试验设计与数据分析

最速下降（上升）法
[email protected] (mailto:[email protected])
2022.04.16

适用场景：
最速下降（上升）法（steepest descent(ascent) algorithm）
用于探索响应面设计的中心点。
避免响应面设计无最优解。

例题，一位化学工程师要确定使过程产率最高的操作条件．影响产率的两个可控变量是反应时间和反应温度工程
师当前使用的操作条件是反应时间为35分钟，温度为155下，产率约为40％，因为此区域不大可能包含最优于是
她拟合一阶模型并应用最速上升法。
l i b r a r y (rsm)
cube.design <- rsm::cube(2, n0=5,
coding =list(x1~(time-35)/5, x2~(temperature-155)/5),
randomize = F)
cube.design$res <- c(39.3,40.9,40,41.5,40.3,40.5,40.7,40.2,40.6)
cube.design

## run.order std.order time temperature res

## 1 1 1 30 150 39.3
## 2 2 2 40 150 40.9
## 3 3 3 30 160 40.0
## 4 4 4 40 160 41.5
## 5 5 5 35 155 40.3
## 6 6 6 35 155 40.5
## 7 7 7 35 155 40.7
## 8 8 8 35 155 40.2
## 9 9 9 35 155 40.6
##
## Data are stored in coded form using these coding formulas ...
## x1 ~ (time - 35)/5
## x2 ~ (temperature - 155)/5

FO.model <- rsm(res~FO(x1,x2),data = cube.design)

summary(FO.model)
##
## Call:
## rsm(formula = res ~ FO(x1, x2), data = cube.design)
##
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 40.444444 0.057288 705.9869 5.451e-16 ***
## x1 0.775000 0.085932 9.0188 0.000104 ***
## x2 0.325000 0.085932 3.7821 0.009158 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Multiple R-squared: 0.941, Adjusted R-squared: 0.9213
## F-statistic: 47.82 on 2 and 6 DF, p-value: 0.0002057
##
## Analysis of Variance Table
##
## Response: res
## Df Sum Sq Mean Sq F value Pr(>F)
## FO(x1, x2) 2 2.82500 1.41250 47.8213 0.0002057
## Residuals 6 0.17722 0.02954
## Lack of fit 2 0.00522 0.00261 0.0607 0.9419341
## Pure error 4 0.17200 0.04300
##
## Direction of steepest ascent (at radius 1):
## x1 x2
## 0.9221944 0.3867267
##
## Corresponding increment in original units:
## time temperature
## 4.610972 1.933633

contour(FO.model, ~ x1 + x2, image = TRUE)

persp(FO.model, x2 ~ x1, zlab = "res")
FOTWI.model <- rsm(res~FO(x1,x2)+TWI(x1,x2),data = cube.design)
summary(FOTWI.model)

##
## Call:
## rsm(formula = res ~ FO(x1, x2) + TWI(x1, x2), data = cube.design)
##
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 40.444444 0.062311 649.0693 1.648e-13 ***
## x1 0.775000 0.093467 8.2917 0.0004166 ***
## x2 0.325000 0.093467 3.4772 0.0177127 *
## x1:x2 -0.025000 0.093467 -0.2675 0.7997870
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Multiple R-squared: 0.9418, Adjusted R-squared: 0.9069
## F-statistic: 26.97 on 3 and 5 DF, p-value: 0.00163
##
## Analysis of Variance Table
##
## Response: res
## Df Sum Sq Mean Sq F value Pr(>F)
## FO(x1, x2) 2 2.82500 1.41250 40.4213 0.0008188
## TWI(x1, x2) 1 0.00250 0.00250 0.0715 0.7997870
## Residuals 5 0.17472 0.03494
## Lack of fit 1 0.00272 0.00272 0.0633 0.8137408
## Pure error 4 0.17200 0.04300
##
## Stationary point of response surface:
## x1 x2
## 13 31
##
## Stationary point in original units:
## time temperature
## 100 310
##
## Eigenanalysis:
## eigen() decomposition
## $values
## [1] 0.0125 -0.0125
##
## $vectors
## [,1] [,2]
## x1 -0.7071068 -0.7071068
## x2 0.7071068 -0.7071068

最速上升（下降）的方向
回归方程的斜率
k = 0.325/0.775 = 0.4193548
x1增加1个单位，x2增加0.4193548
单位圆方向
x1 = 0.9221944 , x2 = 0.3867267, 0.3867267/0.9221944 = 0.4193548
 time = 4.610972, temperature = 1.933633

FO.steepest <- steepest(FO.model, dist = seq(0, 20, by = 1))

## Path of steepest ascent from ridge analysis:

FO.steepest

## dist x1 x2 | time temperature | yhat

## 1 0 0.000 0.000 | 35.000 155.000 | 40.444
## 2 1 0.922 0.387 | 39.610 156.935 | 41.285
## 3 2 1.844 0.773 | 44.220 158.865 | 42.125
## 4 3 2.766 1.160 | 48.830 160.800 | 42.965
## 5 4 3.689 1.547 | 53.445 162.735 | 43.806
## 6 5 4.611 1.934 | 58.055 164.670 | 44.647
## 7 6 5.534 2.321 | 62.670 166.605 | 45.488
## 8 7 6.456 2.707 | 67.280 168.535 | 46.328
## 9 8 7.377 3.094 | 71.885 170.470 | 47.167
## 10 9 8.302 3.482 | 76.510 172.410 | 48.010
## 11 10 9.222 3.867 | 81.110 174.335 | 48.848
## 12 11 10.139 4.252 | 85.695 176.260 | 49.684
## 13 12 11.069 4.642 | 90.345 178.210 | 50.532
## 14 13 11.987 5.027 | 94.935 180.135 | 51.368
## 15 14 12.913 5.415 | 99.565 182.075 | 52.212
## 16 15 13.844 5.806 | 104.220 184.030 | 53.060
## 17 16 14.767 6.192 | 108.835 185.960 | 53.901
## 18 17 15.685 6.578 | 113.425 187.890 | 54.738
## 19 18 16.604 6.963 | 118.020 189.815 | 55.576
## 20 19 17.523 7.348 | 122.615 191.740 | 56.413
## 21 20 18.444 7.735 | 127.220 193.675 | 57.252
cube.design <- rsm::cube(2,
n0=5,
coding =list(x1~(time-85)/5, x2~(temperature-175)/5),
randomize = F)
res <- c(76.5,78,77,79.5,79.9,80.3,80,79.7,79.8)
cube.design$res <- res
cube.design

## run.order std.order time temperature res

## 1 1 1 80 170 76.5
## 2 2 2 90 170 78.0
## 3 3 3 80 180 77.0
## 4 4 4 90 180 79.5
## 5 5 5 85 175 79.9
## 6 6 6 85 175 80.3
## 7 7 7 85 175 80.0
## 8 8 8 85 175 79.7
## 9 9 9 85 175 79.8
##
## Data are stored in coded form using these coding formulas ...
## x1 ~ (time - 85)/5
## x2 ~ (temperature - 175)/5

FO.model <- rsm(res~FO(x1,x2),data = cube.design)

summary(FO.model)
##
## Call:
## rsm(formula = res ~ FO(x1, x2), data = cube.design)
##
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 78.96667 0.45379 174.0156 2.43e-12 ***
## x1 1.00000 0.68069 1.4691 0.1922
## x2 0.50000 0.68069 0.7346 0.4903
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Multiple R-squared: 0.3102, Adjusted R-squared: 0.08023
## F-statistic: 1.349 on 2 and 6 DF, p-value: 0.3283
##
## Analysis of Variance Table
##
## Response: res
## Df Sum Sq Mean Sq F value Pr(>F)
## FO(x1, x2) 2 5.000 2.5000 1.3489 0.3282610
## Residuals 6 11.120 1.8533
## Lack of fit 2 10.908 5.4540 102.9057 0.0003635
## Pure error 4 0.212 0.0530
##
## Direction of steepest ascent (at radius 1):
## x1 x2
## 0.8944272 0.4472136
##
## Corresponding increment in original units:
## time temperature
## 4.472136 2.236068

练习
一位工程师开发双项存货系统的计算机模拟模型，决策变量是订货量（Order Quantity）和再订购点（Reorder
point）。需要最小化的响应是总存货成本（Total Cost）。识别出试验设计，并求最速上升路径。
cube.design <- rsm::cube(2,
n0=3,
reps = 2,
coding =list(x1~(Order-275)/25, x2~(Reorder-60)/20),
randomize = F)
res <- c(625,670,663,648,654,634,692,686,680,674,681)
cube.design$res <- res
cube.design

## run.order std.order Order Reorder res

## 1 1 1 250 40 625
## 1.1 2 2 250 40 670
## 2 3 3 300 40 663
## 2.1 4 4 300 40 648
## 3 5 5 250 80 654
## 3.1 6 6 250 80 634
## 4 7 7 300 80 692
## 4.1 8 8 300 80 686
## 11 9 9 275 60 680
## 21 10 10 275 60 674
## 31 11 11 275 60 681
##
## Data are stored in coded form using these coding formulas ...
## x1 ~ (Order - 275)/25
## x2 ~ (Reorder - 60)/20

FO.model <- rsm(res~FO(x1,x2),data = cube.design)

summary(FO.model)
##
## Call:
## rsm(formula = res ~ FO(x1, x2), data = cube.design)
##
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 664.2727 5.7125 116.2839 3.343e-14 ***
## x1 13.2500 6.6985 1.9781 0.0833 .
## x2 7.5000 6.6985 1.1197 0.2953
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Multiple R-squared: 0.3924, Adjusted R-squared: 0.2405
## F-statistic: 2.583 on 2 and 8 DF, p-value: 0.1363
##
## Analysis of Variance Table
##
## Response: res
## Df Sum Sq Mean Sq F value Pr(>F)
## FO(x1, x2) 2 1854.5 927.25 2.5832 0.1363
## Residuals 8 2871.7 358.96
## Lack of fit 2 1500.0 750.01 3.2807 0.1090
## Pure error 6 1371.7 228.61
##
## Direction of steepest ascent (at radius 1):
## x1 x2
## 0.8702569 0.4925982
##
## Corresponding increment in original units:
## Order Reorder
## 21.756421 9.851964