2. Pareto optimality fundamentals

In occupant protection optimization, we rarely chase a single number. A restraint system must simultaneously reduce head injury criterion (HIC), chest deflection, chest acceleration, neck injury criterion (Nij), and femur loads, while also satisfying packaging, cost, and mass constraints. These objectives conflict: stiffening a restraint system lowers chest deflection but raises chest acceleration; adding pretensioner travel reduces excursion but can increase neck loading. Pareto optimality provides the formal language for reasoning about such trade-offs, and it is the foundation on which every multi-objective evolutionary algorithm in this course is built.

Dominance: The Only Ordering We Can Trust

Let a design be a vector x in the design space, mapped by simulation or surrogate to an objective vector f(x) = [f₁, f₂, …, f_M]. Throughout this course we adopt the convention that all objectives are minimized; a maximization objective such as energy absorption is negated before optimization. With this convention, design a dominates design b (written a ≺ b) if and only if two conditions hold:

• Weak superiority: f_i(a) ≤ f_i(b) for every objective i = 1…M.
• Strict superiority in at least one: there exists at least one j with f_j(a) < f_j(b).

If neither a ≺ b nor b ≺ a, the two designs are mutually non-dominated. This is the crucial difference from single-objective optimization: non-dominated designs are not "tied" — they represent genuinely different engineering compromises, and no information internal to the objective vectors can rank them. A designer's preference, a regulatory threshold, or a robustness criterion must break the tie.

Consider a simplified frontal impact example with two objectives, HIC₁₅ and peak chest acceleration (both minimized). Design A gives (620, 41 g), B gives (700, 38 g), C gives (650, 44 g). A dominates C (better on both), but A and B are non-dominated: B trades 80 HIC points for 3 g of chest acceleration. Whether that trade is acceptable depends on the occupant's injury risk curve, not on the optimizer.

The Pareto Front and Its Geometry

The set of all non-dominated designs in the feasible objective space is the Pareto front (or Pareto frontier). In two objectives it is a curve; in three, a surface; beyond three it is a hypersurface that we can no longer draw but must still reason about. Three geometric properties matter in practice:

• Convexity and concavity: Real crashworthiness fronts are frequently non-convex. Weighted-sum scalarization cannot recover points in concave regions, which is one reason this course relies on Pareto-based evolutionary algorithms rather than weighted aggregation.
• Discreteness: We never obtain the true front, only an approximation from a finite population. All performance indicators below measure the quality of that approximation.
• Scale dependence: HIC values in the hundreds and chest deflection in millimeters differ by orders of magnitude. Without normalization, Euclidean distances in objective space are dominated by the largest-magnitude objective, silently distorting diversity metrics.

Non-Dominated Sorting: Assigning Rank by Layers

NSGA-II and its descendants do not merely identify the front; they peel the population into successive non-dominated layers. The procedure is straightforward and worth implementing once by hand to internalize it:

1. For each individual p, compute the set S_p of individuals it dominates and the counter n_p of individuals that dominate it.
2. All individuals with n_p = 0 form the first front F₁; assign rank 1.
3. For each p in F₁, decrement n_q for every q in S_p. Any q whose counter reaches zero joins the next front F₂.
4. Repeat until every individual is ranked.

The cost is O(MN²) for N individuals and M objectives — acceptable for populations of a few hundred, but the quadratic term becomes painful in many-objective settings where M ≥ 4. Within a front, NSGA-II breaks ties by crowding distance: the perimeter of the cuboid formed by an individual's nearest neighbors along each objective axis, with boundary points assigned infinite distance to preserve extremes. This is why the front in the figure retains well-spread endpoints rather than collapsing to a single cluster.

Performance Indicators for Many-Objective Problems

Once we have an approximated front, we need to judge it. Three indicator families dominate practice, and each answers a different question:

Hypervolume is the workhorse because it is Pareto-compliant: if front A dominates front B in the set sense, HV(A) ≥ HV(B) always. Its weakness is computational cost, which grows exponentially with M. For many-objective crash problems, exact HV becomes intractable beyond roughly six objectives, and Monte Carlo estimators or the WFG / exact HV algorithms are used instead.

IGD reverses the perspective: for each point in a reference set P*, find the nearest point in the approximation A and average those distances. Low IGD means the approximation is close to and covers the reference front. IGD+ modifies the distance to be dominance-aware, removing the pathological case where a dominated point is rewarded for being close to a reference point it should not approach.

Why Many-Objective Problems Break the Classical Machinery

As M grows past three or four, dominance pressure collapses. In a random population, the probability that one individual dominates another falls roughly as (3/4)^M under independence assumptions; at M = 10 almost every individual is non-dominated, and the first front swallows the entire population. Non-dominated sorting then provides no selection gradient, and the algorithm degenerates into random drift.

Two responses are standard. First, reference-point-based selection (NSGA-III, MOEA/D) replaces pure dominance with a structured set of reference directions, maintaining diversity explicitly even when dominance is uninformative. Second, indicator-based selection (SMS-EMOA, HypE) selects directly on hypervolume contribution, which remains discriminative in high dimensions. In occupant protection studies, it is also common to reduce M first: cluster correlated injury criteria (e.g., HIC and head acceleration) or apply principal component analysis to the objective set before optimization, then report the full set post hoc.

With these fundamentals in place — a consistent minimization convention, a correct dominance test, layered non-dominated sorting with normalized crowding distance, and indicators that separate convergence from coverage — we can now evaluate any candidate restraint design set on its merits rather than on a single aggregated score. The next step is to connect these concepts to the specific injury criteria and constraint handling used in occupant protection studies.