Neural Network Architecture

[DailinH]

This neural network architecture is quite different from that in Alphago Zero's paper, for instance, the latter took a resnet approach, using 1 convolutional block and 19 residual blocks. Simply stacking layers may cause certain defects(e.g. speed and accuracy) in network training.

[junxiaosong]

The simple neural network used in this repo is for illustrating purpose, and it also works at least to some extend. If we use the same deep network architecture as in AlphaGo Zero paper, it will be even slower on our personal laptop. And in fact you can easily modify the neural network architecture, for example using resnet or BN, to see how they affect the training speed and overall performance.

[DailinH]

If we attempt training using a 15*15 board, would it be better to enlarge the number of simulations in MCTS? Also, using temp=1may cause too many explorations, especially after early stages, thus weakening the actual playing result.