Reducing VR Delivery Latency in UAV-Enabled Mobile Edge Computing using PSO, GA, and Gradient-Based Approaches

Virtual reality (VR) services delivered through UAV-enabled mobile edge computing (MEC) require joint decisions on user association, caching, computing-resource allocation, and UAV placement. This study benchmarks three CVX-free placement solvers, namely Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and gradient-based L-BFGS-B, within the same alternating-optimization framework for N ∈ {15, 18, 21, 24, 27, 30} VR users served by M = 2 UAV base stations (UBSs). The channel rate follows the outage-constrained logistic approximation of an elevation-angle-dependent Rician fading model. To account for stochastic variability, the evaluation uses 30 independent Monte Carlo system realizations for every N , paired across solvers, and includes random and centroid placement baselines. The optimized maximum aggregate UBS task-completion latency increases from approximately 225 ms at N = 15 to 723 ms at N = 30. PSO and GA achieve numerically identical latency values, while L-BFGS-B yields comparable latency at each tested user load, differing by at most 0.714 ms in any paired comparison; after Holm correction, only the N = 15 difference is statistically significant and its median magnitude is only 0.054 ms. Relative to random placement, optimization reduces mean aggregate latency by approximately 21–33%, whereas the gain over centroid placement is modest (about 0.26–0.47%). L-BFGS-B has the lowest average runtime over all loads, although PSO is faster at several individual load levels, and GA is consistently the slowest. The reported objective is an aggregate task-level latency metric inherited from the reference UAV-MEC model and must not be interpreted as per-frame motion-to-photon latency