An Improved Genetic Algorithm for Hyperparameter Optimization in Predictive Modeling

Abstract

This paper presents an improved genetic algorithm (GenAlg) for hyperparameter optimization in predictive modelling. It addresses common limitations of traditional genetic algorithms (GAs), including slow convergence, high computational cost, and premature convergence. The proposed method incorporates early stopping, flexible mutation, intermediate result caching, parallel computation, and advanced crossover. These enhancements accelerate convergence to the global optimum and improve resilience against local minima. The algorithm’s effectiveness is demonstrated through experiments on benchmark functions such as Rastrigin and Schwefel, as well as a real-world task—predicting real estate rental prices. Comparative analysis shows that the improved GenAlg outperforms PyGAD, DEAP, and Optuna in both accuracy and stability, while also reducing computational demands. This approach is particularly useful for automatic hyperparameter tuning and can be applied to a wide range of optimization problems in machine learning, financial modelling, and engineering.