This study focuses on updating time-varying demand matrices by using real observation counts from advanced traffic management surveillance systems. A machine-learning technique using advanced evolutionary algorithms (EAs) is developed instead of the more conventional approaches in the literature. This EA-based demand estimation framework is implemented into a model called the Dynamic Origin–Destination (O-D) Estimator (DynODE). The potential of EAs in the dynamic O-D estimation problem lies in their powerful global search and optimization capabilities. DynODE is integrated with an existing dynamic traffic assignment platform (e.g., DYNASMART-P). The EA-based methods in this study are further augmented with EA parallelization to improve the quality and efficiency of the solution. DynODE mainly addresses offline O-D estimation problems. However, online O-D estimation can be achieved with the parallel version of DynODE with sufficient multiprocessing and parallel computing. The developed approach is rigorously evaluated on a medium-sized real network to assess the effects of various parallel structures. For all experiments, savings in computation resources as well as enhancement in the quality of solution were realized.