“After playing this game for a few days, I'm totally hooked,” a player wrote on the message board on the game’s Web site. “This game's got my intellectual side by the cerebral cortex and it's squeezing harder by the minute!” In the game, called NERO, players command armies of robot-soldiers and train novice soldiers in the skills needed for battle. They can be trained to go after the enemy, avoid the enemy, run around walls, navigate through mazes and perform other tasks. The soldiers’ learning doesn’t end there, but continues to grow during the game. The best ones reproduce more well trained soldiers while those that perform poorly are eventually replaced. The soldiers can be saved for use in future games. “As they are training, their brains are growing more and more sophisticated and able to figure out the problem you’ve set for them,” said Aliza Gold of the Digital Media Collaboratory (DMC) in the university’s IC2 Institute and the game’s producer. Players act as drill sergeants, training different teams of soldiers and choosing which trainees they want to use in each battle. For example, a player can mix and match soldiers trained for different general skills. That ability to learn and evolve into better soldiers is what sets the game apart from other video games. It’s the result of high-level research in artificial intelligence and machine learning conducted in the Neural Networks Research Group (NNRG) in the Department of Computer Sciences in the College of Natural Sciences.
NERO stands for NeuroEvolving Robotic Operatives. At its heart is an algorithm developed by Dr. Kenneth Stanley, who earned his doctorate as a member of the NNRG, which is headed by Dr. Risto Miikkulainen, professor of computer sciences. Stanley’s area of study is neuroevolution, which is programming a machine to form small simulated brains called neural networks through an evolutionary process. The game itself evolved from a conference the DMC sponsored on artificial intelligence and gaming in 2003. Stanley presented his idea for the game to the conference and a couple of weeks later, the DMC called with an offer to produce the game—to his surprise. “I proposed something that was kind of pie in the sky, I wasn’t certain it was going to work,” he said. “Fortunately, it turned out everything fell into place and worked as a technical matter.” Stanley had developed an algorithm called NeuralEvolution of Augmenting Topologies (NEAT) for controlling such things as robots. But a game required a faster learning process than NEAT afforded. Stanley’s answer was a real time version of NEAT called rtNEAT. Gold took the lead in producing the game, organizing a group of mostly volunteer programmers, artists and others to create it.
After 18 months of development, the game was posted on the Internet and quickly built a following. It has been mentioned on Slashdot.com and a blog on the MIT Technology Review Web site and others. While the result is a game that’s fun to play, it’s based on solid, serious research. “The whole game itself is a scientific experiment,” said Stanley, who calls himself a casual game player. “We want to show off our research to the public and to the scientific community.” The game has won plaudits in both arenas. A paper describing the game won the Best Paper Award at the IEEE 2005 Symposium on Computational Intelligence and Games. The university is seeking a patent for the rtNEAT technology and seeking commercial applications in games and other uses. Stanley will become an assistant professor in computer sciences at the University of Central Florida in January 2006. Tim
Green Related SitesDr. Kenneth Stanley
|
