🔢 mathematics

Would Learning Help? Adaptive CRC-QC-LDPC Selection for Integrity in 5G-NR V2X

This paper proposes and evaluates a Contextual Bandit-based adaptive framework for dynamically selecting CRC polynomials and QC-LDPC coding rates in 5G-NR V2X systems, demonstrating that such learning-assisted configuration significantly reduces undetected error probabilities at low-to-moderate mobility speeds compared to greedy or fixed baselines, while its effectiveness diminishes under high-mobility conditions due to rapid channel decorrelation.

Original authors: Sarah Al-Shareeda, Gulcihan Özdemir, Arouj Fatima, Madalin-Dorin Pop, Bander A. Jabr, Yasser Bin Salamah, Jacques Demerjian

Published 2026-04-07

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Sarah Al-Shareeda, Gulcihan Özdemir, Arouj Fatima, Madalin-Dorin Pop, Bander A. Jabr, Yasser Bin Salamah, Jacques Demerjian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a high-speed car on a busy highway, talking to other cars and traffic lights via a wireless radio. This is V2X (Vehicle-to-Everything) communication. In this world, safety is everything. If your car receives a message saying "Brake Now!" but the message gets garbled by a storm or a glitch, and your car thinks it said "Go," the result could be a crash.

This paper tackles a very specific, dangerous problem: Silent Errors.

The Problem: The "Silent Killer"

In digital communication, errors usually come in two flavors:

Detected Errors: The system realizes, "Hey, this message is broken!" and asks for a redo. This is annoying but safe.
Undetected Errors (The Silent Killer): The system receives a broken message, checks it, and mistakenly thinks, "Looks good!" It accepts the garbage data as truth. In a self-driving car, this is catastrophic.

The paper asks: Can we teach our cars to "learn" on the fly to stop these silent errors, or is it better to just stick to a fixed, conservative rule?

The Setup: The Toolbox

To send messages, 5G cars use a two-step safety net:

CRC (Cyclic Redundancy Check): Think of this as a seal on a letter. It checks if the envelope is intact. There are different sizes of seals (CRC-6, CRC-16, CRC-24A). A bigger seal is harder to fake but takes up more space.
QC-LDPC (Error Correction Code): Think of this as redundant instructions inside the letter. If some words get smudged, the extra words help you guess what they were. There are different "densities" of these instructions (Coding Rates). More redundancy = safer but slower.

Usually, cars pick one specific seal and one specific instruction set and stick with it, regardless of whether they are driving in a calm parking lot or speeding through a blizzard.

The Solution: The "Smart Driver" (Learning Agent)

The researchers proposed a new system where the car acts like a smart driver who constantly adjusts their driving style based on the road.

Instead of a fixed rule, the car uses a Contextual Bandit (a type of AI that learns by trial and error).

The Context: The car looks at the "weather report" (Signal-to-Noise Ratio, interference, deep fades).
The Action: It picks the best combination of Seal Size and Instruction Density for that exact moment.
The Goal: Minimize the chance of a "Silent Error" (accepting a bad message).

They tested three strategies:

The Learner (LinUCB): The smart driver who tries different combinations to see what works best right now.
The Greedy Driver: A driver who picks whatever worked best last time and sticks to it, even if the road changed.
The Conservative Driver: A driver who picks the safest, most boring setting and never changes it, no matter what.

The Findings: When to Learn, When to Freeze

The researchers ran massive simulations (like a video game with millions of driving scenarios) to see who won. Here is what they found, using simple analogies:

1. The Calm Highway (Low to Moderate Speed: 0–120 km/h)

Winner: The Smart Learner.
When the road is relatively stable, the "Smart Driver" shines. It notices, "Oh, the signal is a bit weak right now; let's switch to a bigger seal and more redundant instructions."

Result: It reduced silent errors by 50–70% compared to the "Greedy Driver." It was flexible and adapted perfectly to the changing conditions.

2. The Rollercoaster (High Speed: 180+ km/h)

Winner: The Conservative Driver.
When the car is going super fast, the "weather" changes faster than the driver can react. The signal fades and spikes so quickly that the "Smart Driver" is constantly guessing wrong. It's like trying to catch a butterfly while running a marathon; by the time you decide to grab it, it's already gone.

Result: The learning algorithm got confused and performed no better than the "Greedy Driver." The Conservative Driver, who just stuck to the safest, most robust settings, actually provided the most stable protection because it didn't waste time trying to adapt to chaos.

The Big Takeaway

The paper concludes that learning is a powerful tool, but it has a speed limit.

If you are driving normally: Let the AI learn and adapt. It will find the perfect balance between speed and safety, saving you from silent errors.
If you are driving at the edge of chaos (extreme speed): Turn off the learning. Stick to the "Conservative" settings. Trying to be too clever when the world is moving too fast actually makes things less safe.

In short: Adaptability is great, but only when the world gives you enough time to think.

1. Problem Statement

Vehicle-to-Everything (V2X) communications in 5G New Radio (5G-NR) operate under Ultra-Reliable Low-Latency Communication (URLLC) constraints. A critical challenge in these systems is physical-layer integrity, specifically the risk of Undetected Errors (UE).

The Issue: In short-packet transmissions, if a packet contains errors but passes the Cyclic Redundancy Check (CRC), it is accepted as valid. These "silent" errors corrupt higher-layer data without triggering retransmissions, posing severe safety risks in autonomous driving.
The Limitation: Current 5G-NR configurations (CRC polynomials and Quasi-Cyclic Low-Density Parity-Check [QC-LDPC] coding rates) are largely static and rule-based, determined offline by payload length and service type. They do not adapt to instantaneous channel dynamics caused by high vehicle mobility (Doppler effects, time-correlated fading) and bursty interference.
The Question: Can online, learning-assisted selection of standard-compliant CRC and QC-LDPC parameters reduce the Undetected Error Probability ( $P_{UE}$ ) in dynamic V2X environments, or do fixed configurations remain superior under high non-stationarity?

2. Methodology

The authors propose a Learning-Assisted Framework that treats the joint selection of error detection (CRC) and error correction (QC-LDPC) parameters as a Contextual Bandit (CB) problem.

System Model

Transmission: A single-link 5G-NR V2X physical layer using BPSK modulation.
Encoding: Standard 3GPP CRC polynomials (CRC-6, 11, 16, 24A) concatenated with 5G-NR QC-LDPC codes (Base Graphs 1 & 2).
Channel: Modeled as time-correlated Rayleigh fading (Jakes model) where correlation depends on vehicle speed ( $v$ ). Interference is modeled as a two-state Markov process (bursty interference).
Receiver: Uses soft-input Belief Propagation (BP) decoding. Errors are classified as:
- Detected Error (DE): Decoder fails or CRC fails (packet discarded/retransmitted).
- Undetected Error (UE): Decoder succeeds but output is wrong, and CRC passes (silent corruption).

The Learning Agent (Contextual Bandit)

Context ( $x_t$ ): The transmitter does not have perfect Channel State Information (CSI). Instead, it uses a low-dimensional vector based on delayed receiver feedback:
- Noisy SNR estimate ( $\lambda_{t-1}$ ).
- Burst interference indicator ( $I_{t-1}$ ).
- Deep fade indicator ( $I_{deep, t-1}$ ).
Action Space ( $\mathcal{A}$ ): A discrete set of 12 configurations formed by combining 4 CRC polynomials and 3 LDPC coding rates ( $1/2, 2/3, 3/4$ ).
Reward Function: Designed to prioritize integrity over throughput.
$Reward_t = R_t \cdot \mathbb{I}(\text{No DE} \land \text{No UE}) - \Omega_{DE} \cdot DE_t - \Omega_{UE} \cdot UE_t$
Where $\Omega_{UE} \gg \Omega_{DE}$ , heavily penalizing undetected errors.
Algorithm: A Discounted LinUCB (Linear Upper Confidence Bound) policy.
- It balances exploration and exploitation.
- It uses an exponential discount factor ( $\gamma$ ) to downweight old observations, allowing the agent to adapt to non-stationary channel conditions.

Simulation Environment

Tool: Sionna (NVIDIA's 5G-NR physical layer library).
Parameters: Payload 256 bits, LDPC block length 576, Carrier frequency 5.9 GHz.
Mobility Regimes: Tested at speeds of 0, 60, 120, 180, and 250 km/h.
Baselines:
1. Discounted LinUCB: The proposed adaptive policy.
2. Greedy: Selects the configuration with the best empirical performance so far (no exploration).
3. FixedBest: A single configuration selected offline to minimize $P_{UE}$ for a specific speed, remaining static during operation.

3. Key Contributions

Realistic 5G-NR Simulation: Developed a compliant physical-layer simulation incorporating true 3GPP CRCs, QC-LDPC construction, time-correlated fading, and bursty interference.
Formulation as Contextual Bandit: Successfully mapped the joint CRC-LDPC configuration problem to a lightweight CB task, enabling online adaptation without modifying existing 5G-NR procedures.
Integrity-Aware Objective: Defined a reward function that explicitly minimizes $P_{UE}$ while accounting for $P_{DE}$ , identifying specific mobility regimes where learning helps versus where it hurts.
Regime Analysis: Delineated the boundaries of mobility where adaptive learning is beneficial versus where conservative fixed configurations are superior.

4. Key Results

The study evaluated the Undetected Error Probability ( $P_{UE}$ ) across different Signal-to-Noise Ratios (SNR) and vehicle speeds.

Low to Moderate Mobility (0 – 120 km/h):
- Performance: The Discounted LinUCB policy significantly outperformed the Greedy policy, reducing $P_{UE}$ by 50–70% in the low-SNR regime (-5 to 5 dB).
- Comparison: LinUCB approached the performance of the "FixedBest" configuration at higher SNRs.
- Behavior: The agent successfully adapted to channel conditions, shifting toward stronger CRCs (e.g., CRC-24A) and lower coding rates (e.g., 1/2) when the channel was noisy or fading was severe.
- Reasoning: In these regimes, the temporal correlation of the channel is strong enough that recent feedback remains predictive of future conditions.
High Mobility (≥ 180 km/h):
- Performance: The advantage of online learning diminished. LinUCB and Greedy policies exhibited similar $P_{UE}$ , and both were outperformed by the FixedBest configuration.
- Reasoning: At high speeds, the channel decorrelates rapidly (fast fading). The "temporal predictability" required for the learning agent to make accurate decisions is broken. The agent cannot track the channel fast enough, leading to oscillating or suboptimal configuration choices.
- Optimal Strategy: A conservative, fixed configuration (specifically CRC-16 with Rate 3/4) provided the most stable protection against silent errors.
Decoder Quality: The study confirmed that the Detected Error Probability ( $P_{DE}$ ) remained consistent across all policies, proving that the observed improvements in integrity were due to configuration selection, not changes in decoder behavior.

5. Significance and Conclusion

This paper provides a crucial "reality check" for applying Machine Learning to 5G-NR physical layers:

Conditional Benefit: Learning-assisted adaptation is not universally superior. It is highly effective in low-to-moderate mobility scenarios where the channel is non-stationary but predictable.
The "High Mobility" Trap: In high-speed scenarios (e.g., highway driving >180 km/h), the cost of learning (lag in adaptation, noise in feedback) outweighs the benefits. In these cases, conservative fixed configurations are safer and more reliable for maintaining integrity.
Practical Implication: Future 5G/6G V2X systems should implement mobility-aware switching mechanisms. The system should use adaptive learning when vehicle speeds are moderate but revert to robust, fixed configurations when speeds exceed a threshold where channel decorrelation renders learning ineffective.

The work demonstrates that while AI can optimize physical-layer integrity, it must be deployed with a deep understanding of the underlying channel physics and mobility constraints.