This paper addresses the problem of finding the optimal routing policies with only trajectory information in a stochastic time-dependent network where all link travel times are temporally and spatially correlated. Two equivalent definitions are given for trajectory-adaptive routing policy. The first definition follows the mapping definition in a series of previous studies, where the trajectory information is included in the state variable. With that definition, any subpolicy of an optimal routing policy must also be optimal, and the algorithm developed in previous studies could serve as an exact solution, but the nature of the definition that trajectory information is included in the state variable determines that the number of states is prohibitively huge. Therefore, a second definition is given as a recursive definition, where the trajectory information is not included in the state variable. It is shown that, with the second definition, a subpolicy of an optimal/nondominated routing policy is not necessarily optimal/nondominated, but any subpolicy of a pure routing policy must also be pure which is a new property defined based on nondominance. An exact algorithm is designed to find optimal trajectory-adaptive routing policies based on the aforementioned property of pure routing policy. A comparison is made between the optimal trajectory-adaptive routing policies and the optimal paths in the same test networks to investigate the benefit of being adaptive. The results show that the benefit of being adaptive to trajectory information is more significant with a higher risk aversion, a higher correlation, and/or a larger network.

}, keywords = {adaptive routing, Correlation, nondominance, Risk aversion, stochastic dependencies, trajectory information}, doi = {10.1287/trsc.2016.0691}, author = {Huang, He and Gao, Song} }