Traffic assignment is the process of distributing traffic demand in the form of an Origin / Destination (O/D) matrix onto a network resulting in specific route and link volumes. If a time dimension is considered, the assignment process becomes dynamic and more challenging. This research examines reversing the dynamic traffic assignment (DTA) process by observing real time link data and utilizing the data to estimate dynamic versions of a seed O/D matrix.

The problem is formulated as a Generalized Least Square optimization that minimizes the discrepancy between simulation output and real field data on one hand, and the estimated demand and the apriori demand on the another. A machine learning technique using Evolutionary Algorithms (EA) is developed instead of more conventional approaches in the literature. The potential of EA in the dynamic O/D estimation problem lies in their powerful search and optimization capabilities in high dimensional search spaces.

The EA based demand estimation framework developed in this thesis is implemented into a model that we call DynODE (Dynamic O/ Estimator). DynODE is integrated with an existing DTA package (i.e. Dynasmart-P). Additionally, the EA-based methods in DynODE are further augmented with:​ (1) a hybrid EA search mechanism also known as Memetic Algorithms, (2) EA parallelization, and (3) distributed computing to further improve the quality and efficiency of the solution.

This dissertation mainly addresses off-line O/D estimation problems. However, on-line O/D estimation can be achieved with increased multi processing and parallel computing.

The proposed methodology is applicable to any general network. In addition, the developed model has the ability to combine any type of observation in the network (i.e. counts, densities, travel time) with other sets of apriori information such as:​ outdated survey, probe data, Automatic Vehicle Identification (AVI) data, or previous step estimates or forecasts.

The approach developed in this thesis is rigorously evaluated, using varying network size and congestion levels. Furthermore, sensitivity analysis is conducted to test model performance under several factors. The obtained results, in terms of replicating observed vehicle counts and the closeness to the real demand, are promising and point to the robustness of the proposed framework.