Multi-objective Bayesian optimization from a common descent direction viewpoint
Abstract
Designed for limited evaluation budgets, multi-objective Bayesian optimization (BO) typically aims at representing the entire Pareto front. With limited budget (or based on decision-maker information), it is more realistic to target specific parts of the Pareto front. One such way is to rely on the multiple gradient descent principle, improving all objectives simultaneously from a given starting solution. This is based on estimating a common direction of descent from the gradients. Another local optimization perspective was taken for moderately large input dimensions, where the combination of BO with a trust-region proved to be efficient, but mostly in the single-objective context. In this vein, we propose a local multi-objective Bayesian optimization framework relying on the multiple gradient descent principle and connect it to the multi-objective probability of descent. We also show that the common descent direction viewpoint can help to define new scalarizations and associated acquisition functions for global optimization. We illustrate and show the interest of these approaches on several examples.
Related
- Not All Objectives Are Born Equal: Priority-Constrained Descent for Hierarchical Multi-Objective Optimization
- Simple Extensions of Single-Objective Acquisition Functions and Hedge Strategies for Multi-Objective Bayesian Optimization
- Portfolio-Based Constrained Multi-Objective Bayesian Optimization for Materials Design
- A Feasible-Direction-Type Method for Nonconvex Optimization with Infeasible Starts
- Structure-aware Bayesian full waveform inversion via diffusion-regularized Stein variational gradient descent
BibTeX
@misc{binois2026multiobjective,
title = {{Multi-objective Bayesian optimization from a common descent direction viewpoint}},
author = {Mickael Binois},
howpublished = {HAL},
year = {2026},
month = sep,
url = {https://inria.hal.science/hal-05771806v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Multi-objective Bayesian optimization from a common descent direction viewpoint
AU - Mickael Binois
PY - 2026
DA - 2026/09/30
PB - HAL
UR - https://inria.hal.science/hal-05771806v1
AB - Designed for limited evaluation budgets, multi-objective Bayesian optimization (BO) typically aims at representing the entire Pareto front. With limited budget (or based on decision-maker information), it is more realistic to target specific parts of the Pareto front. One such way is to rely on the multiple gradient descent principle, improving all objectives simultaneously from a given starting solution. This is based on estimating a common direction of descent from the gradients. Another local optimization perspective was taken for moderately large input dimensions, where the combination of BO with a trust-region proved to be efficient, but mostly in the single-objective context. In this vein, we propose a local multi-objective Bayesian optimization framework relying on the multiple gradient descent principle and connect it to the multi-objective probability of descent. We also show that the common descent direction viewpoint can help to define new scalarizations and associated acquisition functions for global optimization. We illustrate and show the interest of these approaches on several examples.
ER -