Forecasting AI Model Computational Requirements Using Random Forest and XGBoost with Entity and Domain Characteristics

Abstrak

This research aims to predict the computational power required by artificial intelligence (AI) models, specifically measured in petaFLOP (Floating Point Operations Per Second), based on their domain and entity characteristics. The study employs Random Forest and XGBoost regression models to predict the amount of computational power needed by AI models. Both models were trained on a dataset that includes features such as the training year, domain (e.g., Language, Vision), and entity characteristics. The results demonstrate that the Random Forest model outperforms XGBoost in terms of prediction accuracy, as indicated by higher R-squared values and lower error metrics. Feature importance analysis revealed that the year of training and domain were the most significant predictors of computational power, with the Language domain emerging as the most influential in both models. The findings highlight the potential for machine learning models to forecast AI computational requirements, which can aid organizations in optimizing computational resources for AI projects. However, the study faces limitations due to data sparsity, particularly in the target variable, and the relatively simple nature of the models employed. Future work should explore incorporating additional features, such as hardware specifications, and leveraging deep learning models to better capture the complexity of AI computational demands. This study lays the groundwork for further research into more precise predictions of AI model resource consumption, helping organizations plan their AI initiatives more effectively.