AI in 5G RAN
Optimization

Reference Signal Received Power (RSRP) is the single most important measurement in any cellular network. Defined in 3GPP TS 38.215 for NR, RSRP is the linear average of the power contributions (in watts) of the resource elements carrying SS/PBCH block reference signals (SS-RSRP) or CSI reference signals (CSI-RSRP), measured in dBm. Unlike RSSI which measures total wideband power including noise and interference, RSRP isolates the reference signal power per resource element, giving a clean view of the serving cell's signal strength at the UE antenna.

In NR, the UE performs SS-RSRP measurement during beam sweeping across up to 64 SSB beams (for FR2 mmWave) or 4-8 SSB beams (for FR1 sub-6 GHz). Each SSB occupies 4 OFDM symbols and 240 subcarriers (20 RBs). The UE measures RSRP per beam index (SSB index), enabling the gNB to track the best beam pair. This per-beam RSRP is the foundation for beam management procedures defined in 3GPP TS 38.321 (L1-RSRP reporting via MAC CE) and TS 38.331 (L3-RSRP reporting via RRC MeasurementReport).

RSRP Ranges and Their Meaning

How RSRP Measurement Works

SINR (Signal-to-Interference-plus-Noise Ratio) is the definitive quality metric for wireless links. While RSRP tells you how strong the signal is, SINR tells you how usable it is. Defined as SINR = S / (I + N), where S is the desired signal power, I is the total interference power from neighboring cells and other sources, and N is the thermal noise power. SINR is measured in dB and directly determines the maximum modulation and coding scheme (MCS) the link can support.

The thermal noise floor N is calculated as: N = kTB, where k = 1.38 x 10^-23 J/K (Boltzmann constant), T = 290K (standard temperature), and B = channel bandwidth in Hz. For a 100 MHz NR carrier: N = -174 dBm/Hz + 10*log10(100e6) = -174 + 80 = -94 dBm. Adding a UE noise figure of 7 dB gives an effective noise floor of -87 dBm. Any signal below this is undecodable.

SINR to MCS Mapping

SINR Calculation Pipeline

The UE reports Channel Quality Indicator (CQI, 0-15) to the gNB, which maps to an SINR range and corresponding MCS index per 3GPP TS 38.214 Table 5.2.2.1-3. CQI 15 maps to 256QAM with code rate 948/1024 (requiring SINR > 22 dB), while CQI 1 maps to QPSK with code rate 78/1024 (requiring SINR > -6 dB). AI optimizes SINR by: (1) beamforming weight optimization to maximize S and null I, (2) inter-cell interference coordination to schedule interferers on different PRBs, and (3) ML-based CQI prediction to preemptively adapt MCS before quality drops.

This is the most misunderstood relationship in RAN optimization. Engineers often assume that if RSRP is good, the user experience is good. Wrong. A UE can have RSRP = -75 dBm (excellent signal) but SINR = 2 dB (terrible quality) because it sits at the intersection of three co-channel cells. Conversely, a UE at RSRP = -105 dBm (poor signal) can achieve SINR = 18 dB if it is the only cell in the area with zero interference. Understanding this duality is critical for any AI RAN optimization strategy.

The Four Quadrants

Decision Flow

AI systems use joint RSRP-SINR analysis on geo-binned data to classify every 50m pixel in the network into one of these four quadrants. This drives fundamentally different optimization actions. A purely RSRP-based system would miss the interference quadrant entirely, while a purely SINR-based system would not distinguish between dead zones (needing new infrastructure) and interference zones (fixable with parameter tuning).

Handover is the most critical real-time decision in a mobile network. A failed handover means a dropped call, a frozen video stream, or a lost gaming session. In 5G NR, the gNB makes handover decisions based on A3 events (neighbor RSRP becomes offset better than serving) with configurable parameters: A3-Offset (0.5-15 dB), Time-to-Trigger (TTT) (0-5120 ms), and Hysteresis (0-15 dB). Getting these wrong leads to four distinct failure types, each with different root causes and different AI solutions.

Four Types of Handover Failure

Handover Process & Failure Points

Traditionally, finding coverage holes required expensive drive testing: $200-500 per cell site per test, repeated quarterly. For a 10,000-site network, that is $2-5 million per year just for measurement. MDT (Minimization of Drive Tests), defined in 3GPP TS 37.320, changes this by collecting geo-tagged RSRP/RSRQ measurements directly from commercial UEs during normal operation. AI ingests these millions of crowd-sourced samples, builds a predictive coverage model using Random Forest or Gradient Boosting, and identifies coverage gaps automatically.

AI Coverage Optimization Pipeline

Interference is the #1 throughput killer in dense 5G deployments. In a typical urban macro network, 40-60% of cell-edge UEs are interference-limited, not noise-limited. This means adding more power does not help — it only creates more interference for neighbors. AI attacks interference through three vectors: (1) ML-based ICIC that coordinates scheduling across cells to avoid PRB collisions, (2) beamforming optimization using Massive MIMO to steer energy toward UEs and nulls toward interferers, and (3) adaptive muting where cells dynamically blank specific PRBs when neighbor UEs are in critical conditions.

AI Interference Mitigation Pipeline

Traditional handover optimization applies the same A3-Offset and TTT to every cell pair in the network. This is like prescribing the same eyeglasses to everyone — it works for nobody perfectly. In reality, each cell pair has unique characteristics: the propagation environment, typical UE speeds, overlap geometry, and interference profile are all different. AI sets different HO parameters for every cell pair, creating thousands of personalized optimization profiles that a human engineer could never manually configure.

AI vs Traditional Handover

AI HO Optimization Pipeline

The Open RAN (O-RAN) architecture, defined by the O-RAN Alliance, introduces the RAN Intelligent Controller (RIC) as the centralized AI engine for RAN optimization. The RIC comes in two flavors: the Non-RT RIC (control loop > 1 second, hosted in the SMO) and the Near-RT RIC (control loop 10ms - 1 second, co-located near the O-DU). Together, they enable third-party AI algorithms to control the RAN through standardized interfaces, breaking vendor lock-in and accelerating innovation.

O-RAN Architecture Stack

Key Interfaces

The xApp runs on the Near-RT RIC and makes per-UE decisions every 10ms-1s. Example xApps: Traffic Steering (move UEs between cells/carriers based on load), QoS Optimization (adjust scheduler weights per slice), Interference Management (coordinate PRB allocation across cells). The rApp runs on the Non-RT RIC and sets high-level policies: "maintain 5th-percentile SINR > 3 dB for eMBB slice" or "prioritize URLLC latency below 5ms." The rApp uses AI/ML model training on historical data, then deploys the trained model to xApps via A1 for real-time inference.

AI RAN optimization is not theoretical anymore. Major operators worldwide have deployed AI-driven systems and published measurable results. These are not lab experiments — they are production networks serving millions of subscribers. The results consistently show that AI outperforms traditional rule-based optimization by significant margins across all key RAN KPIs.

Operator Case Studies

Before/After KPI Comparison (Operator D)

Building an AI RAN pipeline is not just about training a model. It requires a full MLOps stack: data ingestion from OSS/BSS, feature engineering, model training, validation on shadow traffic, A/B testing in production, continuous monitoring, and automated retraining. Most AI RAN projects fail not because the model is bad, but because the pipeline is not robust. Here is the complete architecture used by leading operators.

End-to-End ML Pipeline

Model Selection Guide