Convergence Variability and Population Sizing in Micro‐Genetic Algorithms

Abstract

The issues of micro‐genetic algorithm (micro‐GA) convergence, population size, internal variability, and performance are examined. Procedures are developed to determine best population size when micro‐GAs are used for optimization of real‐world problems. Results show that the best population size cannot be determined before convergence has occurred and that convergence may occur without reaching an optimal solution. To determine the best population size, this article proposes the use of consistency of the final fitness values in addition to the fitness value at convergence. Internal variability must be considered in finding the best population size, and that best population size depends on the level of computational resources used. Smaller populations with generous computational resources performed as good as and sometimes better than larger populations. Midsized populations consistently exhibited lower variation in fitness value. For a given population size, the increase in fitness value may not be worth the added computation cost beyond a certain point. Three approaches are proposed to determine the best population size: (1) a random selection approach, (2) a detailed approach, and (3) a simplified approach. We used the detailed approach with our traffic problem and found that, with sufficient computational resources, the micro‐GA performed best with population size around the square root of the string length. We tested this proposition on standard simple and deceptive problems and found it to hold true.