A Suboptimal Approach to Antenna Design Problems With Kernel Regression

Sang-Woo Lee, Jun Hur, Moon-Beom Heo, Sunwoo Kim, Hosung Choo, Gangil ByunIEEE Access 2019 · DOI

Abstract

Designing high-performance antennas involves navigating complex, non-linear "cost surfaces" where traditional optimization methods often encounter significant hurdles. This paper introduces an iterative optimization algorithm that integrates Kernel regression with K-means clustering to approximate these performance landscapes efficiently. By managing data points through clustering, the algorithm avoids local optima and achieves faster convergence than standard stochastic methods. The approach is validated using the Rastrigin function and a real-world Antipodal Vivaldi Antenna design, demonstrating superior performance in both computational efficiency and design quality.

Introduction

The design of modern antennas is a multifaceted balancing act. Engineers must precisely tune multiple geometric parameters—such as length, width, and curvature—to optimize critical performance metrics like gain and reflection coefficients. This process defines a mathematical "cost surface" where the goal is to find the global minimum, representing the optimal design.

However, two primary challenges persist:

  1. Computational Intensity: Evaluating a single design configuration requires complex electromagnetic (EM) simulations that are both time-consuming and resource-intensive.
  2. Landscape Complexity: These cost surfaces are highly non-linear and frequently contain numerous "local optima"—false peaks that can trap standard optimization algorithms.

While deterministic methods like gradient descent are computationally efficient, they are prone to getting stuck in local traps. Conversely, stochastic methods such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) are more robust but require a prohibitive number of simulations to converge.

Methodology

The proposed "suboptimal" approach aims to find high-quality solutions with significantly reduced computational overhead. The framework relies on three core components:

Kernel Regression

Kernel regression serves as the primary predictive tool. It maps non-linear, high-dimensional data from EM simulations into a simpler linear model using specific kernel functions. This allows the algorithm to predict performance in unexplored regions of the design space without requiring additional simulations for every point. While various kernels were tested, the Tricube kernel was found to provide the lowest Least-Square Error (0.07) for antenna cost surfaces.

K-means Clustering

To prevent the model from becoming computationally bloated or over-fitting to a specific region, the algorithm employs K-means clustering. By partitioning existing data points into a fixed number of clusters ($K$) and using their centroids to update the regression model, the search remains distributed across the design space, effectively smoothing the path toward the global optimum.

Iterative optimization landscapes showing the distribution of sampled points across the search space.

Iterative Refinement

The process begins with a small set of random designs. The algorithm iteratively simulates these points, constructs a Kernel model, identifies the predicted minimum, and updates the model with the new simulation results. This cycle continues until convergence is achieved.

Experimental Validation

The algorithm was benchmarked against two primary challenges: the Rastrigin function, a classic "torture test" for optimization due to its many local minima, and a real-world Antipodal Vivaldi Antenna.

For the Vivaldi antenna, the authors optimized two key parameters: the slope of the inner edges ($s_{ridge}$) and a unit length factor ($\Delta l$). Simulations were conducted using the commercial EM solver FEKO.

Cost surface of the Antipodal Vivaldi Antenna as a function of design parameters $s_{ridge}$ and $\Delta l$.

Results and Discussion

The proposed method was compared against GA and PSO, with all algorithms capped at 300 iterations. The results highlighted several key advantages:

  • Faster Convergence: The proposed method reached its optimal design in 211 iterations, significantly faster than GA (238) and PSO (271).
  • Superior Design Quality: The minimum cost achieved was 0.9458, outperforming both PSO (0.9573) and GA (0.9604).
  • Visual Intuition: Unlike "black-box" methods, the fitted cost surface provides engineers with visual insights into parameter relationships. For instance, the model revealed that steep inner edges require a specific unit length threshold to maintain performance.

Experimental validation was performed by fabricating the optimized Vivaldi antenna on an FR4 substrate. Measured results closely aligned with simulations, confirming the practical viability of the algorithm.

Conclusion

By combining Kernel regression with K-means clustering, this research provides a robust framework for antenna optimization that balances speed and accuracy. While the current study focused on a 2D parameter space, future work will explore higher-dimensional designs and automated methods for determining the optimal number of clusters. This approach offers a powerful tool for engineers to navigate the complex landscapes of modern electromagnetic design.

Summary updated

BibTeX
@article{lee2019suboptimal,
  author = {Sang-Woo Lee and Jun Hur and Moon-Beom Heo and Sunwoo Kim and Hosung Choo and Gangil Byun},
  title = {{A Suboptimal Approach to Antenna Design Problems With Kernel Regression}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2896658},
}

Loading related content...

Figure