Anytime Stochastic Routing with Hybrid Learning

Increasingly massive volumes of vehicle trajectory data hold the potential to enable higher-resolution traffic services than hitherto possible. We use trajectory data to create a high-resolution, uncertain road-network graph, where edges are associated with travel-time distributions. In this setting, we study probabilistic budget routing that aims to find the path with the highest probability of arriving at a destination within a given time budget. A key challenge is to compute accurately and efficiently the travel-time distribution of a path from the travel-time distributions of the edges in the path. Existing solutions that rely on convolution assume independence among the distributions to be convolved, but as distributions are often dependent, the result distributions exhibit poor accuracy. We propose a hybrid approach that combines convolution with estimation based on machine learning to account for dependencies among distributions in order to improve accuracy. Since the hybrid approach cannot rely on the independence assumption that enables effective pruning during routing, naive use of the hybrid approach is costly. To address the resulting efficiency challenge, we propose an anytime routing algorithm that is able to return a "good enough" path at any time and that eventually computes a high-quality path. Empirical studies involving a substantial real-world trajectory set offer insight into the design properties of the proposed solution, indicating that it is practical in real-world settings.