ORACLE: Optimizer-Relative Alignment for Constrained LEarning

Name: ORACLE: Optimizer-Relative Alignment for Constrained LEarning
Uploaded: 2026-10-18T05:50:31.399950+00:00
Duration: 420 s
Description: ORACLE introduces optimizer-relative constrained learning by evaluating constraint compatibility after an optimizer update to align its geometry. This method improves or matches native optimizers in 94% of configurations across eight PDE benchmarks, demonstrating a way to ensure decision-preserving interfaces between optimizers and constraints.

ORACLE introduces optimizer-relative constrained learning by evaluating constraint compatibility after an optimizer update to align its geometry. This method improves or matches native optimizers in 94% of configurations across eight PDE benchmarks, demonstrating a way to ensure decision-preserving interfaces between optimizers and constraints.

Open the full episode