Optimizer relative alignment improves constrained learning for complex problems
ORACLE: Optimizer-Relative Alignment for Constrained LEarning
Machine Learning
Summary
Sometimes, when teaching a computer to solve problems, extra rules (constraints) make training tricky. The authors found that common methods try to satisfy these rules before the computer makes an update, but the way the computer actually updates might change things. They created a method called ORACLE that checks and adjusts rule satisfaction after the computer’s update step, fitting it precisely to the computer’s own way of updating. Their tests on scientific problems and different update methods show ORACLE often helps or matches existing methods better than other approaches.
What this means in practice
- •For machine learning engineers: Improve training of machine learning models that have to obey constraints in physics-based simulations by aligning constraint enforcement to the optimizer’s update step.
- •For engineering simulation teams: Enhance accuracy and stability in solving partial differential equation benchmarks involving constraints by applying post-update constraint alignment.
Authors
Utkarsh Grover, Wyatt Mackey, Kaixun Hua, J. Morris Chang, Xiaomin Lin
Abstract
Constraint handling methods typically intervene before the optimizer acts, by modifying the objective or the gradient. Yet momentum, adaptive scaling, and structured preconditioning can substantially reshape that signal before it becomes a parameter update. We formulate optimizer relative constrained learning, where constraint compatibility is assessed on the post optimizer update. Building on this view, we introduce ORACLE, which evaluates the native optimizer's realized step through a joint endpoint linearization of heterogeneous constraint families, constructs the resulting alignment in the optimizer's own geometry, bounds its authority, and commits it only after validation. We evaluate ORACLE across eight Partial Differential Equation benchmarks and four optimizers spanning Euclidean, diagonal adaptive, and structured preconditioned geometries, where it improves or matches native optimizer in 94% of configurations. Cross model analysis shows the same behavior in 92% of configurations, while matched comparisons show improvements over alternative constraint-handling methods acting at the objective, gradient, and post-optimizer levels.