Swanson School professor using video games to test artificial intelligence


It can be difficult to test artificial intelligence. An untested program controlling a major project could create serious consequences.

To get around this, Daniel Jiang, assistant professor of industrial engineering at the Swanson School of Engineering, uses video games to test the complex algorithms he develops.

Daniel JiangJiang’s research lies in operations research, approximate dynamic programming, reinforcement learning and risk-averse decision making.

These are tools used for AI to help Jiang and other researchers predict future outcomes for decisions made in fields such as health care, energy, supply chains, inventory management and more.

“The general issue with this is that not only do you have to think about the future, but you also don’t really know much about the environment that you’re operating in,” Jiang said of uncertainty.

The future outcomes that may arise in these fields also can evolve over time, and that requires AI to evolve too.

However, “we don't want to test our algorithms in real life for these types of applications,” Jiang said, adding that using untested AI to manage the applications could have serious consequences. But the complex, simulated environments in video games offer a solution.

“At the same time, it’s very, very hard to build an accurate environment in the computer that really represents the real world well,” Jiang said. “Then sometimes we’ll look for these fake simulated environments to test things in, such as these video games, which hold all sorts of challenges that are basically modeled after the real world but without any real consequences if our algorithms fail.”

Jiang uses games similar to “League of Legends” or “Heroes of the Storm” — popular multiplayer online battle arena, or MOBA, games where players use a character to destroy an enemy stronghold.

Jiang said he uses MOBAs, as opposed to other game genres, because they require players to make many decisions to win at a later time. The mathematical term for these decisions where the outcomes aren't known until a later time is a “long horizon” or the “credit assignment problem,” Jiang said.

Jiang used a method called the Monte Carlo Tree search, where an algorithm creates a “decision tree” with several “branches” of decisions and outcomes.

The algorithm essentially tries several random actions until the desired outcome is achieved. Then, the algorithm begins to record and use the decisions that lead to desired results.

“For the MOBA games, you have to think far, far ahead into the future because you’re actually doing something that you get a reward for much later,” Jiang said. “You don't win the game until minutes or hours later. … So that poses a big challenge. And that somehow becomes more interesting for more difficult for the algorithm and more interesting to study.”

Jiang said he’s excited about the increased interest and investment in AI, however, there’s still much work to be done.

“There's a ton more research that we need to do in the area of algorithm design, before we have something that really works reliably, and in the real world,” Jiang said.

Donovan Harrell is a writer for the University Times. Reach him at dharrell@pitt.edu or 412-383-9905.