Source-linked AI summary
Intelligence and Learning in O-RAN for Data-driven NextG Cellular Networks
Leonardo Bonati, Salvatore D'Oro, Michele Polese, Stefano Basagni, Tommaso Melodia
TL;DR
The paper examines how O-RAN can support data-driven, closed-loop control in NextG networks while addressing challenges across control timescales and deployment locations. It combines architectural analysis with large-scale Colosseum experiments using DRL xApps, demonstrating dynamic scheduling control for co-existing network slices. The results show improved spectral efficiency and lower latency than fixed or static scheduling policies.
Problem
Data-driven O-RAN control faces unresolved challenges involving policy deployment, data access, validation, training, and control across different RAN nodes and timescales.
Method
The paper analyzes O-RAN control-loop architectures and deploys DRL agents as near-real-time RIC xApps on Colosseum using open-source RAN and RIC components.
Results
DRL-based dynamic scheduling outperforms fixed policies, improving eMBB spectral efficiency by up to 20% and reducing URLLC average buffer size by 37% versus RR.
Takeaways & Limitations
Closed-control loops provide a demonstrated foundation for data-driven, autonomous, and self-optimizing cellular networks, with scheduling adapted to network state and traffic requirements.
Abstract
from arXiv · showhide
Next Generation (NextG) cellular networks will be natively cloud-based and built upon programmable, virtualized, and disaggregated architectures. The separation of control functions from the hardware fabric and the introduction of standardized control interfaces will enable the definition of custom closed-control loops, which will ultimately enable embedded intelligence and real-time analytics, thus effectively realizing the vision of autonomous and self-optimizing networks. This article explores the disaggregated network architecture proposed by the O-RAN Alliance as a key enabler of NextG networks. Within this architectural context, we discuss the potential, the challenges, and the limitations of data-driven optimization approaches to network control over different timescales. We also present the first large-scale integration of O-RAN-compliant software components with an open-source full-stack softwarized cellular network. Experiments conducted on Colosseum, the world's largest wireless network emulator, demonstrate closed-loop integration of real-time analytics and control through deep reinforcement learning agents. We also show the feasibility of Radio Access Network (RAN) control through xApps running on the near real-time RAN Intelligent Controller, to optimize the scheduling policies of co-existing network slices, leveraging the O-RAN open interfaces to collect data at the edge of the network.
I. INTRODUCTION
NextG networks are moving toward agile, disaggregated architectures based on softwarization, virtualization, openness, and re-programmability. These changes support access to the physical infrastructure through third-party software applications and open interfaces.
- 5G and 6G networks are expected to replace inflexible, monolithic designs with agile, disaggregated architectures.The architectures are based on softwarization, virtualization, openness, and re-programmability.
- Open interfaces allow third-party software applications to access the entire physical network infrastructure in real time.
A. Disaggregation and Programmability in O-RAN
O-RAN combines disaggregated, multi-vendor network functions with standardized interfaces and RAN Intelligent Controllers. Its RICs support closed-control loops across timescales from milliseconds to thousands of milliseconds.
- O-RAN virtualizes and divides base-station functionalities while using standardized interfaces to control multi-vendor infrastructures.
- The RAN Intelligent Controller provides a centralized abstraction for deploying custom control-plane functions and closed-control loops.
- O-RAN envisions control loops ranging from 1 ms transmission control to thousands of milliseconds for slicing and traffic forecasting.
- In Scenario B, cloud-deployed RICs control edge-deployed CU and DU components through standardized interfaces, while the RU remains at the cell site.
B. Contributions
The paper examines how current O-RAN specifications support data-driven control and demonstrates an O-RAN control loop using open, programmable cellular-network components. It combines architectural analysis, large-scale testbeds, and deep reinforcement learning for network-slice optimization.
- The paper investigates limitations of current O-RAN specifications and challenges in deploying data-driven policies across RAN nodes.
- It identifies large-scale testbeds as sources of heterogeneous datasets needed to develop and validate data-driven cellular-network solutions.
- The study demonstrates an O-RAN data-driven control loop on Colosseum using open-source RAN and RIC components.
- DRL xApps optimize performance metrics for multiple network slices through data-driven closed-control loops.
II. INTELLIGENT WIRELESS ARCHITECTURES
O-RAN provides a practical architecture for embedding data-driven intelligence, but its control loops involve tradeoffs in data handling, latency, placement, and standardization. The paper distinguishes supported timescales and discusses extensions beyond current O-RAN coverage.
- Data-driven control requires data collection, processing, training, model deployment, closed-loop control, and testing.
- Data Handling and Training Procedures: Centralized data handling provides a comprehensive network view but incurs overhead and latency, whereas edge processing reduces latency with fewer data sources.
- Data Handling and Training Procedures: O-RAN ML specifications define interfaces and operational guidance for collecting, distributing, training, validating, and deploying AI/ML models.
- Control Loops: O-RAN currently considers non-real-time and near-real-time loops, while real-time loops remain outside its present scope.
- Control Loops: Sub-ms loops can support scheduling and PHY-layer tasks, but their realization requires device- and RU-level standardization.
A. Non Real-time Control Loop
Non-real-time O-RAN control loops operate at infrastructure scale, coordinating resource orchestration and policy decisions across network elements, but require substantial coordination and orchestration.
- Non-real-time control loops operate on timescales of at least one second and coordinate the non-real-time and near-real-time RICs through A1.
- These loops manage infrastructure-level resource orchestration, affecting thousands of devices through data-driven optimization over multiple network sources.
- Typical decisions include instantiating and orchestrating network slices and selecting or placing pre-trained inference models in near-real-time RICs.
- Decisions can account for computational resources, data availability, and service-level performance requirements.
- Practical actuation is difficult because interactions among the non-real-time RIC and network elements require tight coordination, data collection, and orchestration.
B. Near Real-time Control Loops
Near-real-time control loops connect the near-real-time RIC with gNB central and distributed units to make multi-user decisions using aggregated session data and radio-layer KPIs.
- Near-real-time control loops operate between 10 ms and 1 s between the near-real-time RIC and gNB CUs and DUs.
- Because one near-real-time RIC serves multiple gNBs, these loops can affect up to thousands of UEs.
- Their decisions use user-session aggregated data and MAC/PHY-layer KPIs.
- Machine-learning algorithms are implemented as external applications within this near-real-time control framework.
C. Real-time Control Loops
Real-time control loops target sub-10 ms and sub-ms operations within lower RAN layers, where data-driven methods support scheduling and PHY-parameter decisions but face standardization constraints.
- Real-time O-RAN operations target sub-10 ms or sub-ms timescales and mainly involve interactions among DU elements.
- Deploying machine-learning solutions at the DU is not currently supported, leaving these loops for future O-RAN specification extensions.
- Data-driven lower-layer and device approaches can support scheduling decisions and feedbackless detection of modulation, coding, and interference parameters.
- Device- and RU-level standardization required for sub-ms loops makes practical realization challenging and limits applicability.
III. OPEN WIRELESS DATA FACTORIES
Open wireless testbeds and emulators address the need for diverse, large-scale data to develop and evaluate data-driven cellular control, spanning city-scale platforms, indoor deployments, and Colosseum.
- Effective data-driven closed-control policies require large datasets covering diverse scenarios, traffic patterns, requirements, and user behaviors.
- Operator privacy and competition concerns limit open access to real cellular data, leaving researchers reliant on smaller laboratory datasets that often lack deployment scale and variety.
- PAWR platforms provide city-scale scenarios and data-generation tools across SDR/MIMO, high-capacity low-latency, and aerial wireless use cases.
- Arena supports MIMO, cellular, and IoT studies with up to 64 antennas arranged in an 8 × 8 office-space grid.
- Colosseum combines 128 compute nodes with 128 channel-emulator USRPs to model realistic RF effects, while its edge datacenter supports 900 TB storage and 52TB/s RF-data processing.
- Figure 3 concerns O-RAN integration in Colosseum.
IV. USE CASE: SCHEDULING CONTROL IN SLICED 5G NETWORKS THROUGH THE O-RAN RIC
The use case integrates DRL-based closed-loop scheduling control with an O-RAN near real-time RIC and an open cellular stack, evaluating adaptive policies for eMBB and URLLC slices on Colosseum. DRL control outperforms fixed scheduling policies by adapting decisions to slice requirements, available resources, and network state.
- System and control loop: DRL agents running as xApps on the near real-time RIC select scheduling policies for each RAN slice.The selected policy is signaled to the base station through RIC Control messages over the E2 interface.
- Experimental scenario: The experiment emulates a 5G network with 4 base stations and 40 user equipments across eMBB, URLLC, and MTC slices.The emulated Rome scenario covers 0.11 km^2 and uses a 3 MHz channel bandwidth.
- Results: Up to 20% higher downlink spectral efficiency is achieved for eMBB with DRL than with the best-performing static policy.The comparison uses the CDF of eMBB downlink spectral efficiency across RR, PF, WF, and DRL control.
- Results: 37%, 5%, and 17% smaller average URLLC buffer sizes are obtained with DRL than with RR, WF, and PF, respectively.Because larger buffers imply packets wait longer in queues, the smaller DRL buffer indicates lower latency; DRL also significantly outperforms WF from the 50th to 90th percentiles.
- Results: MTC and eMBB agents select WF with 99% probability, whereas URLLC agents frequently select both PF and WF as PRB availability changes.The distributions indicate that agents adapt scheduling choices to current network state and traffic requirements rather than using one static policy.
V. CONCLUSIONS
The article demonstrates a path toward integrating closed-control loops into cellular networks and evaluates their feasibility using O-RAN and large-scale experimental testbeds. Its results position such loops as a foundation for data-driven, autonomous, and self-optimizing future networks.
- Conclusions: The paper reviews O-RAN enablers, analyzes data-driven control loops across timescales, and deploys a DRL-based O-RAN RIC on Colosseum.The deployment demonstrates how large-scale testbeds can develop and validate data-driven cellular-network algorithms.
- Conclusions: Closed-control loops provide a strong foundation toward future-generation data-driven, autonomous, and self-optimizing cellular networks.This conclusion follows the demonstrated feasibility of integrating closed-loop control in cellular networks.
- Conclusions: The paper frames virtualization, disaggregation, openness, and reprogrammability as key enablers of NextG closed-control loops.These architectural properties support the paper’s O-RAN-based path for embedding data-driven control in cellular networks.