tanuki- 2018-02-18 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
- eval_apery_4.5G_limit3000
- 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 8
- 自己対戦を打ち切る評価値の絶対値: 3000
- 生成局面数 20億
- シャフル
- あり
- 機械学習
- NNUE機械学習ルーチン
- 学習手法 SGD ミニバッチ法
- batchsize 1000000
- lambda 1.0 or 0.5
- eta 1.0
- newbob_decay 0.5
- nn_batch_size 1000
- eval_save_interval 100000000
- loss_output_interval 1000000
- mirror_percentage 50
- eval_limit 3000
実験結果
実験1
雑巾絞り
lambda=1.0
03:28:03 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.normal\final 03:28:03 engine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_format 03:28:03 T1,b5000,388 - 37 - 575(40.29% R-68.33) win black: white = 53.79% : 46.21% declaration win engine1=0 engine2=0
11:28:24 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.normal\final 11:28:24 engine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000 11:28:24 T1,b5000,471 - 55 - 474(49.84% R-1.10) win black: white = 54.39% : 45.61% declaration win engine1=0 engine2=0
実験2
lambda=0.5
20:26:43 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\final 20:26:43 engine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_format 20:26:43 T1,b5000,375 - 36 - 589(38.90% R-78.43) win black: white = 52.80% : 47.20% declaration win engine1=0 engine2=0
04:34:56 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\final 04:34:56 engine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000 04:34:56 T1,b5000,476 - 54 - 470(50.32% R2.20) win black: white = 51.90% : 48.10% declaration win engine1=0 engine2=0
実験3
評価値にdiscountを適用してからシャッフルした場合
18:48:15 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.discount\final 18:48:15 engine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_format 18:48:15 T1,b5000,388 - 30 - 582(40.00% R-70.44) win black: white = 51.65% : 48.35% declaration win engine1=0 engine2=0
02:47:15 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.discount\final 02:47:15 engine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000 02:47:15 T1,b5000,455 - 53 - 492(48.05% R-13.58) win black: white = 49.95% : 50.05% declaration win engine1=0 engine2=0
実験4
評価値を手数方向にスライドさせてからシャッフルした場合
スライド量=8
16:49:16 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.slide\final 16:49:16 engine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_format 16:49:16 T1,b5000,382 - 42 - 576(39.87% R-71.34) win black: white = 52.82% : 47.18% declaration win engine1=0 engine2=0
00:48:02 engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.slide\final 00:48:02 engine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000 00:48:02 T1,b5000,474 - 54 - 472(50.11% R0.73) win black: white = 52.75% : 47.25% declaration win engine1=0 engine2=0
