Christopher Clark and Amos Storkey wrote an interesting nine page article titled “Teaching Deep Convolutional Neural Networks to Play Go”. Their deep neural network correctly predicted the moves of experts on a 19×19 Go about 44% of the time. The previous record was 41% by Wistuba and Schmidt-Thieme in 2012. Furthermore, the Clark Storkey network was able to “consistently defeat the well-known Go program GNU Go.” This is the first time that a neural network was able to perform nearly as well as one of the better hand coded programs. It is still not as good at the better UCT programs, but it moves much more quickly than the UCT programs. I imagine that if there were a blitz version of computer Go, the Clark Storkey AI might win a computer competition.
The article reviews other recent attempts to train a neural network to play Go. The Clark Storkey network resembled the Wistuba Schmidt-Thieme network, but it had more 19×19 convolutional layers and the authors added one fully connected layer at the top before the final move decision. Also, known symmetries of the solution were hard-coded. Interestingly, they found that convolution seemed to be required.
“We briefly experimented with non-convolutional networks but found them to be much harder to train, often requiring more epochs of training and the use of approximate second order gradient descent methods, while getting worse results.”
Later they describe their training methods and network architecture as follows
“Networks were trained with mini-batch gradient descent with a batch size of 128, using a learning rate of 0.01 for 7 epochs, and 0.05 for 2 epochs which took about a day on a Nvidia GTX 780 GPU.”
“The best network had one convolutional layer with 64 7×7 filters, two convolutional layers with 64 5×5 filters, two layers with 48 5×5 filters, two layers with 32 5×5 filters, and one fully connected layer.”
They estimate that their AI would probably have a ranking near 4-5 kyu.