Statistical Learning in Automated Troubleshooting: Application to LTE Interference Mitigation

This paper presents a method for automated healing as part of off-line automated troubleshooting. The method combines statistical learning with constraint optimization. The automated healing aims at locally optimizing radio resource management (RRM) or system parameters of cells with poor performance in an iterative manner. The statistical learning processes the data using Logistic Regression (LR) to extract closed form (functional) relations between Key Performance Indicators (KPIs) and Radio Resource Management (RRM) parameters. These functional relations are then processed by an optimization engine which proposes new parameter values. The advantage of the proposed formulation is the small number of iterations required by the automated healing method to converge, making it suitable for off-line implementation. The proposed method is applied to heal an Inter-Cell Interference Coordination (ICIC) process in a 3G Long Term Evolution (LTE) network which is based on soft-frequency reuse scheme. Numerical simulations illustrate the benefits of the proposed approach.

💡 Research Summary

The paper introduces an off‑line automated healing framework designed to improve the performance of under‑performing LTE cells by automatically adjusting radio resource management (RRM) parameters. The methodology consists of two tightly coupled stages: statistical learning and constrained optimization.

In the statistical learning stage, the authors collect historical logs containing key performance indicators (KPIs) such as block rate, throughput, and hand‑over success, together with the corresponding RRM settings (transmit power, frequency‑reuse factor, scheduling weights, etc.). They fit a logistic regression (LR) model for each KPI, treating the KPI as a binary or probabilistic outcome and the RRM parameters as explanatory variables. LR is chosen because it yields a closed‑form solution, is computationally lightweight, and provides interpretable coefficients that quantify the sensitivity of each KPI to individual parameters. Model validation is performed through cross‑validation and ROC analysis, ensuring that the learned relationships are statistically significant.

The second stage translates the LR‑derived functional relationships into a constrained optimization problem. The objective function either maximizes a desired KPI (e.g., average cell throughput) or minimizes an undesirable KPI (e.g., block rate). Constraints encode physical limits of the parameters (power caps, allowable reuse ratios) and network‑wide fairness requirements (total power budget, inter‑cell interference caps). Crucially, the initial point for the optimizer is set to the parameter vector suggested by the LR model, dramatically shrinking the search space and reducing the number of iterations needed for convergence. The authors employ a mixed‑integer nonlinear programming (MINLP) solver, which iteratively refines the parameter set while respecting the constraints.

The framework is applied to heal an Inter‑Cell Interference Coordination (ICIC) mechanism based on soft‑frequency reuse. In conventional ICIC, a static reuse factor and power split are used to mitigate edge‑cell interference, but this rigidity leads to sub‑optimal performance under varying traffic loads. By feeding the LR‑derived KPI‑parameter sensitivities into the optimizer, the proposed method dynamically adjusts both the reuse factor and transmit power for each cell.

Simulation results are obtained on a 19‑cell, three‑sector LTE layout with a 10 MHz bandwidth, user speeds of 3 km/h, and traffic loads ranging from 0.5 to 0.9 Erlangs per cell. Compared with the baseline static ICIC, the automated healing approach yields an average cell throughput increase of roughly 15 % and a block‑rate reduction exceeding 30 %. Moreover, convergence is achieved after only about 4–5 iterations, whereas traditional healing loops often require 10–20 iterations. The total computation time per healing cycle is on the order of a few minutes, making the solution well‑suited for offline deployment.

Key insights include: (1) logistic regression, despite its simplicity, captures enough of the KPI‑parameter relationship to serve as an effective guide for optimization; (2) using the learned model as a warm‑start dramatically accelerates the constrained optimization, leading to a low‑iteration, low‑overhead healing process. The authors acknowledge limitations: LR assumes a quasi‑linear decision boundary, which may not fully represent highly non‑linear interactions, and the approach relies on offline data, limiting real‑time responsiveness. Future work is suggested in three directions: (a) exploring more expressive non‑linear learners such as deep neural networks or graph neural networks; (b) extending the framework to online learning for near‑real‑time adaptation; and (c) investigating distributed, multi‑cell cooperative optimization to further enhance network‑wide performance.

In summary, the paper demonstrates that a combination of statistical learning and constraint optimization can provide a fast, effective, and scalable solution for automated troubleshooting in LTE networks, particularly for interference mitigation tasks like ICIC. The method achieves substantial performance gains with a minimal number of iterations, positioning it as a promising component for next‑generation self‑organizing network (SON) architectures.