Variable-Step Time-Delay Control for Proactive Aperiodic Spacecraft Attitude Control

What the math gives to ML

The paper offers a nonstandard mechanism for jointly selecting a control action and the next update interval: its sliding-manifold time-delay error is quadratically bounded by the realized timestep, enabling an algebraic variable-step law rather than continuous event monitoring. The transferable asset is a controller-aware scheduler for sampled optimization, where the discrepancy between a current-gradient realization and a delayed or extrapolated gradient serves as a measurable time-delay error. Applied to neural-network training, the next optimizer interval can be enlarged in locally smooth regions and shortened near curvature changes, with explicit interval and step-ratio safeguards. The key falsifiable prediction is that the measured delay error scales approximately as the square of the effective interval and that the scheduler maintains it near a prescribed tolerance.