University of Edinburgh researchers have applied machine-learning techniques to the problem of finding the next move in a game of Go.
Conventional Go algorithms play out the entire game after every move, and if the computer wins in the majority of these simulations, then that move is deemed a good one. However, that approach is time-consuming and computationally intensive. In addition, traditional Go algorithms fail to beat human Go experts, who can usually evaluate the state of a Go board with just a glance. Humans are better at Go because of the ability to spot strengths and weaknesses based on the shape the stones make rather than by looking several moves ahead.
Recent advances in pattern-recognition algorithms could help computers do much better at playing Go. The Edinburgh researchers used a vast database of Go games to train a neural network to find the next move. The researchers used more than 160,000 games between experts to generate a database of 16.5 million positions as well as the next move. They then used almost 15 million of these position-move pairs to train an eight-layer convolutional neural network to recognize which move the expert players made next. The researchers found the trained network was able to predict the next move up to 44 percent of the time.
From Technology Review
View Full Article
Abstracts Copyright © 2014 Information Inc., Bethesda, Maryland, USA