Optimizing hyperparameters of deep learning in predicting bus passengers based on simulated annealing

Abstract

Bus is certainly one of the most widely used public transportation systems in a modern city because it provides an inexpensive solution to public transportation users, such as commuters and tourists. Most people would like to avoid taking a crowded bus on the way. That is why forecasting the number of bus passengers has been a critical problem for years. The proposed method is inspired by the fact that there is no easy way to know the suitable parameters for most of the deep learning methods in solving the optimization problem of forecasting the number of passengers on a bus. To address this issue, the proposed algorithm uses a simulated annealing (SA) to find out a suitable number of neurons for each layer of a fully connected deep neural network (DNN) to enhance the accuracy rate in solving this particular optimization problem. The proposed method is compared with support vector machine, random forest, eXtreme gradient boosting, deep neural network, and deep neural network with dropout for the data provided by the Taichung city smart transportation big data research center, Taiwan (TSTBDRC). Our simulation results indicate that the proposed method outperforms all the other forecasting methods for forecasting the number of bus passengers in terms of the accuracy rate and the prediction time.