Multi-Objective Optimization and Multi-Solution Exploration

In real-world engineering, there is almost never a "single optimal solution." Take supercritical airfoil design as an example:lift-to-drag ratio, structural strength, stall margin, and manufacturing/simulation cost are all in tension with one another—improving the lift-to-drag ratio often means structural loads and manufacturing costs rise in tandem. The right way to frame such problems is:multi-objective optimizationthe goal is not to pinpoint a single solution, but to characterize and approximate a Pareto frontacross multiple mutually competing objectives, and to retain a set of candidate solutions with different orientations and trade-offs, leaving the final value judgment to downstream preferences or domain experts.

1. First, confirm: is this really a multi-objective problem?

Multi-solution exploration presupposes that the objectives genuinely conflict. If all objectives can be improved simultaneously without trade-offs, the problem degenerates into a single-objective one, and multi-solution exploration is both unnecessary and a waste of computing power. The criterion is not "more than one objective," but rather:does improving one objective necessarily come at the expense of another. Only opposing directions and mutual trade-offs constitute a true multi-objective problem.

2. Pareto Dominance and the Pareto Front

The core criterion is Pareto dominance: Solution A dominates solution B if and only if A is no worse than B on all objectives (fᵢ(A) ≤ fᵢ(B) for every objective i) and strictly better on at least one objective (there exists j such that fⱼ(A) < fⱼ(B)). Any dominated solution has a replacement that is "at least as good and better somewhere," so it has no value in being kept. All solutions that are not dominated by any other solution form the Pareto front—solutions on the front are mutually non-dominated, and which is more suitable depends on preference, not on mathematics.

3. The Cost of Reducing to a Single Objective

Scalarization methods such as weighted sum and ε-constraint compress a vector objective into a single scalar. This is convenient in engineering, but the costs are concrete:
· Masks the trade-off structure—once weights are fixed, the shape of the front and the alternative directions become invisible, and decision-makers see only a single score.
· Weights must be specified in advance—yet real preferences often cannot be formalized in advance and typically only become clear during the review stage.
· Limited coverage—weighted sum typically hits only isolated points on the front, and portions of a non-convex front are difficult to reach.
Once a later review changes the weights, the entire search must be rerun—the exact opposite of "one exploration for many freedoms in decision-making."

4. Evaluation Is Two-Dimensional: Convergence and Diversity

The quality of a multi-objective solution set must be assessed on two dimensions simultaneously:convergence(whether the solution set approaches the true front) anddiversity(whether the solutions cover different regions of the front). The two often pull against each other: once selection pressure leans too heavily toward "better," the population collapses to a single point or a small segment of the front—appearing to converge quickly while losing the trade-off information. It is therefore necessary to explicitly maintain population diversity and to combine selection pressures such as crowding distance, decomposition, and indicator-driven approaches (e.g., hypervolume).

5. Engineering Implication: The Decision Point Shifts

The engineering meaning of multi-objective optimization is a shift in the division of labor: the optimizer is responsible only for producinga comparable set of candidates, and final selection is made by preferences or domain experts. The optimizer should not prematurely converge on the decision-maker's behalf. The practical rationale for retaining a diverse candidate set is "hedging"—when preferences are unknown, may shift with context, or when constraints get revised during deployment, a diverse set of solutions is more valuable than a single "optimal solution." This is consistent with the sedimentation mechanism in which "candidate modifications must pass compilation, testing, and acceptance before being adopted": first retain a set of candidates, then let verification and preferences decide which one enters the final version.

Corresponding interface-level implementation

At the implementation level, a multi-objective problem should be defined asa vector-valued evaluation function(returning multiple objective values at once) anda set of candidate solutionsas output, rather than a scalar fitness with a single return value. The number of values returned by the evaluation function determines the objective dimensionality, and the output structure determines whether diversity can be preserved.

Pareto Front and Dominance Relations

The scatter plot uses two mutually competing objectives as coordinates, distinguishing dominated solutions from the Pareto front: solutions on the front are mutually non-dominated, so improving one objective necessarily comes at the cost of degrading another. The optimizer therefore outputs a set of candidates with different orientations, leaving the final trade-off to preferences or domain experts.

Question 1

Regarding the determination of Pareto dominance relations, which of the following statements is correct:

Solution A dominates solution B as long as it is better on any single objective.

If two solutions each excel on different objectives, the better of the two dominates the other.

A dominated solution is necessarily infeasible and must be removed from the candidate set.

Solution A dominates solution B if and only if A is no worse than B on all objectives and strictly better on at least one objective.

Question 2

Before committing to a multi-objective solve, the key criterion for determining whether a problem truly qualifies as "multi-objective" is:

The objectives point in opposite directions, so improving one necessarily comes at the cost of degrading another.

Whether weighted sum is used for scalarization.

Whether the problem's scale exceeds single-machine compute capacity.

The number of objectives is greater than 1.

Question 3

Regarding scalarization methods such as weighted sum and ε-constraint, which of the following statements is correct:

Solutions obtained through scalarization are guaranteed to cover all regions of the front.

As long as enough weights are used, scalarization can equivalently replace multi-objective exploration.

Scalarization masks the trade-off structure and requires weights to be specified in advance, yet real preferences often cannot be formalized beforehand.

Scalarization is merely a change of representation and loses no trade-off information.

Question 4

When evaluating the solution set of a multi-objective algorithm, the relationship between "convergence" and "diversity" is:

The two often pull against each other, and overemphasizing convergence easily collapses the population onto isolated segments of the front.

As long as convergence is good, diversity can be disregarded.

Diversity only affects visualization and has nothing to do with solution set quality.

The two are completely independent and can each reach their own optimum simultaneously.

Question 5

According to the engineering meaning of multi-objective optimization, the division of labor between the optimizer and decision-makers should be:

The optimizer should converge the population to a single point early in the search to save compute.

The optimizer outputs a comparable set of non-dominated candidates, and final selection is left to preferences or domain experts.

Decision-makers must specify all weights in advance, otherwise optimization cannot begin.

The optimizer directly outputs a single optimal solution to spare decision-makers from hesitation.

Case Study: Multi-Objective Trade-offs in Supercritical Airfoils

Read the scenario and answer the questions.

An aerospace team needs to design a supercritical airfoil for a next-generation wing. Candidate designs must simultaneously satisfy four metrics: lift-to-drag ratio (higher is better), structural mass (lower is better), stall margin (higher is better), and manufacturing and simulation cost (lower is better). Raising the lift-to-drag ratio often comes with higher structural loads, and enhancing the stall margin often sacrifices the lift-to-drag ratio. Moreover, with a limited simulation budget, the team cannot afford to rerun the search repeatedly. The initial approach was to combine these four metrics into a single total score using a preassigned set of weights and output a single "optimal airfoil" for review.

1. Why do these metrics constitute a genuine multi-objective problem? If a single weighted total score is used to output a single "optimal airfoil," what specific risks arise during the review stage?

Answer:
Because the objectives point in opposite directions and compete with one another: improving the lift-to-drag ratio comes at the cost of higher structural loads and manufacturing costs, while the stall margin and lift-to-drag ratio also compete for the same design space. Any improvement in one metric requires another to give ground, and they cannot be improved simultaneously without trade-offs—so this is a genuine multi-objective problem. The risks are: the weighted total compresses the trade-off structure into a single scalar, masking information about "which metric gave up how much"; weights must be specified in advance, yet the true preferences only become clear during the review itself; and once the review adjusts the weights, the entire search must be rerun, which the simulation budget simply cannot afford. The conclusion: scalarization lets the optimizer make a value judgment on the decision-maker's behalf when information is insufficient.

2. If multi-objective optimization is adopted instead and a set of non-dominated candidates is retained, explain how the evaluation function and output structure should be defined, and how selection should be made during the review.

Answer:
The evaluation function should be changed to a vector-valued form: for each candidate airfoil x, it returns a quadruple (lift-to-drag ratio, structural mass, stall margin, manufacturing and simulation cost) rather than a scalar fitness. The output structure should be a set of candidate solutions—a group of mutually non-dominated Pareto front solutions. At the algorithmic level, population diversity must be explicitly maintained, combined with selection pressures such as crowding distance, decomposition, or hypervolume-driven indicators, to prevent the population from collapsing onto a single segment of the front and to ensure representative solutions exist at both the extremes and the middle of the front. During the review, selection is made directly from the candidate set according to true preferences: if structural reliability is prioritized, choose the candidate with lower structural mass; if aerodynamic efficiency is prioritized, choose the one with a higher lift-to-drag ratio. In this way, a single exploration can support multiple decisions, and when preferences shift with context or constraints get revised during deployment, the team still has viable alternatives on hand.