Decision of Neural Networks Hyperparameters with a Population-Based Algorithm

摘要

This paper proposes a method named Population-based Algorithm (PBA) to decide the best hyperparameters for a neural network (NN). The study focuses on which type of hyperparameters achieve better results in neural network problems. Population-based algorithm inspired from evolutionary algorithms and uses basic steps of genetic algorithms. The distinctive feature of our algorithm from genetic algorithms is fitness evaluation of individuals. To test our approach, we implemented our algorithm to a handwritten digits recognition problem to find the best hyperparameters for a simple neural network and we reached 98.66 accuracy score. Finally, we conclude, how PBA used in neural networks for the best way.