Of Sony and Skynet
A year ago, or just about, I sat in a large banquet hall in Las Vegas as Dave Georgeson outlined the next big project for Sony Online Entertainment. The speech centered on a list of “Holy Grails” in MMO development, and how their EverQuest Next team would be reaching to achieve those lofty benchmarks.
Terrain destruction and the ability to affect permanent change on the game world were two that I’ve been hoping to see for years, and with their voxel-based terrain, it’s something that seemed well within their reach. As expected, SOE demonstrated almost immediate success with both when they rolled out Landmark. There was another goal that I reserved judgment on, however. That was their stated intent to create an emergent AI system.
This year at SOE Live, the AI for EverQuest Next was highlighted by the developers, and once again they were way ahead of where you might have expected them to be. For all of John Smedley’s talk about being open with their development, SOE still knows the importance of showmanship with the big reveal and how to employ it regularly.
It Looks Good
I just have to say that I’m really impressed with where SOE is with their AI development. They demonstrated a number of scenarios this year in Las Vegas, and had more problems getting videos to play than they did getting the AI to behave appropriately. That says a lot about the development and its progress.
One of the first things they demonstrated was that individual AI has actions they want to take, and would dynamically seek to accomplish them without any scripting. Using the newly announced dark elf race for the scenario, they began by showing a priestess going about her daily business of transporting captured life essence to a mana storage (and sort of conversion) station, they called a shadow cage. Shadow energy is what powers magic for the dark elves, so the priestess wants to create as much of it as possible.
That accumulated energy is then used by other AI in the pursuit of accomplishing their tasks. The developers demonstrated this through a dark elf scout during their talk at this year’s conference. The scout or other soldiers will come by to take the shadow energy from storage, and carry it to stations around the walls to power defensive weapons, creating more demand for the priestess to fulfill.
The developers also pointed out how scouts and guards are designed to want to patrol areas, thus they do so without scripts. The importance of that being that if something happens to obstruct their normal patrol route, a scripted AI would typically just stop until it could continue. In EverQuest Landmark it’ll work a little differently as the AI attempt to find another way around. Since their patrols are driven by sort of a digital desire, as opposed to scripted actions, they should be somewhat more capable of dealing with random obstacles.
Another thing demonstrated at SOE Live this year was an example of the visual recognition routines in the AI. To show how it worked, a dark elf priestess was given a set of royal robes. Because the robes indicate royalty, when other AI were within visual range of the priestess they would genuflect out of respect. Apparently the range of recognition is dependent on time of day and weather, which should make it feel less artificial and more immersive.
In the Wild
A point made a year ago by Dave Georgeson and subsequently echoed by John Smedley in other conversations, is that mobs in EQN will be effectively released into the wild where their nature will take over. Orcs will move out to do orcish-type things, and wisps will be all… well, wispy I guess. The point is that mobs and NPCs are all just AI, and effectively there’s not much real difference between how they’ll function, aside from where they spawn.
I find myself with a few questions about how well it might work, though. What impact does spawning a new AI out in the middle of nowhere have on what it decides to do? I’m sure this is something SOE will be working on, but it seems likely to pose a problem for them.
In a city or community what we would normally term NPCs have plenty of stimuli around them, which should make predicting their actions fairly straight forward. In the previous example, there’s a need for shadow energy, so any newly spawned dark elf priestess should immediately set to the task.
The problem is likely to come from the AI living in the wilds, or what we’d normally think of as mobs. With a less structured environment around them to define where they are, it seems like behavior may be a little less predictable. This is a fairly important point in an MMO. Typically designers work to ensure players encounter challenges around their level, and do so in locations that make sense to them. It keeps the pace of the game on track, and prevents player-confusion.
I see that being a potential problem for EQN in the current system. It’s possible that like bands of mobs will combine to form groups too large for co-leveled players to handle. Also, what’s to prevent roving bands of higher level mobs from sweeping through an area attacking lower-level players? Worse, mobs of different types will often attack each other in these sorts of games, so what happens if players can’t find a particular mob because it’s being hunted by roving bands of another mob-type?
I’m simultaneously concerned and thrilled at what the new system means for EQN and the MMOs that follow. On one hand, I feel there’s potential for some major problems when you don’t establish some sort of rails for your mobs. On the other hand, I really hope we don’t see those rails because the idea of not knowing what sort of creature is where really appeals to me and I think would go a long way to creating immersion.
EverQuest Next won’t stop their cunning decent into the complexities of AI interactions at the basic level, though. Like ogres and onions, their system will be composed of layers. Individual AI have certain likes, dislikes, and needs, but they’ll also inherit the same types of parameters from the layers of community above and around them.
The dark elf priestess that we’ve been referencing thus far, for example, could have the attributes of her order, her city, and the values of the dark elf kingdom on top of her personal ones combining for a great deal of additional complexity. Those macro-AI organizations define behavior on an even larger scale.
In their recent talk, the EQN team cited an example involving kobolds, the dark elves, a sect of druids, and the denizens of Kithicor. In their scenario, kobolds push into a mountain range in search of riches until they bump into the Dark Elves. The dark elves don’t like kobolds, so they attack them and drive them back, along the way finding a previously unknown to them sect of druids in another area.
The druids are working to contain the evil of Kithicor through the use of their nature magic. Because their nature magic is a form of life essence, which the dark elves desire for conversion into their own shadow essence, the dark elves go on the attack. At a given point when the energy levels of the nature magic fall low enough, the forces in Kithicor break free.
Up to this point, there’s been no script, and the scenario played out dynamically several times on the screen in front of us at the conference. The interesting thing about it was that it never played quite the same way twice on the tool developed to simulate the interactions of NPC factions.
Apparently, the breakout from Kithicor can become so severe that other factions from around the in-game world could become concerned and join the fray. An event that would generate a whole new list of needs and dislikes shared down to the individual AI and impacting the game world in very unscripted ways. The question is, could it get out of hand?
The Player Effect
A common military anecdote is that “No plan survives contact with the enemy.” A point that’s just as true of games as war, which leads to my second concern over the direction SOE is going with AI in EverQuest Landmark. No matter how hard they plan for it, players will always find a way to hose it up.
Emergent AI has been tried before. One of Richard Garriott’s favorite Ultima Online stories is how they tried an early version of the idea, referred to as the “dragon and sheep conundrum” by a new friend of mine and fellow writer. For those who haven’t heard the story, the UO team was incredibly excited about the AI they’d developed for their new game. Dragons wanted to eat sheep, but villagers would do if the dragons couldn’t find any. Sheep were expected to breed and repopulate over time, so the idea was that there would be ebbs and flows in the rate of dragons attacking villagers. The rate was expected to be based off of whether or not the dragons in question were getting full on sheep.
The problem was that the plan was developed and tested on a small player-base. Once it went live, players quickly exterminated all the sheep in the game, and then proceeded to do the same to the then ravenous dragons. The lesson learned was that emergent behavior is seriously cool, but developers need to cheat a little bit or players will systematically hunt creatures to extinction.
If SOE is planning to cheat, and I really hope they are, then they’re playing their methodology close to the vest at the moment. No matter what they try though, I’m worried about what the player impact will be on their systems. Granted we have a much stronger alpha/beta process for this sort of thing than we once had, but players are also simultaneously very adaptive and often aggressively malicious when it comes to picking on developers. That means the developer response will need to also be very adaptive and equally devious.
It’s not just a matter of players over-hunting various species of creatures, either. The AI behavior on the macro-scale has potential to be impacted by players, as well. If guilds or other masses of players decide to engage in a little in-game AI engineering, there’s no telling what the consequences could be. The SOE team will have to be very careful that they don’t make any assumptions about what players will do. For instance, while all, or most, of the NPC races dislike the undead in the Kithicor Forest example, there’s no guarantee that players won’t take the “wrong” side for the fun of it.
Years of sandbox type games have taught us a few important lessons for dealing with the unleashed player. First, don’t assume they’ll define right and wrong with the same terms you expect them to, and second if even if you do figure out their system of morality, expect them to switch it up occasionally just to torment the developers.
Rise of the AI
Regardless of whether you’re a fan of the EQ franchise or of SOE games, you have to look at what they’re doing with AI and world-building, and just be excited. It’s still too early to tell how successful they’ll be at all of it, but indications are pretty good so far. More importantly, I think they really have a chance to redefine a lot of what we consider standard in MMOs going forward.
Their new AI system is going to require new terms, I think. The terms mob and NPC don’t really hold true under the new model these guys are beginning to define because there’s really no difference between the two. In the past, the difference was fairly significant from a game design and programming standpoint, but that’s no longer really true.
In a way, it’s even analogous for the real world. Some of us seem so foreign that it’s hard to accept how similar to each other we all really are. Human differences are actually more attributable to the overlapping layers of the various communities we belong to than they are a result of any differences on an individual level.
So I can’t help but wonder if EQN will become a new model for studying social interaction, just as EVE Online has become a model for studying economics. If the past year of development is any indication, the odds that it will are certainly good. Just one more reason for me to be really excited about playing this game when it’s finished…as if I needed another!