しっかりとランキング学習

ここまでは、ざっくりと学習が動けばって感じの学習でした。これまでの回帰学習では当たり前のカテゴリ変数の扱いなんかを考慮してみた。イテレーション数3,000、葉255、率0.015、分割数127、葉の最小50、ボーダー500でNDCG@1=0.7789 NDCG@3=.7843となり

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.55% (79.98%)	27.06% (79.05%)	27.89% (79.30%)	30.00% (102.42%)	36.07% (78.26%)	27.69% (76.42%)	26.74% (81.32%)	57.45% (75.59%)
複勝	59.25% (84.40%)	58.85% (83.57%)	58.87% (84.19%)	70.00% (98.83%)	73.52% (91.69%)	63.91% (83.73%)	56.41% (83.96%)	89.67% (83.07%)
枠連	16.16% (88.23%)	16.75% (101.23%)	15.18% (75.71%)	23.23% (102.42%)	-- (--)	18.29% (86.32%)	14.42% (83.28%)	31.34% (78.95%)
馬連	12.62% (85.04%)	13.32% (94.66%)	11.54% (75.12%)	17.50% (85.67%)	23.74% (68.54%)	15.20% (87.70%)	10.75% (85.64%)	25.53% (73.76%)
ワイド	27.64% (87.94%)	28.77% (87.50%)	26.18% (88.71%)	31.67% (83.50%)	51.60% (84.06%)	32.14% (84.23%)	23.98% (89.52%)	49.46% (79.56%)
馬単	6.66% (72.51%)	6.69% (71.44%)	6.62% (72.67%)	6.67% (85.25%)	12.33% (62.42%)	7.17% (81.83%)	5.97% (70.31%)	25.53% (70.05%)
三連複	6.19% (65.65%)	7.28% (67.05%)	5.04% (64.96%)	6.67% (55.33%)	19.18% (77.35%)	7.91% (62.62%)	4.45% (65.60%)	16.47% (70.38%)
三連単	1.33% (51.03%)	1.72% (66.56%)	1.03% (38.82%)	0.00% (0.00%)	2.74% (25.62%)	1.98% (47.61%)	0.99% (54.47%)	16.47% (66.49%)
総合	60.46% (76.69%)	60.27% (80.96%)	59.96% (72.42%)	70.00% (76.10%)	78.08% (69.71%)	64.52% (76.31%)	57.52% (76.76%)	90.01% (70.74%)

これをどうやってNDCG@1をこれまで最高だった0.7824超えや更に0.8以上に持っていくか。少しいじってNDCG@1=0.7778 NDCG@3=0.7970で

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.29% (80.37%)	27.18% (79.54%)	27.40% (80.39%)	27.50% (91.83%)	38.81% (90.09%)	26.95% (73.71%)	26.37% (81.72%)	57.77% (77.58%)
複勝	58.58% (84.31%)	58.67% (84.38%)	58.14% (83.96%)	63.33% (88.33%)	72.15% (89.04%)	63.41% (83.49%)	55.75% (84.16%)	89.70% (82.71%)
枠連	16.13% (84.40%)	16.04% (92.75%)	15.69% (75.57%)	24.24% (105.96%)	-- (--)	16.93% (77.34%)	14.83% (81.39%)	30.82% (77.04%)
馬連	12.36% (81.55%)	12.37% (83.90%)	11.91% (78.05%)	18.33% (96.33%)	22.37% (58.45%)	14.09% (76.18%)	10.88% (85.42%)	25.47% (72.26%)
ワイド	26.74% (83.37%)	27.47% (81.88%)	25.64% (85.46%)	31.67% (75.83%)	47.95% (79.04%)	31.64% (83.45%)	23.20% (83.74%)	49.29% (77.10%)
馬単	6.95% (87.37%)	6.93% (86.23%)	6.87% (87.05%)	8.33% (107.83%)	12.33% (53.29%)	6.30% (66.25%)	6.67% (97.49%)	25.47% (70.36%)
三連複	6.25% (59.12%)	7.16% (56.64%)	5.22% (59.50%)	7.50% (88.92%)	19.63% (79.18%)	7.91% (65.27%)	4.49% (55.26%)	16.27% (73.85%)
三連単	1.42% (52.89%)	1.66% (60.84%)	1.22% (46.49%)	0.83% (28.75%)	3.65% (37.03%)	1.48% (38.34%)	1.19% (59.17%)	16.27% (68.67%)
総合	59.80% (76.57%)	59.86% (77.96%)	59.30% (74.55%)	65.83% (85.02%)	75.80% (69.45%)	64.28% (70.50%)	56.86% (78.54%)	90.01% (71.90%)

これはダメな修正だったorz 元に戻してイテレーション数5,000、葉255、率0.01、分割数127、葉の最小50、ボーダー500でNDCG@1=0.7795 NDCG@3=0.7845でOOM! 分割数63にするとNDCG@1=0.7805 NDCG@3=0.7847とかになるけど、やはりOOM。イテレーション数3,000に落とせばNDCG@1=0.7787 NDCG@3=0.7840で

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.50% (79.37%)	26.88% (78.62%)	27.76% (78.89%)	32.50% (96.58%)	38.81% (88.58%)	29.30% (82.94%)	25.88% (77.35%)	57.40% (76.15%)
複勝	59.16% (84.37%)	58.73% (83.78%)	58.81% (84.14%)	70.00% (95.92%)	71.23% (87.53%)	65.02% (85.91%)	56.12% (83.58%)	89.38% (82.77%)
枠連	15.83% (83.65%)	16.25% (93.43%)	15.18% (75.41%)	20.20% (75.56%)	-- (--)	17.68% (78.73%)	14.22% (79.98%)	31.31% (77.61%)
馬連	12.62% (82.15%)	13.08% (87.43%)	11.91% (77.83%)	15.83% (67.08%)	25.57% (74.79%)	14.96% (81.04%)	10.67% (83.19%)	25.73% (71.42%)
ワイド	27.81% (91.67%)	28.83% (94.68%)	26.61% (89.20%)	30.00% (83.17%)	51.14% (84.47%)	33.75% (90.25%)	23.73% (92.80%)	50.22% (79.93%)
馬単	6.89% (74.28%)	6.93% (75.68%)	6.80% (74.68%)	7.50% (49.17%)	15.07% (76.76%)	7.66% (76.14%)	5.89% (73.44%)	25.73% (67.97%)
三連複	6.11% (74.37%)	6.99% (58.84%)	5.10% (91.42%)	7.50% (59.17%)	18.72% (70.32%)	7.79% (64.45%)	4.41% (78.04%)	16.85% (74.44%)
三連単	1.39% (44.12%)	1.66% (58.04%)	1.22% (33.06%)	0.00% (0.00%)	4.11% (64.61%)	1.85% (42.69%)	0.99% (42.75%)	16.85% (70.26%)
総合	60.46% (76.65%)	60.15% (78.50%)	60.02% (75.58%)	70.83% (65.61%)	76.26% (78.15%)	65.76% (75.27%)	57.27% (76.39%)	89.67% (72.51%)

まあ、これもJV-Linkの32bit縛りの宿命だな。

ちょっと勘違いとかもあり、それらを修正してイテレーション数3,000、葉255、率0.018、分割数63、葉の最小50、ボーダー500でNDCG@1=0.7778 NDCG@3=0.7836となりOOMって最悪です。

その後のランキング予測

継続は力なりってひたすら続けてます。イテレーション数3,000、葉255、率0.015でNDCG@1=0.7798 NDCG@3=0.7862のモデルは

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	26.86% (78.15%)	26.52% (77.47%)	27.10% (77.83%)	28.33% (92.08%)	37.90% (88.63%)	27.94% (78.88%)	25.50% (76.96%)	57.02% (76.73%)
複勝	58.81% (84.73%)	58.38% (83.68%)	58.38% (84.79%)	70.83% (98.75%)	70.32% (88.31%)	63.41% (84.14%)	56.24% (84.61%)	89.64% (82.82%)
枠連	16.16% (86.69%)	15.83% (90.48%)	15.57% (80.98%)	30.30% (123.64%)	-- (--)	17.92% (77.50%)	14.54% (84.26%)	30.92% (80.31%)
馬連	12.50% (84.08%)	12.37% (85.44%)	11.91% (80.52%)	22.50% (113.92%)	20.09% (59.59%)	15.08% (76.58%)	10.96% (88.80%)	25.47% (73.73%)
ワイド	27.24% (86.20%)	27.53% (83.01%)	26.25% (88.95%)	36.67% (93.42%)	46.58% (77.99%)	32.88% (83.20%)	23.61% (87.94%)	48.83% (79.51%)
馬単	6.74% (76.75%)	6.28% (68.85%)	6.93% (83.40%)	10.83% (96.83%)	10.05% (61.37%)	7.66% (68.41%)	6.14% (80.92%)	25.47% (71.05%)
三連複	6.05% (71.01%)	6.99% (66.61%)	5.10% (76.10%)	5.83% (63.08%)	18.72% (76.53%)	7.17% (59.09%)	4.53% (74.48%)	16.38% (69.47%)
三連単	1.22% (45.12%)	1.42% (55.67%)	0.97% (34.76%)	1.67% (38.67%)	1.83% (13.93%)	1.98% (45.28%)	0.91% (47.88%)	16.38% (64.12%)
総合	60.00% (76.45%)	59.56% (76.10%)	59.60% (75.89%)	71.67% (89.30%)	73.97% (66.62%)	63.91% (71.63%)	57.44% (78.23%)	89.99% (69.74%)

これが限界かは不明だけど、もう少し精度上げたい。Copilotはもう限界で、ギブアップさせた。イテレーション数3,000、葉255、率0.012でNDCG@1=0.7810 NDCG@3=0.7861のモデルは

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.09% (81.61%)	26.64% (78.90%)	27.28% (83.09%)	30.83% (99.50%)	36.07% (86.62%)	27.81% (79.67%)	26.04% (81.81%)	56.82% (76.39%)
複勝	58.29% (84.26%)	57.73% (82.79%)	58.02% (84.81%)	70.00% (97.42%)	72.15% (90.73%)	62.42% (84.18%)	55.67% (83.70%)	89.73% (83.36%)
枠連	16.06% (87.13%)	16.04% (93.80%)	15.37% (79.36%)	27.27% (115.45%)	-- (--)	18.05% (81.57%)	14.38% (83.45%)	31.12% (79.57%)
馬連	12.39% (80.45%)	12.31% (83.55%)	11.79% (75.02%)	21.67% (111.17%)	21.92% (68.04%)	14.96% (79.48%)	10.67% (81.89%)	25.56% (72.29%)
ワイド	26.74% (86.70%)	27.18% (86.43%)	25.52% (85.54%)	37.50% (106.42%)	49.77% (81.64%)	32.01% (81.63%)	22.91% (88.85%)	49.06% (78.84%)
馬単	7.09% (81.46%)	7.10% (86.39%)	6.74% (74.71%)	11.67% (104.75%)	12.33% (77.53%)	8.03% (73.15%)	6.30% (84.59%)	25.56% (68.90%)
三連複	5.90% (61.94%)	7.05% (60.56%)	4.74% (64.39%)	5.83% (47.83%)	18.72% (82.01%)	7.42% (60.95%)	4.24% (60.46%)	16.41% (69.09%)
三連単	1.22% (54.29%)	1.60% (67.51%)	0.85% (40.59%)	0.83% (56.08%)	2.74% (112.74%)	1.73% (36.30%)	0.91% (55.01%)	16.41% (64.06%)
総合	59.54% (77.09%)	58.85% (79.70%)	59.36% (73.41%)	71.67% (91.81%)	75.34% (85.62%)	63.04% (72.12%)	56.94% (77.47%)	90.01% (69.25%)

少し精度が上がった。しかし、何が良いって、まあ、相対タイムよりは確実に精度は上で、これ、学習時間が非常に短い。AutoMLでは何時間も学習させてたけど、イテレーション数10,000とかでも30分掛からない。ただ、AutoMLではモニター機能があるので学習中のログが表示出来るけど、ランキング学習ではひたすら待つ(笑)

学習時のオプションを追加してイテレーション5,000、葉255、率0.015でNDCG@1=0.7824 NDCG@3=0.7868のモデルは残念ながらOOM(Out of Memory)。イテレーション数3,500、葉175、率0.01でNDCG@1=0.7812 NDCG@3=0.7862のモデルは

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	26.92% (78.56%)	26.52% (78.58%)	27.16% (77.32%)	29.17% (95.17%)	37.90% (92.47%)	28.31% (82.24%)	25.46% (76.08%)	56.70% (77.24%)
複勝	58.35% (83.43%)	58.67% (83.61%)	57.53% (82.72%)	65.00% (90.58%)	74.89% (95.07%)	63.29% (83.98%)	55.21% (82.19%)	89.46% (83.46%)
枠連	16.06% (85.70%)	16.18% (92.04%)	15.25% (78.93%)	27.27% (103.03%)	-- (--)	17.55% (76.76%)	14.54% (83.24%)	30.86% (77.74%)
馬連	12.36% (74.97%)	12.43% (76.20%)	11.66% (72.45%)	20.83% (92.00%)	22.83% (71.28%)	14.59% (72.45%)	10.67% (76.14%)	25.15% (71.15%)
ワイド	27.06% (85.33%)	28.24% (83.91%)	25.21% (86.53%)	35.83% (88.83%)	50.68% (85.02%)	32.14% (82.97%)	23.24% (86.14%)	49.00% (79.25%)
馬単	6.69% (70.54%)	6.28% (63.92%)	6.80% (77.03%)	10.83% (74.67%)	11.87% (79.13%)	7.29% (59.69%)	6.02% (73.38%)	25.15% (69.00%)
三連複	5.93% (67.47%)	6.81% (63.17%)	4.92% (71.86%)	7.50% (67.75%)	15.98% (61.92%)	8.41% (78.69%)	4.20% (64.23%)	16.18% (66.09%)
三連単	1.39% (65.56%)	1.66% (56.86%)	1.09% (74.39%)	1.67% (67.08%)	3.20% (124.47%)	1.85% (36.46%)	1.07% (69.95%)	16.18% (59.89%)
総合	59.54% (76.32%)	59.86% (74.42%)	58.75% (77.65%)	65.83% (84.48%)	78.54% (87.05%)	63.78% (71.65%)	56.41% (76.42%)	89.81% (66.87%)

これじゃ精度が落ちたorz でNモデルはOOM。ちょっとオプション増やしてイテレーション数3.500、葉255、率0.01、分割数111、葉の最小50だとNDCG@1=0.7815 NDCG@3=0.7858のモデルは

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.06% (80.47%)	27.12% (79.40%)	26.85% (80.06%)	29.17% (101.00%)	40.18% (98.81%)	27.19% (75.19%)	25.83% (80.57%)	56.85% (77.53%)
複勝	58.73% (84.67%)	58.61% (84.06%)	58.14% (84.49%)	68.33% (95.83%)	73.97% (94.38%)	62.30% (82.05%)	56.16% (84.67%)	89.52% (83.36%)
枠連	15.70% (85.17%)	16.11% (92.68%)	14.68% (76.87%)	26.26% (110.30%)	-- (--)	17.55% (78.54%)	14.09% (81.98%)	31.25% (79.81%)
馬連	12.19% (81.42%)	12.73% (87.41%)	11.12% (73.85%)	19.17% (101.00%)	21.92% (70.50%)	14.96% (81.52%)	10.38% (82.37%)	25.38% (78.77%)
ワイド	27.15% (88.37%)	28.54% (89.37%)	25.27% (87.30%)	33.33% (89.08%)	48.40% (81.37%)	33.00% (87.18%)	23.28% (89.40%)	49.09% (80.89%)
馬単	6.66% (76.91%)	6.87% (81.64%)	6.32% (72.32%)	8.33% (73.17%)	12.33% (82.88%)	7.42% (67.03%)	5.89% (79.66%)	25.38% (75.77%)
三連複	5.99% (73.21%)	6.99% (61.69%)	4.92% (85.65%)	6.67% (64.75%)	17.35% (72.05%)	8.41% (76.18%)	4.16% (72.32%)	16.64% (73.68%)
三連単	1.30% (49.74%)	1.60% (67.31%)	1.09% (35.34%)	0.00% (0.00%)	3.20% (119.09%)	1.48% (27.54%)	1.07% (50.88%)	16.64% (67.69%)
総合	59.88% (77.39%)	59.68% (80.19%)	59.48% (74.47%)	68.33% (78.70%)	77.17% (88.44%)	62.79% (71.91%)	57.35% (77.73%)	89.87% (72.89%)

微妙な変化かなぁ。Geminiの助言で評価部分の変更してみた。イテレーション数3.500、葉255、率0.01、分割数127、葉の最小50だとNDCG@1=0.7777 NDCG@3=0.7836と評価自体も落ちてモデルは

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.35% (80.41%)	26.94% (78.50%)	27.76% (81.54%)	27.50% (91.75%)	33.79% (75.11%)	29.17% (81.42%)	26.16% (80.55%)	57.89% (76.61%)
複勝	58.64% (83.74%)	58.44% (83.41%)	58.38% (83.58%)	65.00% (90.42%)	69.41% (86.03%)	63.78% (83.50%)	55.95% (83.61%)	89.46% (82.18%)
枠連	16.22% (84.51%)	16.32% (89.38%)	15.69% (79.20%)	23.23% (99.60%)	-- (--)	17.18% (73.97%)	14.87% (82.66%)	31.93% (80.30%)
馬連	12.42% (79.36%)	12.37% (76.60%)	12.03% (81.54%)	18.33% (88.17%)	22.37% (61.00%)	14.34% (74.59%)	10.88% (82.60%)	25.96% (73.46%)
ワイド	27.26% (88.33%)	27.83% (88.06%)	26.37% (89.24%)	31.67% (79.67%)	47.95% (78.54%)	32.51% (83.82%)	23.65% (90.72%)	49.35% (78.84%)
馬単	6.40% (70.44%)	6.16% (66.72%)	6.62% (73.59%)	6.67% (79.50%)	11.42% (58.49%)	6.80% (68.07%)	5.81% (72.30%)	25.96% (70.83%)
三連複	6.19% (72.52%)	6.63% (58.27%)	5.59% (85.94%)	8.33% (88.92%)	18.72% (66.30%)	7.91% (69.07%)	4.49% (74.23%)	17.02% (70.09%)
三連単	1.27% (34.32%)	1.36% (35.89%)	1.28% (35.21%)	0.00% (0.00%)	3.65% (41.78%)	1.73% (39.21%)	0.91% (32.01%)	17.02% (64.16%)
総合	59.91% (74.06%)	59.80% (71.74%)	59.54% (76.22%)	66.67% (76.75%)	74.89% (66.75%)	64.40% (71.71%)	57.07% (74.84%)	89.75% (69.68%)

当然ながら下げてる感じだけど、微妙な変化で単勝27.35%や枠連、馬連、ワイド、三連複は微増だけど回収率が下がってる。更に評価部を変えてみた所、NDCG@1=0.7789 NDCG@3=0.7843となり、

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.81% (80.32%)	27.53% (80.14%)	27.89% (79.21%)	30.83% (98.17%)	37.44% (87.21%)	29.17% (81.17%)	26.49% (79.42%)	57.48% (75.65%)
複勝	59.83% (85.39%)	59.92% (85.53%)	59.30% (84.90%)	65.83% (90.17%)	73.06% (92.24%)	64.89% (85.77%)	56.94% (84.65%)	89.81% (82.87%)
枠連	16.71% (89.15%)	16.89% (97.10%)	15.95% (80.53%)	26.26% (112.93%)	-- (--)	18.05% (82.99%)	15.20% (85.54%)	31.28% (78.30%)
馬連	13.00% (85.52%)	13.32% (89.44%)	12.15% (80.61%)	20.00% (97.67%)	24.20% (70.27%)	15.08% (82.73%)	11.29% (87.82%)	25.59% (71.88%)
ワイド	27.96% (92.89%)	28.60% (93.57%)	26.91% (91.95%)	33.33% (96.17%)	50.23% (83.06%)	32.26% (84.87%)	24.52% (96.45%)	49.67% (79.13%)
馬単	7.03% (79.00%)	6.75% (77.00%)	7.11% (78.63%)	10.00% (112.33%)	13.24% (76.26%)	7.54% (82.66%)	6.30% (78.03%)	25.59% (68.44%)
三連複	6.19% (75.31%)	7.22% (62.40%)	5.10% (90.72%)	6.67% (45.75%)	19.63% (83.52%)	7.66% (64.50%)	4.49% (78.18%)	16.73% (76.21%)
三連単	1.27% (39.51%)	1.66% (51.78%)	0.97% (29.81%)	0.00% (0.00%)	2.28% (21.19%)	1.85% (45.90%)	0.99% (39.04%)	16.73% (71.93%)
総合	61.13% (78.24%)	61.34% (79.25%)	60.51% (77.03%)	66.67% (80.95%)	77.63% (73.39%)	65.64% (76.32%)	58.14% (78.64%)	90.10% (73.53%)

複勝が59.83は目標の6割にあと少し!評価部のボーダー(400)を変えてNDCG@1=0.7789 NDCG@3=0.7855となると、

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.35% (79.44%)	27.18% (79.07%)	27.46% (78.23%)	28.33% (101.17%)	36.07% (81.42%)	28.68% (81.30%)	26.12% (78.64%)	57.08% (74.68%)
複勝	59.42% (85.35%)	59.38% (84.72%)	58.93% (85.29%)	66.67% (95.00%)	73.06% (91.96%)	63.78% (84.38%)	56.74% (85.08%)	89.49% (82.47%)
枠連	16.06% (86.02%)	15.97% (91.75%)	15.63% (80.02%)	24.24% (99.80%)	-- (--)	17.06% (76.74%)	14.71% (83.65%)	30.92% (75.61%)
馬連	12.45% (80.53%)	12.37% (81.22%)	11.97% (78.70%)	20.00% (95.92%)	21.92% (62.10%)	14.46% (75.64%)	10.92% (83.82%)	25.41% (68.87%)
ワイド	26.98% (87.83%)	27.59% (89.10%)	26.00% (87.31%)	31.67% (76.92%)	46.12% (75.57%)	31.40% (79.21%)	23.77% (91.80%)	49.61% (77.85%)
馬単	7.06% (80.30%)	7.05% (81.92%)	6.87% (77.75%)	10.00% (92.50%)	11.87% (60.91%)	8.03% (81.27%)	6.30% (81.73%)	25.41% (65.45%)
三連複	6.05% (60.61%)	7.10% (61.34%)	4.86% (58.39%)	7.50% (80.92%)	16.89% (70.46%)	7.54% (59.32%)	4.57% (60.16%)	16.67% (72.16%)
三連単	1.36% (38.18%)	1.66% (42.77%)	1.09% (32.16%)	0.83% (56.08%)	4.11% (45.84%)	1.48% (26.95%)	1.07% (41.23%)	16.67% (67.43%)
総合	60.46% (74.63%)	60.45% (76.16%)	59.96% (72.19%)	67.50% (87.01%)	77.17% (69.75%)	64.28% (70.60%)	57.68% (75.76%)	89.78% (70.11%)

これでは逆効果なのでボーダー(600)変えてNDCG@1=0.7774 NDCG@3=0.7819はもっとダメかな?

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.93% (81.85%)	27.47% (80.65%)	28.07% (82.16%)	32.50% (94.58%)	38.36% (85.25%)	30.78% (88.32%)	26.04% (79.39%)	57.74% (76.76%)
複勝	59.28% (84.42%)	58.73% (83.60%)	59.05% (84.44%)	70.00% (95.58%)	70.78% (86.58%)	65.51% (86.42%)	56.16% (83.56%)	89.64% (83.32%)
枠連	16.16% (82.14%)	16.47% (90.69%)	15.31% (72.59%)	25.25% (112.22%)	-- (--)	17.55% (75.33%)	14.67% (79.20%)	31.31% (77.05%)
馬連	12.33% (78.48%)	12.67% (81.98%)	11.42% (72.29%)	20.00% (114.17%)	23.29% (61.64%)	14.59% (79.01%)	10.59% (79.83%)	25.85% (73.76%)
ワイド	27.09% (84.34%)	28.06% (84.19%)	25.64% (84.26%)	33.33% (87.50%)	49.32% (80.50%)	32.14% (83.00%)	23.40% (85.13%)	49.93% (80.21%)
馬単	6.86% (73.91%)	6.69% (70.11%)	6.80% (74.79%)	10.00% (115.33%)	14.16% (66.58%)	7.79% (82.32%)	5.89% (71.76%)	25.85% (71.22%)
三連複	6.08% (66.49%)	6.81% (56.56%)	5.10% (74.69%)	9.17% (93.83%)	19.18% (71.74%)	7.79% (62.92%)	4.33% (67.21%)	16.90% (73.93%)
三連単	1.51% (36.87%)	1.78% (41.30%)	1.22% (31.75%)	1.67% (44.75%)	5.02% (56.76%)	2.10% (49.72%)	0.99% (30.79%)	16.90% (69.50%)
総合	60.43% (73.44%)	60.09% (73.27%)	60.09% (72.12%)	70.00% (94.36%)	74.89% (72.72%)	66.25% (75.88%)	57.19% (72.11%)	89.90% (72.70%)

それでも検証的にはそんなに悪くなった訳じゃない。ボーダー(550)にするとNDCG@1=0.7782 NDCG@3=0.7829で

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.61% (78.89%)	27.00% (78.46%)	27.95% (78.27%)	31.67% (93.42%)	36.99% (85.53%)	28.80% (78.49%)	26.37% (78.42%)	57.92% (77.30%)
複勝	58.70% (83.50%)	58.20% (82.68%)	58.75% (84.08%)	65.00% (87.08%)	72.60% (90.37%)	63.29% (82.03%)	55.91% (83.37%)	89.55% (82.77%)
枠連	15.87% (82.98%)	16.32% (92.95%)	14.99% (72.69%)	23.23% (104.85%)	-- (--)	17.18% (78.76%)	14.42% (79.12%)	31.18% (77.01%)
馬連	12.27% (80.76%)	12.67% (84.65%)	11.42% (75.29%)	18.33% (101.08%)	21.92% (60.18%)	14.34% (79.04%)	10.71% (83.19%)	25.82% (73.47%)
ワイド	27.38% (89.91%)	28.48% (94.23%)	25.88% (85.77%)	32.50% (85.92%)	46.58% (74.47%)	33.00% (87.53%)	23.77% (92.10%)	49.67% (79.86%)
馬単	6.71% (76.38%)	6.69% (82.20%)	6.50% (67.27%)	10.00% (119.42%)	10.50% (51.42%)	7.79% (80.57%)	6.02% (77.24%)	25.82% (71.63%)
三連複	6.37% (68.93%)	7.10% (66.18%)	5.53% (72.48%)	7.50% (59.17%)	18.26% (65.66%)	7.91% (66.98%)	4.78% (69.88%)	17.05% (73.10%)
三連単	1.36% (44.88%)	1.54% (52.05%)	1.22% (39.62%)	0.83% (16.00%)	3.20% (39.45%)	1.85% (41.97%)	1.03% (46.34%)	17.05% (68.05%)
総合	59.80% (75.68%)	59.50% (78.88%)	59.72% (71.93%)	65.00% (82.89%)	76.71% (66.73%)	63.91% (74.42%)	56.90% (76.21%)	89.84% (71.92%)

微妙だ。

WEBカメラ

今の住まいになってから固定電話は契約してないので、まあ、基本通話するのはスマホではあるけど、こちらから掛ける場合、何かとハンズフリーが良いし、スマホでカレンダーとか確認したりもあるのでPCから発信する事が多いのですが、多いと言っても滅多に使わないのも確かです。

少し前にOSのクリーンインストールしてからはここまで利用した事は無かったのですが、歯科定期健診の予約の為にRakuten Link立ち上げて試すもマイクが繋がってないよって感じで発信出来ない。え?って事でこれまで利用していたMicrosoft LifeCam HD-3000を見ると確かに本来はパイロットランプが点灯してたと思うが、消えてる。まあね、これ2013年に2千円未満で購入してます。13年程度利用してれば故障しても不思議ではないです。で、新たなWEBカメラを探すにあたり、滅多にWEBカメラとして使う事はないのでなんでも良い気もしたけど、それでも何かでZoomとか使わなきゃならない事もあるかもだし、そもそも、もう随分前からWindowsでもHelloとかの生体認証はあったりするが、ハードが無かった訳で、先日のOSのクリーンインストールでも、自動ログイン設定で悩んだりもした事もあり、Windows Hello対応で探し、コスト的にもギリなELECOM UCAM-CF20FBBKを楽天で7,660円で購入してみた。2021年発売でメーカー的には既に終売品ですが、何年か壊れずに使えれば良いかなっと。

ユーザーマニュアルは付属せず、ネットで確認して設定すれば、まあ、Windows10とかではドライバーが必要だったりもしたようだけど、Windows 11では特に自動認識してくれて、顔認証設定は当たり前ですが、Windowsの設定から行いました。特に問題なく行けたので良かった。

第76回安田記念

ちと友人の来客とかがあり、時間がないです。まだ勝率は相対予測用のままなので今回のランキングに変更中で判定が真逆なので無視して下さい。

出来たばっかりのランキング予測に丸乗りで

何も考えずに参加(笑)

ランキング学習の初フライト

気合い入れてコーディングしてきましたが、なんとか検証出来る所まで来ました。

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	0.93% (39.29%)	0.77% (22.89%)	1.15% (58.99%)	0.00% (0.00%)	1.83% (11.83%)	1.24% (32.87%)	0.74% (43.91%)	5.30% (50.12%)
複勝	4.11% (41.77%)	4.03% (38.35%)	3.83% (43.00%)	9.17% (72.92%)	6.85% (26.12%)	6.06% (45.75%)	3.21% (41.85%)	19.33% (56.03%)
枠連	1.40% (46.26%)	1.77% (48.20%)	1.14% (47.43%)	0.00% (0.00%)	-- (--)	1.11% (40.06%)	1.40% (45.39%)	3.57% (45.95%)
馬連	0.03% (4.57%)	0.00% (0.00%)	0.06% (9.59%)	0.00% (0.00%)	0.00% (0.00%)	0.12% (19.51%)	0.00% (0.00%)	0.43% (25.44%)
ワイド	0.26% (6.58%)	0.41% (11.24%)	0.12% (2.28%)	0.00% (0.00%)	1.37% (16.12%)	0.25% (4.64%)	0.16% (6.37%)	1.68% (27.69%)
馬単	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.43% (28.34%)
三連複	0.06% (12.41%)	0.00% (0.00%)	0.12% (26.04%)	0.00% (0.00%)	0.00% (0.00%)	0.25% (52.99%)	0.00% (0.00%)	0.26% (29.22%)
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%)	0.00% (0.00%)	0.26% (22.90%)
総合	5.21% (18.48%)	5.39% (14.38%)	4.74% (23.28%)	9.17% (9.32%)	7.76% (7.72%)	6.80% (24.48%)	4.45% (17.19%)	21.22% (29.47%)

で、まあ、単に学習もやっとエラー無く実行出来、検証も出来る所まで持ってくるのが精一杯でしたので、ここからこのモデルの実際の調整になります。しかし、こんなに外せるのかって程酷い予測結果ですね。これ、例えば、18頭立ての単勝って何も考えずに5.5%程度の確率はあるのに0.93%とか狙って外さないと出来ないレベル(笑)

あくまでも参考程度な話ですが、これ500回学習させてNDCG@1=0.7539、NDCG@3=0.7628ってモデルです。

にしても、やはり変だなって事でコードをチェックしてたら決定的なミスがorz 検証部分もそうなんですが、基本的には相対タイム関連を元に派生させる感じでやってたのでやってしまいました。相対タイム予測は予測はタイムなので小さい方が早いので小さい順に順位決めてますが、今回のランキングは数値が大きい方が強いって事で大きい順にしなきゃなのが逆に。どおりで狙って外してるレベルなんですね(笑)

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	24.17% (74.54%)	22.44% (72.05%)	25.70% (76.70%)	27.50% (80.00%)	34.70% (90.50%)	24.85% (73.86%)	22.99% (73.33%)	53.02% (75.61%)
複勝	54.04% (81.66%)	51.81% (78.04%)	55.53% (84.99%)	65.00% (86.83%)	61.64% (79.00%)	57.35% (79.36%)	52.25% (82.66%)	86.34% (81.40%)
枠連	14.05% (76.78%)	13.13% (78.56%)	14.36% (74.49%)	22.22% (87.78%)	-- (--)	14.59% (66.42%)	12.98% (75.36%)	28.88% (78.27%)
馬連	10.88% (74.87%)	10.24% (73.73%)	11.06% (75.90%)	17.50% (76.67%)	21.00% (77.85%)	11.62% (61.97%)	9.72% (78.90%)	22.75% (73.07%)
ワイド	23.94% (76.33%)	23.51% (77.51%)	23.94% (75.66%)	30.00% (68.83%)	43.38% (80.37%)	27.44% (70.63%)	21.01% (77.87%)	44.43% (76.12%)
馬単	5.76% (74.28%)	5.03% (68.57%)	6.20% (79.87%)	10.00% (78.00%)	11.87% (92.92%)	5.69% (55.44%)	5.23% (78.88%)	22.75% (69.29%)
三連複	5.33% (62.21%)	5.51% (48.82%)	5.22% (78.59%)	4.17% (25.92%)	14.61% (74.70%)	6.80% (61.59%)	4.00% (61.28%)	15.02% (72.48%)
三連単	0.93% (50.89%)	0.95% (37.02%)	0.91% (67.19%)	0.83% (22.67%)	2.28% (121.96%)	1.36% (38.58%)	0.66% (48.58%)	15.02% (70.61%)
総合	55.20% (71.37%)	52.93% (66.54%)	56.68% (76.69%)	66.67% (65.35%)	66.21% (88.19%)	58.22% (63.48%)	53.19% (72.11%)	86.63% (72.46%)

こちらが修正した結果です。これなら、ここから学習を進めれば改善の余地がありそうです。

追記 2026.6.7 10:54
2000回にしてNDCG@1=0.7536、NDCG@3=0.7643でしたが、

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	24.63% (75.91%)	23.21% (73.31%)	25.76% (76.85%)	29.17% (99.67%)	35.62% (91.64%)	25.83% (77.53%)	23.24% (73.95%)	53.34% (76.80%)
複勝	54.65% (82.57%)	52.63% (78.89%)	55.71% (85.39%)	68.33% (95.67%)	61.64% (79.27%)	59.33% (82.60%)	52.45% (82.86%)	86.08% (81.27%)
枠連	14.28% (81.49%)	13.34% (86.09%)	14.61% (77.27%)	22.22% (83.13%)	-- (--)	14.96% (77.47%)	13.14% (77.66%)	28.88% (77.72%)
馬連	11.11% (94.45%)	10.54% (112.09%)	11.30% (78.11%)	16.67% (70.42%)	20.09% (61.46%)	11.87% (65.96%)	10.05% (106.93%)	23.07% (76.53%)
ワイド	24.23% (81.99%)	23.80% (87.91%)	24.42% (77.48%)	27.50% (60.58%)	43.38% (85.98%)	27.56% (72.13%)	21.38% (84.92%)	44.34% (77.46%)
馬単	6.31% (79.01%)	5.86% (74.66%)	6.68% (85.04%)	7.50% (57.58%)	12.79% (77.95%)	6.67% (63.07%)	5.60% (84.42%)	23.07% (73.43%)
三連複	5.59% (67.55%)	5.98% (67.66%)	5.29% (70.46%)	4.17% (25.92%)	13.70% (67.53%)	6.80% (63.24%)	4.45% (68.98%)	14.70% (68.27%)
三連単	1.33% (66.18%)	1.42% (82.86%)	1.28% (52.24%)	0.83% (22.67%)	1.83% (42.10%)	1.85% (79.26%)	1.11% (64.00%)	14.70% (66.14%)
総合	55.86% (78.61%)	53.88% (82.87%)	56.93% (75.35%)	69.17% (64.04%)	66.21% (72.28%)	60.20% (72.65%)	53.48% (80.47%)	86.34% (70.75%)

ほんのり改善って感じかな。

新たなモデル

ここまで色々と試した相対タイムなんですが、一定の成果としては出てるとは思うけど、目指しているレベルにはなかなか届かないので、思い切った(?)方向転換をしてみてます。ここまでのモデルはなんにしても回帰(Regression)でした。これ、値の予測です。

今度のモデル学習はModel Builderではサポートされてないので、AutoMLのみです。ざっくりなイメージとして、これまでの順位予測と基本的には似ているのですが、ランキング予測になります。まずはその学習用CSVのコーディングは終わりました。無事に出力出来れば、それを今度は学習させるのですが、順調に進んでくれれば申し分ないのですが、またAIに踊らされて無駄な時間を過ごすかも^^; 週末までにある程度のイメージのものが動いてくれれば良いけど...

追記 2026.6.6 16:13
今回もAIに振り回されてなかなか進展せずに学習が出来ないってか、そもそも学習っていうよりは学習後の評価が出来ないというか、エラーになる現象で苦労しました。CopilotからGeminiに変えてどうにか学習と評価が完了出来るまで来たので、これから予測させて検証って作業になります。良い結果が出ると報われます(笑)

第93回東京優駿(日本ダービー)

今日も色々やりながら、未だ目標には届きませんでした。ダービー終わると新馬戦も始まるので、のんびりしてる暇はない。ダービーはあまり荒れるレースではないけど、基本穴予想なので独自に予想してみました。

無敗の西村君のコンジェスタスやディーのメイショウハチコウも気になりましたが、あっ、書いてて佐々木君より戸崎君の方がとも思いましたが、まあ、だからって馬券には関係ないか(笑)

これ、ここ最近取り掛かってた期待値の導入で変えた奴です。これ含めたリリースに向けて色々やってます。ただ、これ、精度が出ての話なのでちと苦戦してます^^;

新たな試み

またもAIに踊らされてる?(笑) AutoMLやModel Builderの限界?に達したかは別として、精度アップに苦労してるので別のアプローチに挑みました。またPython絡みです。まあ、現在のデータサイエンスの中核なのは間違いないのかもですが、AutoMLやModel Builderとは違うアプローチで最適なLightGBMのパラメータを探すもの? Optunaとかいうものを試し、そのパラメータ使ってLightGBMのモデルを作成して検証してみると

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	24.98% (74.49%)	23.74% (70.17%)	26.37% (79.27%)	23.33% (69.67%)	35.16% (84.38%)	26.82% (75.62%)	23.44% (73.21%)	54.70% (76.27%)
複勝	56.93% (82.66%)	55.77% (81.05%)	57.59% (83.87%)	64.17% (88.58%)	67.58% (84.84%)	61.93% (82.29%)	54.31% (82.58%)	88.08% (81.13%)
枠連	13.53% (70.96%)	12.56% (68.74%)	13.66% (70.17%)	25.25% (115.25%)	-- (--)	16.07% (74.36%)	11.83% (65.32%)	29.04% (75.23%)
馬連	10.33% (62.09%)	9.71% (54.96%)	10.45% (67.31%)	17.50% (90.83%)	23.29% (70.32%)	12.73% (68.03%)	8.36% (59.37%)	23.53% (71.11%)
ワイド	24.31% (75.44%)	23.68% (68.80%)	24.61% (82.36%)	29.17% (73.83%)	47.03% (85.39%)	29.05% (77.35%)	20.68% (73.90%)	45.82% (76.78%)
馬単	5.50% (58.72%)	4.97% (48.41%)	5.83% (67.15%)	8.33% (88.33%)	14.16% (88.58%)	6.18% (55.55%)	4.49% (57.09%)	23.53% (69.60%)
三連複	6.14% (85.96%)	6.63% (61.22%)	5.65% (113.10%)	5.83% (61.83%)	15.07% (61.42%)	9.02% (86.51%)	4.37% (87.99%)	16.03% (68.44%)
三連単	1.48% (48.13%)	1.78% (48.33%)	1.22% (49.78%)	0.83% (22.67%)	5.02% (85.89%)	2.22% (57.97%)	0.91% (41.44%)	16.03% (65.94%)
総合	58.35% (69.79%)	57.31% (62.58%)	58.75% (76.66%)	67.50% (75.51%)	73.52% (80.12%)	63.04% (72.21%)	55.42% (67.61%)	88.39% (69.65%)

こんな感じでした。これRMSEとか\(R^{2}\)値とか不明なんだけど^^; ざっくりとした話、50回が1時間程度で終わって得たパラメータ使ってモデル作成自体は1分程度。次は200回または500回。200回で4時間程度、500回は10時間程度なんだが、ダービーデイにこんな事してるとはorz

追記 2026.5.31 12:13
200回で得たパラメータ使って作成したモデルの検証は

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	25.50% (76.73%)	24.51% (75.65%)	26.31% (77.11%)	28.33% (86.67%)	36.53% (88.22%)	27.94% (82.37%)	23.69% (73.82%)	55.02% (77.50%)
複勝	57.08% (82.30%)	55.77% (80.96%)	58.14% (83.84%)	60.83% (80.00%)	68.95% (84.38%)	61.31% (81.97%)	54.59% (82.22%)	88.31% (81.92%)
枠連	13.76% (76.47%)	12.78% (70.54%)	13.85% (77.16%)	26.26% (149.90%)	-- (--)	16.56% (77.80%)	11.95% (71.17%)	29.79% (78.76%)
馬連	10.56% (67.75%)	9.83% (56.97%)	10.81% (75.95%)	17.50% (107.08%)	22.83% (71.55%)	13.35% (73.13%)	8.53% (65.62%)	23.76% (73.56%)
ワイド	24.31% (75.79%)	23.86% (71.06%)	24.18% (79.82%)	32.50% (87.08%)	46.58% (84.38%)	30.16% (79.09%)	20.35% (73.92%)	46.89% (77.18%)
馬単	5.64% (65.11%)	5.03% (54.04%)	5.95% (73.54%)	10.00% (105.25%)	12.79% (81.32%)	7.05% (69.46%)	4.53% (62.20%)	23.76% (70.63%)
三連複	5.76% (61.08%)	6.45% (58.09%)	4.98% (62.32%)	6.67% (86.17%)	15.53% (58.49%)	8.78% (85.97%)	3.87% (53.02%)	16.32% (68.41%)
三連単	1.19% (48.00%)	1.42% (57.93%)	1.03% (41.32%)	0.00% (0.00%)	3.20% (30.18%)	1.98% (88.89%)	0.74% (35.98%)	16.32% (66.14%)
総合	58.38% (69.05%)	57.31% (65.55%)	59.11% (71.35%)	63.33% (86.38%)	74.43% (71.22%)	61.93% (79.83%)	55.75% (64.74%)	88.54% (70.35%)

確かに改善はしたかもだが、これではまだAutoMLを超えてないorz

追記 2026.5.31 22:13
500回で得たパラメータ使って作成したモデルの検証は

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	25.15% (75.90%)	24.33% (73.64%)	25.88% (77.84%)	26.67% (81.33%)	36.07% (92.56%)	27.19% (76.55%)	23.49% (74.19%)	55.14% (76.69%)
複勝	56.79% (82.29%)	55.12% (79.72%)	57.65% (84.11%)	68.33% (93.42%)	67.58% (83.79%)	61.19% (81.25%)	54.35% (82.50%)	88.48% (81.27%)
枠連	14.05% (76.90%)	12.99% (74.23%)	14.23% (74.88%)	26.26% (146.97%)	-- (--)	16.69% (78.78%)	12.28% (71.39%)	29.23% (75.81%)
馬連	10.74% (68.62%)	9.89% (61.27%)	11.12% (73.53%)	17.50% (104.58%)	21.92% (63.20%)	12.98% (68.74%)	8.98% (69.06%)	23.53% (69.27%)
ワイド	24.40% (75.32%)	23.62% (69.53%)	24.73% (80.54%)	30.83% (85.17%)	47.03% (86.16%)	28.92% (73.62%)	20.85% (74.91%)	46.48% (75.74%)
馬単	5.62% (62.07%)	4.91% (46.65%)	6.20% (76.47%)	7.50% (81.50%)	11.87% (70.37%)	6.92% (60.04%)	4.61% (61.99%)	23.53% (68.64%)
三連複	5.82% (62.30%)	6.39% (50.95%)	5.35% (74.18%)	4.17% (59.00%)	15.53% (61.92%)	8.53% (66.75%)	4.04% (60.85%)	16.32% (67.70%)
三連単	1.42% (49.77%)	1.60% (51.82%)	1.34% (51.30%)	0.00% (0.00%)	4.57% (57.58%)	2.72% (110.58%)	0.70% (28.80%)	16.32% (65.54%)
総合	57.95% (69.04%)	56.54% (63.25%)	58.51% (74.10%)	70.00% (80.03%)	73.06% (73.65%)	62.18% (77.04%)	55.17% (65.46%)	88.71% (69.16%)

これも結局微妙なもので結果論的に言えば今週末も無駄な時間を過ごしたorz

またOSのクリーンインストール

昨夜、AutoMLの12時間×10回ももう少しって段階でしたが、期待出来ないと判断して中断。で、まあ、学習中はWindows Updateをさせない設定にしてるので再開させ、更にNVIDIAのドライバーの更新か?って事でしてみると、見事にクラッシュorz で、基本OSの修復とかはしないので、クリーンインストール。

何度となくクリーンインストールはしてます。で、最近のOSでは自動ログインにちょっと苦労してたりしましたが、前回まではどうにか設定出来てたと思うが、今回は上手く出来ない。で、Copilotに「Microsoftの開発者はバカばっかりだから、OSのクリーンインストール時にはPINしか設定しないのに、自動ログインにはMicroosftアカウントのユーザー名とパスワードとか言われるけど、上手く設定出来ずに、パスワードが違いますとかなるわ」ってぶつけると、ローカルアカウントでのセットアップ方法を教えてくれた。いや、これ、裏技的に情報があったりしたのは確認してますが、アップデート毎に潰されたりしてたので出来ないと思ってました。

Copilotの支持で、まあ、うちは有線でLAN接続してますので、LANケーブルを抜いた状態でクリーンインストールすればローカルアカウントでセットアップが出来、自動ログイン設定もローカルアカウントで可能になるのでって事で試しました。インターネット接続して下さいって所でShift+F10使ってプロンプト起動して OOBE\BYPASSNRO ってコマンド叩くと、OSセットアップがちょっとリスタートする感じで「インターネット接続なし」って選択が可能になりセットアップを継続出来る様になります。で、デスクトップが出る所まで来たら、AutoLogonとか使って自動ログインを有効にして、その後にLANケーブル接続してって感じです。

クリーンインストール後だったのでさらだもインストールしてからフルセットアップしたら、きれいな状態だったからなのか30分で終わった。

いよいよ打つ手が無い?

AutoMLに戻って12時間×10回をしてますが、これまたなんともあまり良い\(R^{2}\)値が出ないのですが、先程終わった学習では0.2549とここまで最高なので検証してみると

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	25.85% (76.80%)	24.63% (72.68%)	26.85% (81.48%)	29.17% (70.75%)	35.62% (77.85%)	28.06% (74.77%)	24.23% (77.38%)	56.27% (76.84%)
複勝	57.42% (84.07%)	56.96% (82.70%)	57.72% (85.85%)	60.00% (78.83%)	71.23% (88.86%)	60.57% (80.41%)	55.13% (84.85%)	89.03% (82.36%)
枠連	13.85% (73.78%)	13.48% (74.61%)	13.60% (71.33%)	23.23% (101.01%)	-- (--)	15.57% (64.55%)	12.40% (72.18%)	30.63% (77.48%)
馬連	11.09% (69.31%)	10.89% (68.80%)	10.87% (69.39%)	16.67% (75.17%)	25.57% (78.90%)	13.47% (70.59%)	8.98% (68.01%)	24.57% (72.13%)
ワイド	25.09% (77.86%)	25.87% (79.44%)	23.75% (75.63%)	32.50% (86.17%)	50.68% (99.00%)	30.90% (80.56%)	20.85% (75.06%)	47.84% (77.30%)
馬単	6.25% (71.70%)	5.98% (68.48%)	6.20% (73.30%)	10.83% (95.00%)	14.61% (82.97%)	7.66% (75.15%)	5.03% (69.53%)	24.57% (69.63%)
三連複	6.11% (75.31%)	7.10% (72.01%)	5.35% (82.13%)	2.50% (28.17%)	14.61% (58.95%)	8.41% (64.45%)	4.57% (80.40%)	16.38% (69.86%)
三連単	1.51% (55.13%)	1.60% (45.30%)	1.46% (67.74%)	0.83% (20.50%)	4.11% (73.88%)	2.35% (59.06%)	0.99% (52.13%)	16.38% (65.61%)
総合	58.61% (72.98%)	58.26% (70.42%)	58.75% (75.88%)	61.67% (68.74%)	74.89% (80.06%)	60.94% (71.19%)	56.37% (72.44%)	89.35% (69.93%)

やはり、もう少しなのは否めないorz まあ、残り4回なんだが、ダメだったら...考えます。

外した

今村聖奈騎手、おめでとうございます。やるならこの子だとは思ってたけど^^;

で、Model Builderですが、今回ももう一つ伸びず、\(R^{2}\)値が0.2726でした。

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	23.94% (74.18%)	23.74% (70.25%)	24.18% (78.91%)	23.33% (64.50%)	36.53% (81.37%)	26.08% (72.09%)	22.08% (74.22%)	53.95% (76.61%)
複勝	55.31% (82.95%)	56.19% (82.75%)	54.37% (83.43%)	55.83% (79.17%)	68.49% (86.16%)	61.31% (83.97%)	52.12% (82.32%)	88.05% (83.33%)
枠連	13.72% (77.15%)	14.12% (89.32%)	13.15% (66.33%)	17.17% (75.96%)	-- (--)	14.83% (78.78%)	12.48% (71.71%)	27.26% (72.36%)
馬連	10.65% (65.83%)	11.07% (73.17%)	10.02% (58.22%)	13.33% (67.00%)	26.48% (99.13%)	11.50% (63.19%)	8.94% (63.71%)	22.32% (65.54%)
ワイド	24.75% (84.09%)	26.35% (86.63%)	23.33% (83.15%)	21.67% (61.33%)	48.86% (86.26%)	29.30% (87.49%)	21.05% (82.76%)	46.11% (82.10%)
馬単	5.50% (56.64%)	5.27% (56.23%)	5.71% (58.04%)	5.83% (43.08%)	13.70% (69.32%)	5.32% (47.26%)	4.82% (58.62%)	22.32% (63.84%)
三連複	6.11% (68.91%)	7.46% (75.63%)	4.56% (59.76%)	8.33% (100.00%)	19.63% (94.84%)	8.28% (81.50%)	4.16% (62.38%)	15.31% (61.81%)
三連単	1.33% (40.02%)	1.36% (40.92%)	1.28% (39.63%)	1.67% (32.67%)	5.02% (75.16%)	1.61% (49.39%)	0.91% (33.72%)	15.31% (57.78%)
総合	56.50% (68.61%)	57.61% (71.49%)	55.16% (65.93%)	59.17% (65.23%)	72.15% (84.61%)	62.18% (70.46%)	53.19% (66.18%)	88.31% (64.34%)

これじゃね^^; 1回目より2回目、2回目より3回目と下がる一方なんですが、これも運任せなので何度か試して行くしかないんですが...辛いです。

第87回 優駿牝馬(オークス)

道半ばでまだまだやる事は沢山残ってますが、地道に進めます。で、相変わらず自力で予想しました。

穴狙いは基本中の基本なので曲げずに

まあ、2冠してもおかしくはないけど、スターアニスは強いのかな? そろそろ歴史的に女性騎手がってのも頭をかすめたけど、まだ今年ではないと判断して印下げた。相対タイムの予測では武だったが、ここはあえて相性良くないので避けた。

色々と手を出したので時間的に間に合ってないのもあるが、ダービー後には新馬戦もまた始まるのでやれる事はやります。    

AutoMLは限界でModel Builderに

AutoMLの6時間×10回は結局進展しないままだったので、Model Builderで24時間やってみました。\(R^{2}\)値は0.2762と前回を上回るもので

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	24.28% (76.72%)	24.27% (76.70%)	24.54% (78.65%)	20.83% (50.42%)	36.07% (88.36%)	26.33% (73.81%)	22.54% (76.64%)	54.62% (77.80%)
複勝	55.57% (82.65%)	56.31% (82.69%)	55.04% (82.94%)	52.50% (78.08%)	65.75% (81.83%)	60.44% (81.64%)	53.03% (83.06%)	87.84% (81.92%)
枠連	13.95% (76.34%)	13.48% (76.93%)	14.10% (76.23%)	18.18% (69.80%)	-- (--)	15.45% (62.31%)	12.57% (76.17%)	28.91% (76.82%)
馬連	10.94% (72.38%)	11.25% (67.60%)	10.51% (78.44%)	12.50% (56.67%)	26.03% (86.58%)	12.48% (58.13%)	9.06% (75.85%)	23.24% (69.53%)
ワイド	24.23% (76.05%)	25.75% (77.04%)	22.30% (73.92%)	29.17% (91.25%)	47.95% (80.50%)	29.54% (79.36%)	20.31% (74.54%)	45.85% (77.20%)
馬単	5.27% (68.31%)	5.03% (48.92%)	5.65% (91.48%)	3.33% (23.42%)	13.70% (66.44%)	5.69% (47.50%)	4.37% (75.41%)	23.24% (67.10%)
三連複	5.93% (69.28%)	6.69% (59.48%)	5.29% (82.85%)	4.17% (20.92%)	15.53% (63.52%)	8.03% (67.31%)	4.37% (70.45%)	15.72% (65.01%)
三連単	1.19% (39.48%)	1.01% (20.14%)	1.46% (62.22%)	0.00% (0.00%)	4.57% (66.62%)	1.73% (37.08%)	0.70% (37.84%)	15.72% (58.95%)
総合	56.87% (70.07%)	57.43% (63.41%)	56.38% (78.35%)	55.83% (48.35%)	70.78% (76.26%)	61.80% (63.39%)	53.98% (71.25%)	88.02% (65.75%)

まあ、何故か前回の0.2746よりも悪い結果ですorz 仕方なく新たに24時間スタートしました。