Accelerating UAV Swarm C2 Optimization via Neural-Enhanced Mixed-Integer Programming

This paper proposes NeuroFix-MIP, a neural-guided optimization framework for adaptive command-and-control (C2) architecture design in heterogeneous UAV swarms. The framework jointly optimizes the spatial deployment of command nodes, the mapping of control relationships, and the configuration of communication links, with the objective of minimizing the overall network latency. The problem is formulated as a large-scale mixed-integer programming (MIP) model that captures the coupling between spatial placement, link selection, and task allocation. Directly solving this model is computationally expensive, especially in dynamic swarm environments. To address this challenge, NeuroFix-MIP introduces a two-stage learning-augmented pipeline: a graph neural network (GNN) is first trained on historical MIP instances to predict the fixation of key decision variables, and the reduced subproblem is then solved by a conventional optimizer to produce a complete feasible solution. Experimental results on multiple UAV swarm scenarios demonstrate that NeuroFix-MIP reduces solving time by more than 85% while maintaining near-optimal latency. These results confirm the effectiveness of learning-based variable fixation in enabling real-time reconfiguration of large-scale heterogeneous C2 networks.