Network Topology Reconfiguration: Optimal Transition Planning

What the math gives to ML

The paper contributes a constructive mechanism for moving between network topologies without violating feasibility during the transition: a receding-horizon planner proposes a sequence, while a higher-fidelity feasibility filter rejects intermediate states and reusable combinatorial cuts prevent repeated infeasible topology choices. The transferable asset is the separation between cheap trajectory planning and expensive certification at every intermediate point. In neural networks, this can implement safe pruning, expert activation, layer-width changes, or parameter-efficient adapter switching while preserving accuracy, activation bounds, latency, and numerical stability throughout reconfiguration. The key falsifiable prediction is that filtered transitions eliminate intermediate constraint violations, whereas direct endpoint interpolation exhibits violations concentrated near topology switches.