In 1997, IBM’s “Deep Blue” defeated Garry Kasparov and became the first machine to win a chess tournament against a world champion.
Deep Blue’s game strategy is a testament to the computer’s ability to sift quickly through huge amounts of information and calculate a successful course of play, using algorithms created by humans.
Héctor Muñoz-Avila, assistant professor of computer science and engineering, wants to take computers a step further.
With a five-year CAREER Award from NSF, Muñoz-Avila hopes to enable computers to “learn” as humans do – by taking a set of inputs, evaluating past outcomes of similar scenarios and making decisions based on experience.
“Many cognitive scientists believe humans begin acquiring knowledge by mastering simple skills and combining them to learn more complex ones,” says Muñoz-Avila. “With chess, we start by learning how to move individual pieces. Then we learn simple strategies, such as opening moves and basic positioning. We build incrementally on our expertise until we master complex game-playing strategies based on the integration of basic skills.”
Muñoz-Avila is seeking to develop a “unified architecture for automated learning of skill hierarchies” from a collection of examples, such as outcomes of previous chess games.
“The architecture incrementally learns simple skills by mimicking a few examples,” he says. “As more examples are given, it learns complex skills that build on the simpler skills it learned previously. The goal, after many examples, is for the computer to be able to generalize these inputs into abstract, complex concepts and apply them to new situations.”
This approach is a departure from the application of computational power and algorithms to a game or problem, says Muñoz-Avila. And it is potentially more powerful than the “search/retrieve/compute” power of systems like Deep Blue.
“Deep Blue and other searchintensive approaches develop strategies differently from humans,” he explains. “Our goal is to build algorithms that resemble the way humans learn and solve problems.
“This line of research will be capable of developing effective strategies and explaining the reasoning behind them. The latter is of crucial importance in areas such as teaching and decision support, where providing justification for the solutions is as important as providing the solutions themselves.”
