RSM: Multi-Response Optimization (MRO)

Topic - In complex industrial systems, optimizing a single process response in isolation is often counter-productive. Improving throughput or chemical yield frequently degrades surface roughness, accelerates machine wear, or elevates operational costs. Multi-Response Optimization (MRO) uses the Derringer-Suich Desirability Function to simultaneously balance multiple competing outputs into a single composite metric.

---

1. The Multi-Response Dilemma

When analyzing a system with multiple dependent variables ($Y_1, Y_2, \dots, Y_m$):

• Single-Response RSM: Maps inputs directly to a single scalar: $(X_1, X_2) \longrightarrow Y$
• Multi-Response RSM: Simultaneously maps inputs to an array of competing criteria: $(X_1, X_2) \longrightarrow \begin{cases} Y_1 \text{ (maximize yield)} \\ Y_2 \text{ (minimize roughness)} \\ Y_3 \text{ (target thickness)} \end{cases} \longrightarrow \text{Composite Desirability } D$

---

2. The Desirability Function Approach

The desirability methodology converts each physical response $Y_i$ with different physical units (e.g., $\%$, $\mu\text{m}$, $\text{MPa}$) into a dimensionless scale bounded strictly between $0$ and $1$:

• $d_i = 0 \implies$ Completely undesirable outcome (unacceptable defect).
• $d_i = 1 \implies$ Fully desirable or ideal target performance.

---

Linear Desirability Formulations

When linear transitions between acceptable boundaries are assumed:

1. Larger-the-Better Response (e.g., Yield, Hardness)

Given lower threshold $Y_{\min}$ and upper ideal target $Y_{\max}$:

$$d_Y = \frac{Y - Y_{\min}}{Y_{\max} - Y_{\min}}$$

• If $Y \le Y_{\min}$, then $d_Y = 0$.
• If $Y \ge Y_{\max}$, then $d_Y = 1$.

2. Smaller-the-Better Response (e.g., Roughness, Defects)

Given lower ideal target $R_{\min}$ and upper intolerable threshold $R_{\max}$:

$$d_R = \frac{R_{\max} - R}{R_{\max} - R_{\min}}$$

• If $R \le R_{\min}$, then $d_R = 1$.
• If $R \ge R_{\max}$, then $d_R = 0$.

---

Non-Linear Exponential Formulations (Derringer-Suich)

When the rate of satisfaction changes non-linearly, shape parameters ($s$ and $t$) are introduced:

Maximization:

$$d_i = \begin{cases} 0 & \text{if } \hat{y} \lt L \\ \left(\frac{\hat{y} - L}{T - L}\right)^s & \text{if } L \le \hat{y} \le T \\ 1 & \text{if } \hat{y} \gt T \end{cases}$$

Minimization:

$$d_i = \begin{cases} 1 & \text{if } \hat{y} \lt T \\ \left(\frac{U - \hat{y}}{U - T}\right)^t & \text{if } T \le \hat{y} \le U \\ 0 & \text{if } \hat{y} \gt U \end{cases}$$

---

3. The Overall Composite Desirability ($D$)

To evaluate the overall system performance across all $k$ responses, the individual desirability scores are combined using the geometric mean:

$$\boxed{D = \left( d_1 \times d_2 \times \dots \times d_k \right)^{\frac{1}{k}} = \left(\prod_{i=1}^k d_i\right)^{\frac{1}{k}}}$$

Or, when responses carry different priority weights ($w_i$):

$$D = \left( d_1^{w_1} \times d_2^{w_2} \times \dots \times d_k^{w_k} \right)^{\frac{1}{\sum w_i}}$$

warning

The Zero-Product Rule of Desirability

The geometric mean possesses a strict mathematical safeguard: $d_j = 0 \implies D = 0$ If even a single response violates specifications ($d_j = 0$), the overall composite desirability collapses to zero regardless of how exceptional the other responses are.

---

4. Comprehensive Solved Example: Yield vs. Roughness Trade-Off

Consider an industrial process evaluated across four candidate factor setpoints:

• Yield ($Y$, in $\%$): Larger-the-better ($Y_{\min} = 80\%, Y_{\max} = 100\%$).
• Surface Roughness ($R$, in $\mu\text{m}$): Smaller-the-better ($R_{\min} = 2.0\,\mu\text{m}, R_{\max} = 4.0\,\mu\text{m}$).

Measured Process Data

Factor Setting	Observed Yield ($Y$)	Observed Roughness ($R$)
Setting A	$88\%$	$3.2\,\mu\text{m}$
Setting B	$92\%$	$2.8\,\mu\text{m}$
Setting C	$95\%$	$3.5\,\mu\text{m}$
Setting D	$90\%$	$2.4\,\mu\text{m}$

Analysis: We cannot simply select Setting C for its highest yield ($95\%$) because its surface roughness ($3.5\,\mu\text{m}$) is substantially degraded.

---

Step-by-Step Desirability Calculation for a Second Production Run

Suppose operating conditions are refined, yielding the following performance metrics across Settings A, B, C, and D:

Setting	Yield ($Y$)	Yield Desirability ($d_Y$)	Roughness ($R$)	Roughness Desirability ($d_R$)	Overall Desirability ($D$)
A	$94$	$\mathbf{0.70}$	$2.7$	$\mathbf{0.65}$	$D_A = \sqrt{0.70 \times 0.65} \approx \mathbf{0.674}$
B	$98$	$\mathbf{0.90}$	$2.3$	$\mathbf{0.85}$	$D_B = \sqrt{0.90 \times 0.85} \approx \mathbf{0.875}$
C	$100$	$\mathbf{1.00}$	$3.0$	$\mathbf{0.50}$	$D_C = \sqrt{1.00 \times 0.50} \approx \mathbf{0.707}$
D	$96$	$\mathbf{0.80}$	$2.1$	$\mathbf{0.95}$	$D_D = \sqrt{0.80 \times 0.95} \approx \mathbf{0.872}$

---

Mathematical Verification for Setting B:

1. Yield Desirability ($d_Y$):

$$d_Y = \frac{98 - 80}{100 - 80} = \frac{18}{20} = \mathbf{0.90}$$

2. Roughness Desirability ($d_R$):

$$d_R = \frac{4.0 - 2.3}{4.0 - 2.0} = \frac{1.7}{2.0} = \mathbf{0.85}$$

3. Overall Desirability ($D$):

$$D = \sqrt{d_Y \times d_R} = \sqrt{0.90 \times 0.85} = \sqrt{0.765} \approx \mathbf{0.875}$$

---

Engineering Decision Analysis

• Setting C: Achieves the maximum yield ($100\%, d_Y = 1.00$), but poor surface finish ($d_R = 0.50$) lowers its overall score to $D = 0.707$.
• Setting D: Achieves extraordinary surface finish ($2.1\,\mu\text{m}, d_R = 0.95$), but slightly lower yield gives $D = 0.872$.
• Setting B (The Global Optimum): Yields $98\%$ with a clean $2.3\,\mu\text{m}$ roughness, producing the highest overall desirability ($D = 0.875$).

Final Recommendation: Select Setting B because it delivers the optimal compromise between productivity and surface quality.

---

5. Exam Traps & Operational Nuances

1. The Arithmetic Average Error

• The Trap: Averaging desirability scores with an arithmetic mean: $\bar{d} = \frac{d_1 + d_2}{2}$.
• The Consequence: If $d_1 = 1.0$ (perfect yield) and $d_2 = 0.0$ (totally defective roughness), the arithmetic average is $0.50$ (acceptable!). In reality, the product is defective scrap. The geometric mean correctly yields $D = \sqrt{1.0 \times 0.0} = 0.00$.

2. Misinterpreting Unconstrained Optima

• The Trap: Optimizing each response's regression polynomial separately and attempting to superimpose the optimum setpoints.
• The Reality: The individual optimum of Response 1 rarely coincides with Response 2. Multi-response optimization must be conducted simultaneously on the composite objective function $D(\mathbf{x})$ over the shared factor space.

---

6. Summary & Cheatsheet

Desirability Formulas

• Larger-the-Better: $d = \frac{Y - Y_{\min}}{Y_{\max} - Y_{\min}}$.
• Smaller-the-Better: $d = \frac{R_{\max} - R}{R_{\max} - R_{\min}}$.
• Range: $0 \le d \le 1$.

Composite Desirability

• Geometric Mean: $D = (\prod d_i)^{1/k}$.
• Zero-Product Safeguard: Any $d_i = 0 \implies D = 0$.
• Goal: Maximize $D$ across the joint factor space.

info

Key Takeaways

• Simultaneous Trade-Off: Multi-Response Optimization systematically resolves competing operational criteria that single-response RSM cannot handle.
• Scale Invariance: Converting physical outputs to dimensionless desirability values ($0$ to $1$) allows apples-to-oranges aggregation of yields, tolerances, and costs.
• Zero Defect Guarantee: The geometric mean prevents any individual critical parameter violation from being hidden by superior performance elsewhere.

Next Section: Global Optimization & Metaheuristics Taxonomy - Components of formal optimization problems, local vs. global extrema, and the master 5-branch taxonomy of derivative-free metaheuristics.