Comparing Scalar Objective Functions for Multi-Criteria Engineering Optimization

Technical Summary: Comparing Scalar Objective Functions for Multi-Criteria Engineering Optimization

Problem Statement
In multi-criteria engineering optimization, problems often involve competing criteria (e.g., mass, drag, cost) that result in a Pareto front of non-dominated solutions. While generating this front is valuable, practical engineering decisions frequently require selecting a single preferred design. This necessitates a scalar objective function to collapse the multi-dimensional problem into a single preference statement. However, the choice of scalarization method significantly influences which compromise is selected, how preference parameters are interpreted, and whether specific regions of the Pareto front (particularly non-supported regions on non-convex fronts) are reachable. The difficulty in interpreting weights and the structural limitations of linear methods on non-convex fronts motivate a comparative analysis of alternative formulations.

Methodology
The paper employs a controlled, diagnostic approach to isolate the effects of scalar objective function formulations from numerical optimizer behavior, discretization errors, and simulation noise. The study utilizes two analytically defined bi-criteria Pareto fronts:

1. Convex Front: Defined by a specific rational function, representing a scenario where linear scalarization can theoretically reach all points.
2. Concave Front: Defined by a quadratic function, representing a scenario where interior points are non-supported by linear scalarization.

Four scalarization formulations are compared under normalized minimization conditions ($c_i \in [0, 1]$):

1. Weighted Sum (WS): A linear combination $f_{WS} = w c_1 + (1-w)c_2$.
2. Achievement Scalarizing Function (ASF): A weighted Tchebycheff formulation with an augmented term to handle tie-breaking, measuring the maximum weighted deviation from an ideal point.
3. Desirability Functions (Des): A two-step process mapping criteria to individual desirability values via nonlinear functions (parabolic, $s=2$), aggregated via a weighted geometric mean.
4. Fuzzy-Logic-Based Formulation (Fuzzy): Maps criteria to graded membership values (desirable, tolerable, undesirable) and aggregates them using explicit rule bases. The study evaluates four specific rule sets (F1–F4) ranging from global exclusion to reference-attraction mechanisms.

The "native" preference parameters for each method ($w$ for WS/ASF, $\alpha$ for Des, $\lambda$ for Fuzzy) are swept uniformly from 0.01 to 0.99. For each parameter value, the optimal point on the Pareto front is identified. The analysis focuses on:

• Reachability: Which regions of the Pareto front can be selected?
• Selection Density: How are parameter values distributed across the front (i.e., does a small change in parameter cause a large jump in solution, or vice versa)?
• Sensitivity: How does the selected solution shift as the preference parameter varies?
• Interpretability: How do the parameters map to engineering preferences?

Key Contributions

• Diagnostic Comparison: A systematic comparison of four distinct scalarization methods using analytically defined test cases to separate formulation behavior from numerical artifacts.
• Distinction of Properties: The demonstration that "reachability" (the ability to select a region) and "parameter-induced selection density" (how uniformly the method explores that region) are distinct properties. A method may reach a region but assign very few parameter values to it, or vice versa.
• Fuzzy Rule-Base Analysis: An investigation showing that fuzzy formulations cannot be characterized by membership functions alone; the explicit rule base (consequents and interactions) fundamentally alters the preference mapping.
• Non-Convex Front Behavior: A detailed analysis of how different methods handle non-supported regions on concave fronts, highlighting the structural limitations of weighted sums versus the capabilities of ASF, desirability, and fuzzy methods.

Results

• Weighted Sum: Simple and smooth but structurally limited. On the concave front, it collapses to endpoint solutions, failing to reach interior non-supported regions. On the convex front, it exhibits higher selection density near endpoints rather than a uniform distribution.
• Achievement Scalarizing Function (ASF): Capable of reaching non-supported regions on the concave front. However, its selection density is not uniform; it shows a mild concentration toward larger values of $c_1$ on the concave front and weak concentration near the balanced compromise on the convex front.
• Desirability Functions: Can reach interior regions on the concave front. The geometric aggregation mechanism strongly penalizes endpoint compromises (where individual desirability becomes zero), leading to a concentration of selected points around the balanced compromise on both fronts.
• Fuzzy Logic:
  • Rule Dependence: The behavior is highly sensitive to the rule base.
    • F1 (Exclusion) & F4 (Exclusion + Attraction): Behave similarly, collapsing to a narrow balanced region on both fronts, dominated by the exclusion rule.
    • F2 (Weak Reference): Produces a broad distribution on the convex front but collapses to an endpoint on the concave front, showing that a reference-dependent premise alone does not guarantee reference-following behavior.
    • F3 (Reference Attractor): On the concave front, this rule set distributes selected points almost uniformly across the front, successfully selecting non-supported regions. On the convex front, it remains centered around the balanced compromise.
  • Sensitivity: The fuzzy formulation's sensitivity to its parameter ($\lambda$) depends on the rule set. F3 exhibits an approximately linear dependence on the concave front, whereas F1 and F4 show little sensitivity due to rule dominance.

Significance and Claims
The paper claims that scalarization is not a neutral numerical post-processing step but imposes a specific preference structure on the Pareto front. The choice of formulation must align with the intended decision logic (e.g., linear compensation vs. rule-based exclusion).

• Reachability vs. Density: A central finding is that reachability alone is insufficient to characterize a scalarization. The mapping from preference parameters to Pareto regions is equally critical; a method might reach a region but assign it low selection density, making it difficult to find via parameter tuning.
• Fuzzy Logic Nuance: The study emphasizes that fuzzy formulations offer a flexible framework for expressing engineering preferences through rules, but this requires careful design. The rule base, not just the membership functions, dictates whether a reference point acts as a weak classifier or a strong attractor.
• Modest Scope: The authors note that while practical engineering problems involve more criteria and constraints, the analytical setting serves as a necessary diagnostic to understand the fundamental assumptions and consequences of these formulations. The paper does not claim one method is universally superior but rather highlights the specific trade-offs in reachability, parameter interpretation, and selection density inherent to each approach.