Does the theoretical development in assignment procedures matter in practice?

This study investigates and compares the behaviour of various formulation of the assignment model on a large-scale network (618-zones, 30,000 links) covering Copenhagen. DTU developed the software, so were not dependent on limitation of specific commercial packages. It was therefore fairly easy to test different model assumptions and solution algorithms on the same case. The first set of experiments compares the theoretical correct solution algorithm of User Equilibrium with incremental loading and iterative assignment. It is shown that the results - as expected - indeed are different, and especially that the iterative assignment is an improper solution method. Following this, different formulations of the error term in stochastic user equilibrium (SUE) are investigated. It is shown, that there is a difference in the results. The core conclusion is that a truncated normal distribution does not approximate a Probit model, and that the Gamma distribution, which is additive in mean and variance and non-negative, is a better choice. It is evident that the User Equilibrium solution is worse in terms of system costs, than the System Optimal Solution (SO) - or equal in very simple cases. However, it is also often claimed that the Stochastic User Equilibrium solution is in practice worse than the User Equilibrium (UE) This is however not necessarily the case, since the error term can make the solution either further away or closer to SO than the UE solution. The interesting implication of this is that providing information - e.g. by traffic informatics - may not improve the overall system performance. In some cases information may make thing worse! A special case of this investigation is calculation of impacts of a new road project in Copenhagen, where Braess paradox appear. The paradox states that new infrastructure projects sometimes make the overall network performance worse, due to the difference between SO and UE. The paper then compares the solutions of the more advanced approaches with random coefficients and multiple classes, with the simple single-class assignment procedure used in most applied models. For the covering abstract please see ITRD E135207.