2026/09/27

第60回スプリンターズステークス

朝からってか、もう随分と前から色々やってますが、全然時間が足りない。相変わらず、独自に予想してみた。

もうさ、全く予想に集中出来ない状態なので、参加してみただけです。

今週中にはモデルまとめてリリース出来ればと思ってます。

on 0 件のコメント:
Labels:

葉の数抑えたランキング学習

ランキング学習では葉の数を多くしてたのですが、ここの所やってる葉の数を抑えたらどうなるかで

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 30
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7670
[Loop 1] NDCG@3 = 0.7721
[Loop 1] NDCG@5 = 0.7887
[Loop 1] NDCG@10 = 0.8497

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.63%
(73.93%)		 24.26%
(76.13%)		 25.04%
(72.77%)		 24.14%
(58.39%)		 32.89%
(70.34%)		 26.99%
(72.75%)		 23.06%
(74.67%)		 54.57%
(75.24%)
複勝 55.66%
(79.18%)		 53.79%
(78.00%)		 57.20%
(80.40%)		 60.92%
(79.20%)		 67.11%
(93.69%)		 61.81%
(80.95%)		 52.41%
(77.30%)		 89.81%
(82.08%)
枠連 13.98%
(78.51%)		 12.56%
(83.73%)		 15.11%
(73.37%)		 16.67%
(83.72%)		 --
(--)		 13.04%
(64.69%)		 13.52%
(79.03%)		 30.04%
(75.87%)
馬連 10.15%
(79.92%)		 9.58%
(100.30%)		 10.72%
(60.93%)		 10.34%
(53.22%)		 21.48%
(94.43%)		 10.58%
(65.67%)		 9.03%
(83.97%)		 23.26%
(72.22%)
ワイド 24.08%
(74.39%)		 23.06%
(77.95%)		 25.29%
(71.70%)		 21.84%
(60.80%)		 40.27%
(82.95%)		 26.84%
(71.30%)		 21.69%
(74.80%)		 46.84%
(74.79%)
馬単 5.70%
(91.56%)		 5.27%
(118.51%)		 6.06%
(65.49%)		 6.90%
(69.54%)		 11.41%
(94.16%)		 6.75%
(70.32%)		 4.83%
(99.20%)		 23.26%
(70.09%)
三連複 6.32%
(65.36%)		 6.30%
(65.83%)		 6.46%
(67.51%)		 4.60%
(28.28%)		 16.11%
(125.57%)		 7.21%
(63.28%)		 5.17%
(61.03%)		 15.96%
(74.27%)
三連単 1.48%
(79.04%)		 1.36%
(96.92%)		 1.55%
(61.09%)		 2.30%
(73.68%)		 4.03%
(226.31%)		 2.15%
(66.70%)		 1.02%
(71.15%)		 15.96%
(70.66%)
総合 57.38%
(77.73%)		 55.55%
(87.23%)		 58.92%
(69.13%)		 62.07%
(63.09%)		 69.80%
(112.49%)		 62.27%
(69.46%)		 54.51%
(77.64%)		 90.05%
(72.50%)

まあ、試しの1回でなので悪くないかな。

NumberOfLeaves: 63
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 30
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7655
[Loop 1] NDCG@3 = 0.7712
[Loop 1] NDCG@5 = 0.7883
[Loop 1] NDCG@10 = 0.8494

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 23.93%
(72.47%)		 23.70%
(74.45%)		 24.14%
(71.29%)		 24.14%
(60.46%)		 32.21%
(68.79%)		 25.92%
(68.27%)		 22.49%
(74.33%)		 54.41%
(74.75%)
複勝 55.54%
(79.34%)		 53.87%
(78.22%)		 56.71%
(80.25%)		 63.22%
(82.76%)		 64.43%
(90.87%)		 61.50%
(80.54%)		 52.58%
(77.93%)		 89.38%
(81.20%)
枠連 13.77%
(77.66%)		 12.46%
(80.48%)		 14.94%
(75.97%)		 14.10%
(64.23%)		 --
(--)		 13.19%
(62.33%)		 13.17%
(78.79%)		 29.48%
(75.34%)
馬連 9.88%
(65.17%)		 9.34%
(70.04%)		 10.47%
(61.50%)		 9.20%
(46.67%)		 22.82%
(100.81%)		 10.89%
(64.57%)		 8.40%
(62.38%)		 23.15%
(71.05%)
ワイド 24.32%
(71.61%)		 23.38%
(69.49%)		 25.45%
(73.78%)		 21.84%
(71.61%)		 41.61%
(79.06%)		 27.61%
(74.05%)		 21.64%
(70.07%)		 46.92%
(75.69%)
馬単 5.23%
(58.90%)		 4.87%
(62.96%)		 5.56%
(54.35%)		 5.75%
(64.14%)		 10.74%
(69.87%)		 6.75%
(71.37%)		 4.20%
(53.35%)		 23.15%
(68.93%)
三連複 6.52%
(69.76%)		 6.30%
(63.72%)		 6.87%
(78.91%)		 4.60%
(28.28%)		 14.09%
(74.70%)		 7.52%
(63.90%)		 5.51%
(71.52%)		 16.12%
(74.53%)
三連単 1.37%
(97.92%)		 0.88%
(64.64%)		 1.80%
(133.76%)		 2.30%
(73.68%)		 2.01%
(30.94%)		 1.99%
(62.06%)		 1.08%
(116.86%)		 16.12%
(72.10%)
総合 57.42%
(74.06%)		 55.95%
(70.31%)		 58.43%
(78.74%)		 64.37%
(61.44%)		 67.11%
(73.58%)		 62.27%
(68.38%)		 54.80%
(75.65%)		 89.62%
(72.96%)

葉の数削り過ぎかな^^; では、葉の数は戻して葉の最小少し増やして

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 50
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7663
[Loop 1] NDCG@3 = 0.7719
[Loop 1] NDCG@5 = 0.7887
[Loop 1] NDCG@10 = 0.8496

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.24%
(72.94%)		 23.86%
(75.23%)		 24.63%
(71.57%)		 24.14%
(59.31%)		 34.23%
(76.78%)		 25.92%
(68.48%)		 22.77%
(74.27%)		 54.72%
(75.58%)
複勝 55.19%
(78.78%)		 53.15%
(77.09%)		 56.79%
(80.25%)		 62.07%
(82.53%)		 64.43%
(88.66%)		 60.89%
(79.60%)		 52.30%
(77.64%)		 89.66%
(81.92%)
枠連 13.98%
(79.62%)		 12.65%
(83.57%)		 15.02%
(75.54%)		 16.67%
(86.67%)		 --
(--)		 13.34%
(62.18%)		 13.40%
(81.43%)		 30.52%
(76.57%)
馬連 10.27%
(80.00%)		 9.74%
(97.04%)		 10.72%
(63.46%)		 11.49%
(66.90%)		 23.49%
(100.94%)		 10.89%
(63.88%)		 8.92%
(84.20%)		 23.54%
(71.50%)
ワイド 24.20%
(72.74%)		 23.30%
(73.81%)		 25.37%
(72.81%)		 20.69%
(56.32%)		 40.94%
(78.66%)		 27.61%
(71.26%)		 21.52%
(72.78%)		 46.99%
(76.75%)
馬単 5.85%
(88.86%)		 5.35%
(117.88%)		 6.22%
(59.71%)		 8.05%
(80.34%)		 13.42%
(108.86%)		 6.90%
(71.72%)		 4.83%
(93.51%)		 23.54%
(69.61%)
三連複 6.83%
(80.21%)		 6.30%
(68.83%)		 7.45%
(90.32%)		 5.75%
(102.07%)		 14.77%
(103.22%)		 8.13%
(85.74%)		 5.68%
(76.22%)		 16.28%
(75.35%)
三連単 1.60%
(118.35%)		 1.20%
(82.79%)		 1.96%
(158.00%)		 2.30%
(73.68%)		 4.03%
(181.95%)		 1.99%
(115.28%)		 1.25%
(114.11%)		 16.28%
(72.53%)
総合 57.14%
(83.99%)		 55.31%
(84.55%)		 58.59%
(84.01%)		 63.22%
(75.84%)		 67.11%
(105.58%)		 61.66%
(77.27%)		 54.63%
(84.27%)		 89.89%
(73.58%)

指数的には落ちた感じだけど、三連系が良くなった。葉の最小を減らすと

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 20
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7643
[Loop 1] NDCG@3 = 0.7715
[Loop 1] NDCG@5 = 0.7885
[Loop 1] NDCG@10 = 0.8492

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.12%
(73.38%)		 23.54%
(73.97%)		 24.80%
(74.19%)		 22.99%
(53.56%)		 31.54%
(68.26%)		 26.23%
(67.41%)		 22.71%
(76.03%)		 54.92%
(76.85%)
複勝 55.85%
(79.85%)		 53.47%
(77.55%)		 57.69%
(81.90%)		 64.37%
(84.25%)		 65.77%
(91.21%)		 61.96%
(80.84%)		 52.75%
(78.52%)		 89.73%
(81.45%)
枠連 13.55%
(75.87%)		 11.62%
(74.15%)		 15.02%
(76.64%)		 17.95%
(88.08%)		 --
(--)		 12.42%
(59.95%)		 13.17%
(77.33%)		 29.70%
(75.77%)
馬連 10.03%
(80.70%)		 8.94%
(93.68%)		 11.05%
(68.65%)		 11.49%
(63.10%)		 22.15%
(102.15%)		 10.28%
(63.60%)		 8.92%
(85.22%)		 23.34%
(72.45%)
ワイド 24.28%
(74.84%)		 22.75%
(74.03%)		 26.02%
(75.86%)		 21.84%
(72.07%)		 37.58%
(70.54%)		 27.76%
(75.38%)		 21.86%
(75.00%)		 46.45%
(75.40%)
馬単 5.39%
(83.65%)		 4.55%
(106.00%)		 6.14%
(62.36%)		 6.90%
(60.80%)		 11.41%
(107.65%)		 5.98%
(58.54%)		 4.66%
(90.92%)		 23.34%
(70.37%)
三連複 6.60%
(74.05%)		 6.46%
(74.87%)		 6.87%
(76.46%)		 4.60%
(28.28%)		 15.44%
(122.62%)		 7.82%
(75.90%)		 5.39%
(69.25%)		 16.47%
(77.54%)
三連単 1.48%
(74.33%)		 1.04%
(52.14%)		 1.96%
(100.14%)		 1.15%
(31.38%)		 4.03%
(226.31%)		 1.84%
(110.41%)		 1.14%
(48.11%)		 16.47%
(73.15%)
総合 57.53%
(77.10%)		 55.39%
(78.38%)		 59.17%
(77.03%)		 65.52%
(59.83%)		 68.46%
(112.67%)		 62.42%
(74.01%)		 54.80%
(75.05%)		 89.93%
(74.14%)

ダメだ(笑) 葉の最小は30～50の間辺りが良いのかな? もしくはそれ以上かも?

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 40
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7653
[Loop 1] NDCG@3 = 0.7717
[Loop 1] NDCG@5 = 0.7882
[Loop 1] NDCG@10 = 0.8495

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.75%
(75.39%)		 24.34%
(76.37%)		 25.20%
(75.30%)		 24.14%
(62.41%)		 32.89%
(70.20%)		 26.38%
(71.35%)		 23.45%
(77.32%)		 54.49%
(74.58%)
複勝 56.01%
(79.90%)		 54.51%
(79.60%)		 57.04%
(79.96%)		 63.22%
(83.33%)		 64.43%
(88.93%)		 61.04%
(80.75%)		 53.44%
(78.82%)		 89.73%
(80.79%)
枠連 13.55%
(75.94%)		 11.90%
(78.15%)		 14.76%
(73.10%)		 17.95%
(88.08%)		 --
(--)		 12.27%
(54.48%)		 13.23%
(79.44%)		 29.61%
(73.36%)
馬連 9.84%
(76.41%)		 8.86%
(92.39%)		 10.72%
(60.96%)		 11.49%
(63.10%)		 20.81%
(94.30%)		 9.97%
(56.46%)		 8.86%
(82.28%)		 23.11%
(69.01%)
ワイド 23.97%
(74.15%)		 22.67%
(75.43%)		 25.53%
(74.19%)		 20.69%
(55.06%)		 38.93%
(76.51%)		 26.07%
(68.62%)		 21.92%
(75.99%)		 46.29%
(73.34%)
馬単 5.62%
(84.50%)		 4.79%
(105.78%)		 6.38%
(63.35%)		 6.90%
(75.29%)		 12.75%
(102.89%)		 6.29%
(63.71%)		 4.77%
(90.65%)		 23.11%
(67.00%)
三連複 6.44%
(70.12%)		 6.07%
(60.97%)		 6.87%
(82.05%)		 5.75%
(34.48%)		 13.42%
(71.74%)		 7.52%
(63.34%)		 5.45%
(72.50%)		 16.35%
(72.69%)
三連単 1.41%
(103.43%)		 1.04%
(72.38%)		 1.72%
(137.39%)		 2.30%
(73.68%)		 2.68%
(86.85%)		 1.53%
(50.09%)		 1.25%
(124.58%)		 16.35%
(69.08%)
総合 57.53%
(80.03%)		 56.19%
(80.17%)		 58.51%
(80.83%)		 63.22%
(66.65%)		 67.11%
(84.49%)		 61.35%
(63.60%)		 55.31%
(85.20%)		 89.85%
(70.67%)

中途半端はダメか? まあ、実際1発勝負だと的確には分からないけど^^; 秋のGIには間に合えばと頑張ったつもりが、既に今日はスプリンターズorz

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 80
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7650
[Loop 1] NDCG@3 = 0.7717
[Loop 1] NDCG@5 = 0.7882
[Loop 1] NDCG@10 = 0.8494

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.16%
(72.86%)		 24.42%
(76.91%)		 24.06%
(70.28%)		 21.84%
(50.92%)		 34.23%
(74.30%)		 25.92%
(70.08%)		 22.66%
(73.78%)		 54.57%
(74.85%)
複勝 55.78%
(79.55%)		 54.11%
(79.02%)		 57.12%
(80.10%)		 60.92%
(79.43%)		 65.77%
(94.03%)		 61.35%
(80.66%)		 52.87%
(77.91%)		 89.11%
(81.29%)
枠連 13.90%
(77.75%)		 12.28%
(77.35%)		 15.11%
(77.69%)		 17.95%
(84.10%)		 --
(--)		 12.73%
(58.79%)		 13.52%
(80.22%)		 30.13%
(76.80%)
馬連 9.88%
(64.86%)		 9.10%
(70.14%)		 10.56%
(59.74%)		 11.49%
(60.69%)		 22.15%
(100.87%)		 9.97%
(57.59%)		 8.80%
(64.50%)		 23.61%
(69.94%)
ワイド 24.16%
(70.86%)		 22.67%
(69.15%)		 25.86%
(72.64%)		 21.84%
(70.69%)		 40.27%
(78.79%)		 26.84%
(70.81%)		 21.81%
(70.21%)		 46.80%
(75.25%)
馬単 5.35%
(61.32%)		 4.95%
(64.60%)		 5.81%
(59.11%)		 4.60%
(44.94%)		 12.75%
(91.81%)		 5.98%
(61.96%)		 4.49%
(58.50%)		 23.61%
(67.96%)
三連複 6.87%
(75.59%)		 6.38%
(64.95%)		 7.53%
(89.86%)		 4.60%
(28.28%)		 15.44%
(95.91%)		 7.98%
(67.75%)		 5.74%
(76.77%)		 16.24%
(66.85%)
三連単 1.44%
(109.13%)		 0.96%
(69.47%)		 1.96%
(155.33%)		 1.15%
(31.38%)		 3.36%
(101.54%)		 1.69%
(51.66%)		 1.19%
(131.05%)		 16.24%
(62.08%)
総合 57.46%
(76.47%)		 56.03%
(71.34%)		 58.59%
(83.12%)		 62.07%
(55.94%)		 68.46%
(91.04%)		 61.81%
(64.91%)		 54.91%
(79.12%)		 89.31%
(67.24%)

結局は葉の最小は50が一番良いのかな? 一応60も試して置くか。

NumberOfLeaves: 79
NumberOfIterations: 1500
LearningRate: 0.02
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 60
FeatureFraction: 0.75
SubsampleFraction: 0.8
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

[Loop 1] NDCG@1 = 0.7653
[Loop 1] NDCG@3 = 0.7719
[Loop 1] NDCG@5 = 0.7884
[Loop 1] NDCG@10 = 0.8495

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.20%
(73.68%)		 23.78%
(75.19%)		 24.71%
(73.46%)		 22.99%
(54.94%)		 32.21%
(68.39%)		 25.77%
(67.12%)		 22.94%
(76.55%)		 54.96%
(75.45%)
複勝 55.58%
(79.21%)		 54.03%
(78.53%)		 56.71%
(79.78%)		 62.07%
(80.92%)		 67.79%
(97.99%)		 61.35%
(79.08%)		 52.41%
(77.67%)		 89.38%
(81.36%)
枠連 13.85%
(77.27%)		 12.00%
(74.07%)		 15.45%
(80.23%)		 15.38%
(76.92%)		 --
(--)		 12.58%
(56.58%)		 13.52%
(80.41%)		 30.13%
(75.59%)
馬連 10.11%
(69.30%)		 8.94%
(67.17%)		 11.29%
(72.38%)		 10.34%
(56.55%)		 20.13%
(91.95%)		 10.12%
(58.30%)		 9.26%
(71.45%)		 23.46%
(73.07%)
ワイド 24.12%
(72.28%)		 22.59%
(67.57%)		 25.70%
(76.69%)		 24.14%
(78.16%)		 38.93%
(72.82%)		 25.31%
(65.89%)		 22.43%
(74.60%)		 46.68%
(76.77%)
馬単 5.35%
(63.47%)		 4.47%
(58.80%)		 6.30%
(69.61%)		 4.60%
(44.60%)		 9.40%
(64.90%)		 5.52%
(55.55%)		 4.94%
(66.29%)		 23.46%
(71.13%)
三連複 6.87%
(80.10%)		 6.62%
(77.24%)		 7.20%
(84.27%)		 5.75%
(62.64%)		 14.77%
(119.33%)		 7.36%
(68.39%)		 6.02%
(81.11%)		 16.16%
(75.00%)
三連単 1.41%
(104.04%)		 0.96%
(77.64%)		 1.88%
(136.28%)		 1.15%
(31.38%)		 2.68%
(170.20%)		 1.69%
(53.77%)		 1.19%
(117.05%)		 16.16%
(71.37%)
総合 57.34%
(77.42%)		 55.87%
(71.99%)		 58.43%
(84.11%)		 63.22%
(60.55%)		 70.47%
(97.94%)		 61.96%
(63.08%)		 54.51%
(80.64%)		 89.62%
(73.17%)

何とも微妙な結果だ。まあ、50で更に攻めるかな。

on 0 件のコメント:
Labels: , ,

本日の通常用走破タイム学習

時間的には12時間と本来やりたい24時間の半分ですがRMSE 1.6106で

2,387R 1点 芝(1,130R) ダート(1,170R) 障害(87R) 8頭以下(121R) 9～12頭(586R) 13頭以上(1,680R) 多点
単勝 20.61%
(82.58%)		 20.35%
(76.86%)		 21.11%
(89.91%)		 17.24%
(58.51%)		 30.58%
(85.87%)		 22.01%
(70.03%)		 19.40%
(86.73%)		 48.30%
(76.59%)
複勝 50.27%
(81.72%)		 47.52%
(76.51%)		 52.99%
(87.10%)		 49.43%
(77.01%)		 66.12%
(101.40%)		 54.61%
(83.62%)		 47.62%
(79.64%)		 85.71%
(81.67%)
枠連 10.41%
(71.55%)		 9.32%
(58.71%)		 11.73%
(86.45%)		 5.13%
(18.85%)		 --
(--)		 10.75%
(59.98%)		 9.70%
(71.54%)		 25.01%
(77.98%)
馬連 7.79%
(67.15%)		 7.08%
(61.12%)		 8.55%
(72.26%)		 6.90%
(76.67%)		 17.36%
(96.45%)		 8.87%
(67.58%)		 6.73%
(64.89%)		 18.43%
(64.58%)
ワイド 20.03%
(70.86%)		 18.76%
(67.65%)		 21.45%
(73.41%)		 17.24%
(78.28%)		 34.71%
(84.46%)		 23.55%
(71.59%)		 17.74%
(69.63%)		 40.09%
(78.79%)
馬単 4.15%
(65.82%)		 3.98%
(65.93%)		 4.44%
(65.19%)		 2.30%
(72.87%)		 10.74%
(117.02%)		 4.44%
(53.57%)		 3.57%
(66.40%)		 18.43%
(62.32%)
三連複 4.82%
(58.62%)		 4.34%
(45.55%)		 5.47%
(73.44%)		 2.30%
(28.97%)		 13.22%
(59.09%)		 4.61%
(36.60%)		 4.29%
(66.26%)		 11.98%
(69.51%)
三連単 1.09%
(49.29%)		 0.88%
(36.97%)		 1.37%
(64.85%)		 0.00%
(0.00%)		 1.65%
(19.59%)		 0.68%
(24.22%)		 1.19%
(60.18%)		 11.98%
(68.50%)
総合 51.32%
(68.41%)		 48.67%
(61.20%)		 54.02%
(76.52%)		 49.43%
(51.82%)		 67.77%
(80.55%)		 55.12%
(58.40%)		 48.81%
(70.66%)		 86.13%
(70.02%)

目指す精度にはまだまだな感じですorz あまりお薦めではないけど、並行して相対タイム学習を3時間程したのは\(R^{2}\) 0.2107で

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.20%
(79.70%)		 23.86%
(81.72%)		 25.12%
(80.77%)		 16.09%
(35.75%)		 31.54%
(68.39%)		 25.92%
(73.25%)		 22.94%
(83.05%)		 52.22%
(76.43%)
複勝 54.45%
(82.13%)		 53.55%
(83.88%)		 55.32%
(80.49%)		 55.17%
(79.89%)		 63.09%
(89.40%)		 58.74%
(81.87%)		 52.13%
(81.61%)		 87.94%
(81.36%)
枠連 12.60%
(76.43%)		 11.34%
(77.25%)		 13.91%
(75.55%)		 10.26%
(78.21%)		 --
(--)		 13.34%
(74.49%)		 11.58%
(72.67%)		 27.53%
(73.74%)
馬連 8.90%
(64.64%)		 8.46%
(62.57%)		 9.33%
(65.77%)		 9.20%
(78.74%)		 18.79%
(78.32%)		 11.04%
(77.44%)		 7.27%
(58.75%)		 21.39%
(65.36%)
ワイド 23.15%
(82.73%)		 22.27%
(82.06%)		 24.30%
(84.44%)		 19.54%
(68.28%)		 36.91%
(81.01%)		 27.61%
(79.29%)		 20.33%
(84.15%)		 44.30%
(78.19%)
馬単 4.92%
(64.09%)		 4.39%
(65.32%)		 5.48%
(64.19%)		 4.60%
(44.94%)		 10.07%
(51.68%)		 6.44%
(77.50%)		 3.92%
(60.17%)		 21.39%
(62.04%)
三連複 6.05%
(78.77%)		 6.46%
(87.04%)		 5.73%
(73.89%)		 4.60%
(28.28%)		 14.09%
(90.74%)		 7.21%
(79.36%)		 4.94%
(77.54%)		 14.99%
(69.95%)
三連単 1.05%
(65.77%)		 1.12%
(69.93%)		 0.98%
(63.96%)		 1.15%
(31.38%)		 2.68%
(58.46%)		 1.84%
(72.78%)		 0.62%
(63.80%)		 14.99%
(64.31%)
総合 55.97%
(74.26%)		 55.23%
(76.20%)		 56.79%
(73.62%)		 55.17%
(55.39%)		 65.10%
(74.00%)		 59.66%
(77.00%)		 53.83%
(72.72%)		 88.25%
(67.71%)

これはまあ、更に学習時間が短か過ぎたかなぁ。

で、ここで手動でパラメーター設定してLightGBMを試すと

NumberOfLeaves: 79
NumberOfIterations: 1,500
LearningRate: 0.020
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 30
SubsampleFraction: 0.025
FeatureFraction: 0.980
L2Regularization: 0.500

通常用走破タイム学習はRMSE 1.6010で

2,387R 1点 芝(1,130R) ダート(1,170R) 障害(87R) 8頭以下(121R) 9～12頭(586R) 13頭以上(1,680R) 多点
単勝 19.02%
(72.64%)		 19.03%
(84.56%)		 19.57%
(63.01%)		 11.49%
(47.47%)		 23.14%
(65.62%)		 21.84%
(74.08%)		 17.74%
(72.65%)		 48.14%
(76.86%)
複勝 48.47%
(79.81%)		 46.46%
(83.96%)		 50.34%
(76.58%)		 49.43%
(69.20%)		 61.16%
(97.36%)		 53.07%
(83.36%)		 45.95%
(77.30%)		 85.13%
(80.85%)
枠連 11.29%
(75.40%)		 10.25%
(85.83%)		 12.44%
(68.62%)		 7.69%
(41.92%)		 --
(--)		 12.12%
(75.53%)		 10.36%
(71.09%)		 26.07%
(76.71%)
馬連 7.79%
(71.94%)		 6.64%
(78.00%)		 8.97%
(67.42%)		 6.90%
(53.91%)		 15.70%
(64.63%)		 9.04%
(80.02%)		 6.79%
(69.64%)		 18.01%
(64.85%)
ワイド 20.70%
(84.82%)		 18.50%
(84.18%)		 22.82%
(83.47%)		 20.69%
(111.26%)		 33.88%
(62.98%)		 23.55%
(85.32%)		 18.75%
(86.21%)		 39.72%
(76.99%)
馬単 3.39%
(68.24%)		 3.19%
(91.26%)		 3.85%
(51.09%)		 0.00%
(0.00%)		 6.61%
(68.68%)		 4.61%
(71.28%)		 2.74%
(67.15%)		 18.01%
(62.55%)
三連複 4.78%
(82.71%)		 3.45%
(104.03%)		 6.07%
(65.74%)		 4.60%
(34.02%)		 6.61%
(33.88%)		 5.29%
(53.75%)		 4.46%
(96.33%)		 12.69%
(60.15%)
三連単 0.88%
(107.72%)		 0.80%
(190.78%)		 1.03%
(35.50%)		 0.00%
(0.00%)		 1.65%
(57.44%)		 1.19%
(38.72%)		 0.71%
(135.40%)		 12.69%
(56.99%)
総合 49.85%
(80.47%)		 48.05%
(100.58%)		 51.62%
(63.90%)		 49.43%
(44.76%)		 62.81%
(64.37%)		 54.61%
(70.26%)		 47.26%
(84.47%)		 85.55%
(63.40%)

相対タイム学習は\(R^{2}\) 0.2290で

2,562R 1点 芝(1,253R) ダート(1,222R) 障害(87R) 8頭以下(149R) 9～12頭(652R) 13頭以上(1,761R) 多点
単勝 24.08%
(78.40%)		 24.02%
(80.99%)		 24.47%
(76.32%)		 19.54%
(70.23%)		 34.90%
(76.71%)		 25.31%
(68.44%)		 22.71%
(82.23%)		 53.32%
(77.31%)
複勝 55.43%
(81.24%)		 53.79%
(80.37%)		 56.79%
(81.72%)		 59.77%
(87.13%)		 64.43%
(86.04%)		 59.82%
(81.52%)		 53.04%
(80.73%)		 88.41%
(81.41%)
枠連 12.38%
(64.41%)		 11.62%
(60.44%)		 13.05%
(67.97%)		 12.82%
(65.51%)		 --
(--)		 13.96%
(69.68%)		 11.07%
(58.69%)		 28.74%
(79.15%)
馬連 9.29%
(60.80%)		 8.86%
(56.62%)		 9.82%
(65.88%)		 8.05%
(49.77%)		 19.46%
(92.75%)		 11.66%
(70.84%)		 7.55%
(54.38%)		 22.29%
(69.34%)
ワイド 22.87%
(70.94%)		 21.23%
(62.50%)		 24.63%
(79.10%)		 21.84%
(77.82%)		 37.58%
(70.07%)		 27.30%
(76.32%)		 19.99%
(69.02%)		 45.36%
(76.88%)
馬単 5.62%
(64.47%)		 5.35%
(64.54%)		 5.89%
(65.22%)		 5.75%
(52.99%)		 12.08%
(92.48%)		 7.21%
(65.09%)		 4.49%
(61.87%)		 22.29%
(68.06%)
三連複 6.21%
(67.11%)		 5.91%
(62.48%)		 6.55%
(73.59%)		 5.75%
(42.64%)		 12.75%
(62.82%)		 6.75%
(56.66%)		 5.45%
(71.34%)		 15.42%
(71.89%)
三連単 1.56%
(61.32%)		 1.28%
(52.16%)		 1.80%
(69.19%)		 2.30%
(82.76%)		 2.68%
(84.77%)		 2.61%
(98.70%)		 1.08%
(45.50%)		 15.42%
(65.67%)
総合 56.67%
(68.64%)		 55.23%
(65.10%)		 57.77%
(72.40%)		 62.07%
(66.11%)		 66.44%
(80.81%)		 60.74%
(73.40%)		 54.34%
(65.47%)		 88.68%
(69.80%)

どちらも10分程度で学習した結果なので、AutoMLで長時間するのが良いのか、パラメーターを探り当てる方が良いのかだなぁ。この辺りの機能追加で最近時間取られてたので、もう少しモデル探しを頑張るか。

on 0 件のコメント:
Labels: ,

2026/09/26

昨夜学習させた新馬戦用モデル

昨夜寝る前に3時間×2回の新馬戦用走破タイム学習では2回とも同じRMSE 2.0408で

171R 1点 芝(120R) ダート(51R) 8頭以下(27R) 9～12頭(65R) 13頭以上(79R) 多点
単勝 25.73%
(110.64%)		 28.33%
(124.50%)		 19.61%
(78.04%)		 33.33%
(75.56%)		 30.77%
(125.85%)		 18.99%
(110.13%)		 47.37%
(71.15%)
複勝 55.56%
(106.08%)		 55.00%
(102.00%)		 56.86%
(115.69%)		 62.96%
(102.96%)		 58.46%
(100.62%)		 50.63%
(111.65%)		 81.87%
(87.43%)
枠連 11.76%
(165.37%)		 13.48%
(230.11%)		 8.51%
(42.77%)		 --
(--)		 13.85%
(153.54%)		 8.86%
(158.35%)		 28.68%
(110.32%)
馬連 8.19%
(130.64%)		 10.00%
(177.42%)		 3.92%
(20.59%)		 22.22%
(142.59%)		 6.15%
(125.08%)		 5.06%
(131.14%)		 19.88%
(92.44%)
ワイド 18.71%
(156.14%)		 21.67%
(170.75%)		 11.76%
(121.76%)		 40.74%
(111.85%)		 20.00%
(224.62%)		 10.13%
(114.94%)		 38.60%
(116.47%)
馬単 4.68%
(213.80%)		 5.83%
(299.42%)		 1.96%
(12.35%)		 11.11%
(32.22%)		 3.08%
(213.69%)		 3.80%
(275.95%)		 19.88%
(84.48%)
三連複 5.85%
(121.93%)		 5.83%
(93.08%)		 5.88%
(189.80%)		 11.11%
(223.70%)		 6.15%
(65.38%)		 3.80%
(133.67%)		 11.11%
(63.11%)
三連単 0.58%
(7.54%)		 0.83%
(10.75%)		 0.00%
(0.00%)		 3.70%
(47.78%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 11.11%
(43.13%)
総合 57.89%
(125.50%)		 57.50%
(148.36%)		 58.82%
(72.92%)		 62.96%
(105.24%)		 63.08%
(126.10%)		 51.90%
(129.48%)		 81.87%
(65.15%)

これ、まあ、想定内なんだが、学習データは2004年～2025年に変え、検証は当然のその後の今年初めから先週末、あっ、今回は末じゃないか、火曜日まで代替開催あったんだった^^;

175R 1点 芝(123R) ダート(52R) 8頭以下(28R) 9～12頭(66R) 13頭以上(81R) 多点
単勝 25.14%
(108.11%)		 27.64%
(121.46%)		 19.23%
(76.54%)		 32.14%
(72.86%)		 30.30%
(123.94%)		 18.52%
(107.41%)		 48.00%
(71.89%)
複勝 55.43%
(105.09%)		 55.28%
(101.54%)		 55.77%
(113.46%)		 64.29%
(103.57%)		 59.09%
(101.06%)		 49.38%
(108.89%)		 82.29%
(87.18%)
枠連 11.51%
(161.80%)		 13.19%
(225.05%)		 8.33%
(41.88%)		 --
(--)		 13.64%
(151.21%)		 8.64%
(154.44%)		 28.06%
(107.75%)
馬連 8.00%
(127.66%)		 9.76%
(173.09%)		 3.85%
(20.19%)		 21.43%
(137.50%)		 6.06%
(123.18%)		 4.94%
(127.90%)		 20.00%
(91.37%)
ワイド 18.86%
(154.23%)		 21.95%
(168.94%)		 11.54%
(119.42%)		 39.29%
(107.86%)		 21.21%
(225.61%)		 9.88%
(112.10%)		 38.86%
(114.90%)
馬単 4.57%
(208.91%)		 5.69%
(292.11%)		 1.92%
(12.12%)		 10.71%
(31.07%)		 3.03%
(210.45%)		 3.70%
(269.14%)		 20.00%
(83.90%)
三連複 5.71%
(119.14%)		 5.69%
(90.81%)		 5.77%
(186.15%)		 10.71%
(215.71%)		 6.06%
(64.39%)		 3.70%
(130.37%)		 10.86%
(61.67%)
三連単 0.57%
(7.37%)		 0.81%
(10.49%)		 0.00%
(0.00%)		 3.57%
(46.07%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 10.86%
(42.14%)
総合 57.71%
(123.04%)		 57.72%
(145.35%)		 57.69%
(71.50%)		 64.29%
(102.09%)		 63.64%
(124.98%)		 50.62%
(126.28%)		 82.29%
(64.21%)

それ程変わらないが、話を正確にね(笑) いや、以前2004年～2024年学習させて2025年で検証したら異常に回収率が良かったって事で、更に2026年になると回収率が落ちてたのを踏まえ、では、なるべく直近まで学習させた後が成績良いんならって事で今回やった訳です。まあ、実戦でどうなるかなんだけど^^;

on 0 件のコメント:
Labels: , ,

DuckDuckGo AI Chat

今朝昨夜の続きの質問をCopilotにぶつけると「1日の上限に達しました。...」とまあ多分課金を促してるんだと思うけど、まあ、だからあえてそれ以上は使うの止めた(笑) 確かにGeminiでも似た感じなので、Geminiに質問として無料で無制限に使えるのはあるか問合わせるとDuckDuckGo AI Chatがあるよって教えてくれた。なので、しばらくはこれにお世話になってみようかと思う。

on 0 件のコメント:
Labels: ,

2026/09/25

直近の相対タイム学習

あれもこれも試してるんですが、昨日やった相対タイム学習の\(R^{2}\)=0.2286は

2,538R 1点 芝(1,241R) ダート(1,211R) 障害(86R) 8頭以下(147R) 9～12頭(645R) 13頭以上(1,746R) 多点
単勝 23.96%
(74.36%)		 23.61%
(74.60%)		 24.77%
(75.52%)		 17.44%
(54.42%)		 36.05%
(77.55%)		 24.96%
(71.75%)		 22.57%
(75.05%)		 53.55%
(77.58%)
複勝 54.89%
(79.39%)		 53.34%
(78.53%)		 56.56%
(80.68%)		 53.49%
(73.72%)		 63.95%
(85.17%)		 59.38%
(81.32%)		 52.46%
(78.19%)		 88.57%
(80.66%)
枠連 12.67%
(65.38%)		 12.03%
(64.91%)		 13.25%
(65.75%)		 12.99%
(66.36%)		 --
(--)		 13.64%
(63.98%)		 11.57%
(62.04%)		 28.85%
(75.76%)
馬連 9.57%
(61.47%)		 9.27%
(60.67%)		 9.99%
(63.09%)		 8.14%
(50.35%)		 21.09%
(97.89%)		 11.16%
(65.05%)		 8.02%
(57.08%)		 22.54%
(68.56%)
ワイド 23.01%
(70.37%)		 21.19%
(61.15%)		 24.94%
(79.17%)		 22.09%
(79.42%)		 40.14%
(77.01%)		 26.20%
(72.05%)		 20.39%
(69.19%)		 45.11%
(74.73%)
馬単 5.36%
(55.20%)		 5.40%
(66.12%)		 5.37%
(44.48%)		 4.65%
(48.49%)		 12.93%
(92.45%)		 6.82%
(68.34%)		 4.18%
(47.21%)		 22.54%
(67.43%)
三連複 6.26%
(70.87%)		 6.04%
(67.89%)		 6.61%
(76.94%)		 4.65%
(28.60%)		 12.93%
(63.81%)		 6.05%
(57.40%)		 5.78%
(76.45%)		 15.25%
(60.63%)
三連単 1.42%
(56.11%)		 1.45%
(66.89%)		 1.40%
(46.79%)		 1.16%
(31.74%)		 2.72%
(85.92%)		 1.86%
(68.50%)		 1.15%
(49.02%)		 15.25%
(57.55%)
総合 56.11%
(66.66%)		 54.71%
(67.65%)		 57.56%
(66.56%)		 55.81%
(53.98%)		 65.31%
(82.83%)		 60.31%
(68.55%)		 53.78%
(64.28%)		 88.77%
(64.39%)

取り立てて良いモデルかは微妙なんですが、酷いわけじゃない。これ、もう少し良くしたいと思い、

NumberOfLeaves: 63
NumberOfIterations: 700
LearningRate: 0.020
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 20

だと\(R^{2}\)=0.2332で

2,538R 1点 芝(1,241R) ダート(1,211R) 障害(86R) 8頭以下(147R) 9～12頭(645R) 13頭以上(1,746R) 多点
単勝 24.07%
(75.62%)		 23.85%
(76.95%)		 24.69%
(75.52%)		 18.60%
(58.02%)		 36.73%
(77.89%)		 25.27%
(70.93%)		 22.57%
(77.16%)		 52.88%
(75.42%)
複勝 55.91%
(81.42%)		 54.47%
(81.48%)		 57.47%
(81.75%)		 54.65%
(75.81%)		 66.67%
(91.56%)		 59.07%
(80.06%)		 53.84%
(81.07%)		 87.90%
(80.39%)
枠連 13.46%
(75.46%)		 12.41%
(73.91%)		 14.20%
(76.16%)		 16.88%
(86.23%)		 --
(--)		 13.95%
(67.13%)		 12.49%
(74.08%)		 28.58%
(73.12%)
馬連 9.65%
(65.71%)		 9.27%
(65.02%)		 9.91%
(66.17%)		 11.63%
(69.07%)		 20.41%
(95.71%)		 11.32%
(65.80%)		 8.13%
(63.14%)		 22.26%
(67.34%)
ワイド 23.72%
(73.49%)		 21.92%
(66.45%)		 25.60%
(80.74%)		 23.26%
(73.02%)		 38.10%
(74.83%)		 27.60%
(76.67%)		 21.08%
(72.21%)		 45.19%
(76.56%)
馬単 5.08%
(55.57%)		 4.83%
(61.13%)		 5.28%
(48.67%)		 5.81%
(72.44%)		 10.88%
(71.63%)		 6.82%
(62.71%)		 3.95%
(51.58%)		 22.26%
(66.41%)
三連複 6.42%
(76.94%)		 6.37%
(72.76%)		 6.52%
(79.36%)		 5.81%
(103.26%)		 12.93%
(92.65%)		 7.60%
(90.82%)		 5.44%
(70.49%)		 15.41%
(72.38%)
三連単 1.38%
(61.52%)		 1.21%
(69.07%)		 1.57%
(55.90%)		 1.16%
(31.74%)		 2.72%
(116.39%)		 2.17%
(82.64%)		 0.97%
(49.10%)		 15.41%
(68.30%)
総合 57.17%
(70.66%)		 55.68%
(70.79%)		 58.71%
(70.50%)		 56.98%
(71.00%)		 67.35%
(88.67%)		 60.16%
(74.59%)		 55.21%
(67.35%)		 88.18%
(70.30%)

少し改善してますね。

ここで、ちょっと気にしてる葉の数なんだけど、極端に少ないのに気が付いた人はここまでの記事をしっかり読んでくれてる方だと思います。今回、Copilotが提案したパラメーターなんです。

on 0 件のコメント:
Labels: , ,

2026/09/18

一旦、新馬戦の学習

通常用走破タイム学習がもう一つ進まないので、2004年～2025年の学習CSVデータ使った新馬戦用走破タイム学習を今朝出勤前にスタート10,800秒×10回にして、ここまで3回終わり、最適RMSE 2.0317は

148R 1点 芝(106R) ダート(42R) 8頭以下(25R) 9～12頭(53R) 13頭以上(70R) 多点
単勝 20.27%
(81.76%)		 24.53%
(73.21%)		 9.52%
(103.33%)		 24.00%
(48.40%)		 28.30%
(106.04%)		 12.86%
(75.29%)		 45.95%
(88.36%)
複勝 37.84%
(61.35%)		 38.68%
(58.21%)		 35.71%
(69.29%)		 40.00%
(62.40%)		 37.74%
(59.25%)		 37.14%
(62.57%)		 80.41%
(76.28%)
枠連 6.96%
(81.74%)		 7.89%
(106.05%)		 5.13%
(34.36%)		 --
(--)		 3.77%
(66.23%)		 8.57%
(84.14%)		 14.78%
(42.80%)
馬連 3.38%
(50.27%)		 2.83%
(53.02%)		 4.76%
(43.33%)		 8.00%
(22.80%)		 1.89%
(18.49%)		 2.86%
(84.14%)		 14.19%
(35.32%)
ワイド 9.46%
(67.70%)		 10.38%
(82.83%)		 7.14%
(29.52%)		 20.00%
(46.40%)		 9.43%
(108.30%)		 5.71%
(44.57%)		 28.38%
(67.82%)
馬単 1.35%
(49.12%)		 1.89%
(68.58%)		 0.00%
(0.00%)		 4.00%
(14.40%)		 0.00%
(0.00%)		 1.43%
(98.71%)		 14.19%
(27.04%)
三連複 1.35%
(9.80%)		 1.89%
(13.68%)		 0.00%
(0.00%)		 4.00%
(25.20%)		 1.89%
(15.47%)		 0.00%
(0.00%)		 6.76%
(36.49%)
三連単 0.00%
(0.00%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 6.76%
(36.80%)
総合 40.54%
(49.31%)		 42.45%
(55.15%)		 35.71%
(34.98%)		 40.00%
(31.37%)		 39.62%
(46.72%)		 41.43%
(56.18%)		 81.76%
(43.34%)

これまでは、普通に2.0未満だったが、なかなかそこまで行かない^^; なんでこんなに悪いってくらい酷い感じです。通常用でもそんな感じがしてるけど、なんだろ? 2025年まで増やした学習データは何かAIを悩ませるものがあるのかな? これさ、例えば2004年～2024年の学習CSVデータで学習させた最適RMSE 1.9055だと、

148R 1点 芝(106R) ダート(42R) 8頭以下(25R) 9～12頭(53R) 13頭以上(70R) 多点
単勝 17.57%
(67.97%)		 18.87%
(61.98%)		 14.29%
(83.10%)		 36.00%
(81.60%)		 20.75%
(111.89%)		 8.57%
(29.86%)		 39.86%
(60.63%)
複勝 42.57%
(69.26%)		 41.51%
(68.49%)		 45.24%
(71.19%)		 56.00%
(71.60%)		 39.62%
(72.08%)		 40.00%
(66.29%)		 78.38%
(79.98%)
枠連 4.35%
(25.39%)		 3.95%
(20.79%)		 5.13%
(34.36%)		 --
(--)		 3.77%
(15.85%)		 4.29%
(29.71%)		 17.39%
(100.16%)
馬連 4.05%
(34.86%)		 4.72%
(39.43%)		 2.38%
(23.33%)		 12.00%
(130.80%)		 3.77%
(23.77%)		 1.43%
(9.00%)		 13.51%
(88.02%)
ワイド 9.46%
(50.27%)		 12.26%
(66.51%)		 2.38%
(9.29%)		 28.00%
(80.80%)		 9.43%
(56.60%)		 2.86%
(34.57%)		 29.73%
(87.09%)
馬単 3.38%
(40.54%)		 4.72%
(56.60%)		 0.00%
(0.00%)		 12.00%
(159.20%)		 1.89%
(10.19%)		 1.43%
(21.14%)		 13.51%
(91.59%)
三連複 2.03%
(77.16%)		 2.83%
(107.74%)		 0.00%
(0.00%)		 8.00%
(51.60%)		 1.89%
(191.13%)		 0.00%
(0.00%)		 10.14%
(69.76%)
三連単 0.68%
(21.49%)		 0.94%
(30.00%)		 0.00%
(0.00%)		 4.00%
(127.20%)		 0.00%
(0.00%)		 0.00%
(0.00%)		 10.14%
(66.48%)
総合 43.24%
(49.03%)		 42.45%
(57.75%)		 45.24%
(27.60%)		 56.00%
(100.40%)		 39.62%
(60.19%)		 41.43%
(23.82%)		 78.38%
(74.49%)

やはり、RMSE値だけでは判断出来ない何かがあるのかな^^; 明朝にはある程度終わってるので、そこまでにどの程度のモデルが出るかかなぁ。

追記 2026.9.19

結局4回目以降7回目まで同じRMSEで終わったので8回目途中で中断しました。

on 0 件のコメント:
Labels: , ,

2026/09/15

簡単には行かないかな

以前通常用走破タイム学習では最適RMSE 1.4程度まで、Model Builderに至っては1.4未満も出てたので、出来ればその辺りを目指いしたんだが、学習させている

NumberOfLeaves: 351
NumberOfIterations: 5000
LearningRate: 0.008
MaximumBinCountPerFeature: 47
MinimumExampleCountPerLeaf: 120

ではRMSE 1.6016～1.6054です。モデルサイズは120MBとまあ大きい部類。まだ10回のうち半分ですが、1回20分程度掛かるしここは中断してWindows Upateもある事だし、再起動してパラメーターを見直してみます。ああ、一応って検証始めたんだが嫌な予感(笑) そもそもモデルが大きいと遅いし、検証開始時は特にスローなんですが、多点買いで三連単２百万馬券当ててる。RMSE以上になんか滅茶苦茶当ててる感じだ。先日のバグった時みたいにエグイ結果が出そうだ。まあ、確かに狙った学習ではあるが過学習って事なのか、何かリークしてるのか、まあ、結果を待つか。

3,151R 1点 芝(1,486R) ダート(1,545R) 障害(120R) 8頭以下(188R) 9～12頭(720R) 13頭以上(2,243R) 多点
単勝 32.34%
(175.07%)		 27.79%
(185.02%)		 36.38%
(167.83%)		 36.67%
(145.00%)		 38.83%
(137.82%)		 32.36%
(141.36%)		 31.79%
(189.01%)		 61.38%
(136.94%)
複勝 64.90%
(132.59%)		 60.70%
(133.13%)		 68.48%
(133.77%)		 70.83%
(110.75%)		 65.96%
(104.73%)		 70.83%
(127.76%)		 62.91%
(136.47%)		 91.08%
(126.12%)
枠連 23.37%
(255.22%)		 19.58%
(238.20%)		 26.23%
(268.36%)		 28.28%
(274.04%)		 --
(--)		 23.61%
(205.25%)		 21.89%
(256.01%)		 39.80%
(182.09%)
馬連 18.85%
(361.52%)		 15.88%
(307.00%)		 21.55%
(414.48%)		 20.83%
(354.83%)		 20.74%
(127.98%)		 22.08%
(258.71%)		 17.65%
(414.10%)		 34.21%
(281.49%)
ワイド 35.39%
(259.36%)		 31.70%
(237.21%)		 38.96%
(283.07%)		 35.00%
(228.42%)		 40.43%
(105.64%)		 40.83%
(218.93%)		 33.21%
(285.23%)		 57.03%
(245.03%)
馬単 12.28%
(434.78%)		 9.42%
(389.99%)		 14.95%
(487.49%)		 13.33%
(310.75%)		 14.36%
(163.19%)		 13.06%
(258.21%)		 11.86%
(514.22%)		 34.21%
(273.98%)
三連複 12.60%
(726.10%)		 10.70%
(641.74%)		 14.30%
(839.50%)		 14.17%
(310.67%)		 15.96%
(190.05%)		 14.86%
(321.61%)		 11.59%
(900.86%)		 25.77%
(542.78%)
三連単 4.67%
(1,070.35%)		 3.50%
(1,019.75%)		 5.70%
(1,141.87%)		 5.83%
(776.17%)		 4.26%
(289.36%)		 6.25%
(562.17%)		 4.19%
(1,298.94%)		 25.77%
(561.52%)
総合 66.30%
(429.09%)		 61.84%
(397.15%)		 70.23%
(468.11%)		 70.83%
(314.72%)		 69.68%
(159.83%)		 71.11%
(261.75%)		 64.47%
(499.36%)		 91.43%
(414.66%)

う～ん、やはりリークしてるんだろか? そんな手違いした覚えないんだけどなぁ^^; でも、明らかに2025年は飛びぬけた回収率。的中率的にはまあ良い事は良いが、リークしてる感じとまでは行かないと思うけど。

2,222R 1点 芝(1,051R) ダート(1,091R) 障害(80R) 8頭以下(106R) 9～12頭(538R) 13頭以上(1,578R) 多点
単勝 17.24%
(68.06%)		 14.18%
(64.76%)		 19.71%
(70.36%)		 23.75%
(80.00%)		 22.64%
(54.62%)		 19.89%
(65.09%)		 15.97%
(69.97%)		 43.97%
(80.47%)
複勝 46.08%
(85.18%)		 40.25%
(80.23%)		 50.60%
(89.58%)		 61.25%
(90.38%)		 54.72%
(88.96%)		 50.00%
(79.46%)		 44.17%
(86.88%)		 82.00%
(82.67%)
枠連 8.84%
(71.26%)		 7.15%
(70.29%)		 9.78%
(67.25%)		 16.22%
(139.86%)		 --
(--)		 10.59%
(75.20%)		 7.73%
(65.86%)		 22.11%
(75.74%)
馬連 5.36%
(56.69%)		 4.47%
(55.08%)		 5.96%
(55.70%)		 8.75%
(91.37%)		 10.38%
(34.34%)		 7.25%
(67.42%)		 4.37%
(54.54%)		 14.76%
(63.27%)
ワイド 16.07%
(78.52%)		 13.51%
(72.80%)		 17.87%
(80.03%)		 25.00%
(133.13%)		 28.30%
(64.34%)		 21.56%
(76.69%)		 13.37%
(80.10%)		 35.37%
(78.41%)
馬単 2.66%
(45.85%)		 1.71%
(34.72%)		 3.30%
(50.23%)		 6.25%
(132.25%)		 5.66%
(44.91%)		 3.72%
(61.34%)		 2.09%
(40.63%)		 14.76%
(60.65%)
三連複 3.20%
(79.00%)		 2.76%
(86.96%)		 3.48%
(68.73%)		 5.00%
(114.50%)		 8.49%
(73.21%)		 4.83%
(63.25%)		 2.28%
(84.76%)		 10.80%
(88.16%)
三連単 0.36%
(28.18%)		 0.10%
(7.98%)		 0.55%
(29.56%)		 1.25%
(274.75%)		 0.00%
(0.00%)		 0.37%
(50.45%)		 0.38%
(22.48%)		 10.80%
(112.22%)
総合 47.39%
(64.01%)		 41.58%
(58.91%)		 51.79%
(63.91%)		 63.75%
(131.96%)		 61.32%
(51.48%)		 50.93%
(67.36%)		 45.25%
(63.15%)		 82.54%
(93.04%)

ここまで同じモデルで違うとねぇ。ただね、Geminiとかに質問しても、確かに学習した年に近い程精度が良い事がって話なんだよ。だとすれば、これ、本当に直前までを学習させて実戦に行けば期待出来るのかな?

追記 2026.9.16 5:30

昨夜CSVを2025年までにして同じパラメーターで学習させてRMSE 1.6004は2026年1月～8月検証すると

2,222R 1点 芝(1,051R) ダート(1,091R) 障害(80R) 8頭以下(106R) 9～12頭(538R) 13頭以上(1,578R) 多点
単勝 17.24%
(68.06%)		 14.18%
(64.76%)		 19.71%
(70.36%)		 23.75%
(80.00%)		 22.64%
(54.62%)		 19.89%
(65.09%)		 15.97%
(69.97%)		 43.97%
(80.47%)
複勝 46.08%
(85.18%)		 40.25%
(80.23%)		 50.60%
(89.58%)		 61.25%
(90.38%)		 54.72%
(88.96%)		 50.00%
(79.46%)		 44.17%
(86.88%)		 82.00%
(82.67%)
枠連 8.84%
(71.26%)		 7.15%
(70.29%)		 9.78%
(67.25%)		 16.22%
(139.86%)		 --
(--)		 10.59%
(75.20%)		 7.73%
(65.86%)		 22.11%
(75.74%)
馬連 5.36%
(56.69%)		 4.47%
(55.08%)		 5.96%
(55.70%)		 8.75%
(91.37%)		 10.38%
(34.34%)		 7.25%
(67.42%)		 4.37%
(54.54%)		 14.76%
(63.27%)
ワイド 16.07%
(78.52%)		 13.51%
(72.80%)		 17.87%
(80.03%)		 25.00%
(133.13%)		 28.30%
(64.34%)		 21.56%
(76.69%)		 13.37%
(80.10%)		 35.37%
(78.41%)
馬単 2.66%
(45.85%)		 1.71%
(34.72%)		 3.30%
(50.23%)		 6.25%
(132.25%)		 5.66%
(44.91%)		 3.72%
(61.34%)		 2.09%
(40.63%)		 14.76%
(60.65%)
三連複 3.20%
(79.00%)		 2.76%
(86.96%)		 3.48%
(68.73%)		 5.00%
(114.50%)		 8.49%
(73.21%)		 4.83%
(63.25%)		 2.28%
(84.76%)		 10.80%
(88.16%)
三連単 0.36%
(28.18%)		 0.10%
(7.98%)		 0.55%
(29.56%)		 1.25%
(274.75%)		 0.00%
(0.00%)		 0.37%
(50.45%)		 0.38%
(22.48%)		 10.80%
(112.22%)
総合 47.39%
(64.01%)		 41.58%
(58.91%)		 51.79%
(63.91%)		 63.75%
(131.96%)		 61.32%
(51.48%)		 50.93%
(67.36%)		 45.25%
(63.15%)		 82.54%
(93.04%)

う～ん、あれ? ここまで全く同じってまだバグってるかな^^; ちとコード確認しなきゃ。

on 0 件のコメント:
Labels: , ,

2026/09/14

大苦戦

やっと解決したと思ったけど、自分のチョンボでまだでしたorz 週末もひたすらエラー回避に奮闘してました。今回元々は単に

TrainTestData trainValidationData = mlContext.Data.TrainTestSplit(data, testFraction: testF);

としてたのは、時系列分割では変えなきゃいけなくて色々とやってますが、それが原因で学習後のモデルでの予測時にエラーが出続けてます。で、結局自前でpipelineを準備しなきゃダメだった。で、その結果、LightGBMのパラメーター指定が出来ない事に。まあ、これは今後また別の方法で実装するかもです。なので若干AutoML5の見た目に違和感がある状態です。

更に、色々変えてたので、ちと原因は不明なんだが、昨夜スタートした24時間学習が今朝起きるとってか、まあ、夜中に目が覚めた時にも、あれ?こんなに早くPCが静かだなぁとは思ったんだが、どうやら1時辺りに吹っ飛んだ模様。なので、今朝起床後に再度24時間スタートさせて出社して、今後半戦がくしゅうそてる最中です。

on 0 件のコメント:
Labels: ,

2026/09/13

Microsoft Rewards 32回目

前回から39日目で本日amazonギフトと交換しました。楽天ポイントは在庫切れ。まあ、どちらでも良かったので。ポイントは色々と持ち合わせてますが、微妙なのはVポイントかな。他のポイ活サイトで貯めたポイントをVポイントに交換する場合に手数料が掛かるとかは問題外だよね。年間で20,000ポイント以上貯まる予定なので有効に利用したいけど、どうかなぁ。

Microsoft Rewards紹介リンク

on 0 件のコメント:
Labels:

2026/09/12

何度も

懲りない自分が情けない(笑) 今週末ってか、まあ、昨日は先日の入院からの経過確認で診察にって事で有給取ってたので、ここの所苦戦してた通常用走破タイム学習に取り組んでました。不覚にもまたCopilotに騙されて本日のつい少し前まで振り回されていました。しかし、今回も奴にギブアップさせて再度自力での作業。まあね、そもそも他力本願とかが悪いのは分かってます。年齢的にもついつい楽をしようとねぇ^^;

人間なめるなよって事で自力で進めました。Copilotとやってた時には本当にダメなのか?って思わされましたが、やはりってか所詮どこかの頭の悪い奴が作ったAIが人間様に勝てる訳はない! こんな事なら最初からって思うけど、この頭の悪いAIも人間よりも優れた部分もあるのも事実。

ここまで何故か以前はRMSE 1.4程度まで出てたのがRMSE 2.15辺りに壁があり、何かが違うんだと思い、AIに頼ってしまったのが原因なんだが、自分がやってきた事が正しかった証明かな? 結局自力で突破して、ってまあね、結局ML.NET 5.0にも戻したし、LightGBMも4.6に戻した。ただね、まだこれをまたML.NET 6.0-previewにしたり、LightGBM 4.7にしたりしたらどうなるかは今後の課題ではあるが^^; ってか、２か月後には.NET 11になり、ML.NET 6.0も正式版になる。その時にLightGBM 4.7が採用されるのかは不明だが、放っておいても着実に進化してく。

少なくとも時間を掛けて学習させても使える状態にはなったので、ここからはまたしばらくは長時間学習させる必要はあります。ただちょっと気になるのは、RMSE的には1.5辺りが出てるんだけど、その検証結果は微妙かも^^; まあ、実際には1.4辺りが期待出来る筈なので、その検証結果を見てみないと何ともではあります。

って事で、実質的に明日朝までにはモデルが揃わないので、最短でも次期バージョンリリースは来週になる感じですm(__)m 参考までにテスト的に学習させた最適RMSE 1.5807は

3,151R 1点 芝(1,486R) ダート(1,545R) 障害(120R) 8頭以下(188R) 9～12頭(720R) 13頭以上(2,243R) 多点
単勝 32.34%
(175.07%)		 27.79%
(185.02%)		 36.38%
(167.83%)		 36.67%
(145.00%)		 38.83%
(137.82%)		 32.36%
(141.36%)		 31.79%
(189.01%)		 61.38%
(136.94%)
複勝 64.90%
(132.59%)		 60.70%
(133.13%)		 68.48%
(133.77%)		 70.83%
(110.75%)		 65.96%
(104.73%)		 70.83%
(127.76%)		 62.91%
(136.47%)		 91.08%
(126.12%)
枠連 23.37%
(255.22%)		 19.58%
(238.20%)		 26.23%
(268.36%)		 28.28%
(274.04%)		 --
(--)		 23.61%
(205.25%)		 21.89%
(256.01%)		 39.80%
(182.09%)
馬連 18.85%
(361.52%)		 15.88%
(307.00%)		 21.55%
(414.48%)		 20.83%
(354.83%)		 20.74%
(127.98%)		 22.08%
(258.71%)		 17.65%
(414.10%)		 34.21%
(281.49%)
ワイド 35.39%
(259.36%)		 31.70%
(237.21%)		 38.96%
(283.07%)		 35.00%
(228.42%)		 40.43%
(105.64%)		 40.83%
(218.93%)		 33.21%
(285.23%)		 57.03%
(245.03%)
馬単 12.28%
(434.78%)		 9.42%
(389.99%)		 14.95%
(487.49%)		 13.33%
(310.75%)		 14.36%
(163.19%)		 13.06%
(258.21%)		 11.86%
(514.22%)		 34.21%
(273.98%)
三連複 12.60%
(726.10%)		 10.70%
(641.74%)		 14.30%
(839.50%)		 14.17%
(310.67%)		 15.96%
(190.05%)		 14.86%
(321.61%)		 11.59%
(900.86%)		 25.77%
(542.78%)
三連単 4.67%
(1,070.35%)		 3.50%
(1,019.75%)		 5.70%
(1,141.87%)		 5.83%
(776.17%)		 4.26%
(289.36%)		 6.25%
(562.17%)		 4.19%
(1,298.94%)		 25.77%
(561.52%)
総合 66.30%
(429.09%)		 61.84%
(397.15%)		 70.23%
(468.11%)		 70.83%
(314.72%)		 69.68%
(159.83%)		 71.11%
(261.75%)		 64.47%
(499.36%)		 91.43%
(414.66%)

うわっ、なんかバグってる(笑) 何をどうするとこんなバグになるのか全く不明orz 明日も忙しくなりそうです。

追記 2026.9.13 8:05

昨夜24時間×5回の学習スタートさせてますが微妙な感じで相変わらずOOMが多い。いや、これ、イテレーション数2,000、葉の数255、分割数47、葉の最小100、学習率0.01でLightGBMパラメーター指定しての学習なんで、もしかすると24時間とかあまり意味ないのかも?まあ、なんにしても、上のモデルを今年1月～8月で検証すると

2,222R 1点 芝(1,051R) ダート(1,091R) 障害(80R) 8頭以下(106R) 9～12頭(538R) 13頭以上(1,578R) 多点
単勝 17.24%
(68.06%)		 14.18%
(64.76%)		 19.71%
(70.36%)		 23.75%
(80.00%)		 22.64%
(54.62%)		 19.89%
(65.09%)		 15.97%
(69.97%)		 43.97%
(80.47%)
複勝 46.08%
(85.18%)		 40.25%
(80.23%)		 50.60%
(89.58%)		 61.25%
(90.38%)		 54.72%
(88.96%)		 50.00%
(79.46%)		 44.17%
(86.88%)		 82.00%
(82.67%)
枠連 8.84%
(71.26%)		 7.15%
(70.29%)		 9.78%
(67.25%)		 16.22%
(139.86%)		 --
(--)		 10.59%
(75.20%)		 7.73%
(65.86%)		 22.11%
(75.74%)
馬連 5.36%
(56.69%)		 4.47%
(55.08%)		 5.96%
(55.70%)		 8.75%
(91.37%)		 10.38%
(34.34%)		 7.25%
(67.42%)		 4.37%
(54.54%)		 14.76%
(63.27%)
ワイド 16.07%
(78.52%)		 13.51%
(72.80%)		 17.87%
(80.03%)		 25.00%
(133.13%)		 28.30%
(64.34%)		 21.56%
(76.69%)		 13.37%
(80.10%)		 35.37%
(78.41%)
馬単 2.66%
(45.85%)		 1.71%
(34.72%)		 3.30%
(50.23%)		 6.25%
(132.25%)		 5.66%
(44.91%)		 3.72%
(61.34%)		 2.09%
(40.63%)		 14.76%
(60.65%)
三連複 3.20%
(79.00%)		 2.76%
(86.96%)		 3.48%
(68.73%)		 5.00%
(114.50%)		 8.49%
(73.21%)		 4.83%
(63.25%)		 2.28%
(84.76%)		 10.80%
(88.16%)
三連単 0.36%
(28.18%)		 0.10%
(7.98%)		 0.55%
(29.56%)		 1.25%
(274.75%)		 0.00%
(0.00%)		 0.37%
(50.45%)		 0.38%
(22.48%)		 10.80%
(112.22%)
総合 47.39%
(64.01%)		 41.58%
(58.91%)		 51.79%
(63.91%)		 63.75%
(131.96%)		 61.32%
(51.48%)		 50.93%
(67.36%)		 45.25%
(63.15%)		 82.54%
(93.04%)

なんか現実的な数値だわ(笑) バグってはいないとなると、たまたま当たり年だった? っていうのか、学習が直前までってか、これもし2004年～2025年まで学習させてると2026年もまた違う感じになるのかも? もしくは、単に秋競馬が得意でこれから化ける?

on 0 件のコメント:
Labels: , ,
登録: 投稿 (Atom)