tanuki- 2018-02-04 NNUE評価関数+雑巾絞り
- 棋譜生成
- tanuki-棋譜生成ルーチン
- 評価関数
- NNUE p-k-halfkp-192-32x2 2018-01-16
- p-k-halfkp-192-32x2.1\55
- 開始局面
- roots.hcpより選択
- 選択された局面からランダムに1手指した局面から開始
- 探索深さ 6
- 自己対戦を打ち切る評価値の絶対値: なし
- 入玉宣言勝ちも学習データに含める
- 生成局面数 10億
- シャフル
- discountあり
- discountには勝率を使用した
- 機械学習
- NNUE機械学習ルーチン
- 学習手法 SGD ミニバッチ法
- 学習率 1.0
- 学習率減衰 0.5
- ミニバッチサイズ 1000
- 元の評価関数
- 検証データ: 学習データとは独立に生成した100万局面から1万局面抽出したものを使用する
- 学習局面数 50億局面
実験結果
実験1
discount絞り
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55engine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b2000,269 - 20 - 711(27.45% R-168.85) win black: white = 50.92% : 49.08% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\p-k-halfkp-192-32x2.7engine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b1000,311 - 14 - 675(31.54% R-134.62) win black: white = 54.56% : 45.44% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\p-k-halfkp-192-32x2.7engine2=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval2=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55T1,b1000,510 - 24 - 466(52.25% R15.67) win black: white = 51.84% : 48.16% declaration win engine1=0 engine2=0
実験2
elmo絞り
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\.11\finalengine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b2000,308 - 13 - 679(31.21% R-137.33) win black: white = 50.76% : 49.24% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\.11\finalengine2=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval2=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55T1,b1000,514 - 21 - 465(52.50% R17.40) win black: white = 52.60% : 47.40% declaration win engine1=0 engine2=0
実験3
slide絞り (Qhapaqさんの言っていた、N手先の評価値を教師信号に使うと良いという話の再現実験)
N=10
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.12\finalengine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b2000,283 - 11 - 706(28.61% R-158.81) win black: white = 47.93% : 52.07% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.12\finalengine2=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval2=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55T1,b1000,497 - 26 - 477(51.03% R7.14) win black: white = 54.62% : 45.38% declaration win engine1=0 engine2=0
実験4
median filter適用後に雑巾絞り 半径=3
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.13\finalengine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b2000,278 - 19 - 703(28.34% R-161.16) win black: white = 51.89% : 48.11% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.13\finalengine2=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval2=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55T1,b1000,497 - 28 - 475(51.13% R7.87) win black: white = 51.23% : 48.77% declaration win engine1=0 engine2=0
実験5
雑巾絞り
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.14\finalengine2=C:\home\nodchip\hakubishin-private\exe\YaneuraOu-2017-early.exe eval2=C:\home\nodchip\hakubishin-private\exe\eval\apery_sdt5_eval_twig_formatT1,b2000,297 - 21 - 682(30.34% R-144.41) win black: white = 49.74% : 50.26% declaration win engine1=0 engine2=0
engine1=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval1=D:\hakubishin-\eval\nn.2018-01-29.14\finalengine2=C:\home\nodchip\nnue\source\YaneuraOu-by-gcc.exe eval2=D:\hakubishin-\eval\p-k-halfkp-192-32x2.1\55T1,b1000,517 - 20 - 463(52.76% R19.16) win black: white = 51.02% : 48.98% declaration win engine1=0 engine2=0