Lower Confidence Bound for Preference Selection in Interactive Multi-Objective Optimization

Arash Heidari, Sebastian Rojas Gonzalez, Tom Dhaene, Ivo Couckuyt

Abstract: In multi-objective optimization, the goal is to find the non-dominated or Pareto-optimal set that reveals the optimal trade-offs among the conflicting objectives. Conventionally, the Decision-Maker (DM) selects their preferred solution from this set post-optimization. Evidently, this approach necessitates the computation of numerous non-dominated solutions, incurring high computational costs when the evaluation of the objectives is expensive. To address these challenges, interactive optimization offers a potentially more efficient approach: sequential interactions with the DM during the optimization process, enabling the judicious allocation of the optimization budget by guiding the process toward the most desirable regions of the Pareto front. In this study, we propose to exploit the well-known Lower Confidence Bound acquisition function in Bayesian optimization, to interactively estimate the DM's preferred solution in a data-efficient manner, even in scenarios where the DM experiences uncertainty in their decision-making process.

External IDs:doi:10.1145/3638530.3654317