Bayesian Optimization Based Neural Architecture Search for Classification of Gases/Odors Mixtures

This work demonstrates the potential of the proposed Bayesian optimization (BO). It is based on neural architecture search (NAS) for the classification of gases/odors for datasets of diverse applications and complexities in terms of the number of analytes, features, and datapoints. NAS is an aspect of automated machine learning (AutoML), which explores and exploits a bounded search space of hyper-parameters to obtain optimal neural network (NN) architectures. Herein, Gaussian distribution has been used to model the search space as a probabilistic distribution and BO has been used as an acquisition function to guide the search process in the modeled search space. NAS has been implemented by iteratively minimizing the loss function i.e., categorical cross entropy (CE). Furthermore, the excellent classification accuracies obtained, in hazardous gas detection, beef quality examination, and vehicular exhaust monitoring are calculated to be 100%, 99.68%, and 99.46% respectively. The innovative approach establishes the robustness of the proposed BO-aided NAS in obtaining optimal gases/odors classification for diverse gaseous mixture datasets.