葉の数抑えた通常用走破タイム学習

葉の数を抑えると何度か回してってのは無意味っぽい。だいたい1回目が一番良い(笑)

パラメータ / 指標	条件 1	条件 2	条件 3	条件 4	条件 5	条件 6	条件 7
NumberOfLeaves	63	79	47	63	63	63	63
NumberOfIterations	1,500	1,500	1,500	1,500	1,500	1,500	1,500
LearningRate	0.020	0.020	0.020	0.020	0.020	0.020	0.020
MaximumBin	63	63	63	79	95	111	127
MinimumExample	50	50	50	50	50	50	50
SubsampleFraction	0.050	0.050	0.050	0.050	0.050	0.050	0.050
FeatureFraction	0.800	0.800	0.800	0.800	0.800	0.800	0.800
L2Regularization	0.500	0.500	0.500	0.500	0.500	0.500	0.500
RMSE	1.6048	1.6072	1.6217	1.6167	1.6073	1.6205	1.6061

追記 2026.10.4 21:22

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

先月中旬頃の自分の記事確認してたら、こんな感じが良いモデル出てるって事で改めて5回程試すと確かにRMSE 1.5910～1.5931でした(笑) これでは4回目が一番だった。

2,427R	1点	芝(1,149R)	ダート(1,189R)	障害(89R)	8頭以下(129R)	9～12頭(597R)	13頭以上(1,701R)	多点
単勝	19.90% (69.19%)	19.15% (71.42%)	20.94% (66.73%)	15.73% (73.26%)	27.13% (69.38%)	21.61% (71.07%)	18.75% (68.51%)	48.62% (72.17%)
複勝	49.44% (78.79%)	45.43% (75.31%)	53.32% (82.72%)	49.44% (71.24%)	59.69% (86.82%)	52.60% (80.44%)	47.56% (77.61%)	85.04% (78.90%)
枠連	11.31% (78.43%)	10.29% (94.24%)	12.38% (66.64%)	8.86% (48.86%)	-- (--)	10.55% (59.58%)	10.93% (80.57%)	27.53% (83.22%)
馬連	7.91% (67.61%)	7.22% (76.01%)	8.75% (61.81%)	5.62% (36.85%)	15.50% (77.29%)	7.71% (61.73%)	7.41% (68.95%)	19.61% (70.77%)
ワイド	21.43% (81.88%)	18.10% (75.92%)	24.73% (88.39%)	20.22% (71.80%)	31.78% (65.35%)	23.62% (81.16%)	19.87% (83.39%)	39.93% (75.69%)
馬単	4.12% (60.97%)	3.57% (74.06%)	4.79% (51.49%)	2.25% (18.65%)	8.53% (78.68%)	4.86% (55.14%)	3.53% (61.67%)	19.61% (68.60%)
三連複	5.23% (86.02%)	3.92% (96.59%)	6.48% (78.85%)	5.62% (45.39%)	8.53% (44.73%)	6.03% (75.01%)	4.70% (93.02%)	13.84% (76.44%)
三連単	0.95% (84.14%)	0.87% (149.14%)	1.01% (26.02%)	1.12% (21.35%)	3.10% (113.80%)	1.17% (64.24%)	0.71% (88.87%)	13.84% (71.43%)
総合	50.97% (75.85%)	47.26% (88.99%)	54.67% (65.32%)	49.44% (48.42%)	62.02% (76.58%)	54.10% (68.54%)	49.03% (77.82%)	85.37% (72.84%)

AIに言わせると、この辺りが純粋に走破タイムを予測させる限界だからと何度も言われ続け、いやね、既にその為の対処として相対タイムやランキングもためしてますって(笑) ただ、新馬戦用走破タイム学習はもう少し良い感じになってるから諦めきれてないんだよね。

SubsampleFractionを0.025→0.050にして5回試すとRMSE 1.5909～1.5932だった。

2,427R	1点	芝(1,149R)	ダート(1,189R)	障害(89R)	8頭以下(129R)	9～12頭(597R)	13頭以上(1,701R)	多点
単勝	19.78% (68.64%)	19.76% (73.91%)	20.19% (64.75%)	14.61% (52.58%)	27.91% (70.70%)	22.11% (66.65%)	18.34% (69.18%)	48.95% (76.39%)
複勝	49.65% (79.57%)	46.13% (78.97%)	52.73% (79.98%)	53.93% (81.80%)	63.57% (93.80%)	52.26% (76.67%)	47.68% (79.51%)	85.00% (80.47%)
枠連	11.72% (74.64%)	10.90% (83.85%)	12.64% (68.21%)	8.86% (51.27%)	-- (--)	10.55% (55.73%)	11.46% (77.03%)	27.31% (81.11%)
馬連	8.16% (76.46%)	7.40% (80.46%)	9.00% (75.14%)	6.74% (42.36%)	17.05% (77.52%)	7.54% (60.32%)	7.70% (82.04%)	19.78% (71.76%)
ワイド	21.63% (89.17%)	18.62% (84.53%)	24.56% (93.92%)	21.35% (85.62%)	34.88% (68.53%)	22.95% (80.85%)	20.16% (93.65%)	41.08% (81.10%)
馬単	4.33% (69.63%)	4.18% (79.96%)	4.63% (62.13%)	2.25% (36.63%)	7.75% (51.55%)	5.36% (72.13%)	3.70% (70.13%)	19.78% (70.25%)
三連複	5.27% (75.96%)	3.74% (76.70%)	6.73% (72.09%)	5.62% (117.98%)	8.53% (34.65%)	6.20% (58.73%)	4.70% (85.14%)	13.97% (77.80%)
三連単	0.78% (31.33%)	0.52% (31.98%)	1.01% (30.77%)	1.12% (30.67%)	0.78% (9.69%)	1.51% (74.77%)	0.53% (17.73%)	13.97% (77.01%)
総合	51.30% (70.63%)	48.30% (73.62%)	53.99% (68.37%)	53.93% (62.52%)	66.67% (58.06%)	53.43% (68.23%)	49.38% (71.80%)	85.37% (76.56%)

RMSEの差はたったの0.0001なんだが、随分と違う感じだ。ひとつひとつでは判断難しいけど、総合的には的中率は向上してるのはその差かな?