Dynamic Interference Prediction for In-X 6G Sub-networks

Technical Summary: Dynamic Interference Prediction for In-X 6G Sub-networks

Problem Statement
The paper addresses the challenge of Interference Management (IM) within In-X 6G Sub-networks (SNs), which are designed to support hyper-reliable, low-latency communication (HRLLC) with latency requirements of 0.1–1 ms and reliability up to 99.99999 percentile. In ultra-dense SN environments, interference varies rapidly due to channel characteristics, mobility, and resource limitations.

Traditional Link Adaptation (LA) schemes, which rely on Channel Quality Indicator (CQI) reports and ACK/NACK feedback, face significant hurdles in this context:

1. Rapid Interference Variation: HRLLC traffic utilizes short packets occupying only a fraction of physical resources per Transmission Time Interval (TTI), causing interference to fluctuate faster than traditional tracking methods can handle.
2. Reporting Delays: The inherent delay in CQI reporting mechanisms renders the interference information outdated by the time it reaches the Access Point (AP).
3. Retransmission Limits: Extreme latency constraints often limit retransmissions to a single attempt, making accurate initial Modulation and Coding Scheme (MCS) selection critical.

Existing solutions often require explicit channel estimates from interfering devices (increasing overhead) or rely on supervised learning models that need ground-truth interference labels (difficult to obtain in real-time). Furthermore, many existing methods focus only on the tail probabilities of interference distributions, oversimplifying the LA problem.

Methodology
The authors propose an online, unsupervised, AP-side interference prediction method that relies solely on available CQI reports without requiring explicit feedback or ground-truth labels. The core methodology involves three steps:

1. System Modeling:

   • The authors model inter-SN interference using a spatially consistent 3GPP channel model and an ON-OFF traffic model.
   • Interference Power Values (IPVs) are defined as the sum of contributions from interfering SNs, incorporating path loss, small-scale fading (Rician for LOS, Rayleigh for NLOS), and shadowing.

2. Discrete-Time State Space Model (DSSM):

   • A Gaussian DSSM is formulated where the IPV is treated as a latent (hidden) state variable, and the CQI report serves as the observation.
   • Process Model: Describes the dynamic evolution of the IPV over TTIs, incorporating process noise to account for modeling errors and transmission delays.
   • Measurement Model: Relates the observed CQI (mapped to SINR) to the underlying IPV, accounting for errors from Effective SNR Mapping (ESM) compression and CQI quantization.

3. Extended Kalman Filter (EKF) Approximations:

   • To efficiently apply the EKF to this non-linear system, the authors introduce specific approximations:
     • Transition Dynamics: The correlation of small-scale fading is approximated using a zeroth-order Bessel function. The transition function is linearized using a correlation-matching criterion, creating a convex combination of the two most recent IPV estimates.
     • Measurement Noise: The measurement noise is approximated as the product of two marginal Gaussian variables representing ESM error and CQI quantization error.
   • The EKF operates in a standard two-step process (prediction and update) to estimate the IPV at every TTI, which is then used to adjust the SINR estimate for optimal MCS selection.

Key Contributions

1. Formal Characterization: The paper formally characterizes inter-SN interference using a spatially consistent 3GPP channel model combined with an ON-OFF traffic model.
2. Novel DSSM Formulation: It presents a discrete-time Gaussian DSSM where IPVs are modeled as latent variables and CQI reports as observations. Unlike previous methods, this approach models the full range of IPVs rather than just tail probabilities.
3. EKF-Based Prediction: The work introduces necessary approximations to the DSSM to enable the efficient use of an Extended Kalman Filter for online IPV prediction using only CQI reports.
4. Performance Validation: Numerical results demonstrate that the proposed method outperforms conventional baselines and achieves performance comparable to supervised learning methods (LSTM) and an ideal "Genie" baseline, despite operating without explicit interference labels.

Numerical Results
Simulations were conducted in a 20x20 m² area with 16 SNs, using 6 GHz carrier frequency and 0.1 ms TTIs.

• Prediction Error: The proposed EKF predictor achieved a Relative Absolute Error (RAE) approximately 8 dB lower than a Moving-Average (MA) predictor. The EKF's performance was nearly comparable to a pre-trained LSTM predictor, despite the EKF being an unsupervised, model-based approach.
• Link Adaptation Performance: In terms of Block Error Rate (BLER), the EKF-based LA achieved the target BLER of $10^{-5}$ for 95% of the cases.
• Comparison: The EKF approach closely matched the performance of the "Genie" LA (which assumes perfect SINR knowledge) and the LSTM baseline. It significantly outperformed the Moving-Average predictor, which struggled to track non-linear interference dynamics.

Significance and Claims
The paper claims that its proposed approach offers a low-complexity, computationally efficient solution suitable for small form-factor SN devices. By leveraging the inter-dependencies between CQI reports and IPVs through a state-space model, the method achieves performance comparable to complex machine learning baselines (LSTM) and optimal theoretical limits (Genie) without the need for training data or explicit interference feedback. The authors conclude that with appropriate modeling and approximations, low-complexity state-space modeling is a viable alternative to machine learning methods for dynamic interference prediction in 6G HRLLC scenarios. They note that extending this to vector-valued DSSMs for multiple UEs per AP remains an open area for future investigation.