Computer programs for Chess are already quite adept, playing at a level that exceeds the best human grandmasters. Currently, though, they work on a brute-force approach, using raw processing power to evaluate all possible subsequent moves. That may change.
Matthew Lai, a graduate student at the Imperial College of London, has developed a Chess engine that, using neural networks, applies pattern recognition techniques to narrow the decision tree. The program, which Lai named Giraffe [PDF], honed its pattern recognition skills by playing against itself for a period of 72 hours and comparing the results of each decision to a database of Chess games.
Now after considering the global state of the game, piece-specific factors, and a map of the spaces threatened and defended, Giraffe is able to predict the best move 46 percent of the time and select a top-three that includes that best move 70 percent of the time, all without looking ahead. This rates Giraffe at about the level of a FIDE International Master.
The beauty of this approach is in its generality. While it was not explored in this project due to time constraint, it is likely that this approach can easily be ported to other zero-sum turn-based board games, and achieve state-of-art performance quickly, especially in games where there has not been decades of intense research into creating a strong AI player.
[via MIT Technology Review]