Leela Zero is a green-field effort to reproduce and explore the methods used by Alphago zero which I've been following.
http://zero.sjeng.org/ It seems to be having some success, and it's open source. The general archecture it uses is
a NN similar to the one described for Alphago zero, which guides the candidate selection for a MCTS search. Training
for the network is being done by crowd sourcing volunteers to run self play games, which are uploaded and batched
into the training algorithm, and as better networks are trained, they are downloaded and used by the clients.
It's all in the sources somewhere, but it's not clear exactly how the training works. What is uploaded ? Just the game
record of the self-play games? Presumably some back-propagation is done through the network for each move of the
game, but exactly what is not clear.
1 post • Page 1 of 1