“One of the last bastions of human mastery over computers is about to fall to the relentless onslaught of machine learning algorithms,” according to a December 15 report in the MIT Technology review. Why Neural Networks Look Set to Thrash the Best Human Go Players for the First Time reviews the work of Christopher Clark and Amos Storkey at the University of Edinburgh in Scotland, who “have applied the same machine learning techniques that have transformed face recognition algorithms to the problem of finding the next move in a game of Go.”
“The question that these guys have trained a deep convolutional neural network to answer is: given a snapshot of a game between two Go experts, is it possible to predict the next move in the game?…Clark and Storkey used over 160,000 games between experts to generate a database of 16.5 million positions along with their next move. They used almost 15 million of these position-move pairs to train an eight-layer convolutional neural network to recognize which move these expert players made next…the trained network was able to predict the next move up to 44 percent of the time, ‘surpassing previous state of the art on this task by significant margins.’”
After just a few days training, Clark and Storkey’s neural network beat GNU Go almost 90 percent of the time in a run of 200 games, but against Fuego 1.1, it fared less well, winning only just over 10 percent of its games.
“There is no suggestion from Clark and Storkey that this approach will beat the best Go players in the world,” the report concludes. “But surely, it is only a matter of time before even Go players will have to bow to their computerized overlords.”
Thanks to John Goon for passing this along.