Deep Learning with Shogi AI on Mac and Google Colab Chapter 9
Learning technique
SGD
lr = 0.01
Momentum SGD
lr = 0.01
Learning stopped on the way. Was lr too big? (In the book, lr = 0.005.)
SGD , Batch Normalization
lr = 0.01
Accuracy was improved by Batch Normalization.
The details of Batch Normalization are not written in the book, so I studied by looking at the following article. https://qiita.com/omiita/items/01855ff13cc6d3720ea4 -Batch nom is a normalization of the value before passing through the activation function to an average of 0 and a variance of 1 for each mini-batch and for each same channel. ・ The reason for improvement with batch nom is unknown. I understood.
Surprisingly, shogi is shit weak. What is the cause?
Game video https://youtu.be/9YBImGLzm1w
Final figure like shit
Recommended Posts