GitHub Source, executable files. GitHub(Japanese top page)
2019-07-09 v1.4 Update required. Random seed for visit count sampling was constant.
2019-07-08 v1.3 Tree reuse. PV and score is available on Shogidokoro.
2019-05-29 v1.2 Update required. MCTS initial value is not draw(0), but loss(-1).
|In past hour,||number of clients are 16,||576 games.|
|In past 24 hours,||number of clients are 30,||14679 games.|
|In past 1000 games,||Average of moves 123.9,||Sente winrate 0.550|
|In past 500,000 games,||Average of moves 126.6,||Sente winrate 0.538|
AobaZero 800playouts/move vs Kristallweizen 100k/move. 800 match games.
These are self-play games for training. It often plays blunder for the first 30 moves.
And sometimes it choose a move that is not a best by adding noise on root node.
From no000000000000.csa to no000000121031.csa are generated by not using neural network, but random function. The first game that is generated by neural network is no000000121032.csa no000001017999.csa. Up to here, 64x15block, window is past 100,000 games. no000001018000.csa. From here, 256x20block, window is past 500,000 games.Weights
w001 ... 64x15b, 64 minibatch, learning rate 0.01, wd 0.00005, momentum 0.9, 120000 games w156 ... 64x15b, 64 minibatch, learning rate 0.001, wd 0.00005, momentum 0.9, 430000 games w449 ... 256x20b, 64 minibatch, learning rate 0.01, wd 0.0002, momentum 0.9, 1018000 games w465 ... 256x20b, 64 minibatch, learning rate 0.001, wd 0.0002, momentum 0.9, 1180000 games Weights are updated each 2000 games ( 4000 iterations) up to w448. Weights are updated each 10000 games (20000 iterations) from w449.