4. Bayesian multi-objective optimization

Classical multi-objective optimization assumes the objective functions are cheap to evaluate and deterministic. In engineering practice—tuning a controller, designing a heat sink, selecting hyperparameters for a model that takes hours to train—each evaluation is expensive, noisy, and often only available through simulation or experiment. Bayesian multi-objective optimization (BMOO) addresses this setting by placing a probabilistic surrogate (typically a Gaussian process) over each objective and using an acquisition function to decide where to sample next. The central question is no longer "which point is optimal?" but "which candidate, once evaluated, is expected to improve our knowledge of the Pareto front the most?"

This chapter develops the three ideas that make BMOO work in practice: expected hypervolume improvement (EHVI) as the dominant acquisition function, explicit modeling of the Pareto front as a set-valued object, and scalarization-free methods that avoid collapsing multiple objectives into a single weighted sum.

Why Scalarization Breaks Down

The simplest way to handle multiple objectives is to minimize a weighted sum f(x) = Σ wᵢ fᵢ(x). This is attractive because it reduces to single-objective BO, but it has two well-known failure modes. First, a fixed weight vector can only recover points on the convex portion of the Pareto front; concave regions are unreachable regardless of the weights chosen. Second, the mapping from weights to front points is highly non-uniform—small weight changes near the extremes can produce large jumps, so a uniform sweep of weights does not give a uniform sampling of the front. Scalarization-free approaches instead reason directly about the Pareto dominance relation and the geometry of the front.

Hypervolume as a Quality Indicator

The hypervolume indicator (also called the S-metric) measures the region of objective space dominated by a solution set and bounded by a reference point r. For a minimization problem with a set of non-dominated points P, the hypervolume is

HV(P, r) = Λ( ∪_{p ∈ P} [p, r] )

where Λ denotes the Lebesgue measure and [p, r] is the axis-aligned box with corners p and r. Hypervolume is Pareto-compliant: if set A dominates set B, then HV(A) ≥ HV(B). This property is what makes it suitable as the target of an acquisition function. The reference point must be chosen to dominate all observed points; a common heuristic is to set r slightly beyond the worst observed value in each objective.

Expected Hypervolume Improvement

EHVI is the expected increase in hypervolume obtained by adding a candidate point x to the current Pareto set P:

EHVI(x) = E[ HV(P ∪ {y}, r) − HV(P, r) ]

where y ~ N(μ(x), σ²(x)) is the surrogate's posterior predictive distribution at x. The expectation is taken over the joint uncertainty in all objectives. For two objectives, EHVI has a closed-form expression involving the bivariate normal CDF; for three or more objectives, exact computation becomes expensive and practitioners use Monte Carlo estimation or the decomposed EHVI (DG-EHVI) formulation, which partitions the improvement region into disjoint cells and sums their expected contributions.

A practical implementation loop looks like this:

1. Fit one GP per objective on the current dataset D = {(xᵢ, yᵢ)}.
2. Compute the current Pareto set P from observed points and fix the reference point r.
3. Maximize EHVI over the design space using a multi-start L-BFGS-B or a genetic algorithm, since EHVI is multimodal.
4. Evaluate the true objectives at the selected x*, append to D, and repeat.

Modeling the Pareto Front Directly

An alternative to acquisition-function-driven search is to model the Pareto front as a structured object. Two families are common. The first fits a Pareto front model—for example, a monotone spline or a Gaussian process over the front's parameterization—and uses the posterior over the front to guide sampling. The second treats the front as a random set and uses Bayesian set estimation, where the posterior over the front is represented by the collection of non-dominated draws from the surrogate posteriors. This set-based view is useful for reporting uncertainty: instead of a single estimated front, you obtain a credible band that widens in sparsely sampled regions.

In practice, the choice between EHVI and front modeling depends on the evaluation budget and the dimensionality of the objective space. EHVI scales well to 2–3 objectives and moderate budgets; front modeling is preferable when the front has known structure (e.g., a smooth trade-off curve in a physical design problem) and when interpretability of the front's uncertainty matters.

Worked Example: Two-Objective Design

Consider a heat-sink design with two objectives to minimize: thermal resistance R (K/W) and pressure drop Δp (Pa). Suppose after 12 evaluations the observed Pareto set contains three points: (0.42, 18.0), (0.38, 24.5), and (0.35, 31.0). Set the reference point to r = (0.50, 40.0). The current hypervolume is the union of three rectangles:

HV = (0.50−0.42)(40−18) + (0.50−0.38)(40−24.5) + (0.50−0.35)(40−31)
   = 1.76 + 1.86 + 1.35 = 4.97

A candidate x* has posterior μ = (0.40, 22.0) and σ = (0.02, 1.5). Its expected improvement is the probability-weighted area it would add if it lands in the gap between the first two observed points. Monte Carlo estimation with 10,000 draws gives an EHVI of roughly 0.31, which exceeds the EHVI of all other candidates in the pool, so x* is selected for the next evaluation. After evaluation, if the true point is (0.39, 23.1), it enters the Pareto set and the hypervolume rises to about 5.12.

Practical Considerations

Three issues dominate real deployments. First, noise: when objectives are stochastic, use a GP with a nugget term and compute EHVI against the posterior mean, or use the noisy EHVI formulation that integrates over the predictive distribution. Second, constraints: feasibility is often as important as optimality; multiply EHVI by the probability of feasibility from a classifier GP, or use constrained EHVI. Third, scaling: normalize each objective to [0, 1] before computing hypervolume, otherwise the objective with the larger numeric range dominates the indicator.

The figure on the left shows the acquisition loop; the right panel shows the geometry that EHVI evaluates. The shaded union is the current hypervolume, and the dashed box marks the region a candidate x* would newly dominate if it lands near its posterior mean. EHVI integrates the area of that box over the predictive distribution, which is why candidates near the front's "gaps" receive high scores even when their mean prediction is not itself Pareto-optimal.

Summary

EHVI turns the multi-objective search into a sequence of well-posed single-objective maximization problems while preserving Pareto compliance. Scalarization-free methods avoid the convexity and weight-sensitivity pitfalls of weighted sums, and explicit Pareto front modeling provides uncertainty-aware reporting. The practical recipe is: normalize objectives, fit independent GPs, recompute the reference point each iteration, maximize EHVI with a global optimizer, and handle noise and constraints through the appropriate EHVI variants.