Abstract
Wireless ad hoc networks rely on distributed coordination under stringent energy, mobility, and spectrum constraints, and optimization plays a central role in allocating resources, tuning protocols, and adapting to environmental variation. In large-scale deployments with heterogeneous nodes, centralized solvers are often impractical due to the cost of broadcasting global state and the fragility of single points of failure. Nonconvexity arises naturally from interference coupling, fading-driven link adaptation, discrete scheduling primitives, and nonlinear protocol parameters. This study examines hybrid swarm–gradient methods that combine exploration-driven population dynamics with descent-based refinement, designed for decentralized operation over time-varying graphs. The approach maintains per-agent positions, velocities, and personal exemplars while tracking local gradients and exchanging compressed messages through gossip. The method tolerates asynchronous clocks, lossy links, and delays by leveraging consensus corrections and mixing properties of the communication graph, and it employs event-triggered communication to moderate overhead. A stability analysis relates swarm parameters, stepsizes, and spectral gaps to descent of a smoothed Lyapunov functional, while characterizing convergence to critical points and escape from strict saddles under mild noise. Under error-bound conditions, the mechanism achieves nonasymptotic stationarity rates compatible with communication budgets typical of low-power radios. Experiments on synthetic and network-inspired objectives illustrate behavior under mobility, quantization, and adversarial perturbations. The results indicate that the hybridization strikes a balance between global search and local refinement in settings where curvature is irregular and global coordination is constrained, without presupposing a shared clock or reliable broadcast.