PaperlayerView exact revision on arXiv

Source-linked AI summary

Programmable and Customized Intelligence for Traffic Steering in 5G Networks Using Open RAN Architectures

Andrea Lacava, Michele Polese, Rajarajan Sivaraj, Rahul Soundrarajan, Bhawani Shanker Bhati, Tarunjeet Singh, Tommaso Zugno, Francesca Cuomo, Tommaso Melodia

arXiv:2209.14171v3cs.NIcs.AI

TL;DR

The paper addresses the lack of a sandbox for developing data-driven xApps and collecting large-scale training data, alongside the need for user-level RAN control. It connects a near-RT RIC to an ns-3 RAN simulation and develops an RL traffic-steering xApp; evaluations report 30%–50% improvements across key performance indicators.

  • Problem

    The research community lacks a sandbox for building data-driven xApps and collecting large-scale datasets, while traditional RAN architectures do not support fine-grained UE-level control.

    Source: §1 · p. 2

  • Method

    The paper connects a near-RT RIC to an ns-3 simulated RAN and uses an RL traffic-steering xApp to select target cells for individual UEs.

    Source: §1 · p. 2, §4.2 · p. 7, §6 · p. 12

  • Results

    30%–50% improvements across UE throughput, spectral efficiency, and mobility overhead were reported for the large-scale evaluation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    UE-level control through the near-RT RIC supports throughput maximization by selecting the serving NR cell in an EN-DC setup.

    Source: §6 · p. 12

    The authors identify a possible reward trade-off: logarithmic throughput gain can outweigh handover cost, increasing mobility overhead H_u versus all baselines.

    Source: §5.2 · p. 12

Abstract

from arXiv · show

5G and beyond mobile networks will support heterogeneous use cases at an unprecedented scale, thus demanding automated control and optimization of network functionalities customized to the needs of individual users. Such fine-grained control of the Radio Access Network (RAN) is not possible with the current cellular architecture. To fill this gap, the Open RAN paradigm and its specification introduce an open architecture with abstractions that enable closed-loop control and provide data-driven, and intelligent optimization of the RAN at the user level. This is obtained through custom RAN control applications (i.e., xApps) deployed on near-real-time RAN Intelligent Controller (near-RT RIC) at the edge of the network. Despite these premises, as of today the research community lacks a sandbox to build data-driven xApps, and create large-scale datasets for effective AI training. In this paper, we address this by introducing ns-O-RAN, a software framework that integrates a real-world, production-grade near-RT RIC with a 3GPP-based simulated environment on ns-3, enabling the development of xApps and automated large-scale data collection and testing of Deep Reinforcement Learning-driven control policies for the optimization at the user-level. In addition, we propose the first user-specific O-RAN Traffic Steering (TS) intelligent handover framework. It uses Random Ensemble Mixture, combined with a state-of-the-art Convolutional Neural Network architecture, to optimally assign a serving base station to each user in the network. Our TS xApp, trained with more than 40 million data points collected by ns-O-RAN, runs on the near-RT RIC and controls its base stations. We evaluate the performance on a large-scale deployment, showing that the xApp-based handover improves throughput and spectral efficiency by an average of 50% over traditional handover heuristics, with less mobility overhead.

1 INTRODUCTION

The introduction argues that O-RAN enables user-level RAN control, but effective xApps need large-scale data and safe testing. It presents ns-O-RAN and a DRL traffic-steering xApp, reporting performance gains over traditional handover heuristics.

1 INTRODUCTION

  • Traditional handover heuristics use channel-quality and load thresholds, while RIC-based data-driven control can use centralized information to target each UE's QoS requirements.Traffic steering manages individual UEs' mobility; use cases such as eMBB and URLLC have different handover requirements.Source: §1 · p. 2
  • The paper also builds a standard-compliant near-RT RIC platform with O-RAN open interfaces and service models.The introduction frames O-RAN as an open, virtualized, programmable alternative to monolithic RAN functions that are difficult to reconfigure on demand.Source: §1 · pp. 1–2
  • ns-O-RAN bridges large-scale ns-3 5G simulations with a real-world near-RT RIC, supporting data collection, xApp development, and closed-loop testing without large-scale deployments.It streams simulated RAN data to the RIC and returns control actions to the simulation; the same xApps can subsequently be tested on a real RAN without additional development effort.Source: §1 · p. 2
  • The paper proposes a UE-level traffic-steering xApp that models the problem as an MDP and uses CQL, REM, and a custom CNN with a single RL agent for multi-UE control.The introduction identifies the need for large-scale training data and an isolated environment for testing and online exploration, since poorly trained solutions on commercial RANs may degrade service.Source: §1 · p. 2
  • 30% to 50% improvements across UE throughput, spectral efficiency, and mobility-overhead KPIs are reported for the xApp on a large-scale ns-O-RAN deployment.The evaluation uses a large-scale RAN network created by ns-O-RAN.Source: §1 · p. 2

2 BACKGROUND

The background introduces O-RAN architecture and RIC-based intelligence, then situates traffic steering in dual-connectivity networks and describes ns-3 as a 5G simulation platform. It contrasts the paper's practical O-RAN integration with prior approaches.

2.1 O-RAN Cellular Architecture

  • O-RAN splits gNB functions across CU-CP, CU-UP, DU, and RU nodes; the near-RT RIC hosts xApps that make RRM decisions over E2 while E2 nodes execute procedures.KPM carries RAN measurements, including UE-level data, and RAN Control supports RIC control actions.Source: §2.1 · p. 3

2.2 Intelligence in the RIC

  • The RIC's centralized data supports large datasets for data-driven RRM, while reinforcement learning maps environment states to actions to maximize long-term reward.The paper focuses on DQN, whose Q-value estimates expected reward for an action at a given state and whose learned policy can be refined online.Source: §2.2 · p. 3

2.3 Dual connectivity and traffic steering

  • In EN-DC, a UE connects to LTE and NR nodes; traffic steering manages per-UE primary-cell handover and dual-connectivity node and secondary-cell choices.Existing approaches generally do not optimize individual UEs; the paper uses a broader set of RAN KPMs than the per-UE RSRP input described for.Source: §2.3 · pp. 3–4
  • The paper presents its standard-compliant near-RT RIC and xApps as practically deployable in O-RAN and 3GPP networks, unlike a prior simulator that lacked 3GPP stochastic channels.The authors also distinguish their focus on xApp performance from work centered on O-RAN capabilities and orchestration.Source: §2.3 · p. 4

2.4 5G and AI in ns-3

  • The ns-3 5G module supports large-scale simulations with 3GPP channel models, while ns3-gym and ns3-ai enable ML interaction but not production-ready O-RAN xApp development.The paper uses the 5G and mmWave module, which models channel dynamics, antenna patterns, and beamforming.Source: §2.4 · p. 4

3 SYSTEM DESIGN AND ARCHITECTURE

The system design connects a near-RT RIC to a simulated 5G NSA RAN and places a traffic-steering xApp in the RIC. ns-O-RAN links the RIC and ns-3 so simulated telemetry can support xApp development and control testing.

3.1 System Model

  • The system model assumes a 5G NSA EN-DC option 3X network, with LTE eNBs as master nodes, NR gNBs as secondary nodes, and a near-RT RIC controlling UE handovers.E2SM-KPM delivers measurements and E2SM-RC carries handover control actions and UE-specific L3 RRC measurements.Source: §3.1 · p. 5

3.2 Near-RT RIC Software Architecture

  • The RIC platform aggregates KPM data for xApps, which process measurements, infer whether UEs should hand over, and send control actions to the RAN.The platform services manage E2 messaging, subscriptions, KPM collection, and ETL aggregation; the implementation supports streaming 40 UE-, cell-, and node-level KPMs.Source: §3.2 · p. 5

3.3 Connecting O-RAN with ns-3: ns-O-RAN

  • ns-O-RAN connects a real-world, O-RAN-compliant near-RT RIC to ns-3 for large-scale RAN KPM collection and closed-loop testing of simulated networks.It extends e2sim with multiple independent E2 endpoints and uses timestamps based on simulation elapsed time plus a baseline Unix time to synchronize with the RIC.Source: §3.3 · pp. 5–6
  • The ns-3 environment provides 3GPP-compliant propagation and fading models and a full-stack 5G EN-DC model for testing xApps across RAN setups.The RIC used in the framework can communicate with real hardware, so xApps developed and tested with ns-O-RAN can subsequently be activated on real-world RANs.Source: §3.3 · p. 6

4 TRAFFIC STEERING OPTIMIZATION

The traffic-steering xApp formulates UE handover as a throughput-utility optimization constrained by handover cost, then uses data-driven reinforcement learning to choose target cells. Its offline-learning design combines REM and CQL with a CNN Q-function approximator.

4 TRAFFIC STEERING OPTIMIZATION

  • Up to 50% gains in throughput and spectral efficiency are reported for the CQL-based optimization.The paper says these gains follow from embedding the optimization in a data-driven RL framework; it does not give a baseline comparison in this passage.Source: §4 · p. 6

4.1 Problem Formulation

  • The objective maximizes weighted cumulative log-throughput across UEs and time while constraining UE-specific handover costs to limit frequent handovers.The handover cost decays exponentially with time since the UE’s previous handover, and must stay below a predefined threshold.Source: §4.1 · pp. 6–7

4.2 Algorithm Design

  • The MDP agent uses per-UE RAN measurements and serving or neighboring-cell information as state, with missing parameters recoverable from a configurable recent-history window.State data includes UE-specific measurements and cell-level utilization and transmission metrics.Source: §4.2 · p. 7
  • The action space allows a handover to a candidate cell or no handover, and reward is log-throughput improvement minus handover cost.Reward is positive when throughput improvement exceeds the cost and negative otherwise.Source: §4.2 · p. 7
  • A CNN approximates the Q-function, estimating a Q-value for each state-action pair and allowing one set of weights to be used across cells.This replaces a Q-value table with learned CNN weights.Source: §4.2 · p. 8
  • Offline DQN training uses REM to combine Q-value estimates and CQL to counter overestimation for actions outside the offline dataset’s behavior distribution.The trained model is then deployed for online interaction and iterative Q-function updates.Source: §4.2 · p. 9

5 PERFORMANCE EVALUATION

The evaluation tests the TS xApp in dense-urban 850 MHz and C-band deployments with varied users and traffic, then compares it with handover baselines. The xApp improves throughput and spectral efficiency, generally without increasing mobility overhead, and provides higher cell-edge throughput than the agent.

5.1 Simulation and Scenario Design

  • The dense-urban scenario uses one eNB and seven gNBs, with 850 MHz and 3.5 GHz configurations, and evaluates randomly moving dual-connected users under four traffic models.The baselines include a 3 dB SINR-threshold heuristic, fixed-TTT SON1, and dynamic-TTT SON2.Source: §5.1 · pp. 9–10, Figure 5 · p. 10
  • The study collects data from more than 2,000 simulations and performs 40 million offline training steps before online evaluation.The training steps are organized as 400 iterations of 100K steps each.Source: §5.1 · pp. 10–11

5.2 Results

  • 30% improvement at the 10th-percentile user throughput and around 60% at the 95th percentile are reported for RIC RL.The results also report 52% and 35% improvements in spectral efficiency, respectively.Source: §5.2 · p. 11, Figure 7 · p. 11, Figure 8 · p. 11
  • RIC RL generally does not increase mobility overhead over the baselines, except at 42 users, where it triggers more handovers.The paper identifies the throughput-gain reward compensating for handover cost as one possible reason for the exception.Source: §5.2 · pp. 11–12, Figure 8 · p. 11
  • The RIC RL-to-SON2 performance ratio shows improvement in most KPMs across 850 MHz and C-band, while average PRB utilization falls below the baseline.Fig. 9 defines ratios greater than 1 as better performance than SON2.Source: §5.2 · p. 12, Figure 9 · p. 12
  • RIC-enabled RL provides consistently higher cell-edge user throughput than the handover-control xApp from in 850 MHz and C-Band deployments, across different numbers of users.Fig. 10 compares 10th-percentile user throughput.Source: §5.2 · p. 12, Figure 10 · p. 12

6 CONCLUSIONS

The paper presents an O-RAN-compliant framework and a REM- and CQL-based xApp for traffic steering in 3GPP networks. Across multiple frequency bands and UE combinations, the solution improves average throughput and spectral efficiency by 30% to 50%.

6 CONCLUSIONS

  • 30% to 50% improvements in average throughput and spectral efficiency demonstrate performance gains across different combinations of UEs.The agent was profiled on a large-scale deployment in multiple frequency bands.Source: §6 · p. 12
  • The cloud-native near-RT RIC connects through open O-RAN interfaces to a simulated RAN environment in ns-3.The framework supports traffic-steering optimization in 3GPP networks.Source: §6 · p. 12
  • The custom xApp uses REM and CQL for data-driven handover control, selecting the NR serving cell in an EN-DC setup to maximize throughput.Source: §6 · p. 12

Related papers

Loading 2209.14171v3…