Releases · optuna/optuna · GitHub

This is the release note of v4.5.0. Highlights GPSampler for constrained multi-objective optimization GPSampler is now able to handle multiple objective and constraints simultaneously using the newly introduced constrained LogEHVI acquisition function. The figures below show the difference between GPSampler (LogEHVI, unconstrained) vs GPSampler (constrained LogEHVI, new feature). The 3-dimensional version of the C2DTLZ2 benchmark problem we used is a problem where some areas of the Pareto front of the original DTLZ2 problem are made infeasible by constraints. Therefore, even if constraints are not taken into account, it is possible to obtain the Pareto front. Experimental results show that both LogEHVI and constrained LogEHVI can approximate the Pareto front, but the latter has significantly fewer infeasible solutions, demonstrating its efficiency. Optuna v4.4 (LogEHVI) Optuna v4.5 (Constrained LogEHVI) Significant speedup of TPESampler TPESampler is significantly (about 5x as listed in the table below) faster! It enables a larger number of trials in each study. The speedup was achieved through a series of enhancements in constant factors. The following table shows the speed comparison of TPESampler between v4.4.0 and v4.5.0. The experiments were conducted using multivariate=True on a search space with 3 continuous parameters and 3 numerical discrete parameters. Each row shows the runtime for each number of objectives and each column shows each number of trials to be evaluated. Each runtime is shown along with the standard error over 3 random seeds. The numbers in parentheses represent the speedup factor in comparison to v4.4.0. For example, (5.1x) means the runtime of v4.5.0 is 5.1 times faster than that of v4.4.0. n_objectives/n_trials 500 1000 1500 2000 1 1.4 $\pm$ 0.03 (5.1x) 3.9 $\pm$ 0.07 (5.3x) 7.3 $\pm$ 0.09 (5.4x) 11.9 $\pm$ 0.10 (5.4x) 2 1.8 $\pm$ 0.01 (4.7x) 4.7 $\pm$ 0.02 (4.8x) 8.7 $\pm$ 0.03 (4.8x) 13.9 $\pm$ 0.04 (4.9x) 3 2.0 $\pm$ 0.01 (4.2x) 5.4 $\p