A generative adversarial network framework for joint computation and communication slice orchestration reduces request-level deadline-miss rates in cloud-edge networks subjected to heavy traffic bursts. Researchers Yuanhe Qiu and Xihui Zhang developed the system, designated GAN-JCSO, to coordinate multi-access edge computing and bandwidth allocations during unpredictable spikes in demand across fifth-generation and sixth-generation network infrastructure.
The proposed framework pairs conditional demand sampling with queue correction and isolation-aware placement routines to schedule incoming tasks without violating slice isolation boundaries.
Under controlled heavy-burst evaluation scenarios, the standard GAN-JCSO framework achieved an 11.16 percent request-level deadline-miss rate while maintaining a 90.56 percent request acceptance rate. In comparison, a baseline JointHeuristic allocation strategy recorded a higher deadline-miss rate of 18.96 percent and accepted 83.55 percent of incoming workload requests under identical testing conditions.
To manage extreme fluctuations, the researchers configured a risk-aware variant designated GAN-QCVaR-JCSO. This configuration draws 256 conditional samples and establishes a resource envelope by reserving a fixed blend of the 0.90 quantile and the 0.90 conditional value-at-risk, hedging against severe overload states in edge computing nodes.
The risk-aware model was compared against several baseline controllers, including a Quantile Long Short-Term Memory network, a Quantile Transformer, and a discrete reinforcement learning system named DQN-JCSO. The Quantile LSTM and Quantile Transformer controls predicted the same 0.90 quantile target, while DQN-JCSO learned discrete reservation envelopes directly.
Across ten paired random seeds evaluated at a matched 40 percent over-reservation level, GAN-QCVaR-JCSO produced the lowest mean request-deadline miss rate under rare extreme demand events at 13.14 percent. Corresponding miss rates reached 14.81 percent for the LSTM baseline, 14.04 percent for the Transformer model, and 16.30 percent for the DQN approach.
In experiments testing an elevated burst-intensity parameter of 2.6, GAN-QCVaR-JCSO similarly recorded the lowest mean miss rate at 12.49 percent. Comparative miss rates under this burst setting were measured at 13.08 percent for the LSTM baseline, 12.96 percent for the Transformer model, and 13.70 percent for the DQN architecture.
Statistical evaluation confirmed that the rare-extreme performance difference between the GAN-assisted model and DQN-JCSO remained significant following Holm correction, yielding an adjusted p-value of 0.031 from a raw p-value of 5.61×10⁻⁴. However, the evaluation also revealed performance trade-offs, as the DQN controller recorded lower tail latency despite its higher overall rate of missed deadlines.
Subject of Research: Joint computation and communication slice orchestration in cloud-edge networks using generative adversarial networks
Article Title: GAN-Assisted joint computation-communication slice orchestration for 5 G/6 G cloud-edge networks
Article References: Qiu, Y., & Zhang, X. (2026). GAN-Assisted joint computation-communication slice orchestration for 5 G/6 G cloud-edge networks. Journal of Cloud Computing. https://doi.org/10.1186/s13677-026-00985-4
Image Credits: AI Generated
DOI: 10.1186/s13677-026-00985-4
Keywords: network slicing, multi-access edge computing, cloud-edge networks, generative adversarial networks, resource orchestration, conditional value-at-risk, traffic bursts, deadline-miss rate, 5G networks, 6G networks, deep Q-networks, quantile transformer, quantile LSTM, bandwidth allocation, tail latency, queue correction
News Source: Denise Maddox. (November 19, 2026). GAN Slicing Controller Reduces Cloud Edge Network Deadline Misses Under Heavy Bursts. Scienmag.



