Reference-Point-Free Multi-Objective Bayesian Optimization with Learned Pareto-Front Information

Many engineering and machine-learning problems trade off several objectives at once, such as brake mass against stopping time, or accuracy against fairness. The goal is the Pareto front: the set of trade-offs where no objective can improve without another getting worse. The most common way to steer a sample-efficient search toward that front is the hypervolume, which is the volume of objective space a set of solutions dominates, measured up to a user-chosen corner called the reference point. Where that corner sits quietly decides which part of the front the search favors. When objectives live on very different scales, a good corner cannot be chosen without already knowing the answer. We propose to remove the reference point from the decision of where to evaluate next. The idea is to choose each new query by how much its outcome would tell us about the Pareto front, and to estimate that information with a small trained neural network (a "critic") instead of with closed-form formulas.

Information-based search of this kind already exists and is reference-point-free. Its current forms compute the information analytically and therefore require independent Gaussian-process models for each objective and a partition of objective space into boxes, a partition that grows exponentially with the number of objectives. A learned critic needs only the ability to draw plausible functions from the surrogate model. That makes the same criterion available for correlated objectives, many objectives, and neural surrogates in high-dimensional inputs. It also has a property hypervolume lacks: its value does not change under any strictly increasing rescaling of any objective. We test that property directly, and the proposal stands or falls on it.

Core Claim

Let $\mathcal{D}$ be the data observed so far, $x$ a candidate input, $y_x$ its noisy multi-objective outcome under the surrogate's posterior, and $F$ the random Pareto front implied by that posterior. The acquisition we target is

$$\alpha(x) = I(y_x; F \mid \mathcal{D}),$$

the mutual information (the expected reduction in uncertainty about $F$ from observing $y_x$). Strictly increasing per-objective maps $\phi$ preserve dominance, so the front of the transformed problem is exactly $\phi(F)$. Mutual information is unchanged under bijections of either argument, so $\alpha$ is therefore exactly invariant to such rescalings once the posterior is expressed consistently. Expected hypervolume improvement is not invariant, because its value depends on where the reference point falls in the rescaled space.

We estimate $\alpha$ with a contrastive bound. A critic $T_\psi(y, x, e(F))$ scores how well an outcome matches a front, where $e$ is a learned permutation-invariant encoder of a set of front points. The bound is

$$\alpha(x) \ge \mathbb{E}\left[ T_\psi(y_x^{s}, x, e(F_s)) - \log \frac{1}{S}\sum_{s'=1}^{S} e^{T_\psi(y_x^{s}, x, e(F_{s'}))} \right],$$

where $s$ indexes $S$ functions sampled from the surrogate posterior, $F_s$ is the front of sample $s$, and $y_x^s$ is the outcome at $x$ under that same sample. Matched pairs are joint draws and mismatched pairs are marginal draws. This bound saturates at $\log S$. That ceiling is harmless here, because one query typically carries well under a nat of information about the front.

The claim is specific. On problems with disparate or nonlinearly warped objective scales, correlated objectives, or five or more objectives, this acquisition should match or beat hypervolume-based and analytic entropy-based methods on reference-free quality measures. Its performance should also stay flat under objective rescalings that change the ranking of hypervolume-based methods.

Relation to Prior Work

Three groups of papers play different roles here.

The first group motivates the problem. Tobias Friedrich et al. showed that the reference point dictates where hypervolume-driven search concentrates along the front. Ke Shang et al. proved that an even spread of solutions requires stringent conditions on its placement. Qiuyi Zhang tied misplacement to distorted scalarized acquisition and regret. Mingqian Li et al. found that batch Bayesian optimization recovers incomplete fronts when the point is wrong. The sharpest practical evidence comes from GPSL, which degraded on Disc Brake Design because one objective, total constraint violation, had a scale for which no reference point worked. The hypervolume-driven set learners (Hypervolume Maximization: A Geometric View of Pareto Set Learning, Pareto Set Learning for Expensive Multi-Objective Optimization) all inherit this dependence. We do not modify their training losses. We replace the query-selection step that feeds them data.

The second group is the closest prior work and constrains our design. PFES maximizes information about the front in closed form. It requires independent Gaussian processes and a box partition whose cost explodes beyond about ten objectives. It also reports that 10 to 50 sampled fronts suffice for its estimator, which tells us the front-sampling budget we need is modest. {PF}²ES (see also; PF2ES: Parallel Feasible Pareto Frontier Entropy Search for Multi-Objective Bayesian Optimization; PF2ES: Parallel Feasible Pareto Frontier Entropy Search for Multi-Objective Bayesian Optimization) extends this to constraints and batches through an analytic variational bound over box partitions, with a small corrective shift for clustering. It is reference-point-free and was more stable than expected hypervolume improvement on disjoint Disc Brake fronts. However, it is limited to Gaussian-process surrogates, inputs under about twenty dimensions, and batches of two to four. Our method keeps their target quantity and drops both the box partition and the independence assumption. PFES also shows why a simple summary of the front is not enough: MESMO, which keeps only each objective's maximum, fails on concave fronts such as DTLZ3 and DTLZ4. That result motivates our learned set encoder and supplies our key ablation.

The third group supplies the method. MINE showed that neural critics give differentiable information estimates that scale where the Kraskov nearest-neighbor estimator underestimates at 20 dimensions. MINEBED used such critics to choose experiments for simulators, scaling to 100-dimensional designs, but only for single-objective parameter identification and one-shot designs. We move that machinery to sequential multi-objective search, where the "parameter" being learned about is the Pareto front, a set-valued object.

Research Program

Surrogate and front sampling. At each iteration we fit a surrogate that can produce function samples. We use correlated multi-output Gaussian processes with random-feature sampling in low dimensions and deep ensembles in high dimensions. For each of $S$ function samples, a cheap evolutionary solver run on the sampled function produces an approximate front $F_s$. Outcomes at many candidate inputs under the same sample form the matched pairs. The pair count is therefore $S$ times the number of candidates, which gives the critic plenty of data from few front samples.

Critic and acquisition. The critic conditions on $x$, so one training run amortizes across all candidates. It is warm-started across iterations to control cost. Candidates are ranked by the bound, and gradients with respect to $x$ refine the top candidates. For batches, the critic takes the set of $q$ joint outcomes in place of $y_x$, which yields a joint batch information criterion with no box decomposition. Constraints enter by sampling feasibility along with the objectives, so $F_s$ is the feasible front.

Analysis. We will characterize two things. The first is when critic error is small relative to differences between candidates, since ranking only needs differences to be right. The second is how invariance degrades when the surrogate itself is refit on rescaled data. In that case the posterior changes and exact invariance no longer holds.

Baselines

The naive reference point is expected hypervolume improvement in its noisy batch form, qNEHVI, run the way practitioners run it: min-max normalize objectives on observed data and put the reference point slightly beyond the observed worst values. This is the default most users meet, and our claim is that its quality depends on this heuristic. As an upper-bound probe, we also run it with an oracle reference point computed from the true nadir. The PSL paper reports qEHVI at 75.9 seconds per iteration on DTLZ2 (Pareto Set Learning for Expensive Multi-Objective Optimization), which gives a wall-clock reference our critic training must stay within.

The strongest existing competitor is {PF}²ES (PF2ES), together with PFES where its independence assumption holds. These methods already answer the question "can one avoid the reference point" analytically. The file reports only qualitative results for them: {PF}²ES matched or beat expected hypervolume improvement in log-hypervolume difference on DTLZ4 and C2-DTLZ2, and PFES converged fastest on three of four synthetic benchmarks. No reported numerical log-hypervolume value was found on Lacuna for {PF}²ES or PFES on DTLZ4, VLMOP2, Disc Brake, or Marine Design. We will therefore rerun both ourselves. Fairness rules apply to all methods. Every method gets the same initial designs, evaluation budget, surrogate family where the method permits it, and twenty seeds. Every method's acquisition optimizer gets the same wall-clock allowance per iteration. Beating {PF}²ES means statistically better final quality at equal budget on correlated-objective or many-objective problems, while staying within noise of it on the low-dimensional, independent-objective benchmarks where it is designed to excel.

Random-scalarization Thompson sampling with the same surrogate, in the style of ParEGO with Tchebycheff weights, is the cheap scalable baseline. It runs on deep ensembles in high dimensions where box-based methods cannot. Beating it on concave fronts is expected, given the linear-scalarization failures reported in Pareto HyperNetworks. Beating it on convex, well-scaled fronts is not guaranteed, and we will report either outcome.

Two ablations isolate the key ingredients. The first replaces the learned front encoder with per-objective maxima, which turns our estimator into a neural MESMO. If the set encoder matters, this variant should lose on DTLZ3 and DTLZ4 as MESMO does in PFES. The second replaces the critic with a Kraskov estimator on the same samples, which isolates the value of the neural estimator.

Why not just…?

Why not just normalize the objectives and set the reference point at 1.1 times the observed nadir? This handles affine scale differences. It does not handle heavy-tailed or nonlinearly spread objectives like a constraint violation spanning orders of magnitude. That is exactly where GPSL reported failure, and normalization is shared by all hypervolume methods in the file. The threat is real on well-behaved problems, which is why this heuristic is our first baseline. The warp experiment below shows directly whether its ranking moves while ours does not.

Why not just use PFES or {PF}²ES, which are already reference-point-free? In their regime they may be as good or better, and we expect a tie there. Their stated limits are independent outputs for PFES, box partitions exponential in the number of objectives, inputs under about twenty dimensions, and batches of two to four. Our gain is claimed only beyond those limits. If {PF}²ES wins on correlated three-objective Marine Design at equal wall-clock, the proposal's claim shrinks to many-objective and neural-surrogate settings, and we will say so.

Why not just use random Tchebycheff scalarizations, which need no hypervolume and scale to any surrogate? Tchebycheff scalarization needs an ideal point and per-objective weights whose meaning depends on scale, so it carries its own reference and its own scale sensitivity. The warp experiment tests this. Linear scalarization also provably misses concave regions (Pareto HyperNetworks). If Tchebycheff Thompson sampling matches us under warps on high-dimensional problems, that is a strong negative result, and it would favor the simpler method.

Why not just estimate the information with nearest neighbors and skip training a critic? The front embedding is high-dimensional, and MINE found Kraskov estimates collapse by 20 dimensions. A nearest-neighbor estimate is also not differentiable in $x$, so it cannot drive gradient refinement of candidates. The Kraskov ablation settles this empirically.

Proposed Experiments and Decisive Checks

The first check is estimator fidelity. On ZDT1, ZDT2, DTLZ2, and DTLZ4 with independent Gaussian processes, where PFES's value is the reference, we compute the rank correlation between our critic's acquisition values and PFES's over shared candidate pools. High correlation is required before anything else counts. Low correlation means the critic is too noisy to rank queries, and the project must fix estimation before claiming benefits.

The second, central check is a controlled warp test. We hold a fitted surrogate fixed and re-express the problem through strictly increasing per-objective maps: scaling one objective by a thousand, applying exponential and logarithmic warps, and inflating a constraint-violation objective. We record which query each method would pick. The prediction is that our choice is unchanged, while expected hypervolume improvement's choice and Tchebycheff's choice move. We then repeat the test end to end with refitting, on ZDT and DTLZ problems under the same warps and on Disc Brake Design, Four Bar Truss, Marine Design, and Rocket Injector. Quality is scored by inverted generational distance to a dense reference front, a measure that needs no reference point. Hypervolume is reported at several reference points, so that no single choice of corner favors any method.

The third check covers the regimes the analytic methods cannot reach. These are DTLZ2 and DTLZ5 with five and eight objectives, correlated-output variants of BraninCurrin built by mixing objectives, and the 1,119-candidate, 2,185-feature LLTO materials set from PFES using deep-ensemble surrogates. Batch sizes go up to sixteen, against {PF}²ES's two to four.

Wall-clock per iteration is reported throughout against the 75.9-second qEHVI figure on DTLZ2. The claim is supported only if quality holds under warps, beats {PF}²ES in the many-objective and correlated regimes, and stays within noise of it elsewhere, all at comparable cost.