实验 84:A78 LOO Observation Budget Router

V1:A77 规则族在 leave-one-case-out 下 2/3 匹配 oracle,说明状态变量有信号但训练覆盖不足

实验定位 V1 完成 / 泛化仍不足

A78 是用户要求的“三个版本”里的第 1 版:把 A77 v2 的回顾诊断规则压进一个小阈值网格,做 leave-one-case-out。训练时只在两个 train case 上用 GT joint 拟合阈值;held-out case 选择时只用 observation features,不看 GT。

结果:A78 LOO 选中 2/3 oracle,mean joint=-0.0010410,比 A66=-0.0009299 稍好,也远强于 A76 post-hoc best=-0.0001506;但仍弱于 A77 retrospective/oracle=-0.0014752。失败点是 gso008 held-out:train split 里没有 visible-recovery 类型样本,规则退回 saturated damp iter1 λ=0.125,而 oracle 是 iter8。

Only 3 active cases have A65 sampler proxy trajectories, so this is a falsification/stability check rather than a publishable generalization result.

实验设计(Image-2 风格)

实验设计图:A78 LOO observation budget router
实验设计图:A78 LOO observation budget router
模块设计图:diagnostic rule to learnable router
模块设计图:diagnostic rule to learnable router

模块设计(Image-2 风格)

Rule Grid

  • late-safe:后期 MV improve 且 final local damage 为 false。
  • visible recovery:visible residual 强烈下降,用来保留 recall/coverage-friendly step。
  • MV low-spread:MV improve 且 region spread 低。
  • saturated fallback:高 damage ratio 时只做第一步阻尼。

LOO Protocol

  • 每折 2 个 case 训练阈值,1 个 case held-out。
  • 训练目标是 joint score,held-out 选择不使用 GT。
  • 和 A66、A77 retrospective、GT oracle、A76 post-hoc 统一对比。
  • 重点观察状态类型覆盖,而不是追求 3-case 高分。

实验结果(表格)

policy comparison

policy性质mean ΔChamfermean ΔF@5jointiterslambdas
GT joint oracleGT oracle-0.00055810.011463-0.0014752{'gso_000_input4': 3, 'gso_002_input4': 8, 'gso_008_input4': 8}{'gso_000_input4': 1.0, 'gso_002_input4': 1.0, 'gso_008_input4': 1.0}
A77 v2 retrospective回顾诊断-0.00055810.011463-0.0014752{'gso_000_input4': 3, 'gso_002_input4': 8, 'gso_008_input4': 8}{'gso_000_input4': 1.0, 'gso_002_input4': 1.0, 'gso_008_input4': 1.0}
A78 LOO observation routerLOO held-out-0.00061380.005340-0.0010410{'gso_000_input4': 3, 'gso_002_input4': 8, 'gso_008_input4': 1}{'gso_000_input4': 1.0, 'gso_002_input4': 1.0, 'gso_008_input4': 0.125}
A66 uncertainty damped预注册 no-GT-0.00055180.004726-0.0009299{'gso_000_input4': 8, 'gso_002_input4': 8, 'gso_008_input4': 1}{'gso_000_input4': 1.0, 'gso_002_input4': 1.0, 'gso_008_input4': 0.125}
carrier control预注册 no-GT0.00000000.0000000.0000000{'gso_000_input4': 0, 'gso_002_input4': 0, 'gso_008_input4': 0}{'gso_000_input4': 1.0, 'gso_002_input4': 1.0, 'gso_008_input4': 1.0}

LOO held-out details

held-outtrain casesselectedoracle itermatchΔChamferΔF@5jointreason
gso_000_input4['gso_002_input4', 'gso_008_input4']iter3 / λ=1.0003-0.00041590.000512-0.0004568LOO route: min region-spread fallback; picked iter003
gso_002_input4['gso_000_input4', 'gso_008_input4']iter8 / λ=1.0008-0.00160530.015600-0.0028533LOO route: min region-spread fallback; picked iter008
gso_008_input4['gso_000_input4', 'gso_002_input4']iter1 / λ=0.12580.0001798-0.0000940.0001873LOO route: saturated damage fallback; damage_ratio=1.000 >= 0.875; lambda=0.125

observation-state features

casedamage ratioMV improvelate-safe itersvisible recovery minbest MV low-spreadmin spread iter
gso_000_input40.7503[]-0.00014733
gso_002_input40.7508[8]0.00000088
gso_008_input41.0005[]-0.00729711

可视化结果

A78 joint comparison
A78 joint comparison
A78 oracle gap
A78 oracle gap
A78 oracle match count
A78 oracle match count

实验结论

下一步想法