91�Թ���

Zach Sunberg

Zach Sunberg’s research developing better artificial intelligence systems is getting a major boost from two federal grant awards.

Sunberg is receiving a $599,000, five-year and is a partner on a related $4 million multi-university initiative from the U.S. Office of Naval Research.

Both projects focus on advancing game theory algorithms so AI systems can better solve problems in real-time in the field.

“I’m excited to receive the recognition that I’m looking at important problems to solve in both of these areas,” Sunberg said. “ONR shows the relevance for defense applications, and the NSF award focuses on making our nation and our world a better place.”

Sunberg is an assistant professor in the Ann and H.J. Smead Department of Aerospace Engineering Sciences at the 91�Թ���. His research focuses on autonomous systems and AI, with an emphasis on game theory.

Although game theory has origins in solving tabletop board games and card games, it is a broad field of research that studies problem-solving in complex, real-world situations.

“You might think of a game as a board game, but any situation where multiple agents are interacting and have their own goals can be mathematically represented as a game. Poker has clear rules, but so does driving a car; there are just a lot more variables,” Sunberg said.

AI systems like those used in self-driving cars typically rely on offline reinforcement learning. In such a system, automakers use historical data collected from a fleet of vehicles to optimize an algorithm to react to future situations. Sunberg is seeking to develop online decision-making systems, where an AI can think in real time to interact with situations that do not match historical data.

“This has previously been considered a computationally intractable problem” Sunberg said. “But our lab recently had a breakthrough with single-agent planning where we proved we did not need a lot of computation even in a large state space. What we want to do next is work on more complex multi-agent situations.”

The research focuses on a framework used by scientists and engineers to model possible outcomes when full data is not available, called partially observable Markov stochastic games.

“An application is airborne collision avoidance. In the past, the other pilot’s actions would be modeled as a probability distribution. That wasn’t satisfying to me. The other pilot is a decision maker themselves, so it would be better to model as a multi-agent game, but we don’t know how to solve partially observable games like that using online systems,” Sunberg said.

The research from the NSF grant has applications across an array of autonomous systems, from search and rescue robots, to self-driving cars, to how satellites navigate while orbiting the Earth. The Navy award is focused more on AI deception and counter deception in the military realm.

“An enemy is going to try to deceive you in some way, so we want to focus on how that can happen and how do we make AI resistant to it. We’re also looking at developing AI that can deceive an adversary. If you have a drone that you want to avoid enemy air defenses, how can it use bluffing to help it do that,” Sunberg said.

The ONR award is brings together four universities. The project is being led by the Georgia Institute of Technology with partners at the University of California Santa Barbara, 91�Թ���, and the University of Texas at Austin. Sunberg’s portion of the $4 million grant is worth roughly $1 million.

“The current most widely used methods for AIs, these offline systems, are really a function approximater. It’s kind of an intuitive reaction or instinct. We want an AI that can go further, like people, and think and deliberate about a situation. There’s huge potential with this work,” he said.