故事 · Box 与 Wilson 的「爬山」
Origin Story · The "Hill Climb" of Box and Wilson
1951 年,统计学家 George Box 与 K. B. Wilson 在英国一家化工厂面对一个老问题:
怎样找到让产率最高的温度和压力组合?他们的洞见是分两步走 —— 先用全因子的一次模型拟合一小片平面,
看坡度朝哪走,沿最陡上升路径大步逼近高产区;一旦逼近顶部、平面开始「弯曲」,就切换到二阶模型,
在中心点周围加密布点(这就是后来的中心复合设计 CCD),拟合出一座局部的小山,
再用微积分直接解出山顶坐标。Box 把整套方法叫响应曲面法(RSM)。
它的精髓就一句话:一次项告诉你往哪走,二次项告诉你停在哪 —— 没有二次项,你永远不知道自己是不是已经到顶了。
In 1951, statisticians George Box and K. B. Wilson faced an old problem at a British chemical plant:
what temperature-and-pressure combination maximizes yield? Their insight was a two-step march — first fit a small plane with a full-factorial first-order model,
read which way the slope rises, and stride toward the high-yield region along the steepest-ascent path. Once the response starts to "curve" near the top, switch to a second-order model,
sprinkle extra points around the center (this became the central composite design, CCD), fit a local hill,
and use calculus to solve for the summit coordinates analytically. Box called the whole framework response surface methodology (RSM).
Its essence fits on one line: first-order terms tell you which way to walk; second-order terms tell you where to stop. Without the quadratic, you can never tell whether you've actually reached the top.
1 响应曲面:把二次项拨进来,平面就隆成了山 The Response Surface: Dial In the Quadratic and the Plane Rises Into a Hill
极大(峰)Maximum (peak)2 二阶模型逐项拆解 The Second-Order Model, Term by Term
β₀ · 常数项:曲面整体的高度基线,平移全图上下。
β₀ · the intercept: the baseline height of the whole surface — shifts everything up or down.
Σ βᵢxᵢ · 一次项:平面的坡度,告诉你哪个方向上升 —— 只有这一项就是全因子的平面。
Σ βᵢxᵢ · first-order terms: the slope of the plane, telling you which direction climbs — with only these terms you get the flat full-factorial plane.
Σ βᵢᵢxᵢ² · 二次项:让曲面弯曲。同号→峰或谷,异号→鞍点。没有它就没有最优点。
Σ βᵢᵢxᵢ² · quadratic terms: these curve the surface. Same sign → a peak or valley; opposite signs → a saddle. Without them there is no optimum.
Σ βᵢⱼxᵢxⱼ · 交互项:让山脊斜着长,主轴不再对齐坐标轴。
Σ βᵢⱼxᵢxⱼ · interaction terms: rotate the ridge off-axis so the principal axes no longer line up with the coordinate axes.
3 现实里的响应曲面法 RSM in the Real World
工艺优化:注塑温度+保压、焊接电流+速度,RSM 在最优区拟合曲面,精修到产率/强度的峰顶。
Process optimization: injection-mold temperature plus hold-pressure, weld current plus speed — RSM fits a surface in the high-yield region and tunes the process right up to the peak of yield or strength.
中心复合设计 CCD:在 2^k 角点外加轴点+中心点,正好够估全部二次项,是 RSM 的标准布点。
Central composite design (CCD): layer axial points and a center point on top of the 2^k corners — just enough to estimate every quadratic term. It's the standard RSM layout.
寻找最优点:对二阶模型求导令 ∇ŷ=0,直接解析算出山顶坐标,不必盲目试参数。
Locating the optimum: differentiate the second-order model, set ∇ŷ = 0, and solve the summit coordinates analytically — no need to keep probing blindly.
鞍点警觉:典型分析(特征值)能识别驻点是峰、谷还是鞍 —— 鞍点意味着没有唯一最优,需换方向。
Saddle alert: canonical analysis (eigenvalues of the Hessian) tells whether a stationary point is a peak, valley, or saddle — a saddle means there is no single optimum, and you need to redirect.
一句话In One Line
为什么非要二次项?因为平面没有极值 —— 一张斜平面,最高点永远在你布点范围的边界角落,
它只告诉你「再往这边走更好」,却从不说「停,这里就是顶」。只有当响应面弯曲,
中间才可能凸起一座峰、凹下一个谷,梯度归零的那个点,才是真正的最优。
全因子(一次模型)负责在远处「找方向、爬大坡」,RSM(二阶模型)负责在近处「拟合曲面、精修登顶」 ——
这正是 Box-Wilson 两步法的分工。一次项管方向,二次项管收敛,缺了后者,优化永远到不了家。
Why insist on quadratic terms? Because a plane has no extremum — on a tilted plane, the high point is always in some corner of the boundary.
It only ever says "this way is better," never "stop, this is the top." Only when the response surface curves
can a peak rise or a valley dip in the interior, and only then does a zero gradient mark the true optimum.
Full-factorial (the first-order model) handles the "find a direction, climb the long slope" part from afar; RSM (the second-order model) handles the "fit a curve, fine-tune to the top" part up close —
that's the division of labor in the Box–Wilson two-step. First-order terms point the direction, second-order terms drive convergence — without the latter, optimization never quite arrives home.
常见误用Common Mistakes
离最优区还远就上 RSM。先用最陡上升爬到高产区附近,再上二阶模型,否则曲面只是局部噪声。
Jumping to RSM while still far from the optimum region. Climb to the neighborhood of the high-yield region with steepest ascent first, then fit the second-order model — otherwise the surface is just local noise.
用 2^k 角点直接拟二次项。角点估不出 xᵢ²,必须加轴点+中心点(CCD/Box-Behnken)才能拟合曲面。
Trying to estimate quadratic terms from 2^k corner points alone. Corners can't identify xᵢ². You need axial and center points (CCD or Box–Behnken) to fit a curved surface.
解出驻点就当最优。先做典型分析判断是峰、谷还是鞍 —— 鞍点不是最优,且要确认它落在设计区内。
Treating any stationary point as the optimum. Run canonical analysis first to classify it as a peak, valley, or saddle — saddles are not optima, and confirm the point falls inside the design region.