Conflict Detection in AI-RAN: Efficient Interaction Learning and Autonomous Graph Reconstruction

TL;DR

The paper tackles conflict detection in AI-native mobile networks by decomposing the problem into three stages: interaction learning, graph reconstruction, and conflict identification. It introduces a lightweight two-tower encoder to map parameters and KPI into a shared latent space, enabling efficient learning of cross-entity interactions, and a data-driven sparsemax-based mechanism to autonomously reconstruct conflict graphs without manual tuning. Empirical results show the two-tower approach outperforms GNNs in learning efficacy and conflict-detection speed, while sparsemax dramatically accelerates graph reconstruction and conflict identification. Together, these contributions offer a scalable, autonomous framework for detecting and classifying conflicts in O-RAN-like AI-enabled networks, with potential for real-time deployment and extension to temporal dynamics.

Abstract

Artificial Intelligence (AI)-native mobile networks represent a fundamental step toward 6G, where learning, inference, and decision making are embedded into the Radio Access Network (RAN) itself. In such networks, multiple AI agents optimize the network to achieve distinct and often competing objectives. As such, conflicts become inevitable and have the potential to degrade performance, cause instability, and disrupt service. Current approaches for conflict detection rely on conflict graphs created based on relationships between AI agents, parameters, and Key Performance Indicators (KPIs). Existing works often rely on complex and computationally expensive Graph Neural Networks (GNNs) and depend on manually chosen thresholds to create conflict graphs. In this work, we present the first systematic framework for conflict detection in AI-native mobile networks, propose a two-tower encoder architecture for learning interactions based on data from the RAN, and introduce a data-driven sparsity-based mechanism for autonomously reconstructing conflict graphs without manual fine-tuning.

Figures (6)

• Figure 1: Our proposed conflict detection framework for AI-native mobile networks, where independent AI agents control the RAN to achieve distinct and possibly conflicting objectives.
• Figure 2: Our two-tower encoder architecture, showing the operations for independently encoding parameters and KPI into the latent space. We calculate the similarity of their node embeddings to infer interactions between parameters and KPI.
• Figure 3: Example score matrices capturing interactions between parameters and KPI, highlighting the different metrics and scales associated with correlation (left) and similarity (right).
• Figure 4: Classification performance of different models in terms of accuracy and AUCAUC across 100 independent training batches.
• Figure 5: Graph reconstruction and conflict identification performance using the interactions learned from different models in terms of F1 score across 100 independent training batches.