A Simple and Extremely Efficient Predictive Control for Power Converters

What the math gives to ML

The paper offers a concrete computational mechanism for finite-control-set predictive control: algebraically rearrange the one-step cost so that the minimizing switch vector is obtained from precomputed decision maps and a fitted selection law, rather than enumerating all admissible vectors online. The transferable asset is the observation that, for affine dynamics and quadratic tracking costs, each discrete action owns a region of state space defined by pairwise cost inequalities. This can become a low-latency finite-action prediction head for neural controllers, model-based reinforcement learning policies, or learned-dynamics systems, with an explicit testable boundary against exhaustive argmin selection.