Abstract:
The purpose of this research was to compare the prediction performance between a standard algorithm and an optimally-parameterized Deep Learning model. Data used for prediction were road traffic status in the Bangkok Sathorn Road area. A literature review found 21 related concepts and theories indicating that deep learning has superior prediction performance. This research developed a method for collecting traffic status from the Google Maps website by comparing the accuracy of the traffic flow with the CCTV cameras of the Department of Traffic and Transport, Bangkok. A Pearson correlation coefficient of 0.96 was found, indicating that the traffic flow as measured had a high level of consistency with reality. A tool was developed and used to continuously collect image data every half hour for two months, resulting in comprehensive 1,762 images data for machine learning process.
The results can be summarized as follows: 1) the number of neurons was 1,024; 2) the activation function was ReLU; 3) the optimizer was Adam; 4) the number of training epochs was 1,000, and 5) the batch number was 30. An accuracy of 82% with the standard deviation of accuracy equal to 2% in predicting traffic status with 10 folds cross validation. Using the training data set by a random 80%, predicting the traffic status with the test data 20% divided, it was found that the percentage of prediction accuracy obtained in each algorithm was: Decision Tree: 71%; Random Forest: 71%; K Nearest Neighbors: 73%; Naive Bayes: 63%; Gradient Boosting: 72%, and Deep Learning: 75%. The most predictive model was found to be deep learning.