Adaptive Random Neural Architecture Search
Mohammed R. AL-Matari, Ahmad Mustafa · 2025
Network Architecture Search is the science of searching for the best network structure to provide the optimal solution to a problem. It attempts to propose architecture formulated by the appropriate hyperparameters for the neural network. It has been a debatable matter as no fixed rules have been discovered. This paper investigates the usage of adaptive random search for Neural Architecture Search precisely with main focus on two hyperparameters, layers and neurons. The paper will utilize the random search with a modification that supports the direction in which better results are obtained. Adaptive random search is superior to random search and automated specified search as it uses previous model guidance to construct the new model architecture. This method has been tested on two benchmark datasets FashionMNIST-10 and CIFAR10 and comparing random search, adaptive random search, and grid search. Adaptive random search achieved higher average validation accuracy on CIFAR-10 dataset and the lowest average training time up to obtaining the best model with values of 56.74 seconds.