Abstract
Optimization in complex, noisy environments remains a significant challenge for traditional evolutionary algorithms. This paper presents the Divergence Differential Evolution Algorithm (DDEA), which replaces standard discrete mutation operators with a Gaussian-based approach. By implementing dynamic population management through divergence and assimilation operators, DDEA improves stability and accuracy in finding global optima within complicated fitness landscapes, as demonstrated through both benchmark functions and real-world telecommunications data.
Introduction
Optimization algorithms are essential for tasks ranging from neural network tuning to signal interference tracking. However, standard Differential Evolution (DE) faces two primary hurdles: the tendency to become trapped in local optima and the computational overhead of discrete vector-based mutation operators. In large-scale problems, these limitations can prevent the discovery of global peaks and hinder real-time performance.
The Divergence Differential Evolution Algorithm (DDEA) addresses these issues by introducing a "divide and conquer" strategy. It aims to provide a robust framework for navigating "complicated functions"—landscapes characterized by numerous peaks and valleys—while maintaining the efficiency required for dynamic environments like millimeter-wave telecommunications.
Methodology
The core innovation of DDEA lies in its departure from traditional mutation strategies toward a statistical model of population evolution based on the Central Limit Theorem.
Gaussian Mutation Operator
Standard DE generates candidates using weighted differences between random population members. DDEA, however, posits that as a population converges, its distribution naturally approximates a Gaussian curve. By replacing the traditional three-vector mutation with a single Gaussian density function, the algorithm reduces the number of required random variables, increases computational efficiency, and enables a continuous search range.
Divergence and Assimilation Operators
To maintain diversity and prevent premature convergence, DDEA monitors population density. If the population becomes multimodal—clustering around multiple distinct peaks—the divergence operator splits it into independent subspecies. This allows for simultaneous exploration of different regions of the fitness landscape.
Conversely, to conserve resources, the assimilation operator merges subspecies that have become redundant or too similar. This dynamic population management ensures that computational effort is focused where it is most likely to yield results.
Experimental Evaluation
The performance of DDEA was validated using two distinct benchmarks:
- Michalewicz Function: A highly non-linear optimization benchmark with numerous local optima, designed to test an algorithm's ability to find global minima.
- Rain-Attenuation Series: Real-world time-series data from Xi’an, China, representing signal loss in millimeter-wave communications. This dataset tests the algorithm's robustness against noise and its ability to track dynamic changes.
Results and Discussion
Experimental results indicate that DDEA significantly outperforms standard DE in several key areas:
- Stability and Accuracy: In Michalewicz function tests, DDEA demonstrated tighter convergence and superior stability compared to standard DE.
- Computational Efficiency: Within a fixed budget of 1,500 calculations, DDEA consistently achieved better objective values than standard DE, showing that its statistical mutation model is more efficient than discrete vector math.
- Real-World Application: When applied to rain-attenuation tracking, DDEA provided higher accuracy and smoother transitions, proving its effectiveness for MIMO (Multiple Input Multiple Output) channel characterization at millimeter-wave frequencies.
Conclusion
DDEA represents a significant advancement in evolutionary optimization, offering a robust solution for complex, noisy landscapes. While the algorithm shows great promise, future research should address the computational overhead of the underlying clustering algorithms and explore the sensitivity of its new parameters, such as the divergence threshold and assimilation constant. Expanding testing to higher-dimensional antenna arrays will further validate its utility in advanced telecommunications.
Summary updated