故事 · DOE 教学最经典的弹射器
Story · The Catapult That Taught a Generation DOE
几乎每一门六西格玛黑带课,都会摆上一台弹射器(catapult):一个能调臂长、调橡皮筋挂位、调释放角度的小弹弓,
要求学员把小球稳稳射中某个目标距离。新手的本能是「一次只调一个参数」(OFAT)——
固定其它,单独把角度从小扫到大,记下最好的;再单独扫挂位…… 试验做了一大堆,却始终差一口气。
为什么?因为 OFAT 看不见参数之间的交互:臂长一变,最优角度也跟着变,单独扫出来的「最优角度」一换臂长就失效了。
直到老师亮出实验设计(DOE):用一张精心安排的表,同时变动几个参数,少数几次试验就同时估出每个参数的主效应和它们的交互 ——
又快又准。这台小弹射器,讲透了 DOE 相对 OFAT 的全部优势。
Almost every Six Sigma Black Belt classroom rolls out a tabletop catapult: a small wooden trebuchet with adjustable arm length, rubber-band hook position and release angle.
The assignment: land the ball reliably at a chosen distance. The beginner's instinct is one-factor-at-a-time (OFAT) —
hold everything else fixed, sweep angle from low to high, note the best, then sweep the hook position… Lots of runs, never quite there.
Why? OFAT is blind to interactions: change the arm and the best angle moves with it. Whatever "optimal angle" you found at one arm length stops being optimal at the next.
Then the instructor reveals a designed experiment (DOE): a single planned matrix that moves several factors together, estimating main effects and interactions in just a handful of runs —
faster and sharper. That little catapult crystallises every advantage DOE has over OFAT.
1 弹射器靶场 · 调参数 → 发射 → 看落点 The Range · tune → fire → watch where it lands
瞄准 30maim for 30 m绿色虚线是目标 30 米。手动拖参数发射,或点上面两个按钮:OFAT 要一个一个参数地扫,DOE 用一张表几次就摸到最优组合。 The dashed green line is the 30 m target. Drag the knobs and fire by hand, or use the two buttons above: OFAT sweeps one factor at a time, DOE reaches the optimum in a handful of runs.
2 OFAT vs DOE · 谁更省试验、谁更准 OFAT vs DOE · fewer runs, sharper answers
OFAT(一次一因子):固定其它、单独扫一个参数。直觉、好记,但试验次数随参数线性堆,且测不出交互。
OFAT (one factor at a time): hold the rest fixed, sweep one knob. Intuitive and easy to remember, but runs pile up linearly with the number of factors — and interactions are never measured.
DOE(实验设计):按正交表同时变多个参数。少数几次试验就估出主效应+交互,更少试验、更早收敛。
DOE (designed experiment): an orthogonal matrix changes several factors together. A handful of runs estimate main effects and interactions — fewer runs, faster convergence.
现实里,DOE 用在哪Where DOE Actually Earns Its Keep
工艺参数优化:注塑温度×压力×保压时间、焊接电流×速度……DOE 用最少试验找到良率最高的参数窗口,是制造业改进的主力工具。
Process parameter tuning: injection-mould temperature × pressure × hold time, welding current × speed… DOE locates the highest-yield window in the fewest runs and is the workhorse of manufacturing improvement.
交互作用:催化剂×温度、施肥量×灌溉量常有交互 —— 单独看每个都没用,合起来才出效果。只有 DOE 能直接量出这种「1+1≠2」。
Interaction effects: catalyst × temperature, fertiliser × irrigation often interact — each alone does nothing, the combination is where the gain lives. Only DOE measures that "1 + 1 ≠ 2" directly.
OFAT 的局限:参数一多,OFAT 试验次数爆炸还测不出交互,容易停在「局部最优」。这正是 Fisher 当年发明 DOE 的动机。
The limit of OFAT: with several factors, OFAT runs explode while interactions stay invisible — you settle on a local optimum. That was exactly the pain Fisher set out to solve when he invented DOE.
配方与实验:药物配比、农业育种、A/B 测试矩阵、机器学习超参搜索,本质都是 DOE —— 在有限预算里高效探索多维参数空间。
Formulations & experiments: drug dosing, plant breeding, A/B test matrices, ML hyper-parameter search — at heart they are all DOE: exploring a multi-dimensional parameter space efficiently on a fixed budget.
一句话In One Line
OFAT 的致命假设是「参数之间互不影响」—— 现实里几乎从不成立。臂长一变,最优角度就跟着变,
单独扫出来的最优一换条件就失效。DOE 的全部价值,就是用一张正交表把交互一并测出来,
同时还省试验。它是六西格玛 Improve 阶段的核心工具:当你要在多个可调参数里找最优组合、且怀疑它们彼此纠缠时,
别再一个一个扫,设计一次实验,让数据一次告诉你答案。
OFAT's fatal assumption is that factors are independent — in the real world they rarely are. Change the arm and the optimum angle moves with it;
whatever you swept out alone collapses the moment conditions shift. The entire value of DOE is that one orthogonal matrix captures the interactions
and saves runs in the process. It is the workhorse of the Six Sigma Improve phase: when you need the best combination across several knobs and suspect they are entangled,
stop sweeping one at a time — design one experiment, let the data answer in one pass.
常见误用Common Mistakes
多参数还一个一个 OFAT 地试 → 慢且漏掉交互。参数 ≥3 个、怀疑有交互,就上 DOE。
Sticking with OFAT when you have multiple factors → slow and blind to interactions. Three or more factors with suspected interaction? Reach for DOE.
因子取值范围太窄 → 主效应淹没在噪声里看不出。水平拉开到工艺允许的边界,让效应显著。
Factor levels set too tight → main effects vanish into noise. Spread the levels as wide as the process allows so the effects ring out clearly.
只做一次没有重复/中心点 → 分不清是效应还是随机波动,也测不出曲率。加重复与中心点估误差。
Running each combination once with no replication or centre points → you cannot tell signal from noise and you miss any curvature. Add replicates and centre points to estimate error and curvature.