The perfect challenge

Computer science and engineering researcher seeks to develop an optimal opponent for educational simulations

Sushil Louis

On first glance, natural selection and computer simulations may seem to be about as far apart as you can get in STEM fields.

After all, one is a biological process that happens without human intervention or direction and the other relies heavily on man-made design and manipulation.

But Sushil Louis, professor of computer science and engineering, is using inspiration from biological evolution, mixed with a heavy dose of computer science theory, to design games and simulations that can evolve along with their opponents.

"If you think of biological evolution as searching for fit structures that are good at performing certain tasks, then life on earth is good at doing all the niche tasks that they've evolved to do," Louis said. "We try to model that."

Louis' current research focuses on evolving tactics that can be used in Navy training simulations. The Navy is already using software developed by his research group at the Surface Warfare Officer School in Newport Rhode Island, where naval officers are put through a series of high-fidelity simulations that test their ability to handle various situations.

"It teaches them good decision making," he said. "The idea is to put them as close as possible to the real thing. Hopefully when the real situation arises it's something they've been trained to do so often that they know exactly what to do."

Louis designed an interface and the navigation artificial intelligence, or AI, that allows instructors at the school to easily manage complex and varied training scenarios that change depending on how a student in the simulator reacts to a given set-up. That interface is used daily in naval training scenarios. And the navigational AI provides a near-optimal solution to a previously unsolved research problem in optimal control

The next step is to develop a more autonomous program that can mimic the tactical approach an enemy might take.

"The research part of what we're doing now is coming up with tactics" he said. "So you just tell a group of small boats to attack, and they'll figure out how to do it best."

That research relies on a particular theoretical approach within artificial intelligence called co-evolution.

"Co-evolution is well-known in the genetic algorithm community, but co-evolution in games is a much smaller community, and we are one of the leaders in that area," Louis said.

Louis' research evaluates the evolutionary fitness of randomly generated tactics in certain pre-determined scenarios by playing tactics against each other and probabilistically weeding out lower-performing tactics. As better and better tactics develop on one side, the other side must respond.

"It's like an arms race," Louis said. "We breed tactics for one side, which results in good tactics for the other side to learn to beat, and then the first side learns to beat the other side's new tactics, and then the other side learns to beat the first side's even better tactics and so on. The two sides keep going until you come up with a set of tactics that seem to work well in a variety of situations for both sides."

Because the games are used for training and education purposes, Louis' research also draws on education theory, which suggests that for the most learning to occur, an exercise must be optimally targeted to an individual's competence, providing the right amount of challenge without being overly frustrating or too easy.

"If you play somebody who is about equal or just slightly better, it challenges you to the maximum," Louis said. "That's the place where you learn the most, it motivates you the most."

Louis hopes to continue his work in educational simulations by developing games for children.

"I'm interested in applying our research for teaching STEM concepts to kids, because games and simulations are attractive to kids and motivational to kids," Louis said. "We may not even be playing a game, it might be something else, but the idea of adapting to your opponent, in this case a child's level of knowledge to optimize learning, can remain the same."

Take the next step...