A conceptual investigation of the effect of random numbers over the performance of metaheuristic algorithms

A lot of research studies focus on the development of a new algorithm or the techniques which improve the performance of the original algorithm. Very few studies conduct the research on the effect of the initial population on the solution quality of algorithms. However, in these studies, one or two algorithms have been used, and a limited number of problems have been handled. To fill in the gap in the literature, this study presents a comprehensive analysis of the five algorithms on the effect of the initial population on their final results including both the numerical and real-world problems along with a wide variety of types of distributions. The study consisted of three rounds and followed the strategy for determining the candidate algorithms to be participated in the next rounds, supported by the statistical tests. Rather than using popular random numbers, fourteen different distributions are used to imitate the random numbers in the initial population generation mechanisms of the algorithms. Two different numerical benchmark sets along with nine real-world problems are used to evaluate the performance of the algorithms. The results are compared with the original ones and other distribution-integrated algorithms. Since knowledge of the appropriate random number source is not available a priori, this study could be a good foundation for future studies not only on the matter of the effect of several distributions on the performances of the algorithms but also introducing an alternative way in generating an initial population.