JoWUA

Abstract

Evolutionary algorithms are optimization techniques based on biological and natural evolution mechanisms. These algorithms are a subset of evolutionary computation and fall under unsupervised learning. The Genetic Algorithm (GA) is one of the most common types of evolutionary algorithms. It begins with an initial set of candidate solutions and starts the evolutionary process by applying certain operators to generate new solutions. The newly produced solutions are expected to outperform the previous ones. Premature convergence is a problem encountered by most evolutionary algorithms, particularly genetic algorithms. It occurs when parental solutions fail to generate better offspring or children with superior traits. Self-adaptive mutations and Panmictic populations are the main factors contributing to premature convergence. Several approaches can be applied to avoid premature convergence and sustain population diversity, including the crowding method, incest prevention algorithm, scheduled sharing approach, cooperation-based approach, syntactic analysis of convergence, random offspring generation, selective mutation, and dynamic reproduction operators. The lack of population diversity leads directly to convergence, forcing the evolutionary algorithm to stop evolving and return the dominant value as the candidate solution. In most cases, this is not an optimal solution. One approach to sustaining population diversity is applying dynamic reproduction genetic operators. The main objective of this research is to propose an enhancement to the standard genetic algorithm to overcome premature convergence. A dynamic reproduction mutation operator is proposed to vary the probability of mutation based on the fitness value in each iteration. The methodology employed by the researcher involves conducting experiments to demonstrate the results achieved after applying the enhanced genetic algorithm (Rowe, 2008). Three different experiments with varying population sizes and mutation probability values were carried out to identify the best solution for an optimization problem. A total of 100 generations were produced by applying 10,000 iterations, and a binary genetic algorithm was used for running iterations with 16-bit chromosome lengths to represent candidate solutions. The results show that improvements in fitness scores were achieved, which enhanced the performance of the genetic algorithm for the produced generations (offspring). Moreover, population diversity was maintained.