Simulation-Optimization of Systems of Optimizers: Exploiting the Inner Optimization's Geometry

What the math gives to ML

The paper’s key transferable asset is the critical-region geometry of parametric linear programs: away from measure-zero basis changes, an LP solution is an affine function of its state and parameters, and its dual variables identify the local sensitivity. This gives a practical way to differentiate through repeated optimization decisions inside a neural network without unrolling an iterative LP solver or using noisy finite differences. The same pathwise sensitivity can be propagated through stochastic state dynamics, while basis changes provide an explicit diagnostic for when gradients become unreliable. The strongest ML application is a basis-aware differentiable optimization layer for policies whose outputs parameterize costs, capacities, or right-hand sides of embedded LPs.