HEC-NAS-FDS: hybrid expert-conditioned exhaustive neural network architecture search over finite design space

This article presents a new proof-of-concept method called Hybrid Expert-Conditioned Exhaustive Neural Network Architecture Search over Finite Design Space (HEC-NAS-FDS), which aims to find a suitable deep neural network (DNN) architecture with lower computational and time requirements thanks to parallel processing. The method is based on a combination of deep learning and machine learning (ML) techniques. The method is guided by an expert who defines the design space from which a list of all possible combinations is generated (Stage II). The optimal solution is found by training all these combinations. A sub-optimal solution is found, when the R parameter is used, a subset of the randomly selected combinations is trained in parallel (according to the data split R). The remaining combinations are entered into machine learning models (Stage III, Random Forest, XGBoost, etc.) to predict the performance metrics achieved through deep learning. The output of the HEC-NAS-FDS method is an optimal/sub-optimal DNN structure from an expert-defined finite space that is directly related to the input dataset. This hybrid approach enables efficient architecture evaluation without exhaustive training. The main innovation is the combination of DNN and ML approaches, which enables significant savings in both time and computational power within a comprehensive framework. The proposed method was tested on a public dataset for evaluation. The mean absolute percentage error (MAPE) reached 0.9377 %.