Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Deep learning techniques have achieved tremendous success in many real applications in recent years and show their great potential in many areas including transportation. Even though transportation becomes increasingly indispensable in people’s daily life, its related problems, such as traffic congestion and energy waste, have not been completely solved, yet some problems have become even more critical. This dissertation focuses on solving the following fundamental problems: (1) passenger demand prediction, (2) transportation mode detection, (3) traffic light control, in the transportation field using deep learning. The dissertation also extends the application of deep learning to an embedding system for visualization and data retrieval.
The first part of this dissertation is about a Spatio-TEmporal Fuzzy neural Network (STEF-Net) which accurately predicts passenger demand by incorporating the complex interaction of all known important factors, such as temporal, spatial and external information. Specifically, a convolutional long short-term memory network is employed to simultaneously capture spatio-temporal feature interaction, and a fuzzy neural network to model external factors. A novel feature fusion method with convolution and an attention layer is proposed to keep the temporal relation and discriminative spatio-temporal feature interaction. Experiments on a large-scale real-world dataset show the proposed model outperforms the state-of-the-art approaches.
The second part is a light-weight and energy-efficient system which detects transportation modes using only accelerometer sensors in smartphones. Understanding people’s transportation modes is beneficial to many civilian applications, such as urban transportation planning. The system collects accelerometer data in an efficient way and leverages a convolutional neural network to determine transportation modes. Different architectures and classification methods are tested with the proposed convolutional neural network to optimize the system design. Performance evaluation shows that the proposed approach achieves better accuracy than existing work in detecting people’s transportation modes.
The third component of this dissertation is a deep reinforcement learning model, based on Q learning, to control the traffic light. Existing inefficient traffic light control causes numerous problems, such as long delay and waste of energy. In the proposed model, the complex traffic scenario is quantified as states by collecting data and dividing the whole intersection into grids. The timing changes of a traffic light are the actions, which are modeled as a high-dimension Markov decision process. The reward is the cumulative waiting time difference between two cycles. To solve the model, a convolutional neural network is employed to map states to rewards, which is further optimized by several components, such as dueling network, target network, double Q-learning network, and prioritized experience replay. The simulation results in Simulation of Urban MObility (SUMO) show the efficiency of the proposed model in controlling traffic lights.
The last part of this dissertation studies the hierarchical structure in an embedding system. Traditional embedding approaches associate a real-valued embedding vector with each symbol or data point, which generates storage-inefficient representation and fails to effectively encode the internal semantic structure of data. A regularized autoencoder framework is proposed to learn compact Hierarchical K-way D-dimensional (HKD) discrete embedding of data points, aiming at capturing semantic structures of data. Experimental results on synthetic and real-world datasets show that the proposed HKD embedding can effectively reveal the semantic structure of data via visualization and greatly reduce the search space of nearest neighbor retrieval while preserving high accuracy.
Liang, Xiaoyuan, "Applied deep learning in intelligent transportation systems and embedding exploration" (2019). Dissertations. 1422.