Evolutionary Acquisition of CNN Architecture by Thermodynamical Genetic Algorithm

Convolutional Neural Network (CNN) is a method that has been used for many years in the field of computer vision. In recent years, CNN has become huge and complex in proportion to the performance of computers and the difficulty of problems. As a result, the expertise and time required to build a better CNN architecture have increased, making it difficult to easily find a good CNN by hand or design for each problem. Therefore, it is important to optimize the CNN architecture automatically. On the other hand, the Genetic Algorithm (GA) is known as a method to optimize various problems based on the evolutionary process efficiently, and the evolutionary acquisition approach to CNN architecture using GA has attracted much attention. In this study, we propose the Thermodynamical Genetic Algorithmic Convolutional Neural Networks (tdgaCNN) based on two methods: gaCNN, one of the architecture search methods for CNN by GA, and Thermodynamical Genetic Algorithm (TDGA), one of the methods for GA. This method achieves better CNN architecture acquisition by emphasizing the importance of maintaining population diversity during the GA search phase. tdgaCNN achieved a prediction accuracy of up to 94.64% in experiments on the Fashion MNIST dataset. 12 existing algorithms' predictions compared to the accuracy, tdgaCNN outperformed 9 other methods, including other CNN architecture search methods.