Near-RT RAN Intelligent Controller

• Near-RT RIC is a central component in O-RAN that enables sub-second closed-loop control and data-driven optimization through containerized xApps.
• It utilizes standardized A1, E2, and O1 interfaces for policy integration, telemetry, and configuration management across distributed RAN elements.
• Key features include digital twin validation, hierarchical reinforcement learning for policy decomposition, and robust security measures to ensure network stability.

A Near-Real-Time RAN Intelligent Controller (Near-RT RIC) is a central architectural component in Open Radio Access Network (O-RAN) systems. Its principal role is to enable closed-loop, low-latency, data-driven control and optimization of RAN behavior via pluggable applications (“xApps”). The Near-RT RIC operates at timescales between 10 ms and 1 s, sitting logically between the Service and Management Orchestration (SMO) plane and the data-plane elements such as O-Centralized Units (O-CUs) and O-Distributed Units (O-DUs). It exposes standardized open interfaces (A1 for northbound policy/model integration, E2 for southbound real-time control and telemetry, O1 for configuration and management) which allow the orchestration and automation of advanced RAN functionalities including mobility, slicing, collision avoidance, security, and more. The following sections provide a detailed technical analysis of the Near-RT RIC, covering its architectural foundations, control loop mechanics, integration of closed-loop learning, digital twin-based validation, and scaling/transfer mechanisms.

1. Architectural Placement and Interface Specification

The Near-RT RIC is logically situated between the SMO/non-RT RIC and distributed data-plane elements (O-CUs/O-DUs) within the O-RAN architecture (Sun et al., 2023, Lacava et al., 2023). Its function is to execute closed-loop control with sub-second latencies (10 ms–1 s). The platform exposes three principal open interfaces:

Interface	Direction	Function
A1	SMO ⇄ Near-RT RIC	Policy/ML model provisioning, guidance
E2	Near-RT RIC ⇄ O-CU/DU	Telemetry, KPI streaming, control
O1	SMO ⇄ Near-RT RIC	Configuration, FCAPS management

This configuration enables modular deployment and integration of varying control logic via containerized xApps, while enforcing cycle and actuation deadlines to ensure network stability. The E2 interface consists of the E2 Application Protocol (E2AP) for session/subscription management and E2 Service Models (E2SM) for granular C-plane/U-plane/MAC control and KPM (Key Performance Measurement) reporting (Lacava et al., 2023). Closed-loop feedback under the physical constraints of current cloud/edge hardware enables large-scale deployments with E2 per-node bitrates in the sub-kbps to few-kbps range and round-trip latencies well within the 1 s (and often sub-100 ms) envelope.

2. Closed-Loop Control and xApp Mechanics

Within the Near-RT RIC, xApps are implemented as microservices that subscribe to real-time RAN streaming telemetry (typ. KPMs every 10–100 ms) via E2, process these data streams (potentially with embedded AI/ML), and issue control commands back to the O-DU/O-CU on timescales commensurate with observed RAN dynamics (Sun et al., 2023, Xavier et al., 2023). For instance, a Traffic Steering (TS) xApp subscribes to per-cell features including PRB utilization, RSRP/RSSI distributions, and UE counts. It issues E2 control messages to update mobility and offloading parameters such as max residual capacity, target cell headroom, ΔRC, RSRP filter, and A5-event CIO offset.

The control workflow can be formalized as a Markov Decision Process (MDP):

• State space: e.g., $$s_t = [f_\text{serv}, f_\text{tgt}, \text{tod}, \text{dow}, u_\text{serv}, u_\text{tgt}, \text{thr}_\text{serv}, \text{thr}_\text{tgt}] \in \mathbb{R}^8$$
• Action space: $a_t =$ tuple of mobility and steering parameters
• Reward: $r_t =$ spectral efficiency metric, e.g.,

$$r_t = \frac{\text{Total DL bits transferred}}{\text{Total TX time}} \ \text{in interval} \ [t,t+1]$$

Closed-loop inference and actuation must fit within the cumulative delay budget set by E2 transport, data plane, and xApp computation, with empirical end-to-end control loops demonstrated at $\sim$10–50 ms per cycle for modern Near-RT RIC solutions (Lacava et al., 2023).

3. Hierarchical Policy Decomposition and Learning Workflow

To address the scalability and sample complexity challenges intrinsic to high-dimensional state/action RIC tasks, advanced policy architectures leverage factorization and modularization. The Cascade Reinforcement Learning (CaRL) framework, for example, decomposes monolithic policies into ensembles of lightweight sub-policies via state space factorization (Sun et al., 2023):

1. Factorized state space: the full state $s$ is partitioned into $M$ subspaces, each corresponding to distinct traffic regimes (e.g., high-load morning, interference-limited periods).
2. Mixture-of-experts selection: a classifier $C(s;\theta_C) \rightarrow [c_1, ..., c_M]$ distributes control responsibility across sub-policies with weighted gating, such that the final action is $a = \sum_{i=1}^M c_i a_i$.
3. Policy update (AWAC): all parameters (classifier and sub-policy weights) are updated jointly according to an advantage-weighted objective,

$$\max_{\theta_C, \{\theta_i\}} \mathbb{E}_{(s,a) \sim \beta}\left[ \log \pi(a|s) \exp\left(\frac{1}{\lambda}A^Q(s,a)\right) \right]$$

This architecture reduces memory/schedule complexity and enhances transferability to new network regions, as federated-style initialization and fine-tuning can rapidly adapt to novel traffic statistics.

4. Digital Twin Validation and Experimental Performance

Data-driven digital twins (DT) play an essential role in validating Near-RT RIC logic prior to deployment. High-fidelity simulation platforms reproduce:

• Actual cell layouts and operator configurations
• Geo-referenced UE mobility distributions
• Time-varying, class-based traffic models
• Detailed PHY abstractions, TTI-level MAC, per-QCI PRB allocation
• Full RRC/RLC mobility and handover logic, load balancing

Calibration against real-world operator KPIs (e.g., minute-by-minute downlink volume, RRC-connected UE counts) ensures simulation traceability (demonstrated to within 5–10% error on average) (Sun et al., 2023). Evaluations across multiple clusters have shown that advanced xApp designs (such as CaRL) can increase average cluster-aggregated downlink throughput by 18–24% over business-as-usual baselines, with factorized policies yielding an additional 10–12% relative gain. Fast knowledge transfer capabilities via seeded sub-policy weights yield "few-shot" adaptation to new domains, dramatically lowering cold-start latency.

5. Security, Privacy, and Zero-Trust Implications

The integration of programmable intelligence and third-party xApps in the Near-RT RIC creates new attack vectors and privacy risks. Mitigation architectures include multi-layer defense frameworks combining (Alimohammadi et al., 1 Dec 2025, Lin et al., 2024):

• Signature-based message inspection: validation of structural and semantic integrity of E2/AP signaling.
• Telemetric poisoning detection: LSTM-based detectors identify anomalous KPM streams by predicting expected temporal patterns and classifying large deviations as suspicious.
• Runtime xApp attestation: challenge-response protocols using cryptographically-secure hash functions (e.g., SHA-256) to check binary integrity of running xApp processes at runtime.
• Zero-trust RIC (ZT-RIC): all RAN and UE KPM vectors are encrypted at the source using functional encryption (e.g., Inner Product Functional Encryption, IPFE), allowing xApps to perform inference tasks over ciphertexts without direct access to plaintext data. Such schemes can achieve round-trip times below RIC timing requirements and classification accuracy equivalent to baseline unencrypted approaches.

6. Scalability, Deployment, and Resource Optimization

Disaggregated RIC deployments enable flexible placement of Near-RT RIC components across the cloud-edge continuum, trading off control-loop latency, resource utilization, and OPEX/CAPEX (Almeida et al., 2023). The RIC-O orchestrator formulates a mixed-integer optimization problem, deciding on optimal placement of E2 terminators, xApps, shared data layers, and state databases under per-node CPU/memory/storage constraints and per-E2 node end-to-end control loop latency ($\leq$10 ms for latency-critical xApps). Dynamic triggers (e.g., latency violation or node failure) prompt real-time re-optimization and re-deployment via Open RAN O2 APIs, with open-source implementations on Kubernetes delivering sub-10 ms control window restoration while reducing edge resource duplication.

7. Future Directions: Hierarchical Control and Adaptive Orchestration

Next-generation Near-RT RIC platforms are adopting hierarchical, two-tiered control frameworks. The non-RT RIC (operating at $>1$ s) executes global policy synthesis, strategic ML guidance generation (potentially LLM-driven), and provides policy input over A1. The Near-RT RIC (10 ms–1 s) specializes in rapid, localized, data-driven adaptation (Bao et al., 25 Apr 2025). Task formulation as (multi-agent) MDPs, deep RL (e.g., DDPG, PPO, SAC) policy training, and hybrid offline/online pipelines (leveraging federated transfer and domain adaptation) promote robustness in nonstationary environments.

As RAN and AI workloads increasingly converge in shared cloud-native infrastructure, Near-RT RIC frameworks are being extended with resource orchestration capabilities (e.g., through interfaces such as Y1), balancing real-time RAN strictness against opportunistic AI/ML batch workloads (Shah et al., 10 Mar 2025). State-of-the-art solutions integrate graph-convolution and sequence-modeling components for scalability (e.g., in resource slicing xApps) (Yan et al., 17 Sep 2025), and lightweight, state-space sequence predictors to substitute memory/latency-bounded transformer networks (Rezazadeh et al., 6 Oct 2025), aiming for sub-100 ms inference latencies on commodity edge hardware.

---

References:

• (Sun et al., 2023) Cascade Reinforcement Learning with State Space Factorization for O-RAN-based Traffic Steering
• (Lacava et al., 2023) ns-O-RAN: Simulating O-RAN 5G Systems in ns-3
• (Xavier et al., 2023) Machine Learning-based Early Attack Detection Using Open RAN Intelligent Controller
• (Almeida et al., 2023) RIC-O: Efficient placement of a disaggregated and distributed RAN Intelligent Controller with dynamic clustering of radio nodes
• (Alimohammadi et al., 1 Dec 2025) Towards a Multi-Layer Defence Framework for Securing Near-Real-Time Operations in Open RAN
• (Lin et al., 2024) ZT-RIC:A Zero Trust RIC Framework for ensuring data Privacy and Confidentiality in Open RAN
• (Bao et al., 25 Apr 2025) LLM-hRIC: LLM-empowered Hierarchical RAN Intelligent Control for O-RAN
• (Shah et al., 10 Mar 2025) The Interplay of AI-and-RAN: Dynamic Resource Allocation for Converged 6G Platform
• (Yan et al., 17 Sep 2025) Near-Real-Time Resource Slicing for QoS Optimization in 5G O-RAN using Deep Reinforcement Learning
• (Rezazadeh et al., 6 Oct 2025) Rivaling Transformers: Multi-Scale Structured State-Space Mixtures for Agentic 6G O-RAN