[Paper Review] Intelligence and Learning in O-RAN for Data-driven NextG Cellular Networks

Leonardo Bonati, Salvatore D’Oro+3 more|arXiv (Cornell University)|Dec 2, 2020

Advanced MIMO Systems Optimization14 references282 citations

TL;DR

This paper demonstrates the first large-scale, end-to-end integration of data-driven deep reinforcement learning (DRL) agents within an O-RAN-compliant, software-defined cellular network using the Colosseum emulator. By deploying DRL xApps on the near real-time RAN Intelligent Controller (RIC), the system achieves dynamic, closed-loop optimization of network slicing, improving spectral efficiency by up to 20% and reducing buffer occupancy by up to 37% compared to static scheduling policies.

ABSTRACT

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.

Motivation & Objective

• To validate the feasibility of deploying data-driven, closed-loop control in NextG cellular networks using the O-RAN architecture.
• To address challenges in implementing intelligent, real-time RAN control through open interfaces and distributed intelligence placement.
• To demonstrate scalable, topology-independent DRL agents that optimize network slicing across diverse traffic types (eMBB, URLLC, MTC).
• To evaluate the performance of DRL-based scheduling policies against fixed policies (RR, WF, PF) in a realistic, large-scale wireless emulation environment.
• To provide a reproducible, open-source testbed framework for future research in autonomous, self-optimizing 6G networks.

Proposed method

• Deployed O-RAN-compliant software components on Colosseum, a large-scale wireless network emulator, to simulate a full softwarized cellular network with 4 SDR base stations and 40 SDR user equipments.
• Implemented DRL agents as xApps on the near real-time RIC, using Proximal Policy Optimization (PPO) for offline training and online inference.
• Collected real-time performance metrics (e.g., spectral efficiency, buffer size, SINR) via O-RAN E2 interface using RIC Indication messages (72 bytes/s per UE overhead).
• Applied dimensionality reduction via neural network encoders to process variable-sized inputs (e.g., multiple UEs per slice) into fixed-size state representations.
• Mapped DRL actions to scheduling policies (RR, WF, PF) and used slice-specific rewards: throughput maximization for eMBB/MTC, latency minimization for URLLC.
• Trained agents in the non-real-time RIC and deployed them in the near real-time RIC for closed-loop control, enabling real-time adaptation to dynamic network states.

Experimental results

Research questions

• RQ1 Can data-driven, closed-loop control via DRL agents be effectively implemented in an O-RAN-compliant, open-architecture NextG cellular network?

Key findings

• The DRL-based control system achieved up to a 20% improvement in downlink spectral efficiency for the eMBB slice compared to the best static scheduling policy (WF).
• Buffer occupancy in the URLLC slice was reduced by 37% on average with DRL control compared to the RR policy, and by 17% compared to the PF policy.
• DRL agents dynamically selected scheduling policies based on resource availability and network state, with eMBB and MTC agents selecting WF with 99% probability under most conditions.
• The DRL agents demonstrated topology independence, as they controlled a single slice per base station without requiring retraining for different network topologies or UE counts.
• The system achieved low control loop overhead (72 bytes/s per UE) through efficient E2 interface messaging and state encoding, enabling real-time adaptability.
• The experimental results confirm that dynamic, data-driven scheduling outperforms fixed policies across all evaluated KPIs, validating the feasibility of autonomous, self-optimizing RANs.