Genetic algorithm with adaptive freezing for the simultaneous optimization of selected feature subsets and hyperparameters in backpropagation neural networks

Abstract

This study proposes a framework that integrates a genetic algorithm with adaptive freezing (GAAF) to simultaneously optimize the feature subset, network architecture, and hyperparameters of a backpropagation neural network (BPNN) model; the framework is denoted as GAAF–BPNN. In contrast to conventional sequential optimization pipelines, the proposed framework optimizes the aforementioned elements simultaneously by using a multichromosome encoding scheme. Moreover, it incorporates a novel architecture-triggered freezing mechanism that dynamically excludes genes corresponding to inactive layers from genetic operations, enabling the construction of variable-depth network architectures while maintaining chromosome consistency. The GAAF–BPNN framework was evaluated in a vibration-signal-based tool wear prediction task, in which it achieved a substantially lower prediction error than did seven baseline models, including standard BPNN models and widely used tree-based ensemble models. The results of this study demonstrate that the proposed framework can effectively improve modeling accuracy while eliminating the need for manual feature selection and architecture tuning.