⚡ electrical engineering

Adaptive Selection of Codebook Using Assistance Information and Artificial Intelligence for 6G Systems

This paper proposes an AI-driven, UE-assisted adaptive codebook selection framework for 6G downlink systems that utilizes reported statistical channel properties to optimize precoder quantization, thereby reducing CSI reporting overhead while maintaining target system throughput.

Original authors: Denis Esiunin, Alexei Davydov

Published 2026-02-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Denis Esiunin, Alexei Davydov

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 trying to send a complex, high-definition video stream to a friend's house over a shaky internet connection. If the connection is perfect, you can send the full, uncompressed file. But if the connection is spotty, you have to compress the file so it doesn't get stuck or lost. The trick is knowing exactly how much to compress it: compress it too much, and the video looks pixelated and unwatchable; compress it too little, and the file is too big to send, causing delays.

This paper is about solving that exact problem for 6G mobile networks, but instead of video files, we are talking about radio signals used to beam data to your phone.

Here is the breakdown of the paper using simple analogies:

1. The Problem: One Size Does Not Fit All

In the current mobile world (5G), the cell tower (Base Station) sends data to your phone using a "Codebook." Think of a Codebook as a dictionary of pre-approved signal shapes.

The Issue: The tower doesn't know exactly what your phone's connection looks like right now. Is your phone in a crowded stadium (lots of signal bouncing around)? Is it in a straight line of sight? Is your car moving fast?
The Old Way: The tower usually picks one "standard" dictionary for everyone. This is like a chef cooking the same meal for everyone, regardless of whether they are hungry, full, or have a specific allergy. It works okay, but it's not perfect. Sometimes the signal is too heavy (wasting battery and data), and sometimes it's too weak (dropping calls).

2. The Solution: Asking the Phone for Help

The authors propose a new system where the phone (User Equipment) acts as a "weather reporter" for the signal.

The Report: Instead of just saying "I have a signal," the phone measures how the signal behaves in three directions:
1. Space: How does the signal bounce off buildings? (Spatial)
2. Frequency: How does it change across different radio channels? (Frequency)
3. Time: How fast is the signal changing as you move? (Time)
The Analogy: Imagine you are driving. The phone tells the tower: "The road is bumpy (Space), the traffic is heavy (Frequency), and I'm driving at 60mph (Time)."

3. The Brain: Artificial Intelligence (AI)

Once the tower gets this "weather report" from the phone, it doesn't just guess. It uses an AI brain (Neural Network).

The Prediction: The AI looks at the report and predicts: "If I use Dictionary A, the signal will be clear but heavy. If I use Dictionary B, it will be light but might get pixelated."
The "Aging" Factor: The AI also knows that by the time the phone sends the report and the tower sends the signal back, a tiny bit of time has passed. If you are moving fast, the "weather" might have changed. The AI accounts for this delay (called "channel aging") to make sure the prediction is still accurate.

4. The Decision: Picking the Perfect Dictionary

Based on the AI's prediction, the tower picks the best Codebook for that specific moment.

Strategy 1 (The Budget Saver): "Find the lightest dictionary that still gives us a clear enough picture." (Good for saving data usage).
Strategy 2 (The Balancer): "Is the extra clarity of the heavy dictionary worth the extra data cost?" If the answer is no, pick the lighter one. If yes, pick the heavy one.

5. The Results: Faster and Smarter

The authors tested this in a computer simulation of a busy city.

The Outcome: Their new method was like having a smart traffic controller. It reduced the amount of "traffic" (data overhead) needed to send the signal by nearly 50% compared to the old fixed methods.
The Benefit: You get the same fast internet speeds, but the network is much more efficient, saving battery life and allowing more people to use the network at once without it getting clogged.

Summary Metaphor

Think of the old system as a mailman who always delivers packages in the same size box, regardless of what's inside. If you send a letter, it goes in a huge box (waste of space). If you send a piano, it might not fit.

This new paper proposes a smart mailman who:

Asks you what you are sending (The Phone's Report).
Uses a computer to figure out the perfect box size (The AI Prediction).
Packs it efficiently so it arrives safely without wasting cardboard (Optimal Codebook Selection).

In short: By letting phones tell the network exactly how the signal is behaving and using AI to interpret that, 6G can send data more efficiently, faster, and with less waste.

1. Problem Statement

In 6G and advanced 5G NR systems utilizing Massive MIMO (M-MIMO), accurate Channel State Information (CSI) is critical for downlink (DL) precoding. However, a "one-size-fits-all" approach to Codebook (CB) selection is suboptimal because:

Channel Variability: Propagation conditions (e.g., Line-of-Sight vs. Non-Line-of-Sight, mobility levels, scattering environments) vary significantly between User Equipments (UEs).
Trade-off Dilemma: High-precision CBs (using more Discrete Fourier Transform (DFT) basis vectors) improve precoding accuracy but incur high UpLink (UL) control overhead. Low-overhead CBs reduce signaling but may cause severe performance degradation in complex channels.
Limitations of Existing Methods: Previous approaches often rely on simple thresholds (e.g., LoS/NLoS classification) or single-domain metrics (e.g., Time-Domain Channel Properties only), which fail to capture the joint spatial, frequency, and time characteristics required for optimal 6G generalized precoding.
Channel Aging: CSI feedback delays cause "channel aging," where the reported channel no longer matches the actual channel, particularly in high-mobility scenarios.

2. Methodology

The paper proposes a UE-assisted, AI-driven adaptive CB selection framework deployed at the Base Station (BS).

A. Generalized Precoding Model

The system utilizes a generalized precoding matrix $W$ that spans spatial, frequency, and time domains. The quantization accuracy is controlled by the number of DFT basis vectors used in each dimension:

$L$ : Spatial domain basis vectors.
$M$ : Frequency domain basis vectors.
$T$ : Time domain basis vectors.
Different combinations of $(L, M, T)$ correspond to different CB types (e.g., Type-1, eType-2, eType-2a).

B. Extended UE Assistance Information

Instead of relying on a single metric, the UE reports statistical channel properties across three dimensions to the BS:

Spatial-Domain Channel Properties (SDCP): Normalized autocorrelation across antenna ports.
Frequency-Domain Channel Properties (FDCP): Normalized autocorrelation across Resource Blocks (RBs).
Time-Domain Channel Properties (TDCP): Normalized autocorrelation across time slots (accounting for Doppler/mobility).
Note: These metrics are derived from existing CSI-RS signals, minimizing additional signaling overhead.

C. AI-Based Quantization Error Prediction

The BS employs a Neural Network (NN) to predict the quantization accuracy for various CB configurations based on the reported assistance information.

Metric: The paper introduces Aging-Aware Generalized Cosine Similarity (AGCS). Unlike standard GCS, AGCS incorporates the CSI measurement and reporting delay ( $\delta$ ) to account for channel aging effects caused by UE mobility.
Architecture: A Fully Connected (FC) Neural Network is trained as a regression model to map the input vector $[\bar{c}_p, \bar{c}_f, \bar{c}_t]$ to the predicted AGCS ( $\hat{\rho}_\delta$ ) for each candidate CB.
Training: The model is trained using Mean Squared Error (MSE) loss against ground-truth AGCS values derived from system-level simulations.

D. CB Selection Strategies

Two strategies are proposed for the BS to select the optimal CB based on the NN's predictions:

Threshold-Based Selection: Select the CB with the lowest overhead that still meets a minimum AGCS threshold ( $\rho_{min}$ ). This threshold is adjusted based on the transmission scheme (e.g., higher for MU-MIMO than SU-MIMO).
Trade-off Optimization: Select the CB that maximizes the AGCS gain over a reference CB (e.g., Type-1) while satisfying a CB-specific overhead threshold ( $\rho_0$ ). This explicitly balances the gain in DL precoding performance against the cost of increased UL overhead.

3. Key Contributions

Multi-Dimensional Assistance: Extends UE assistance beyond time-domain correlation to include spatial and frequency correlation, enabling optimal selection of generalized precoders.
Aging-Aware Metric: Proposes AGCS, a metric that explicitly models the degradation of CSI due to feedback delays and mobility, making it suitable for dynamic 6G environments.
AI-Driven Adaptation: Demonstrates the feasibility of using a lightweight NN at the BS to predict quantization errors for multiple CB types simultaneously, replacing complex heuristic rules.
Feature Pruning Analysis: Identifies that while full correlation data is ideal, approximately 20-40% of the assistance information features are sufficient for high-accuracy prediction, offering a path to reduce reporting overhead.

4. Simulation Results

The authors evaluated the approach using system-level simulations in a 6G M-MIMO scenario (7 GHz, 100 MHz bandwidth, 256 antenna ports) with 3GPP Urban Macro (UMa) channels.

Prediction Accuracy:
- The NN accurately predicts AGCS using the proposed assistance information.
- Low Mobility: Spatial (SDCP) and Frequency (FDCP) correlations are sufficient.
- High Mobility: Incorporating Time-Domain (TDCP) correlation is critical; omitting it significantly degrades prediction accuracy for high-mobility users.
System Performance:
- Throughput: The proposed UE-specific CB selection outperforms fixed CB schemes and even "ideal" LoS/NLoS-based switching by ~5% in average User Perceived Throughput (UPT).
- Overhead Reduction: Compared to the baseline high-precision CB (Case 4), the proposed methods achieve comparable throughput while reducing CSI overhead by 45%–53%.
- Edge Users: The 5%-tile UPT (cell-edge performance) loss is kept below 3% compared to the high-overhead baseline.
Strategy Comparison: The second strategy (Trade-off Optimization) yielded better overhead savings than the first (Threshold-Based) by adaptively managing CB selection even in low-AGCS regions.

5. Significance

This paper addresses a critical bottleneck in the evolution toward 6G: the efficient management of CSI feedback in highly dynamic and complex M-MIMO environments.

Scalability: By moving from static, network-wide CB configurations to dynamic, UE-specific adaptation, the system can scale to support diverse channel conditions without sacrificing spectral efficiency.
AI Integration: It provides a concrete, practical example of how AI/ML can be integrated into the physical layer (PHY) to optimize control signaling, moving beyond theoretical concepts.
Future-Proofing: The framework is designed for "generalized" precoding, making it directly applicable to future 6G standards that will likely rely heavily on compressed, multi-dimensional CSI reporting.

In conclusion, the proposed method enables a "smart" feedback loop where the network intelligently selects the most efficient quantization scheme for each user, maximizing system throughput while minimizing control channel congestion.