A Review of the Use of Quasi-random Number Generators to Initialize the Population in Meta-heuristic Algorithms

Different computational tools require random numbers to operate; this is the case with Meta-heuristic Algorithms (MA's). Many studies in the literature have demonstrated that the spatial distribution of the first generation of candidate solutions has a significant influence on the effectiveness of the MA’s. This article tests and analyzes the effect of various types of initializations, specifically contrasting two classes: Low Discrepancy Sequences (LDS) such as Halton, Sobel, Hammersley, and Latin Hypercube and Pseudo-Random Number Generators (PRNG) initialization processes of popular state−of-the−art algorithms as Particle Swarm Optimization (PSO), Differential Evolution (DE), Genetic Algorithms (GA), and Stochastic Fractal Search (SFS). Experimental results are compared and analyzed, showing that, like the butterfly effect of chaos theory, small changes in the initialization of an optimization algorithm with random processes of different nature can induce significantly different results.