NumberOfLeaves: 511
NumberOfIterations: 4000
LearningRate: 0.005
MaximumBinCountPerFeature: 255
MinimumExampleCountPerLeaf: 150
SubsampleFraction: 0.8
FeatureFraction: 0.75
CustomGains: 0～500 = 0点, 501～1000は0.95乗に0～1000点
これ20回スタートして寝た。最適は
[Loop 9] NDCG@1 = 0.7667
[Loop 9] NDCG@3 = 0.7735
[Loop 9] NDCG@5 = 0.7895
[Loop 9] NDCG@10 = 0.8498
| 2,610R | 1点 | 芝(1,278R) | ダート(1,243R) | 障害(89R) | 8頭以下(157R) | 9～12頭(665R) | 13頭以上(1,788R) | 多点 |
| 単勝 | 24.56% (73.47%) |
24.49% (76.02%) |
24.94% (72.60%) |
20.22% (49.10%) |
32.48% (65.54%) |
26.92% (68.96%) |
22.99% (75.84%) |
55.67% (76.39%) |
| 複勝 | 55.13% (78.55%) |
53.05% (77.52%) |
56.80% (79.32%) |
61.80% (82.47%) |
61.15% (77.13%) |
60.15% (77.85%) |
52.74% (78.93%) |
89.73% (81.68%) |
| 枠連 | 13.76% (73.34%) |
12.68% (83.58%) |
14.39% (61.96%) |
18.99% (102.41%) |
-- (--) |
13.23% (56.08%) |
13.14% (75.45%) |
30.20% (75.27%) |
| 馬連 | 10.15% (63.67%) |
9.31% (70.53%) |
10.78% (55.27%) |
13.48% (82.36%) |
21.66% (81.27%) |
10.68% (54.83%) |
8.95% (65.41%) |
23.87% (76.31%) |
| ワイド | 24.44% (73.67%) |
23.08% (72.22%) |
25.91% (75.54%) |
23.60% (68.31%) |
42.04% (75.22%) |
27.07% (72.27%) |
21.92% (74.05%) |
47.20% (78.67%) |
| 馬単 | 5.59% (53.77%) |
4.85% (51.88%) |
6.44% (56.43%) |
4.49% (43.93%) |
12.10% (68.47%) |
6.62% (59.23%) |
4.64% (50.45%) |
23.87% (73.75%) |
| 三連複 | 6.97% (108.15%) |
6.73% (136.04%) |
7.40% (85.23%) |
4.49% (27.64%) |
14.01% (105.48%) |
7.22% (61.22%) |
6.26% (125.83%) |
17.43% (81.10%) |
| 三連単 | 1.69% (103.69%) |
1.41% (72.53%) |
2.01% (140.97%) |
1.12% (30.67%) |
3.18% (176.50%) |
1.80% (54.14%) |
1.51% (115.73%) |
17.43% (79.55%) |
| 総合 | 57.24% (78.60%) |
55.71% (79.97%) |
58.49% (78.52%) |
61.80% (60.27%) |
64.97% (92.80%) |
60.90% (63.07%) |
55.20% (82.71%) |
89.96% (78.42%) |
後はloop2とloop14辺りが続いた感じかな。流石にこの設定ではモデルも44.5MBと大きくなった。モデルが大きくなった割にはそこまで改善されてないのは残念。まあ、これはこれでサンプルとしてありだが、採用はしないかな。
今日仕事から戻ったらL2追加してるのでそれを試します。
0 件のコメント:
コメントを投稿