Abstract
In 5G small cell deployments, Inter-Cell Interference (ICI) significantly degrades network performance when adjacent cells share the same frequency spectrum. While the Open RAN (O-RAN) architecture introduces the Near-Real-Time RAN Intelligent Controller (Near-RT RIC) to manage such resources, a critical time-scale mismatch exists between the RIC's control loop and the real-time operations of Distributed Units (DUs). This paper introduces xDiff, a novel xApp that leverages a Conditional Denoising Diffusion Probabilistic Model (DDPM) integrated into a reinforcement learning (RL) framework. xDiff generates "soft" preference values for resource allocation, allowing DUs to maintain local scheduling agility while adhering to global interference management strategies. Evaluated on a real-world 5G O-RAN testbed, xDiff demonstrates significant improvements in throughput and queueing delay over traditional RL and rule-based methods.
Introduction
The deployment of 5G small cells often utilizes a frequency reuse factor of one to maximize spectrum efficiency. However, this leads to severe Inter-Cell Interference (ICI), particularly at cell edges. The O-RAN architecture addresses this through a disaggregated approach, where a Near-RT RIC hosts third-party applications (xApps) to optimize radio resource management.
A primary challenge in this architecture is the time-scale mismatch: the Near-RT RIC operates on a scale of 10ms to 1s, whereas the DUs handle radio hardware at a 1ms scale. This latency prevents real-time coordination between DUs. Traditional solutions using "hard" binary constraints often fail to account for rapid channel fluctuations, leading to sub-optimal performance.
The xDiff Framework
To overcome these challenges, xDiff employs a generative approach to policy creation. Instead of discriminative models that output a single "best" action, xDiff uses a diffusion process to capture the complex, multi-modal distribution of effective resource allocation policies.
Diffusion-based Reinforcement Learning
At the heart of xDiff is a Conditional DDPM. The model starts with Gaussian noise and iteratively "denoises" it to produce a high-quality policy. This policy is expressed as Preference Values—continuous weights between -1 and 1 assigned to each Resource Block (RB). These values act as "soft" guidance for the DU's local scheduler, preserving its ability to perform fast, local optimizations like Proportional Fairness while following the RIC's global coordination.
Joint Training and Stability
The framework is trained using a dual-objective loss function. It minimizes denoising loss to learn the policy distribution and maximizes a Q-value to ensure performance gains. To prevent overestimation bias, a Double Q-Learning approach is implemented with two separate critic networks to evaluate the diffusion model's output.
Experimental Evaluation
The performance of xDiff was validated on a real-world 5G O-RAN testbed using OpenAirInterface (OAI) and the Mosaic5g Flexric platform. The setup included three cells and 10 commercial smartphones (e.g., Google Pixel, OnePlus).
Key findings include:
- Performance Gains: xDiff significantly outperformed standard RL methods like DDQN and DDPG. It maintained stable connections in high-interference scenarios where other methods caused UE disconnections.
- Latency Compliance: The inference time for xDiff is approximately 21.8 ms, fitting well within the Near-RT RIC's operational window (10ms - 1s).
- Throughput and Delay: xDiff achieved higher aggregate throughput and lower queueing delays compared to cell-independent resource allocation, effectively mitigating the impact of ICI.
Conclusion
xDiff demonstrates the potential of generative diffusion models for complex network optimization tasks in 5G O-RAN. By bridging the time-scale gap with "soft" preference values and leveraging the expressive power of diffusion processes, it provides a robust solution for inter-cell interference management. Future work will focus on scaling the framework to larger deployments and adapting to high-mobility scenarios.
Summary updated