Multi-objective Bayesian optimisation (MOBO) seeks a high‑quality Pareto front under limited evaluation budgets, requiring both convergence (closeness to the front) and diversity (spread across the front). Existing approaches try to achieve the two goals simultaneously, which demands careful acquisition‑function design and often fails when the budget is very tight because insufficient nondominated solutions are generated. To tackle this, we propose a converge‑then‑diversify (CTD) framework that separates convergence and diversity into two stages. The first stage focuses solely on convergence, rapidly pulling the search toward a single point on the Pareto front. The second stage switches to diversity, using acquisition functions to expand around the converged point and populate the entire front. We instantiate CTD with two widely used acquisition functions. Across 446 pairwise benchmark comparisons, CTD statistically outperforms state‑of‑the‑art methods in 72.9% of cases, ties in 21.1%, and is worse in only 6.1%, with the advantage especially pronounced under very tight budgets or in high‑dimensional problems.
Review