Heuristic operational network for deep learning search and machine learning model search, design and development

Abstract

Computer-automated tools to assist machine learning model development may be suitable for users without domain expertise or specialized expertise in particular model types. An initial grid search via gradient boosting is applied to a source data set and target to generate a reduced search space including prioritized set of features ranked by importance value. A genetic algorithm searches the supervised learning classification problem space with various model choices and feature engineering. AI model training techniques may include machine learning test models, test hyperparameters, and network architectures. Feature engineering generates new features, and model development promotes features with fitness scores satisfying a defined threshold. Feature engineering may employ feature combinations, scaling, normalization, and feature convolutions. ML model development may apply a genetic operator including one or more of crossover, mutation, and selection. Iterative test ML model development generates pre-trained models that can yield code for deployment of production models.