OREGON STATE UNIVERSITY

Beyond computerized chess – teaching computers more complex strategies

02/26/2010

CORVALLIS, Ore. – Someday when military leaders are planning a battle or air traffic controllers are trying to land a dozen aircraft at once, they may benefit from studies now being done at Oregon State University – experts explaining how they play a cool video combat game in order to help artificial intelligence researchers build better virtual combatants.

The idea is to have experienced players of real-time strategy games, such as computerized combat, explain to a novice what they are doing, and why, and how it might work. These “think out loud” approaches allow researchers to design better interfaces and knowledge representations for computers.

Findings on this work were just published in Knowledge Based Systems, an academic journal, by Ron Metoyer, an associate professor in the OSU School of Electrical Engineering and Computer Science, and Simone Stumpf at City University London.

The research is being supported by the Defense Advanced Research Projects Agency, the research and development office for the U.S. Department of Defense which – among other things – helped create the Internet.

“We had groups of people playing Wargus, an open-source video game that’s very involved,” Metoyer said. “As they played, we asked them to explain their actions to a novice, while we did both audio and video recordings. Our goal was to find out the tricks the best human players use and let machines learn from them.”

Such “real time strategy,” researchers say, is applicable to military operations, air traffic control, emergency response team management, or other demanding tasks in which many different elements have to be considered and decided quickly. While playing the computer games, people have to determine how to allocate their resources, where to place armies, how to time their battles, deal with uncertainty, what units to create, and other challenges.

In this exercise, the expert players were soon chatting about how to place their archers, build a farm, anticipate the enemy attacks, be aware of surroundings. The player warns, “I wanna not cut down those trees.” Translated, the challenge is how to tell the computer which objects are part of the environment that should be left alone, as well as when and where to do something in a specific game situation.

“These are pretty complex games and there’s a lot going on at once,” Metoyer said. “Trying to build an army, feed it, form communities and carry on battles demands a lot of strategy and coordination. Very good game players can do that, and we think we can tap into these human insights to help machine learning.”

Previous approaches, he said, were often based on computers playing the games multiple times and trying to figure out strategy on their own. The new concept uses traditional human-computer interaction techniques to inform machine intelligence, and helps the computer learn more quickly.

And it should be possible, Metoyer said, to even pay back the favor a little - and create more fun video games.