Skip to content

MANE 6313

Week 15, Module B

Student Learning Outcome

  • Select an appropriate experimental design with one or more factors,
  • Select an appropriate model with one or more factors,
  • Evaluate statistical analyses of experimental designs,
  • Assess the model adequacy of any experimental design, and
  • Interpret model results.

Module Learning Outcome

Employ method of steepest descent (MSD) to locate the region of a minimum


Method of Steepest Descent Demonstration

  • Provide demonstration of method of steepest descent
    • Similar to Example 11.1 from textbook but simplified
    • Only dealing with natural (uncoded variables) and ignoring coded variables
  • Model description
    • Second order model with 3 inputs and one output
    • Range for \(x_1\) is 100 - 200
    • Range for \(x_2\) is 0.5 - 0.75
    • Range for \(x_3\) is 25 - 75
  • The true model will only be used to generate response variable values and not used in the method of steepest descent
  • The true model was created to contain a minimum with the range of the input variables

Steps in Steepest Descent Analysis

  1. First Order DOE
  2. Selection of Minimum Design Point
  3. Construction of Gradient Vector
  4. Gradient Vector at Minimum Point
  5. Selection of Step Size
  6. Construction of Search Path
  7. Implementation and Results of MSD

First Order DOE

  • A single replicate of a \(2^3\) full factorial was conducted around the point (110, 1.0, 30).
  • The data for building the model is
x1 x2 x3 y
100 0.5 25 94.575
100 0.5 35 82.825
100 1.5 25 85.625
100 1.5 35 74.375
120 0.5 25 63.875
120 0.5 35 51.325
120 1.5 25 55.125
120 1.5 35 43.075
  • The fitted linear model is
\[ \hat{y}=273.05-1.44x_1-11.20x_2-0.8x_3+0.01x_1x_2-0.004x_1x_3+0.05x_2x_3 \]

Minimum Point

  • Select a starting point for the steepest method search
  • Since minimum is desired, select design point with smallest value for y

(120, 1.5, 35)


Gradient Vector

  • The gradient vector is found by taking partial derivatives with respect to each variable
\[ \begin{aligned} \frac{d\hat{y}}{dx_1}&= -1.44+0.01x_2-0.004x_3\\ \frac{d\hat{y}}{dx_2}&= -11.20+0.01x_1+0.05x_3\\ \frac{d\hat{y}}{dx_3}&=-0.80-0.004x_1+0.05x_2 \end{aligned} \]
  • Notice that the gradient vector is a function of \(x_1\), \(x_2\), and \(x_3\).

Gradient Vector at DSM Starting Point

  • Recall from the first order design of experiments, (120, 1.5, 35) had the smallest response value and will be starting for the deepest descent method
  • The gradient vector at the starting point is
\[ \begin{aligned} \frac{d\hat{y}}{dx_1}&= -1.44+0.01(1.5)-0.004(35)=-1.565\\ \frac{d\hat{y}}{dx_2}&= -11.20+0.01(120)+0.05(35)=-8.25\\ \frac{d\hat{y}}{dx_3}&=-0.80-0.004(120)+0.05(1.5)=-1.205 \end{aligned} \]
  • The gradient defines the direction of steepest ascent; for the steepest descent, move in the opposite direction

  • Therefore, the search direction is \((1.565, 8.25, 1.204)\)


Step Size

  • Selection of the step size is very subjective
  • Montgomery recommends selecting the largest gradient component and fixing the step size at a convenient number
\[ \Delta x_2=0.5 \]
  • The other step sizes are scaled proportionally to \(x_2\) for this example
\[ \begin{aligned} \Delta x_1&= \Delta x_2\left(\frac{-\frac{d\hat{y}}{dx_1}}{-\frac{d\hat{y}}{dx_2}}\right)=0.5\left(\frac{1.565}{8.25}\right)=0.09485\\ \Delta x_3&= \Delta x_2\left(\frac{-\frac{d\hat{y}}{dx_3}}{-\frac{d\hat{y}}{dx_2}}\right)=0.5\left(\frac{1.204}{8.25}\right)=0.07297\\\end{aligned} \]

Search Path

  • Combining the starting point and step size, the search path is
\[ \begin{aligned} x_1&=120+0.09485t\\ x_2&=1.5+0.5t\\ x_3&=35+0.07297t \end{aligned} \]

Search Results

t x1 x2 x3 y
0 120 1.5 35 43.075
1 120.09485 2 35.7297 39.11371494
2 120.1897 2.5 36.4594 35.82161728
3 120.28455 3 37.1891 33.198707
4 120.3794 3.5 37.9188 31.24498412
5 120.47425 4 38.6485 29.96044862
6 120.5691 4.5 39.3782 29.34510052
7 120.66395 5 40.1079 29.3989398
8 120.7588 5.5 40.8376 30.12196648
  • The minimum was found at step 6 with values (120.57,4.5,39.38)

  • This value should be in the neighborhood of the minimum

  • A full second-order RSM design should be build centered at (120.57,4.5,39.38)