tanuki- 2017-10-14 depth 10 epoch 2
- 棋譜生成
- 評価関数
- 2017-10-20-00-18-32 評価関数
- 開始局面
- まふ氏提供定跡学習用定跡データベースから選択
- 1～71手目からランダムに局面を選択
- 選択した局面から0～6手、MultiPV5の中から等確率に指し手を選択
- MultiPVで選択した指し手も教師局面に含める
- 探索深さ 12
- 自己対戦は途中で打ち切らず詰みの局面まで学習データに含める
- 入玉宣言勝ちも学習データに含める
- 生成局面数 1.3億
- 機械学習
- 学習手法 Adam ミニバッチ法
- 学習率 0.5
- ミニバッチサイズ 100万
- 元の評価関数 2017-10-20-00-18-32
- 検証データは100万局面生成したものをそのまま使っている
- 学習局面数1.3億 + epoch 1の1.3億 = 2.6億局面
実験結果
elmo lambda = 0.3
03:51:47.867 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-10-21-54\260000000 03:51:47.867 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:51:47.867 T1,b2000,478 - 34 - 488(49.48% R-3.60) win black: white = 53.31% : 46.69% declaration win black=0 white=0
06:44:32.484 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-10-21-54\260000000 06:44:32.484 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:44:32.484 T1,b2000,644 - 11 - 345(65.12% R108.43) win black: white = 55.11% : 44.89% declaration win black=0 white=0
elmo lambda = 0.4
03:45:26.573 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-25-21-10-45\260000000 03:45:26.573 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:45:26.573 T1,b2000,496 - 29 - 475(51.08% R7.52) win black: white = 52.73% : 47.27% declaration win black=0 white=0
06:33:58.254 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-25-21-10-45\260000000 06:33:58.254 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:33:58.254 T1,b2000,659 - 16 - 325(66.97% R122.80) win black: white = 51.32% : 48.68% declaration win black=0 white=0
elmo lambda = 0.5
03:57:44.220 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-25-14-26-54\260000000 03:57:44.220 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:57:44.220 T1,b2000,495 - 47 - 458(51.94% R13.50) win black: white = 52.26% : 47.74% declaration win black=0 white=0
06:48:36.333 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-25-14-26-54\260000000 06:48:36.333 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:48:36.333 T1,b2000,651 - 12 - 337(65.89% R114.38) win black: white = 53.34% : 46.66% declaration win black=0 white=0
elmo lambda = 0.6
03:47:47.585 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-03-44-42\260000000 03:47:47.585 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:47:47.585 T1,b2000,492 - 40 - 468(51.25% R8.69) win black: white = 51.35% : 48.65% declaration win black=0 white=0
06:37:11.549 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-03-44-42\260000000 06:37:11.549 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:37:11.549 T1,b2000,630 - 15 - 355(63.96% R99.64) win black: white = 51.57% : 48.43% declaration win black=0 white=0
elmo lambda = 0.7
03:50:26.137 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-17-06-28\260000000 03:50:26.137 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:50:26.137 T1,b2000,478 - 52 - 470(50.42% R2.93) win black: white = 51.27% : 48.73% declaration win black=0 white=0
06:42:06.023 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-17-06-28\260000000 06:42:06.023 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:42:06.023 T1,b2000,673 - 15 - 312(68.32% R133.54) win black: white = 51.78% : 48.22% declaration win black=0 white=0
elmo lambda = 0.8
03:49:33.456 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-27-08-15-33\260000000 03:49:33.456 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:49:33.456 T1,b2000,493 - 40 - 467(51.35% R9.41) win black: white = 51.88% : 48.13% declaration win black=0 white=0
06:39:04.051 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-27-08-15-33\260000000 06:39:04.051 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:39:04.051 T1,b2000,642 - 15 - 343(65.18% R108.90) win black: white = 50.05% : 49.95% declaration win black=0 white=0
elmo lambda = 0.9
03:51:50.971 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-27-14-54-37\260000000 03:51:50.971 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:51:50.971 T1,b2000,443 - 55 - 502(46.88% R-21.72) win black: white = 50.90% : 49.10% declaration win black=0 white=0
06:45:11.312 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-27-14-54-37\260000000 06:45:11.312 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:45:11.312 T1,b2000,641 - 16 - 343(65.14% R108.63) win black: white = 48.48% : 51.52% declaration win black=0 white=0
elmo lambda = 1.0
03:50:28.542 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-23-48-32\260000000 03:50:28.542 engine2=YaneuraOu-2017-early.exe eval2=eval\2017-10-20-00-18-32\130000000 03:50:28.542 T1,b2000,437 - 46 - 517(45.81% R-29.20) win black: white = 49.79% : 50.21% declaration win black=0 white=0
06:39:09.289 engine1=YaneuraOu-2017-early.exe eval1=eval\2017-10-26-23-48-32\260000000 06:39:09.289 engine2=YaneuraOu-2017-early.exe eval2=eval\elmo_wcsc27 06:39:09.289 T1,b2000,637 - 16 - 347(64.74% R105.52) win black: white = 53.25% : 46.75% declaration win black=0 white=0
elmo lambda vs レーティング差
