Researchers presented the CTD method for optimizing multi-criteria tasks

Researchers have proposed a new approach called Converge-then-Diversify (CTD), which splits the process of multi-objective Bayesian optimization (MOBO) into two consecutive stages. The method allows for efficient finding of solutions under limited computational budget conditions, where traditional algorithms often fail to simultaneously search for convergence and diversity. On the first stage, CTD focuses exclusively on achieving convergence, directing the search towards a single point on the Pareto front. On the second stage, the algorithm switches to ensuring diversity, distributing the found solutions along the entire front. This two-stage process allows avoiding premature scattering of resources when working with black boxes and complex functions. The effectiveness of the method was confirmed during large-scale testing, including 446 paired comparisons. According to the results, CTD surpasses modern analogs in 72.9% of cases, demonstrating equivalent performance in 21.1% of situations. Statistical lag was recorded in only 6.1% of experiments. The greatest advantage of CTD is manifested when working with strict limitations on the number of iterations or in tasks with high-dimensional parameter spaces. The authors of the study published the research under the CC BY 4.0 license, providing a detailed description of the mathematical apparatus and benchmark results. The implementation of this method allows optimizing work with expensive computational models that require precise parameter tuning with a minimum number of launches. This reduces the time spent searching for optimal configurations in engineering and scientific tasks, where each call to the target function requires significant resource costs.