A hybrid learning approach to stochastic routing

Increasingly available trajectory data enables detailed capture of traffic conditions. We consider an uncertain road network graph, where each graph edge is associated with a travel time distribution, and we study probabilistic budget routing that aims to find the path with the highest probability of arriving within a given time budget. In this setting, a fundamental operation is to compute the travel cost distribution of a path from the cost distributions of the edges in the path. Solutions that rely on convolution generally assume independence among the edges' distributions, which often does not hold and thus incurs poor accuracy. We propose a hybrid approach that combines convolution and machine learning-based estimation to take into account dependencies among distributions in order to improve accuracy. Next, we propose an efficient routing algorithm that is able to utilize the hybrid approach and that features effective pruning techniques to enable faster routing. Empirical studies on a substantial real-world trajectory set offer insight into the properties of the proposed solution, indicating that it is promising.