Optuna v5.0.0 introduces significant default behavior changes

Optuna v5.0.0, as detailed in the project's GitHub release notes, introduces several major updates that change default behaviors. The TPESampler now enables multivariate TPE and the constant liar strategy by default for single-objective optimization, while replacing NSGAIISampler with TPESampler as the default for multi-objective optimization. Additionally, PED-ANOVA becomes the default parameter importance evaluator, replacing f-ANOVA, and GPSampler introduces stable Monte Carlo-based q-batch acquisition functions for parallel optimization. Constrained optimization is now directly supported in the core Trial API, with deprecated constraints_func arguments phased out. Several breaking changes include the removal of deprecated modules and APIs, normalization of trial timestamps to UTC, and adjustments to plotting and storage functionality.

Operators should carefully evaluate the impact of these changes on existing workflows, particularly those relying on the previous default samplers or importance evaluators. The shift to multivariate TPE and constant liar strategies may alter optimization results, while the removal of deprecated modules and APIs could break integrations with older codebases. Before upgrading, it is essential to test the new configurations in a staging environment to ensure compatibility and assess performance improvements. This release reflects a broader trend in optimization frameworks toward default settings that prioritize advanced methodologies and computational efficiency.