boston.com Business your connection to The Boston Globe
ENTERTAINMENT 2.0

Game designers test the limits of artificial intelligence

Growing up in Gardner in the 1980s, Michael Zarozinski used to stop by the bowling alley after school to play Pac-Man, Q*bert, and Donkey Kong. Once he'd fed a few hundred quarters into the machines, though, he started noticing a central limitation.

"As soon as you learned the pattern the computer characters moved in, the game was beatable," he says. "The experience didn't grow with you."

More than two decades later, videogame developers are still hacking away at that same problem: how to create more intelligent, more realistic characters, whether they're foes to fight or allies who'll assist a player. And Zarozinski has joined their ranks, working for a Shrewsbury company called All inPlay as the company's "head geek."

"A lot of the most interesting work in artificial intelligence is being done by game developers," says Bruce Blumberg, senior scientist at Blue Fang Games in Waltham, and formerly a professor at MIT's Media Lab. "You have really bright kids who are dealing with problems they don't realize are insoluble. They're very motivated."

But those bright people writing software that better mimics human intelligence are running smack into an interesting new trend: online virtual worlds like Second Life, and "massively multiplayer online games" such as World of Warcraft, in which characters are inherently pretty intelligent because they're being controlled by other humans. Players in those games design their own characters and then enter the game realm to interact with one another. Their strategies and dialogue seem authentic -- because they are.

"In some ways, all of these massively multiplayer games have shone a light on the deficiencies of artificially intelligent characters in games to this point," says Hank Howie, chief executive of Blue Fang. Howie's company makes Zoo Tycoon, distributed by Microsoft, in which players build and operate a virtual zoo. Now the company's focus is on using artificial intelligence to create more believable animals for future products.

In Empire Earth III, a strategy game designed by Mad Doc Software in Lawrence, the objective is to guide a young civilization to greatness. Of course, that involves a lot of vanquishing, and artificial intelligence (AI) software helps shape the strategy of the opposing army.

"It looks at how the player is playing, what he is doing, and it comes up with a counter strategy," says Mad Doc CEO Ian Lane Davis. "Instead of just reacting to the units it sees on the ground, it thinks, 'What is the human trying to do?"'

But Davis says the limitations of today's AI can put creative shackles on game designers. One major barrier is using language to interact with a character. "You can't talk to characters and expect a response that feels real," Davis says. "So there are no games that are like detective stories, or romances, which are popular genres in the movies, because you can't interview suspects or talk to other people."

The challenge game designers are grappling with today is the same one that the British mathematician Alan Turing posed in 1950: writing software that could easily be mistaken for a human. "You're always aiming to create something where the players won't be aware that they're dealing with an AI character," says Michael Gesner, founder of Dragonfly Game Design in Westborough.

One way to get there is by having humans "train" the AI software. That's the approach that game designer Jeff Orkin, now a grad student at the Media Lab, is taking. With a project called The Restaurant Game, Orkin invites players to assume the role of a restaurant's wait staff. His plan is to capture their behavior and dialogue, and use it to build more realistic software-driven characters, in the same way that designers sometimes use motion capture cameras to record and replicate human movement.

"Ideally, AI systems in the future will observe as designers directly control characters, and learn to play roles and even converse," Orkin writes via e-mail.

Of course, it's hard to ignore the success of massively multiplayer games. World of Warcraft, one of the oldest, said in March that it has 8.5 million paladins, priests, and druids pursuing quests and casting spells. Market research firm Screen Digest estimated that the global market for massively multiplayer games hit $1 billion last year. Fueling that growth is the fact that massively multiplayer games demand an ongoing monthly subscription fee rather than a one-time software purchase.

"It's like the gym subscription model," says Brett Close, CEO of 38 Studios in Maynard. Once players have built a character and "leveled it up," acquiring skills and weapons, they're unlikely to cancel a subscription -- and sometimes they even forget about it.

But even though most game developers may be eager to shift to that kind of business model, there isn't a sense in the industry that massively multiplayer games, with their hordes of human-guided characters, will necessarily render the pursuit of better AI irrelevant.

Close and others envision a hybrid, in which AI-driven characters help advance a story. While massively multiplayer games have been successful, he says, "nobody is successfully delivering episodic content, and a compelling story, so that you understand that what you're doing in the game is affecting the story." More realistic AI characters could serve as guides, he suggests, "taking you through certain situations," or act as wily opponents. 38 Studios, with 40 employees, has as its major investor Red Sox pitcher Curt Schilling, an avid gamer. The company's first game will be based on the stories of fantasy writer R.A. Salvatore.

"You need AI characters in tandem with yourself and your few human party members," Davis says. But those AI characters, he acknowledges, will need to "have a mood, a goal, and something they're focused on. We need to give them the right body language and facial expressions so they seem more human."

In Newburyport, Muzzy Lane Software said recently that it is partnering with Harvard historian Niall Ferguson to create a new game, due out next year, that will allow players to simulate modern-day conflicts like the wars in Iraq and Afghanistan, or clashes involving Iran or North Korea. CEO Dave McCool says the AI software will observe how humans play, and "learn how to beat you."

Truly evolved AI, of course, would figure out a way to avoid taking up arms in the first place. But that'd be better-than-human AI -- still a few years off.

Scott Kirsner is a freelance writer in San Francisco who maintains a blog on entertainment and technology, cinematech.blogspot.com. He can be reached at kirsner@pobox.com.

SEARCH THE ARCHIVES