Exploration of Parameter Configurations in Evolutionary Algorithms

Abstract

Genetic algorithms (GA) have proven to be highly effective solvers, outperforming other methods by efficiently navigating complex search spaces and converging to high-quality solutions. Investigating the dynamics of genetic algorithms, this work focuses on key factors such as mutation rate, crossover method, selection strategy, population size, and number of elites to understand their influence on algorithm performance, particularly in solving the traveling salesperson problem (TSP). Rather than merely pursuing optimal solutions, the research uncovers unexpected behaviors emerging from various configurations, offering insights into the mechanisms that drive these outcomes. The study highlights that specific parameter combinations, such as EdgeRecombinationCrossover, GeneralizedRankSelector, and CosaCrossover, can consistently deliver strong performance when properly configured. However, this effectiveness often comes at the expense of increased runtime. Alternatively, configurations involving RandomSelector are generally not effective. In addition to this some it is shown that some configurations converge very fast while others do not. Anotable finding is the role of elitism in evolutionary algorithms. Increasing the number of elites accelerates convergence and improves solution quality, at the expense of reduced population diversity. The research also demonstrates that effective solutions can still be achieved without mutation, provided the setup leverages other parameters effectively. Interestingly, it was observed that RandomSelector, when paired with a high number of elites, can yield competitive solutions. However, such configurations may still face challenges in maintaining diversity, increasing the likelihood of premature convergence. In addition, a machine learning model was developed to predict the behavior of genetic algorithms, focusing on population size as the primary parameter. The model demonstrated that it is feasible to forecast solution quality with a good accuracy.

Date of Award	2024
Original language	English (American)
Supervisor	Stefan Wagner (Supervisor) & Francesco Calimeri (Supervisor)