GDC: What's next for video game AI?

Three next-gen AI demos at Wednesday's Game Developer's Conference showed off things that would seem like every day occurrences. But if their creators can get them to work, this distinct mundanity is a crowning achievement.

Richard Evans of Maxis and EA fame was the AI lead on the Sims 3. His demoed two new types of AIs, the first--called "Sim Tribe" allows developers to create their own societies. These societies can work on the same social rules as real cultures, meaning that if a player ventures into a different location, the other nonplayable characters (NPCs) will change their habits accordingly.

In terms of the demo, this paradigm shift was with eating, which became a social taboo. This acted out comically as characters could not eat until everyone else was outside of viewing distance. Evans joked that you could also reverse the rules so that going to the bathroom in public became an every day activity. What made the entire system more interesting was that there were also societal punishments built in, so that if players or other NPCs alike disobeyed the social norms, it would change how other characters interacted with them.

Evans' second demo was something equally ambitious that gave NPCs "very long term plans." The idea is that developers can give NPCs hundreds of actions that the characters can (and want) to do in their every day lives. This is as opposed to the three to four actions Evans said most developers will program out for an NPC.

In terms of the demo, that kind of scale panned out to a character named "Bob" whose character type was a workaholic. Not only did he work long hours at a nearby restaurant, he also wanted to be promoted. In order to get that goal, Bob proceeded to befriend his boss, buy cookbooks, take cooking classes, and practice cooking. When his apartment got dirty, he even hired a maid so as to save time he would have had to spend doing it himself, leaving him free to learn more skills.

What made the demo really impressive though, is that Bob's actions were working in a world full of other NPCs doing the same thing and with different programmed personalities. As Evans explained it, all that was needed to make this really work was to figure out some of the steps needed in getting from a desire to a goal and mapping those things out within a social structure. From there the NPCs will simply go about their lives doing these actions and interactions to reach whatever goals are set.

Computational biologist Ian Holmes from UC Berkeley had a very different layer of AI to show off from Evans. Instead of going for social normality with people, he went with animals and the environment they live in. Holmes demoed something called Zoo Gas, which has players trying to create and maintain a zoo. This involves keeping certain animals apart, as well as breeding them, and keeping them away from tourists with a fence that begins to decay over time.

Where Holmes sees the title going is toward a massive multiplayer online game (MMO) in which players get their own chunk of land in a larger online world. It will be their job to breed certain types of creatures, as well as work with others to create new types. All the while they'll be fighting off things like disease, mutation, overpopulation, and equality with a human population.

Though AI isn't just about creating people going to work and fornicating animal life. UC Santa Cruz professor Michael Mateas and his colleagues are attempting to create enemy AI for the popular video game StarCraft, that's less of a pushover than what comes with the game. "With current real time strategy AI, moderately skilled humans can kill the computer," Mateas said. "We're interested in getting close to human level competence not because it's interesting but to start thinking about creating StarCraft agents who play with a certain style."

Mateas wants to make it possible for the computer to emulate the playing style of specific people, as has been done classic games like Go and Chess. This would let someone play a champion right out of the box and maybe even learn a few of their tactics. All without ever having to hop online, or enter in a tournament. Mateas even wants to create a system where users can dial down that competitor's AI, so as not get beaten in just a few minutes.

In order to do all this Mateas and two doctoral students have created an AI system that runs specific parts of StarCraft's game functions but works as one so that each component is aware of what the other one is doing. This keeps any one resource from burning the other, be it displacing troops, or not gathering enough resources.

This is all being done behind the scenes in a program that runs in the background. According to Mateas, they recently got the legal approval from StarCraft-maker Blizzard to use it in competition. Thus far, they've been running it on various StarCraft servers and seeing if it can beat human opponents and are winning one of of every five matches. Mateas wants to get that higher and is using a mix of machine learning and data mining to figure out more complex strategies.

Mateas' second demo showed a system that could have big implications on social gaming. In this case, it was three characters planning a "counterculture prom." Each of the characters had specific relationships with one another. These could be affected by conversation that would change dynamically based on how other characters had interacted with each other in the past. For the demo, it was a love triangle of sorts. One character wanted to ask the other one out, but he knew his friend did not like that person.

In order to get over the love triangle problem, Mateas demonstrated two solutions. One with the player attempting to make the two characters who did not like each other, like each other enough through conversation so that a date could be made without offending any parties. The second had the same set of characters, with one of them simply insulting the other character into not being their friend any longer, so that asking that other character out would not cause a problem.

Mateas said he wants to scale the system up enough to be workable in even the most basic social games. The brunt of the work is actually building out each character's knowledge base with things that happened to them in the past and enough dialogue options. The reward, Mateas said, is "endlessly ramifying social possibilities." In other words, a game that's different every time and that learns as it goes.

So what are the chances you'll see any of these AI systems in games of the future? Pretty good. Holmes has shared the source code for Zoo Gas here. And Mateas' StarCraft AI bots are a part of a tournament being held at UC Santa Cruz in October that could end up making them even better. As for Bob the workaholic, considering Evans' Sims cred, the quirky character could be one of the many NPCs you end up competing with in a future edition of the game.