Parallel predictive entropy search for multi-objective Bayesian optimization with constraints
Algorithms AIMech:ParallelPredictiveEntropySearchForMultiObjectiveOptimizationWithConstraints DRAFT
A batch Bayesian optimization algorithm for multiple objectives and unknown constraints that jointly selects parallel evaluation locations using approximate information gain about the feasible Pareto set, computed from sampled Pareto representatives and Gaussian EP conditional batch covariance log determinants.
Identity Notes & Concept Scope
PPESMOC is a joint-batch information construction for a constrained Pareto set, retaining batch covariance terms within its stated approximations. It is not independent repetition or summation of the single-query PESMOC acquisition. It also differs from single-objective PPES and from PESC-F, which schedules inexpensive acquisition updates rather than defining this joint multi-objective batch information target.
Claims & Grounded Evidence
6 Claims
Mechanism
#joint-batch-feasible-pareto-information
PPESMOC selects a batch X of B input locations for parallel evaluations of multiple objectives and unknown constraints. In the source minimization convention, feasibility requires c_j(x) >= 0 for every constraint, and dominance means no worse in every objective and strictly better in at least one. Its ideal target is I(X_star; Y_X | D), information about the feasible Pareto set in input space from the joint observation matrix of the whole batch. Assume a well-defined posterior over attained, nonempty feasible Pareto sets, explicit tie handling and finite mutual information. By mutual-information symmetry, this equals joint predictive entropy minus its expected value conditional on the Pareto set; the practical acquisition approximates this target.
Mechanism
#within-task-batch-covariances
Model each objective and constraint with an independent Gaussian process and Gaussian observation noise. At fixed hyperparameters, each task provides a B-dimensional predictive distribution with a full covariance matrix across the batch. Independence across tasks does not mean independence across batch locations. The ordinary joint entropy is a sum of taskwise Gaussian log determinants, with observation-noise covariance included. Conditioning on a shared feasible Pareto set generally couples tasks; the subsequent EP approximation discards that cross-task dependence while retaining within-task batch covariance.
Mechanism
#finite-sampled-feasible-pareto-sets
Approximate the expectation over feasible Pareto sets with samples constructed from random-feature approximations to the objective and constraint GP posteriors. Solve each sampled constrained multi-objective problem approximately and retain a finite representation of its feasible nondominated locations. The journal manuscript describes a grid of 1000 times the input dimension and at most 50 retained Pareto points; these are reported approximation choices, not exact whole-domain sampling. A failed finite search and a posterior draw with no feasible solution are distinct cases; the inspected method does not specify a general acquisition policy for either.
Mechanism
#reused-ep-sites-and-batch-updates
Restrict the conditioning calculation to previous evaluation locations, the sampled Pareto representatives and all candidate batch locations. Use feasibility factors for Pareto representatives and factors coupling competitor feasibility with objective comparisons. EP projects these onto univariate constraint factors and bivariate factors within each objective for pairs of locations. Reuse the fitted sites for observations and Pareto representatives, and refine candidate-batch sites once for each acquisition evaluation. The resulting proper Gaussian approximation supplies each task’s conditional batch covariance. Finite sites and taskwise projection do not implement exact feasible Pareto-set conditioning.
Mechanism
#joint-log-determinant-acquisition
For positive-definite ordinary and approximate conditional observation covariances, score a batch by one half of the sum over tasks of log det Sigma_task(X), minus one half of the Monte Carlo average of the corresponding sums of log det Sigma_task(X | X_star_sample). Include each task’s observation noise in both covariance terms. These are B-by-B determinants, not sums of independent single-location acquisitions; at B = 1 they reduce to log-variance terms. Permuting batch locations and all corresponding matrix indices leaves this score unchanged. The printed Eq16 omits the common one half, preserving its uncosted maximizers but doubling Gaussian information units. Task additivity is an approximation and does not make location scores additive.
Mechanism
#joint-numerical-batch-selection
Optimize the approximate acquisition jointly over the B times input dimension coordinates of the batch. The reported procedure uses automatic differentiation and L-BFGS, with previously fitted EP sites reused and candidate sites refined once. These are derivatives of the chosen numerical approximation, not certified derivatives of exact mutual information, and local optimization does not guarantee a global best batch. This joint construction differs from repeatedly applying single-query PESMOC with hallucinated pending outcomes. The paper also describes task-specific batch scores for decoupled evaluation but explicitly leaves their empirical assessment for future work.
⚠️ Curated Limitations & Negative Results
- A bounded input domain alone does not ensure an attained, nonempty measurable feasible Pareto target. Distinguish absence of an observed feasible point, failure of a finite feasible-point search and an empty feasible region in a posterior function draw. The inspected primary method does not supply a general failure or empty-draw convention. Pareto ties and self comparisons require explicit treatment.
- The journal source is an institutional author manuscript with a cover plus printedpp.1–28, associated with the2023 publication DOI10.1016/j.eswa.2022.119328. The official publisher supplement is separately pinned. The two-author arXivv2 dated July1,2021 is an earlier version; it is not the three-author journal layout. The inaccessible publisher main layout and external author code are not claimed as inspected or executed.
- Main PDFp.13 incorrectly describes the observation vector as K+J dimensional rather than B(K+J); Eq11 on PDFp.14 omits log on constraint determinants, while Eq16 on PDFp.15 omits the common one half and uses k inside constraint sums indexed by j. The descriptive PESMO Eq8 on PDFp.12 is malformed. The record uses the supported Gaussian entropy construction without certifying those literal displays.
- Main Eq12 on PDFp.14 omits the normalization needed for conditional distributions. With Theta(0)=1, Eq13 includes self comparisons and annihilates every nonempty feasible supplied set; removing self comparisons still requires tie handling and does not enforce equality to the complete Pareto set. The nearby prose also confuses non-domination with dominance. These are source caveats, not an asserted author-issued erratum or a complete repaired EP implementation.
- The supplement contains algebraic and notation defects: Eq7 on p.3 gives zero feasibility log-normalizer curvature at zero instead of minus2/pi for unit cavity variance; Eq12 on p.5 reverses the minimizing difference for the declared observation-first ordering; Eqs22 and25 place logarithmic terms outside exponentials; Eqs27–30 on p.6 invert a rank-one difference Hessian product; Eq35 on p.8 has inconsistent N versus N+M block bounds; p.9 describes a damping loop ending with non-positive definite covariance. Literal execution of these equations is not validated, and convergence is not guaranteed.
- Random features, finite feasible searches and Pareto representations, EP projection, one candidate-site pass, local acquisition optimization and sampled hyperparameters introduce approximation. Averaging fixed-hyperparameter information scores is not generally entropy of the predictive mixture. Duplicate noiseless latent evaluations can make covariance determinants singular; independent observation noise can make repeated evaluations informative. The source’s plotted diversity behavior is not a universal prohibition on repeated locations.
- The original bounded checks in research/discovery/ppesmoc-boundaries.py concern rational covariance matrices, finite information channels and selected source-formula counterexamples. They are not GP training, full EP inference, author-code execution, a numerical benchmark or independent scientific approval. No universal runtime, approximation-accuracy, convergence, optimal-scheduling or superiority claim is made.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ActiveLearning
Active Learning
|
related_match | PPESMOC adaptively queries informative batches of objective and constraint evaluations. This relates to querying an information source, but constrained multi-objective optimization is not identical to the pinned task of labeling data for predictive learning. This qualified curator mapping asserts neither exact identity nor an upstream subclass relation. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
garrido2023-ppesmoc
|
Eduardo C. Garrido-Merchán, Daniel Fernández-Sánchez, Daniel Hernández-Lobato
|
primary paper
full_text
|
sha256:eeb6e41bf7ae…
|
garrido2023-ppesmoc-supplement
|
Eduardo C. Garrido-Merchán, Daniel Fernández-Sánchez, Daniel Hernández-Lobato
|
primary paper
full_text
|
sha256:10ca9c713d68…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-15T04:14:51Z
Updated: 2026-10-15T04:14:51Z