Vehicle Routing with an Environment Model

We consider the problem of vehicle routing in the fully general setting of a transportation network with stochastic, time-dependent travel times. To forecast future system state, we adapt a traffic flow model to provide probabilistic predictions. This becomes a generative model for the Markov decision process formalizing the planning problem. We propose an efficient planning algorithm to take advantage of such a predictive asset. The evaluation on actual traffic flow data shows that the largest improvement is to be had from forecasting the traffic flow better, followed by a smaller improvement from the new planning algorithm.