NEWS · RESEARCH · #162
Converge-Then-Diversify (CTD): Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation
An arXiv cs.AI paper proposes Converge-Then-Diversify (CTD), a two-stage approach to multi-objective Bayesian optimisation that first focuses on rapid convergence to a point on the Pareto front and then focuses on spreading solutions to improve diversity. The authors present two simple CTD instantiations using common acquisition functions and report that across 446 pairwise comparisons CTD statistically outperforms state-of-the-art methods in 72.9% of cases, ties in 21.1%, and is worse in 6.1%, with gains especially under very tight evaluation budgets or in high-dimensional problems.
KEY POINTS
- An arXiv cs.AI paper proposes Converge-Then-Diversify (CTD), a two-stage approach to multi-objective Bayesian optimisation that first focuses on rapid convergence to a point on the Pareto front and then focuses on spreading solutions to improve diversity.
- The authors present two simple CTD instantiations using common acquisition functions and report that across 446 pairwise comparisons CTD statistically outperforms state-of-the-art methods in 72.9% of cases, ties in 21.1%, and is worse in 6.1%, with gains especially under very tight evaluation budgets or in high-dimensional problems.
- Decoupling convergence and diversity can substantially improve MOBO performance when evaluations are very limited or problems are high-dimensional, which matters for expensive black-box optimization in ML and engineering.
WHY IT MATTERS
Decoupling convergence and diversity can substantially improve MOBO performance when evaluations are very limited or problems are high-dimensional, which matters for expensive black-box optimization in ML and engineering.