2024/10/21

LightGBMでの学習

昨日の菊花賞は▲→△→◎決着で◎の複勝のみと残念な結果でした。三連複の4頭ボックス程度は買っておけばねぇってのは後の祭り😣

まっ、そんなのを見ながらも色々やってます。先日書いた3つの値を指定する学習ですが、AutoMLやModel Builderも学習する度に得られる学習モデルは違います。これ、機械学習の基本的な部分なのかもだけど、AutoMLは自分でコーディングして実現してるので

MLContext mlContext = new MLContext(seed: 0);

このMLContextの作成時に引数のseedを渡さずにおくとint型の値がランダムに使われるって認識なんですが、デバックの際に毎回同じ動作をさせる時に、このseedに数字を指定するって事だった筈。なのでザックリとプラスマイナス21億なのがint型なのでそんな中でトライアル開始させてモデルの学習がされるので毎回違う結果になっているんだったかと。LightGBMでは

int boosterResult = LGBM_BoosterCreate(dataset, boosterParameters, out booster);

としてるんだけど、こちらはこのboosterParametersにseedの指定がある。こちらもdefault=Noneで指定してませんので多分ランダムなんだと。

って事で、AutoML同様に1サイクルの学習を指定回数回す指定を追加。新馬戦CKは1サイクルも早いので既に50回程度試したかな。最適RMSEは1.6306とかが得られてます。でも、相変わらずRMSEは目安的なのかどうなのかって感じで各モデルを実際に検証(2024.7.1~2024.10.14)で比べてどれを採用するか吟味。現在の候補はRMSE 1.6627の

124R 1点 芝(95R) ダート(29R) 8頭以下(19R) 9~12頭(42R) 13頭以上(63R) 多点
単勝 32.26%
(97.82%)
31.58%
(105.26%)
34.48%
(73.45%)
52.63%
(233.68%)
28.57%
(68.57%)
28.57%
(76.35%)
52.42%
(73.52%)
複勝 55.65%
(79.44%)
55.79%
(81.68%)
55.17%
(72.07%)
63.16%
(98.42%)
52.38%
(70.95%)
55.56%
(79.37%)
83.87%
(73.31%)
枠連 10.75%
(182.47%)
10.61%
(231.21%)
11.11%
(63.33%)
--
(--)
14.29%
(326.43%)
6.35%
(51.75%)
10.75%
(98.69%)
馬連 6.45%
(161.61%)
7.37%
(180.63%)
3.45%
(99.31%)
10.53%
(70.00%)
9.52%
(351.90%)
3.17%
(62.38%)
21.77%
(103.25%)
ワイド 16.94%
(83.71%)
17.89%
(91.47%)
13.79%
(58.28%)
42.11%
(91.58%)
14.29%
(108.81%)
11.11%
(64.60%)
37.90%
(64.38%)
馬単 4.03%
(226.45%)
4.21%
(245.37%)
3.45%
(164.48%)
10.53%
(203.68%)
4.76%
(462.86%)
1.59%
(75.71%)
21.77%
(86.71%)
三連複 3.23%
(45.08%)
4.21%
(58.84%)
0.00%
(0.00%)
15.79%
(117.37%)
2.38%
(80.00%)
0.00%
(0.00%)
8.87%
(38.21%)
三連単 0.81%
(75.40%)
1.05%
(98.42%)
0.00%
(0.00%)
5.26%
(492.11%)
0.00%
(0.00%)
0.00%
(0.00%)
8.87%
(29.60%)
総合 58.06%
(116.95%)
57.89%
(132.86%)
58.62%
(66.39%)
68.42%
(186.69%)
54.76%
(183.69%)
57.14%
(51.27%)
83.87%
(52.67%)

これだけ見るとまずまずなモデルではあります。でも、自分本来のスタンスのワイドからって事だと微妙ではあります。通常用CKを今回してる最中なんですが、既に数十回はしてますが、こちらは1サイクルが長い時は1時間半とか掛かるので昨日10回以上回して今朝30回スタートさせたんですがやっと10回程度終わってます。ここまでの最適RMSEは1.3677ですが、

835R 1点 芝(411R) ダート(389R) 障害(35R) 8頭以下(62R) 9~12頭(255R) 13頭以上(518R) 多点
単勝 23.35%
(74.13%)
22.14%
(76.23%)
23.39%
(70.64%)
37.14%
(88.29%)
45.16%
(95.32%)
27.45%
(68.78%)
18.73%
(74.23%)
50.06%
(71.52%)
複勝 55.09%
(86.72%)
51.34%
(83.84%)
58.10%
(89.56%)
65.71%
(88.86%)
66.13%
(81.94%)
66.67%
(94.59%)
48.07%
(83.42%)
88.14%
(84.40%)
枠連 13.23%
(100.36%)
11.61%
(107.53%)
14.60%
(95.04%)
15.79%
(75.26%)
--
(--)
14.90%
(73.29%)
11.00%
(103.03%)
26.32%
(82.80%)
馬連 9.46%
(64.73%)
7.79%
(42.04%)
10.80%
(89.72%)
14.29%
(53.43%)
20.97%
(75.97%)
11.76%
(66.90%)
6.95%
(62.32%)
19.76%
(59.32%)
ワイド 22.04%
(80.43%)
19.22%
(62.00%)
24.68%
(101.72%)
25.71%
(60.29%)
45.16%
(79.19%)
27.84%
(87.96%)
16.41%
(76.87%)
43.23%
(82.10%)
馬単 4.79%
(59.84%)
4.14%
(42.58%)
5.14%
(79.02%)
8.57%
(49.43%)
14.52%
(99.03%)
5.88%
(60.71%)
3.09%
(54.73%)
19.76%
(56.32%)
三連複 5.15%
(67.86%)
5.11%
(54.62%)
4.88%
(84.47%)
8.57%
(38.57%)
16.13%
(53.06%)
9.02%
(95.10%)
1.93%
(56.22%)
14.13%
(72.69%)
三連単 1.32%
(67.94%)
1.22%
(14.23%)
1.29%
(127.30%)
2.86%
(38.86%)
6.45%
(58.06%)
1.57%
(66.67%)
0.58%
(69.75%)
14.13%
(68.89%)
総合 57.25%
(74.80%)
54.01%
(59.28%)
59.64%
(92.16%)
68.57%
(60.80%)
74.19%
(77.51%)
66.67%
(76.75%)
50.58%
(72.57%)
88.62%
(69.68%)

それ程インパクトのあるモデルではないですが、RMSE 1.3878の

835R 1点 芝(411R) ダート(389R) 障害(35R) 8頭以下(62R) 9~12頭(255R) 13頭以上(518R) 多点
単勝 25.51%
(81.44%)
24.57%
(86.01%)
24.68%
(72.11%)
45.71%
(131.43%)
41.94%
(96.29%)
30.20%
(83.33%)
21.24%
(78.73%)
51.98%
(73.22%)
複勝 57.25%
(91.65%)
54.01%
(89.90%)
59.13%
(92.47%)
74.29%
(103.14%)
69.35%
(103.39%)
67.45%
(100.08%)
50.77%
(86.10%)
85.27%
(79.74%)
枠連 12.53%
(85.57%)
10.71%
(86.58%)
13.22%
(79.61%)
31.58%
(181.58%)
--
(--)
15.69%
(82.27%)
9.65%
(78.11%)
26.74%
(84.45%)
馬連 8.98%
(74.10%)
7.06%
(75.50%)
9.25%
(65.66%)
28.57%
(151.43%)
20.97%
(76.61%)
12.16%
(71.10%)
5.98%
(75.27%)
21.68%
(83.98%)
ワイド 22.28%
(81.78%)
19.71%
(71.31%)
22.88%
(89.51%)
45.71%
(118.86%)
48.39%
(90.00%)
26.67%
(80.59%)
16.99%
(81.39%)
42.63%
(77.11%)
馬単 5.03%
(68.86%)
3.65%
(48.32%)
4.88%
(73.88%)
22.86%
(254.29%)
12.90%
(68.55%)
6.67%
(68.78%)
3.28%
(68.94%)
21.68%
(79.63%)
三連複 5.15%
(79.35%)
6.08%
(118.22%)
3.60%
(41.57%)
11.43%
(42.86%)
17.74%
(70.81%)
7.45%
(49.61%)
2.51%
(95.02%)
13.65%
(81.54%)
三連単 0.96%
(50.54%)
0.73%
(73.70%)
0.51%
(17.58%)
8.57%
(144.86%)
1.61%
(21.94%)
2.35%
(54.71%)
0.19%
(51.91%)
13.65%
(89.52%)
総合 58.44%
(76.50%)
55.72%
(81.07%)
59.90%
(66.44%)
74.29%
(138.60%)
74.19%
(75.37%)
67.45%
(73.81%)
52.12%
(76.93%)
85.63%
(84.65%)

こちらと選択に悩む感じですが、まだ残り20回あるので全く別のモデルが出てくるかも?

追記 2024.10.22.7:35
RMSE 1.3924とそれ程でもない感じだけど

835R 1点 芝(411R) ダート(389R) 障害(35R) 8頭以下(62R) 9~12頭(255R) 13頭以上(518R) 多点
単勝 23.83%
(78.68%)
22.87%
(78.32%)
23.39%
(77.20%)
40.00%
(99.43%)
35.48%
(62.74%)
30.98%
(86.24%)
18.92%
(76.87%)
52.46%
(80.05%)
複勝 57.84%
(91.37%)
56.20%
(92.19%)
58.61%
(90.67%)
68.57%
(89.43%)
70.97%
(87.90%)
66.27%
(96.75%)
52.12%
(89.13%)
88.98%
(82.37%)
枠連 15.32%
(112.23%)
16.37%
(125.60%)
13.77%
(93.80%)
26.32%
(227.89%)
--
(--)
18.82%
(105.65%)
11.97%
(103.55%)
25.91%
(81.31%)
馬連 11.38%
(111.78%)
12.17%
(140.61%)
10.03%
(83.01%)
17.14%
(93.14%)
17.74%
(49.35%)
16.47%
(119.57%)
8.11%
(115.42%)
21.44%
(78.82%)
ワイド 22.75%
(89.86%)
22.87%
(92.85%)
21.85%
(86.76%)
31.43%
(89.14%)
41.94%
(70.00%)
28.63%
(88.63%)
17.57%
(92.84%)
42.28%
(78.81%)
馬単 5.15%
(95.28%)
4.14%
(112.65%)
5.40%
(74.42%)
14.29%
(123.14%)
8.06%
(29.03%)
7.84%
(78.67%)
3.47%
(111.39%)
21.44%
(74.74%)
三連複 5.63%
(77.80%)
6.33%
(101.07%)
4.11%
(52.24%)
14.29%
(88.57%)
17.74%
(56.77%)
8.63%
(81.18%)
2.70%
(78.65%)
14.97%
(94.90%)
三連単 1.32%
(88.29%)
0.97%
(90.44%)
1.03%
(74.60%)
8.57%
(215.14%)
3.23%
(30.00%)
1.96%
(54.27%)
0.77%
(112.01%)
14.97%
(91.18%)
総合 59.16%
(92.82%)
57.91%
(103.72%)
59.38%
(78.96%)
71.43%
(122.20%)
74.19%
(55.12%)
67.06%
(88.87%)
53.47%
(97.48%)
89.10%
(86.19%)

なかなかなパフォーマンスかな?

0 件のコメント:

コメントを投稿