Preferred Networks changed the default method for estimating hyperparameter importance—the influence of each setting on optimization results—in Optuna v5 and Rustuna from f-ANOVA to PED-ANOVA. The company says PED-ANOVA is faster, requires no additional dependencies, and supports conditional search spaces and multi-objective optimization. Preferred Networks developed its approach for conditional search spaces and presented it in a paper accepted at KDD 2026.

Users can calculate importance by installing “optuna>=5.0.0” or “rustuna>=0.1.0” and passing a study, including completed trials, to “optuna.importance.get_param_importances.” Optuna can display importance in a graph, but that requires separately installing Plotly or Matplotlib. The estimator’s “target_quantile” and “region_quantile” settings specify how much each parameter matters for reaching higher performance within the top-performing region of observed results.

The performance evaluation used instance 1053 of YAHPO Gym’s rbv2_super scenario. The task selected among six algorithms—approximate k-NN, regularized linear models, random forests, decision trees, support vector machines, and XGBoost. The categorical “learner_id” setting created a conditional search space in which only settings relevant to the selected algorithm were active. The evaluation used N=1000 samples to maximize accuracy, with PED-ANOVA settings of target_quantile=0.1 and region_quantile=1.0. The environment used Python 3.13.11 and Optuna v5.0.0.

Across 10 runs with different random seeds, “trainsize,” “learner_id,” and some settings related to SVM and XGBoost showed relatively high importance. SVM and XGBoost also outperformed the other algorithms on the objective-function values, and the article says PED-ANOVA captured the importance of algorithm selection. The correlation between each algorithm’s best objective-function value and its estimated importance was higher than that of a baseline that simply extended an existing method for conditional search spaces. PED-ANOVA also avoided unstable behavior such as the negative correlation seen with the baseline.

In runtime comparisons across the tested trial counts, Optuna v5’s PED-ANOVA ran about 180–1,600 times faster than f-ANOVA. Rustuna’s PED-ANOVA ran about 2–4 times faster than Optuna’s implementation. These comparisons were based on five runs with different random seeds. Preferred Networks said PED-ANOVA evaluates importance quickly while supporting conditional search spaces and multi-objective optimization. The company also announced an Optuna meetup in Tokyo on October 16, with sessions planned on Optuna 5.0 features, Rustuna, and a talk by creator Takuya Akiba. The article identifies the event day as Friday.

Preferred Networks changed the default method for estimating hyperparameter importance—the influence of each setting on optimization results—in Optuna v5 and Rustuna from f-ANOVA to PED-ANOVA. The company says PED-ANOVA is faster, requires no additional dependencies, and supports conditional search spaces and multi-objective optimization. Preferred Networks developed its approach for conditional search spaces and presented it in a paper accepted at KDD 2026.

Users can calculate importance by installing “optuna>=5.0.0” or “rustuna>=0.1.0” and passing a study, including completed trials, to “optuna.importance.get_param_importances.” Optuna can display importance in a graph, but that requires separately installing Plotly or Matplotlib. The estimator’s “target_quantile” and “region_quantile” settings specify how much each parameter matters for reaching higher performance within the top-performing region of observed results.

The performance evaluation used instance 1053 of YAHPO Gym’s rbv2_super scenario. The task selected among six algorithms—approximate k-NN, regularized linear models, random forests, decision trees, support vector machines, and XGBoost. The categorical “learner_id” setting created a conditional search space in which only settings relevant to the selected algorithm were active. The evaluation used N=1000 samples to maximize accuracy, with PED-ANOVA settings of target_quantile=0.1 and region_quantile=1.0. The environment used Python 3.13.11 and Optuna v5.0.0.

Across 10 runs with different random seeds, “trainsize,” “learner_id,” and some settings related to SVM and XGBoost showed relatively high importance. SVM and XGBoost also outperformed the other algorithms on the objective-function values, and the article says PED-ANOVA captured the importance of algorithm selection. The correlation between each algorithm’s best objective-function value and its estimated importance was higher than that of a baseline that simply extended an existing method for conditional search spaces. PED-ANOVA also avoided unstable behavior such as the negative correlation seen with the baseline.

In runtime comparisons across the tested trial counts, Optuna v5’s PED-ANOVA ran about 180–1,600 times faster than f-ANOVA. Rustuna’s PED-ANOVA ran about 2–4 times faster than Optuna’s implementation. These comparisons were based on five runs with different random seeds. Preferred Networks said PED-ANOVA evaluates importance quickly while supporting conditional search spaces and multi-objective optimization. The company also announced an Optuna meetup in Tokyo on October 16, with sessions planned on Optuna 5.0 features, Rustuna, and a talk by creator Takuya Akiba. The article identifies the event day as Friday.