AI-Driven Management Framework for Multi-Access Edge Computing on 5G Standalone Non-Public Network Testbed

Achmad Basuki, S.T., MMG., Ph.D.Dosen
•
Dr. Kasyful Amron, S.T., M.Sc.Dosen
•
Widhi Yahya, S.Kom., M.T., M.Sc., Ph.D.Dosen
•
Hubert CendanaMahasiswa
•
Alexander William Sinjaya Mahasiswa
•
Yoga Raditya NalaMahasiswa
•
Ahmad Nafi MubarokMahasiswa
•
Vincentia Melody VivianneMahasiswa

21 Agustus 2026

Brawijaya University is actively advancing telecommunications with its currently ongoing project, "AI-Driven Management Framework for Multi-Access Edge Computing on Standalone Non-Public Network Testbed in Indonesia". Led by a team of experts including Achmad Basuki, Kasyful Amron, Rizal Setya Perdana, and Widhi Yahya, the project aims to integrate Artificial Intelligence (AI) with 5G Radio Access Networks (RAN) to enhance network efficiency and user experience.

A Robust Two-Tier Architecture

At the core of this research is a 5G Standalone Non-Public Network (SNPN) testbed designed specifically for university environments. The infrastructure leverages a cloud-native, two-tier Multi-access Edge Computing (MEC) architecture:

  • RAN-MEC: Distributed computing nodes located closely to the next-generation NodeB (gNB) to support edge applications with low latency.
  • CN-MEC: Centralized computing resources located at the Core Network site.

To simulate real-world conditions, the team uses UERANSIM to emulate user equipment (UE) traffic and task rates. This traffic represents diverse service requirements, such as Ultra-Reliable Low-Latency Communication (uRLLC), enhanced Mobile Broadband (eMBB), and massive Machine Type Communications (mMTC).

Key Research Pillars

The research tackles several complex challenges in orchestrating distributed computing and communication services across the 5G network.

Dynamic Service and VNF Placement

The management plane must dynamically decide where and what AI models and Virtual Network Functions (VNFs) are deployed across the network. To optimize this, the research utilizes a heuristic-based approach, specifically proposing a Resource-aware Best-Fit algorithm with Service Level Objective (SLO) constraints. This heuristic method is highly efficient for online decision-making, offering low computational costs while maintaining placement quality and strict SLO support. The placement controller gathers infrastructure telemetry (like CPU, memory, and latency) via Prometheus and Node Exporter, filters feasible nodes, scores them, and assigns the workload to the best available node through the Kubernetes API.

Adaptive Task Offloading

In the control plane, the network must determine where and how much of a computing task to offload (e.g., to a local RAN-MEC, a neighbor RAN-MEC, or the CN-MEC). To address this dynamically, the research explores two distinct variants:

  • Heuristic Approach: An initial CPU-based task offloading controller uses HAProxy as a traffic steering point, enabling rapid runtime backend switching between MEC nodes based on a heuristic decision function.
  • Deep Reinforcement Learning (DRL) Approach: A more advanced DRL model is utilized to adaptively learn and make offloading decisions that minimize latency and deadline misses while optimizing resource usage across the distributed MEC sites.

ML-Assisted Network Slicing

To ensure diverse services coexist without interference, the project incorporates Machine Learning-assisted dynamic network slicing configuration. This system observes throughput, latency, and drops to update slice configurations and QoS parameters, minimizing SLO violations for mixed traffic across the Open5GS-based core.

MLOps-Based AI Model Deployment and Optimization

The management plane must dynamically manage the lifecycle and deployment of AI models across distributed MEC and core infrastructures to support intelligent 5G services. To optimize this process, the research proposes an MLOps-based framework that integrates automated model training, continuous monitoring, model versioning, and adaptive deployment strategies. The framework collects telemetry data from network infrastructure and AI services, including resource availability, inference latency, and model performance, to enable efficient model placement and runtime optimization. Through the integration of MLOps pipelines with edge computing and local breakout-enabled architectures, AI workloads can be dynamically deployed closer to users, reducing latency, improving resource efficiency, and ensuring reliable AI-driven service delivery in heterogeneous 5G environments.

Global Impact and Future Outlook

This ambitious project establishes a replicable testbed blueprint for the Asia-Pacific region. By providing a secure 5G SNPN for hosting university AI services, the initiative plays a crucial role in upskilling local personnel in edge computing, 5G RAN, and AI-driven network management. The advancements and findings from this testbed will be shared globally, contributing significantly to the worldwide evolution of AI-enhanced telecommunication infrastructures.