Using Different Probability Distributions Instead of Uniform Random Variables in Fick’s Law Algorithm
2026 10th International Artificial Intelligence and Data Processing Symposium (IDAP), Malatya, Türkiye, 6 - 07 Eylül 2026, ss.1-10, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/idap71869.2026.11727924
- Basıldığı Şehir: Malatya
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1-10
- İnönü Üniversitesi Adresli: Evet
Özet
In metaheuristic algorithms, the uniform distribution is generally chosen for random number generation. However, the effect of this preference on performance is rarely questioned. In this study, all uniform random number generations in the physics-based Fick’s Law Algorithm (FLA) were replaced with nine different distributions (Uniform, Gaussian, Cauchy, Lévy, Exponential, Rayleigh, Weibull, Kumaraswamy, and a logistic chaotic map) through a single parametric generator. On twenty benchmark functions, each distribution was run for 30 repetitions and 1000 iterations. The fact that the Uniform-distribution version produced statistically equivalent results to the original code shared by the algorithm’s authors on all twenty of the twenty functions confirmed that the observed differences originated solely from the choice of distribution. At the middle of the search process (at iteration 500), a Friedman test was applied, and a notable difference among the distributions was observed. The Exponential and Weibull distributions proved significantly superior to the Uniform distribution on 14 and 10 functions, respectively. At the end of the process (at iteration 1000), however, since most functions were fully solved under all distributions, the difference became statistically invisible. In this case, the comparison was instead carried out through the number of exact solutions, the speed of convergence, and the Rastrigin function, which resisted saturation. The findings show that the random number distribution is a parameter overlooked in metaheuristic design yet possessing a measurable effect, and that the Exponential and Weibull distributions can be preferred over the Uniform distribution for general-purpose use.