Deep Learning Techniques for Optimization of Roadside Unit Placement
Smart cities have the potential to improve the safety and efficiency of a number of different fields, including transportation. Intelligent transportation systems (ITS) use data collected by the sensor infrastructure in a smart city to allow agents to make more informed decisions while traveling along a road network. This requires the use of a communication network to pass information from sensors to the agents. Some researchers propose decentralized systems which would exchange information locally between vehicles and infrastructure without any central controller. While this is cheaper to implement, these decentralized networks are susceptible to some issues such as the hidden node problem and limited bandwidth. Because of these issues, other researchers suggest a centralized approach to ITS communication focusing on the roadside unit (RSU). RSUs can support more powerful equipment with longer range and more bandwidth, but are more expensive to implement. Due to the increased cost, a feasible RSU network must minimize the number of RSUs in order for it to be affordable while maximizing the effect that the network has on the ITS. This is called the RSU placement problem, and as components of an ITS become more common on the roadway solving this problem will become more important. Researchers have applied a number of traditional optimization algorithms to the RSU placement problem, but there are no works that apply more modern machine learning based approaches. This dissertation aims to address this gap by applying different types of neural networks to the RSU placement problem, each of which provide unique benefits over traditional optimization approaches. First, a convolutional neural network (CNN) is presented to optimize the placement of an RSU in a road network by analyzing different images. The network analyzes images of a map, the building traces, and the road traces in order to determine the optimal intersections to place an RSU. The methods needed to create these images are developed, and the resulting dataset is fed into the CNN. The performance of the network is analyzed, and based on these results several shortcomings of the method are identified and discussed. Based on the limitations of the CNN based approach, a second algorithm is developed using pointer networks to optimize an RSU network. The pointer network is trained using data recorded from a simulated road network, and the environment used to create this data is presented. The model is trained using the resulting dataset with a supervised learning approach. The results are analyzed and some improvements to the model are proposed.