Learning-Based Robust Resource allocation for D2D Underlaying Cellular Network

Weihua Wu, Runzi Liu, Qinghai Yang, Tony Q. S. Quek

Abstract

In Device-to-Device (D2D) networks underlaying cellular systems, managing interference is critical but complicated by uncertain Channel State Information (CSI). This paper proposes a robust resource allocation framework that utilizes Support Vector Clustering (SVC) to learn data-driven uncertainty sets for channel noise. By moving beyond traditional symmetric geometric assumptions like boxes or ellipsoids, the SVC-based approach provides a tighter fit to actual noise distributions, minimizing performance loss while ensuring high reliability. We further introduce a low-complexity bisection search algorithm for efficient power allocation. Experimental results show that the proposed method significantly outperforms non-robust and standard robust baselines in terms of throughput and outage probability.

Introduction

Modern cellular networks increasingly leverage D2D communication to enhance spectrum efficiency by allowing direct communication between devices. However, when D2D pairs reuse the spectrum of cellular users (CUEs), they introduce mutual interference. Effective power control is necessary to protect CUE performance while enabling D2D links.

System model of D2D underlaying cellular network showing interference links between CUE, DUE, and BS.

A primary challenge in this optimization is the lack of perfect CSI. Factors such as feedback delays and estimation errors lead to CSI uncertainty. Traditional robust optimization often relies on simple geometric sets (e.g., boxes or ellipsoids) to bound this uncertainty. While mathematically convenient, these sets are often overly conservative, encompassing large regions of "empty space" where noise is unlikely to occur, thereby unnecessarily restricting network throughput.

Methods

The proposed framework addresses CSI uncertainty through a three-stage approach:

1. Geometric Baselines

The authors first establish robust counterparts using standard uncertainty sets—Polytope, Ellipsoidal, and Box sets. These serve as benchmarks for evaluating the proposed learning-based method.

2. Support Vector Clustering (SVC) Innovation

To overcome the limitations of fixed geometric shapes, the authors employ Support Vector Clustering. SVC maps CSI samples into a high-dimensional feature space where a minimal enclosing sphere is found. When mapped back to the original space, this sphere forms a tight, asymmetric boundary that accurately reflects the data's distribution.

Comparison of uncertainty sets in (a) 2D original space and (b) 3D feature space, highlighting the tighter fit of the SVC-based set.

As shown in the figure above, the SVC-based set (red dots enclosed by a tight boundary) avoids the excessive "safety margins" of the Box (purple dotted) or Ellipsoidal (black dashed) models. This data-driven approach bridges the gap between black-box learning and mathematically rigorous convex optimization.

3. Bisection Search Algorithm

The resulting optimization problem, involving chance constraints, is transformed into a tractable convex form. A bisection search-based algorithm is then used to iteratively find the optimal transmit powers for CUEs and D2D users. This algorithm typically converges in fewer than 17 iterations, making it suitable for real-time applications.

Key Results

The framework was evaluated using synthetic datasets with Exponential and Bivariate Gaussian noise distributions. Key findings include:

  • Reliability: While non-robust methods suffered outage probabilities over 50%, the proposed robust methods consistently maintained outages below the 5% target.
  • Throughput: The SVC-based approach achieved higher CUE throughput than Polytope or Box methods by reducing unnecessary power margins.
  • Adaptability: The data-driven nature of SVC allowed it to adapt to different noise distributions more effectively than fixed-shape models.

Conclusion

This paper demonstrates that machine learning can be effectively integrated into robust optimization by defining tighter, data-driven uncertainty sets. The SVC-based framework provides a robust solution that minimizes the performance waste typical of "worst-case" models. Future work will explore extending this model to dynamic environments through online learning and multi-cell coordination scenarios.

Summary updated

BibTeX
@article{wu2022learning,
  author = {Weihua Wu and Runzi Liu and Qinghai Yang and Tony Q. S. Quek},
  title = {{Learning-Based Robust Resource allocation for D2D Underlaying Cellular Network}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2022},
  doi = {10.1109/twc.2022.3152260},
}
Figure