Creating Intelligent Creatures
-Computer Graphics World
Game developers are turning to AI to give their characters personalities and to distinguish their titles from the pack.
What are looks without brains? That`s the question many a game developer is posing these days. With incredible graphics becoming more the norm than the exception, developers are turning to Artificial Intelligence (AI) to distinguish their titles from the hundreds vying for consumers` attention.
True, AI in games is nothing new. But the amount and sophistication of the AI showing up in games is, as is the growing trend to use AI to make characters appear more interesting and believable as "living" creatures. Driving these trends is the ever-increasing processing power of computers, in both microprocessors and 3D accelerator chips. As more processing power becomes available at lower costs, computing the graphics portion of games becomes less taxing, freeing up processing power for other features, such as AI.
"Generally, AI hasn`t gotten a lot of attention in the past, but I`ve seen that change in the last few years," confirms Steven Woodcock, a software engineer for Real3D (Orlando, FL) who also has an extensive background in AI and acts as a consultant. Additionally, he has run sessions on AI at the Computer Game Developers Conference (CGDC) for the past two years. "I ask a lot of the participants at my AI sessions at the CGDC how much CPU they get and how many people are working on the AI in their game. Just last year, the answers were often `less than a percent of CPU` and `Joe works on the AI half time.` This year the CPU percentage was pretty decent, and many companies have dedicated AI developers."
Perhaps one of the most intriguing games in development that`s using a hefty amount of AI is Galapagos from Anark (Boulder, CO). In this game, players guide a four-legged creature called Mendel through the 3D texture-mapped worlds of Galapagos, solving puzzles and helping Mendel eventually escape from this world.
Here`s the amazing part: Mendel is an artificial organism with the ability to learn, adapt, and interact with his environment and the player. Some refer to this type of AI as artificial life; Scott Collins, development engineer at Anark, prefers to call it by its more scientific name, adaptive technology. "A neural network is an example of an adaptive technology," notes Collins. "Genetic algorithms are a form of adaptive technology as well."
But Anark didn`t use either of these technologies for its AI. For Mendel`s "brain," Collins and his coworkers developed their own version of an adaptive controller called Non-stationary Entropic Reduction Mapping (NERM). The NERM controllers accept inputs and produce outputs that are translated into Mendel`s behavior. It`s also self-organizing, which is how it learns. Collins explains, "Self organizing means that there`s nothing outside of it that dictates its final form. For example, take a dog and give it praise and Scooby snacks, or scold it, and based on your [input], the dog will figure out how to behave to minimize its scoldings and maximize the affection. It`s self-organizing. Our adaptive controller receives feedback error and, based on this error, what it tries to do is converge in a way to minimize this error."
In fact, the player doesn`t have to interact at all with Mendel for him to learn; he`ll do it all by himself. "If you take a newborn Mendel, it`s actually very amusing. He`s flexing his limbs and falling off ledges and dying all the time. But if you watch him during his first hour, you`ll notice how he dramatically changes from infant-like movements to more mature actions."
Actually, says Collins, Anark set out to create an adaptive-controller technology that could be used in many applications, including industrial. Galapagos is just the first. "Frankly, more of NERM`s commercial value is outside [of entertainment]. There`s a variety of methods we could`ve applied to this game, but the engineering world is more critical," says Collins, explaining that to compete on that level, Anark had to take a different approach, which resulted in the NERM technology.
That`s not to say the technology doesn`t fit well in games. There are definite benefits to using adaptive AI versus a more traditional rules-based AI in games, notes Collins. "You end up with a more human-like opponent," he says. "A quick example would be when you have an enemy in a shoot `em up, and it uses traditional AI. It`s rules-based, as in `if within a certain range, then do this.` The thing is, these behaviors are very easy to uncover. A human can explore those patterns very quickly and figure out how they work and how to beat the machine.
The reason an adaptive technology would be interesting here is there`d be enemies observing the behavior of the player and changing in response to the player`s play-style."
Adds Collins, "That`s why network games are becoming so popular. The human opponent is so interesting."
Naturally, Collins finds Galapagos interesting, too. (The title should ship sometime this fall.) And because of the nature of Galapagos, it`s never the same game twice. "It`s quite addictive," says Collins.
Already on the market in the UK and expected any day in the US is Creatures from CyberLife (Cambridge, England), a spin-off of Millennium Interactive. This title lets players breed and raise Norns--cute little two-legged creatures with big eyes, long ears, and tails.
CyberLife`s technology, also named Cyberlife, is a proprietary system that uses neural networks. "What we`re doing is modeling biological building blocks that are then combined to create biological systems," says Toby Simpson, producer. "Our neural net is not like textbook neural nets. Neural nets were originally a biological metaphor, but they were highjacked by mathematicians. We stayed with the biological approach, where you have a huge set of chemicals, and no one individual neuron knows the big picture."
Norns begin as children and then progress to adolescents, adults, and finally, seniors. During their lifetime, they can breed, creating offspring with genetically different systems. Norns also make choices and learn from their mistakes. These creatures learn best and live longest, though, when assisted by the player, who can teach them right from wrong as well as help them with a number of tasks, such as breeding, finding food, and avoiding danger. If cared for, a Norn has a typical lifespan of about 15 hours. But because players can help create generations of Norns, the "game" can last much longer.
Upon hearing various stories about the Creatures, it`s easy to understand how people can get so attached to these simulated organisms. Relating one story, Simpson tells how two Creatures followed each other around all the time, much like best friends. Then one died. "The other just stood next to the body until it died, too," says Simpson. "To this day we have no idea what happened. We can`t, obviously, get inside their heads. But we can logically come up with an idea, and we think it committed suicide. The links in that Creature`s head between being alive and the other object were so substantial that it killed itself when the other Creature died. Those are the sort of things--the glimmers of life--that make this worthwhile."
Your Pet Computer
Fin Fin--a charming "animal" that`s half dolphin, half bird and was created by engineers at Fujitsu--is another character recently introduced to the market that uses AI to appear more lifelike. The AI in Fin Fin on Teo, the Magic Planet is based on what Michael Pontecorvo, director of technology at Fujitsu Interactive (San Francisco), calls believable agents. "It`s good for modeling internal-state and decision-making processes that lead to externally believable behavior. The engine executes a set of productions that work from a goal to a solution through activating multiple plans. It doesn`t do the planning; it uses them. We pre-build the plans, and it applies those in an expert-system-like fashion," he says.
Technically, Fin Fin does not learn; the technology used for Fin Fin is not adaptive. Pontecorvo explains, "The only thing Fin Fin really does [that`s learning] is there`s a variable that`s familiarity. He becomes more familiar with you over time as you attend to him. There`s a sensor that`s in your range; it`s a smart sensor and an audio and proximity detector. So up to a meter away, it can detect the presence of a person. As much as you attend to that system, Fin Fin will become more familiar with you."
Fin Fin also has been programmed with other "human" characteristics, such as thirst, fear, happiness, and sadness. He also gets hungry and sleepy.
Fin Fin has at least a couple more releases to come, adds Pontecorvo. Plans are to add to Teo, the planet on which Fin Fin lives, and to include more creatures. He also notes that they are working on a completely new game, which isn`t slated for introduction for at least two years. (Although Pontecorvo did not specify that this game would use adaptive technology, he did say that Fujitsu is working on such an implementation.)
NPCs With Character
Creating commercial titles in which the star players are based on AI is a recent development in the gaming world. A more traditional use for AI is to give non-playing characters (NPCs) intelligence and, on some level, personalities. However here, too, developers are beginning to push AI in new directions, employing more sophisticated AI than ever before and getting some fascinating results.
Dungeon Keeper from Bullfrog (Austin, TX) is an excellent example. In this game, the player takes the role of a bad guy, says Peter Molyneux, designer, producer, and programmer. (Note: Molyneux recently left Bullfrog to start his own company, which was unnamed at press time.) During the game, players build an elaborate, complex dungeon and fill it with monsters.
In Dungeon Keeper, AI is used to give the monsters character. Molyneux says it`s difficult to categorize the type of AI employed. "What we`ve had to do is take the traditional science methods, such as rules-based and neural nets, and throw away about 50% of that because it`s too processor-intensive and memory hungry to use. [The game can] have upwards of 300 creatures all processing intelligence at once. I`ve been working on AI routines since 1988, since my first game, called Populous. And all the games I`ve done since then have been hybrids and improvements to that AI. So it`s based on that."
Each monster has its own intelligence as well as individual personality. There are about 15 features designed into each monster, says Molyneux, including sight, memory, curiosity, fear, anger, sadness, happiness, and sociability. "The interaction of those variables then creates the individual personalities," says Molyneux.
Molyneux uses one of the monsters, a giant fly, as an example. "Flies are very curious," he says. "They will go out and explore, and they`ll remember where they`ve explored and communicate that to the other creatures and formulate maps of their environment. But they`re also very cowardly. So if there`s an enemy creature nearby or a corridor that`s completely dark, they will literally run away--and they will communicate that fear and caution to the other creatures and modify how they think of that area of the dungeon."
As the "boss," the player can affect the monsters` behavior by scolding and praising them. "For example, you can slap these creatures, and that causes them pain, and they remember those slaps," says Molyneux.
The monsters also will interact with each other, regardless of whether the player is present or not. Says Molyneux, "You can turn on the game and have a dungeon full of monsters, go out, and come back after an hour, and find they`ve explored areas and built things. They will get angry with you, too, if you don`t pay them or give them enough food. They might put graffiti on the wall or destroy the dungeon. If one of their buddies dies, they`ll get sad. There are a fairly broad range of emotions."
This May, Psygnosis (Foster City, CA) introduced Sentient, another title that uses AI to make its NPCs seem more human-like. Sentient is a sci-fi adventure in which the player is a medical technician sent to a space station to investigate an outbreak of radiation sickness.
On the space station are 60 NPCs, notes Julian Hicks, producer. "They all have objectives and missions and things they want to do. You try to persuade them to help you do what you want to do. But the NPCs tend to carry on about their own personal goals. And whether you`re there or not, they will have conversations and pass on bits of gossip and knowledge to each other."
Within the NPCs, there are a range of factions, notes Hicks, citing engineers and scientists as examples. Then there are groups with similar moral allegiances. ("The engineer who believes in what the scientists stand for," notes Hicks.) And then each NPC has a personal style and attributes.
The AI used to create these characters is based on a combination of techniques, says Hicks. Some is decision-based, he says, but it`s not that simple. "The decisions are derived based on stats, but we`ve also introduced a random element. Again, we`re trying to capture that human essence.
"The fact is, in most computer role-playing games, the NPCs don`t give any semblance of intelligence," he continues. "Say you`re in a bar and you throw your beer at the bartender one day. The next day you go back, and he`s just as happy to see you. That shouldn`t happen."
It won`t in Sentient. Each NPC has its own memory in terms of where it has been and what it has learned, says Craig Grounsell, one of the programmers. So if the player isn`t nice to a character, that character will remember that and treat the player with equal disdain. In fact, the NPC may pass on ill feelings about the player to the other members of its division, who also will then treat the player poorly. Additionally, NPCs have "feelings" about each other, which will determine whether or not they pass information to various characters.
The ultimate, adds Hicks, will be when a person is playing an online game and comes upon a character and can`t figure out whether it`s a real person or an NPC.
Dinos That Think
At DreamWorks Interactive (Los Angeles, CA), inspiring fear in the player is the ultimate goal. These folks are currently developing a game called Trespasser: Jurassic Park based on The Lost World: Jurassic Park movie and are using AI for the dinosaur characters. "We want you to look at the animal and feel like it`s thinking," says Seamus Blackley, producer. (Note: this title isn`t expected to ship until sometime 4th QTR. For more details on the CG in Trespasser, see the June CGW feature, "Interacting with Dinosaurs.")
Like most of the others, Andrew Grant, the title`s AI programmer, has used a "mishmash" of AI techniques to achieve this goal. "The way we work it is, every animal has its own little brain, which consists of a bunch of sub-brains that are each in charge of one important activity--eating, drinking, showing excitement, running away. These sub-brains are competing for attention in the way the animal expresses itself. So depending on how the dinosaur is feeling and its environment, different [sub-brains] will be dominant. For example, if there`s a T-Rex and it sees a goat, then its eating brain will come to the fore. If it`s hungry, then that will take over and the T-Rex will try to eat the goat. It`s a complicated system of balances, sort of a fuzzy logic sort of system."
Grant goes on to explain that if a Raptor sees a goat but also a T-Rex, its eating brain may want the goat, but its run-away brain is probably scared of the T-Rex and will cause the Raptor to scram. "But let`s say the Raptor hasn`t eaten in days and is really hungry. And it hasn`t had a bad experience with the T-Rex lately. If the T-Rex is farther away than the goat is, the Raptor might go for it," says Grant. "That`s really hard to do. I`d rather have it err on the side of running away usually, but what I`m shooting for--and what`s the hard part--is getting to the case when in the right circumstances, it will go after the goat anyway."
Sound amazing? As game developers see it, they`re only beginning to explore what`s possible with AI. In the next few years, 3D accelerators will take over not only the rendering process but also the geometry pipeline, freeing up more processing cycles without even counting the increased power of the microprocessor itself.
"It is inevitable that AI will get a lot more popular," says Molyneux, creator of Dungeon Keeper. "What we`ve done is to create these graphically rich worlds, and we`ve done quite an amazing job of it. And that`s all well and good, creating beautiful graphics. But unless you get some good AI routines that fill the game with reasonably intelligent characters, you`ll have a rather hollow experience."
Woodcock, AI consultant, agrees. "The last few years we`ve been working on the big 3D problem. Well, we pretty much have that solved now--basically--we know how to do it and do it well. So now the question is, what do you do to make your game different from everybody else`s?"
Mendel, star of the upcoming game Galapagos from Anark.
Mendel, star of the upcoming game Galapagos from Anark.
Pictured here is Mendel, star of the upcoming game Galapagos from Anark. Mendel is based on an adaptive technology developed by Anark called Non-stationary Entropic Reduction Mapping (NERM), which gives Mendel the ability to learn and adapt to his surroundings. Interestingly, Anark developed its NERM technology with more precise engineering applications in mind, such as industrial control.
DreamWorks Interactive is using AI to make it seem like the dinosaurs in its upcoming title, Trespasser: Jurassic Park, are thinking.
There are 60 non-playing characters (NPCs) in Sentient from Psygnosis.
There are 60 non-playing characters (NPCs) in Sentient from Psygnosis. And thanks to the AI in this game, these NPCs tend to carry on about their business, with or without player interaction.
It`s easy to see why kids love Fin Fin, the cute creature depicted here. AI gives him such real-life characteristics as happiness, sadness, and hunger.
Donna Coco is a senior associate editor of CGW.
: Computer Graphics World July, 1997