故事 · 二战、稀疏性,与「先抓大鱼」
Origin Story · WWII, Effect Sparsity, and "Catch the Big Fish First"
二战期间,盟军要在极有限的弹药、原料和时间里同时优化大量参数,全因子根本做不起 ——
Finney、Box、Hunter 等人把 Fisher 的因子设计推进了一大步:
既然资源不够把所有组合都试,那就只试一个聪明挑选的子集。
支撑这个冒险的是两条经验法则:效应稀疏原理(真正重要的因子总是少数)和
效应遗传原理(高阶交互通常可忽略)。于是把宝贵的试验「省」在那些大概率不重要的高阶交互上,
让它们与主效应混杂——只要那些交互本就接近零,混杂就不损害结论。
这就是部分析因的智慧:先用最少的实验抓住大鱼(关键因子),再回头精修。
During WWII, the Allies needed to optimize many parameters at once under brutal limits on ammunition, raw materials, and time — full factorials were simply unaffordable.
Finney, Box, Hunter, and others pushed Fisher's factorial design one big step further:
if you cannot test every combination, test only a cleverly chosen subset.
Two empirical pillars made the bet pay off — the effect sparsity principle (the truly important factors are always few) and
the effect heredity principle (high-order interactions are usually negligible). So you spend your precious runs by aliasing main effects
with the high-order interactions that were almost certainly zero to begin with — and the conclusion holds.
That is the wisdom of fractional factorial design: land the big fish (the vital few factors) with the fewest possible runs, then come back and polish.
1 试验数对比:全因子 vs 部分析因 Runs Compared: Full vs Fractional Factorial
Res Ⅴ2 别名结构:谁和谁被绑在一起分不开 Alias Structure: Which Effects Get Bound Together
连成一组的效应是别名(混杂):试验只能算出它们的和,无法各自分开。 分辨度 = 「被定义关系混在一起的最短效应字长」 —— Res 越高,混杂的效应阶数越高、越能忽略。 Effects connected on the same row are aliases (confounded): the experiment can only estimate their sum, never the pieces. Resolution = "the shortest word length in the defining relation" — the higher the resolution, the higher the order of confounded effects, and the safer it is to ignore them.
3 现实里的部分析因 Fractional Factorials in the Real World
因子筛选:开发初期七八个候选因子,先用 Res III 筛选设计几次实验,揪出关键少数。
Factor screening: seven or eight candidate factors at the start of development — run a few Resolution III screens to flush out the vital few.
分辨度选择:Res III 最省(主效应混二阶交互),Res IV(主效应净但二阶互混),Res V(二阶全净)。
Picking resolution: Res III is cheapest (main effects aliased with 2-factor interactions), Res IV keeps main effects clean (2-factor interactions still alias each other), Res V keeps 2-factor interactions clean too.
混杂/别名:靠定义关系 I=ABC… 推出哪些效应同名 —— 设计前就知道自己将分不开哪些项。
Confounding / aliasing: the defining relation I = ABC… tells you which effects share a column — you know upfront which terms the design cannot separate.
折叠/续做:筛选后若发现混杂干扰,可用折叠设计或加做半组,把关键交互解混。
Fold-over / augmentation: if aliasing muddies the screen, fold the design over or add the complementary half-fraction to break the key interactions free.
一句话In One Line
部分析因的核心交易是一句话:用混杂换效率。
全因子能把每个主效应、每个交互都干干净净地分开,但代价是试验数随因子数指数爆炸。
部分析因故意「省掉」一半甚至更多组合,让那些大概率为零的高阶交互去和主效应共用一个名额 ——
这就是别名。分辨度就是给这笔交易标价:阶数越高的效应才被混在一起,混杂越「安全」。
所以聪明的做法不是一上来就追求干净,而是两步走:
先用低分辨度、少试验从一大堆因子里筛出关键少数,再把资源集中到这几个因子上做高分辨度或全因子的精修。
这正是六西格玛 Improve 阶段「先广撒网、再精打捞」的实验哲学。
The core trade of fractional factorial design is one line: aliasing in exchange for efficiency.
A full factorial separates every main effect and every interaction cleanly, but the run count blows up exponentially with the number of factors.
A fractional design deliberately "skips" half — or more — of the combinations, parking the almost-certainly-zero high-order interactions in the same column as the main effects.
That's aliasing. Resolution is the price tag on that trade — the higher it is, the higher the order of effects that get mixed together, and the safer the confounding becomes.
So the smart move is not to chase clean estimation upfront, but to walk it in two steps:
screen a forest of factors with a low-resolution, few-run design, then concentrate the resources on the vital few with a high-resolution or full factorial design.
That is the experimentation philosophy of the Six Sigma Improve phase — cast wide, then haul in precisely.
常见误用Common Mistakes
无视别名直接下结论。报告显著效应前先查它和谁混杂 —— 可能真正起作用的是被它别名的那个交互。
Reporting significant effects without checking aliases. Before declaring an effect significant, look up its alias chain — the real driver could be the interaction it shares a column with.
怀疑有交互还用 Res III。Res III 把主效应和二因子交互混在一起,要研究交互至少用 Res IV/V。
Using Resolution III when you already suspect interactions. Res III aliases main effects with 2-factor interactions — if you want to study interactions, step up to Res IV or V.
把筛选设计当最终优化。筛选只为找关键因子,最优工艺要回头用全因子/RSM 在关键因子上精修验证。
Treating a screening design as the final optimization. Screening only flags the vital few — come back with a full factorial or RSM on those survivors to lock in the optimum process.