Computer programs that play the game of Go have advanced to the point where they can beat professional human players, as demonstrated by recent victories at the Taiwan Open and a Chicago exhibition. These programs employ variations on mathematical methods originally devised by Manhattan Project physicists to shape order from pure randomness. However, the programs' reliance on number crunching ensures that few insights into the function of the human mind can be derived from their performance, says Dartmouth College programmer Bob Hearn.
The strategy that artificial intelligence experts historically followed in the development of their Go programs was to attempt to tap pattern recognition principles, but David Doshay with the University of California at Santa Cruz says that guiding computers with human-rules patterns was erroneous from the very start. By harnessing the Monte Carlo method, which consists of random simulations repeated over and over until patterns and probabilities emerge, Go programs became more capable. Crunching the accumulated statistics enables probabilities to take shape, and this information allows the programs to channel more processing power to promising branches and less power to alternatives with less promise.
Hearn and others predict that Monte Carlo-based Go programs will continue to improve now that they have started to beat professional players. They project that within no more than a few decades such programs will be capable of trouncing the top players.
From Wired News
View Full Article