GAN-Assisted joint computation-communication slice orchestration for 5 G/6 G cloud-edge networks

Abstract

Network slicing and multi-access edge computing (MEC) require coordinated bandwidth and edge-computing allocation under bursty demand. This paper presents Generative Adversarial Network (GAN)-assisted Joint Computation-Communication Slice Orchestration (GAN-JCSO), which combines conditional demand sampling with queue correction and isolation-aware placement. Under controlled heavy bursts, GAN-JCSO achieves an 11.16% request-level deadline-miss rate and 90.56% acceptance, compared with 18.96% and 83.55% for JointHeuristic. The risk-aware GAN-QCVaR-JCSO configuration draws 256 conditional samples and reserves a fixed blend of the 0.90 quantile and 0.90 conditional value-at-risk (CVaR). Quantile LSTM and Quantile Transformer controls predict the same 0.90 quantile, whereas DQN-JCSO learns a discrete reservation-envelope action. At 40% matched over-reservation over ten new paired seeds, GAN-QCVaR-JCSO records the lowest mean miss rate under rare extremes: 13.14%, compared with 14.81%, 14.04%, and 16.30% for LSTM, Transformer, and DQN, respectively. It also records the lowest mean miss rate under the burst-intensity setting of 2.6: 12.49%, compared with 13.08%, 12.96%, and 13.70%. The rare-extreme GAN–DQN contrast remains significant after Holm correction (adjusted \(p=0.031\) , raw \(p=5.61\times10^{-4}\) ), whereas DQN records lower tail latency. These results support a targeted reliability advantage for the distributional controller under matched-resource tail stress.