggers."

I did get a kick out of what Matt Kane said... he looked at the reason for aggro from a slightly different angle:

"The real culprit here is the design of typical raid encounters. 1 big raid boss sitting still for 5 years picking his nose waiting for a group of 4-50 random guys to come and hack at his shins for 10 minutes and 24 seconds til he suddenly dies of congestive heart failure.

The boss has no goals. If you wouldn't have come along with your group of buddies, he would have probably continued to pick his nose for another 5 years.

So designers compensate for the fact that they develop enemies and encounters with no purpose by imposing standardized, shallow enemy motivation.

Instead of attempting to conquer the world, your terrifying end boss to the dungeon you've been trudging through for hours is actually playing whack-a-mole with one of your UI elements.

A good start to solving boss mechanics would be endowing bosses with motivation and purpose (beyond dialogue text)."

To sum up, he suggests aggro is a bastard child of the simplistic design of MMOs to begin with. That is definitely beyond the scope of my column... but I do want to suggest that there may be a "chicken and egg" phenomenon going on here. Designers may not even be aware of the fact that we can do nifty things. I would suspect that they design their worlds and encounters in this way because they haven't been aware that there are other things that we can do with AI to enable their creativity. That is largely the purpose of my lecture in Austin, by the way... not to "solve" anything but rather to get the creative juices flowing.

That being said, I hope that those of you who have commented or even simply read the referenced threads and are also going to be Austin will stop by and take in my lecture. Please make a point to introduce yourself. I believe you will find it entertaining as well as informative. If my audience is even remotely as spirited as these discussions, I'm going to very interested in taking questions at the end although I should only have 5 minutes or so. Regardless, because I am on day one of the main conference we can take the rest of the week to hash it out.

See you in Austin!

Dave Mark
President & Lead AI Consultant
Intrinsic Algorithm LLC

Read more about:


About the Author(s)

Daily news, dev blogs, and stories from Game Developer straight to your inbox

You May Also Like