User Preference Meets Pareto-Optimality in Multi-Objective Bayesian Optimization

Joshua Hang Sai Ip, Ankush Chakrabarty, Ali Mesbah, Diego RomeresAAAI 2025 · DOI

Abstract

Multi-Objective Bayesian Optimization (MOBO) often struggles to balance mathematical Pareto-optimality with specific user preferences. We present PUB-MOBO (Preference-Utility-Balanced MOBO), a framework that integrates global utility-driven search with local multi-gradient descent. By utilizing a novel Preference-Dominated Utility Function (PDUF) and a gradient-refining acquisition strategy, PUB-MOBO identifies solutions that are both high-utility and Pareto-optimal. Experimental results on synthetic and real-world engineering benchmarks demonstrate superior efficiency and solution quality compared to state-of-the-art methods.

Introduction

In Multi-Objective Optimization (MOO), practitioners frequently encounter conflicting goals, such as maximizing engine performance while minimizing fuel consumption. Traditional methods aim to identify the entire Pareto-front—the set of solutions where no objective can be improved without degrading another. However, for human users, much of this front is irrelevant; they typically possess specific preferences that prioritize certain trade-offs over others.

Existing preference-based Bayesian Optimization (BO) methods utilize pairwise comparisons to learn a latent utility function. A significant drawback of these methods is their tendency to settle for solutions that are "preferred" by the model but are technically dominated, meaning a nearby solution exists that is superior in every objective. Furthermore, mapping the entire Pareto-front is computationally prohibitive and often ignores the user's specific region of interest.

Methodology: The PUB-MOBO Framework

PUB-MOBO bridges the gap between user preference and mathematical optimality through a three-stage iterative process:

  1. Preference Exploration (PE): The system presents the user with two candidate solutions. The user's choice updates a latent utility model, typically represented via Gaussian Processes (GPs).
  2. Outcome Evaluation (EXP): A global search stage utilizes the qEIUU acquisition function to identify new points in the decision space likely to yield high utility.
  3. Local Gradient Descent (GD): This core innovation uses the high-utility point from the EXP stage as a starting point for a local search using the Multiple Gradient Descent Algorithm (MGDA). Since objective functions are often black-boxes, gradients are estimated using the GP surrogate.

Preference-Dominated Utility Function (PDUF)

A critical component of this framework is the PDUF, a novel utility function designed to strictly respect the principle of dominance. Unlike standard distance-based utilities (e.g., $\ell_1$ distance from an ideal point), which may treat dominated and Pareto-optimal points as equally desirable, PDUF ensures that if solution A dominates solution B, A must have a higher utility score.

Comparison between standard $\ell_1$ distance utility and the proposed PDUF, illustrating how PDUF aligns with dominance principles.

Refining Gradient Estimates

In black-box optimization, estimated gradients are prone to high uncertainty. PUB-MOBO addresses this by incorporating a Gradient Information (GI) acquisition function. This step allocates a portion of the evaluation budget to specifically sample points that reduce uncertainty in the gradient direction, allowing the algorithm to accurately "slide" solutions toward the Pareto-optimal front.

Experimental Evaluation

The method was evaluated against state-of-the-art baselines, including EUBO+qEIUU, across various benchmarks:

  • Synthetic Problems: DTLZ1, DTLZ2, and DH1.
  • Real-World Engineering: Vehicle Safety (crashworthiness), Conceptual Marine Design, and Car Side Impact optimization.
Performance comparison on synthetic benchmarks (DTLZ1, DTLZ2, DH1) showing utility regret (R) and distance to the Pareto front ($d_{Pareto}$).

Key Results

  • Superior Efficiency: PUB-MOBO consistently outperformed baselines across all benchmarks, achieving lower utility regret (the difference between the user's true ideal and the found solution) and a significantly smaller distance to the Pareto-front ($d_{Pareto}$).
  • Ablation Insights: The full PUB-MOBO implementation (including GI and local refinement) performed better than versions that skipped gradient refinement, proving that the computational cost of refining gradients is a necessary investment for reaching the true Pareto front.

Conclusion

PUB-MOBO introduces an efficient "shortcut" to high-quality solutions: instead of mapping the entire Pareto-front, it uses user preferences to locate a promising neighborhood and then employs local gradients to ensure the final solution is mathematically optimal. While the local gradient descent stage requires additional evaluations, the resulting gains in solution quality and adherence to user preferences make it a robust framework for complex, multi-objective engineering challenges. Future work may explore scaling these gradient-based techniques to many-objective problems beyond the four-objective scenarios tested here.

Summary updated

BibTeX
@inproceedings{ip2025user,
  author = {Joshua Hang Sai Ip and Ankush Chakrabarty and Ali Mesbah and Diego Romeres},
  title = {{User Preference Meets Pareto-Optimality in Multi-Objective Bayesian Optimization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i19.34230},
}
Figure