tanuki- 2018-03-05 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 0.1
- 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
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\finalengine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_formatT1,b5000,348 - 33 - 619(35.99% R-100.04) win black: white = 48.40% : 51.60% declaration win engine1=0 engine2=0
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\finalengine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000T1,b5000,419 - 62 - 519(44.67% R-37.18) win black: white = 53.20% : 46.80% declaration win engine1=0 engine2=0
実験2
elmo絞り lambda=0.5
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\finalengine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_formatT1,b5000,410 - 30 - 560(42.27% R-54.16) win black: white = 55.57% : 44.43% declaration win engine1=0 engine2=0
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.elmo\finalengine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000T1,b5000,490 - 67 - 443(52.52% R17.52) win black: white = 51.02% : 48.98% 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
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.discount\finalengine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_formatT1,b5000,376 - 36 - 588(39.00% R-77.68) win black: white = 51.76% : 48.24% declaration win engine1=0 engine2=0
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.discount\finalengine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000T1,b5000,474 - 62 - 464(50.53% R3.70) win black: white = 50.75% : 49.25% 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
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.slide\finalengine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_formatT1,b5000,397 - 28 - 575(40.84% R-64.35) win black: white = 50.62% : 49.38% declaration win engine1=6 engine2=1
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.slide\finalengine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000T1,b5000,466 - 64 - 470(49.79% R-1.48) win black: white = 49.47% : 50.53% declaration win engine1=0 engine2=13
実験5
評価値に手数方向にメディアンフィルタをかけてからシャッフルした場合 フィルタ半径=3
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.median\finalengine2=hakubishin-private\build\2017Early\YaneuraOu-2017-early.exe eval2=F:\hnoda\hakubishin\exe\eval\apery_sdt5_eval_twig_formatT1,b5000,337 - 36 - 627(34.96% R-107.86) win black: white = 52.70% : 47.30% declaration win engine1=0 engine2=0
engine1=nnue\source\YaneuraOu-by-gcc.exe eval1=F:\hnoda\nnue\eval\zoukin2.2018-02-11.median\finalengine2=nnue\source\YaneuraOu-by-gcc.exe eval2=F:\ynasu\nnue-v0.2\p-k-halfkp_96w-32\eval_apery_4.5G_limit3000T1,b5000,396 - 59 - 545(42.08% R-55.48) win black: white = 52.28% : 47.72% declaration win engine1=0 engine2=0
