Abstract
In swarm intelligence, Particle Swarm Optimization (PSO) and Differential Evolution (DE) have been successfully applied in many optimization tasks. However, the proliferation of bespoke variants has led to a fragmented landscape. This paper proposes a modular framework, PSODE, which decomposes PSO and DE into interchangeable components. By systematically combining these modules, the authors generate 800 novel hybrid algorithms. Testing on the COCO/BBOB benchmark suite across 5-D and 20-D spaces reveals that specific hybrid configurations can significantly outperform standard PSO and remain competitive with state-of-the-art DE variants.
Introduction
Black-box optimization seeks to find the global minimum of a function without internal knowledge of its structure. Two dominant paradigms in this field are Particle Swarm Optimization (PSO), inspired by social behavior like bird flocking, and Differential Evolution (DE), rooted in evolutionary principles. While PSO excels at rapid exploitation, it often suffers from premature convergence. Conversely, DE is robust and maintains population diversity but can be slower to converge.
Historically, improvements to these algorithms have been developed in isolation. A new "inertia weight" for PSO is rarely evaluated alongside a novel "mutation strategy" from DE. This paper addresses this fragmentation by treating algorithm components as modular "LEGO bricks," allowing for the systematic exploration of the hybrid algorithmic space.
The PSODE Framework
The researchers first modularized both algorithms into their constituent parts. PSO was decomposed into velocity update strategies and neighborhood topologies, while DE was broken down into mutation and crossover operators. To bridge these two paradigms, a Selection Module was introduced to determine how individuals from the PSO-updated and DE-updated populations survive into the subsequent generation.
A technical challenge in this hybridization is the management of "velocity" (momentum) in PSO when a DE operator moves a particle. The authors implemented a velocity re-computation step, calculating a new velocity based on the displacement caused by DE operators to maintain swarm logic consistency.
Experimental Evaluation
The framework was benchmarked using the COCO/BBOB (Black-Box Optimization Benchmark) suite, which includes 24 functions across five categories (e.g., separable, multi-modal). The study evaluated 830 algorithm instances: 800 hybrids, 20 pure PSO variants, and 10 pure DE variants. Performance was assessed in both 5-D and 20-D search spaces and normalized against the state-of-the-art CMA-ES (Covariance Matrix Adaptation Evolution Strategy).
Analysis of Results
The experimental results highlighted several key findings regarding the interaction of modular components:
- Selection Strategy is Critical: The choice of selection operator was the most significant factor in performance. The "pairwise/3" strategy—which compares the original individual, its PSO-updated version, and its DE-hybrid version—consistently yielded the best results.
- Hybrid Performance: Top-performing hybrids, such as
H_I_I_T1_B_P3, consistently outperformed all standard PSO variants. While pure DE variants often remained at the top of the leaderboard, the hybrids proved highly competitive. - Dimensionality and Operators: In 5-D, the Target-to-pbest/1 mutation strategy was most effective, whereas the greedier Target-to-best/1 strategy performed better in 20-D. Binomial crossover generally outperformed exponential crossover as dimensionality increased.
Interestingly, neighborhood topology had a minimal impact on hybrid performance, suggesting that the diversity introduced by DE operators compensates for simpler social structures.
Conclusion
By modularizing PSO and DE, this work transforms algorithm design from a manual, bespoke process into a systematic configuration task. The PSODE framework not only identifies high-performing hybrid configurations but also provides a foundation for Automated Algorithm Selection. Future work will focus on using machine learning to dynamically select the optimal combination of modules based on the specific characteristics of a given optimization landscape.
Summary updated