RL-CNN: Reinforcement Learning-designed Convolutional Neural Network for Urban Traffic Flow Estimation

Accurate prediction of urban traffic flows brings enormous advantages to big cities. Therefore, many urban traffic flow predictors have been developed in recent years. Urban traffic flow predictors aim to identify complex mobility patterns and capture urban traffic flow characteristics from large-scale historical datasets. Afterward, trained models are used to predict the future traffic volume in terms of the number of moving objects (e.g., vehicles). Convolutional Neural Networks (CNN) and other deep learning approaches are brilliant choices because of their ability to learn Spatio-temporal dependencies. Nevertheless, the extensive set of hyper-parameters tends to make these neural networks overly complex and challenging to design. In this work, we introduce RL-CNN, a framework based on Reinforcement Learning whose objective is to autonomously discover high performance CNN architectures for the given traffic prediction task without human intervention. We examine the proposed RLCNN model as a traffic flow estimator on a real-world and large scale vehicular network dataset. We observe improvements of 5% - 10% in the average traffic flow prediction accuracy over the state-of-art approaches.