Evaluating GAN-enhanced deep deterministic policy gradient for semantic-aware resource allocation in 6G network slicing

Abstract This study evaluates semantic prioritisation and conditional generative replay for continuous bandwidth allocation in a simulated five-slice network. The framework combines a DDPG allocator, a packet-level semantic mechanism, and a conditional WGAN-GP traffic augmentation module. Ten methods are evaluated over ten random seeds: three non-learning heuristics, four DDPG ablations, and TD3, SAC and PPO. All methods use immutable seed-matched training and test workloads and frozen-policy deterministic evaluation. Each DDPG configuration is trained for 2,000 environment interactions; TD3, SAC and PPO were trained for 6,000 interactions and are therefore reported as reference comparisons rather than budget-matched competitors. Relative to base DDPG, the semantic configuration increases delivered spectral efficiency from 1.673 to 1.739 bps/Hz and semantic efficiency from 2.579 to 2.711, while reducing the demand-shortfall latency proxy from 12.03 to 4.44 ms and the unmet-demand ratio from 1.9% to 0.8%. These effects are directionally consistent with large paired effect sizes (Cohen’s d = 1.07, 1.00, − 1.26 and − 1.33), but do not remain statistically significant after Holm correction across the full family of 60 tests (adjusted p = 0.115 to 0.570). Conditional GAN augmentation does not improve DDPG: both cGAN-containing configurations have lower mean delivered spectral efficiency and semantic efficiency and higher shortfall than their non-GAN counterparts. A demand-proportional heuristic achieves the strongest delivered-service results, highlighting the importance of simple baselines and metric selection when evaluating learning-based network-slicing policies. The findings support semantic prioritisation as a promising direction but do not establish a benefit from generative replay at the evaluated training budget.