Intelligent Network Slicing in 6G Networks Using Generative Reinforcement Learning and Diffusion Models

Abstract

Network slicing in 6G environments necessitates resource orchestration that is adaptive, priority-aware, and capable of managing heterogeneous QoS requirements and dynamic traffic conditions. Conventional reinforcement learning techniques struggle with limited data, delayed convergence, and inadequate generalization when interacting with unforeseen traffic demands. This paper introduces a Generative Reinforcement Learning (GRL) framework integrating a Denoising Diffusion Probabilistic Model (DDPM) to enable priority-aware network slicing in 6G networks. The diffusion model functions as a scenario broker, producing synthetic edge-case traffic demands from real Uni-Cauca network flow traces using an experience mixing coefficient of λ = 0.6. Under a 150% load stress test with per-slice demands of D = [50, 50, 50] MHz, the proposed framework achieves a URLLC SLA satisfaction index of 1.0, meeting the 6G ultra-reliability target, compared to 0.667 for both Vanilla PPO and A2C baselines — a reliability gain of +33.33% over each baseline. Under URLLC-heavy demand (D = [30, 70, 20] MHz), GRL-Diffusion achieves a weighted reward of R = 0.771, compared to 0.760 for both baselines. Under balanced load (D = [50, 50, 50] MHz), the reward gain reaches +0.093 over both baselines, representing the largest improvement across all demand profiles. Over 151 training epochs, the diffusion model converged to a stable MSE loss of 1.019 from an initial peak of 2.72, with a computational overhead of T = 10 diffusion steps per synthetic sample generation. The overall weighted reward improved by 9.33% over both baselines across all evaluated load profiles. These findings confirm that combining priority-weighted allocation with diffusion-based generative modeling produces a slice orchestrator that is reliable, data-efficient, and well-suited for dynamic 6G environments.