Iterative State- and Control-Dependent Model Predictive Control: A Jacobian-Free Formulation for Constrained Nonlinear Systems

What the math gives to ML

The paper's transferable asset is an exact, derivative-free pseudo-linearization of nonlinear dynamics, followed by repeated solution of structured linear-quadratic subproblems. Rather than differentiating a learned world model through a long horizon, a neural controller can construct secant-based state- and control-dependent matrices from function evaluations, freeze them along the current predicted trajectory, and solve a constrained quadratic planning problem. The contraction and terminal-cost arguments suggest a practical stopping rule and a way to retain stability when only a few planning iterations are affordable, although the guarantees require local regularity and may not hold for an unconstrained neural model.