Deep learning for… Go · Erik Bernhardsson

A 2014 blog post discussing a paper by Christopher Clark and Amos Storkey on using convolutional neural networks to play Go. The paper demonstrates that a move prediction model with no deep search beyond one level can beat certain Go engines. The author speculates on how to extend this approach by combining move prediction with an evaluation function (as done in chess experiments) and plugging it into a Monte Carlo tree search framework, predicting this could produce the world's best Go engine.