Hyper-parameter Tuning using Genetic Algorithms for Software Effort Estimation

Research on software effort estimation has investigated the impact of hyper-parameter tuning approaches for machine learning. Many automated tuning approaches have been proposed in the literature to improve the performance of their base models. In this study we compare the Dodge, Grid, Harmony, Tabu, and Random Search algorithms against Standard Genetic, 1+1 Genetic, and Compact Genetic Algorithms in terms of their impact to model accuracy and stability. This evaluation was performed using the ISBSG R18 dataset and on the Support Vector Regression, Classification and Regression Trees, and Ridge Regression techniques. Results of the Scott-Knott analysis show that the genetic algorithms perform similar or better to the other tuners, achieving an improvement of standardized accuracy of up to 0,21 and final values of up to 0,53. Genetic algorithm-tuned Support Vector Regression achieves the highest accuracy while improving estimation stability relative to other tuners.