Targeting Evaluation, Computation and Interaction Costs in Expensive Black-Box Multi-Objective Optimization

Arash Heidari, Sebastian Rojas Gonzalez, Ivo Couckuyt

Abstract: Multi-objective optimization algorithms aim to identify a set of non-dominated solutions that capture the complex and often conflicting trade-offs between objectives. In black-box settings, where the input-output relationship is usually analytically intractable, these algorithms face two major challenges: limited evaluation budgets, especially when objective evaluations are costly, and computational difficulties in proposing new candidates. While all non-dominated solutions hold equal mathematical value, the Decision-Maker is typically interested in only the most preferred trade-offs. This adds an additional challenge for preference-driven optimization: the cost of human interaction. To address these challenges, this paper proposes leveraging surrogate-assisted evolutionary multi-objective optimization methods, combined with advanced discretization techniques and strategic filtering of preferred solutions to mitigate the overall computational burden. After evaluating against similar methods in the literature, the experimental results suggest the method to be competitive in enhancing both the efficiency and precision of the decision-making process.

External IDs:doi:10.1145/3712255.3726735