Learning Route Planning from Experienced Drivers Using Generalized Value Iteration Network

Traffic congestion has long been a serious problem in cities, and route planning can improve traffic efficiency. The existing route planning approach relies on current and future traffic status. However, because traffic prediction and route planning interact with each other, the actual driving results deviate from expectations, and the performance is not satisfactory. In order to solve this problem, considering the topology of road networks, this paper proposes a route planning algorithm based on generalized value iteration network (GVIN), which uses graph convolution to extract the features of traffic flow, and then imitates human routing experience under various traffic status. Finally we evaluate the performance of the proposed network on real map and trajectory data in Beijing, China. The experimental results show that GVIN can simulate the human’s routing decisions with high success rate and less commuting time.