Parallelization of Travel Demand Models with Dependent Agents: Comparing Model Simplification and Agent Synchronization for Shared Household Vehicles

The increasing complexity and scale of modern agent-based travel demand models pose significant computational challenges in terms of required memory and especially computation time. Improving the efficiency and scalability of these simulations is crucial to produce statistically reliable forecasts for large-scale studies of transportation systems and the effects of policies and measures. Modern approaches from parallel computing have been applied in other agent-based models to evaluate agent decisions in parallel and thereby reduce computation time. In this paper, we present two approaches to enhance the efficiency of agent-based travel demand models through parallelization. We use the dependency of shared household vehicles and therefore dependent mode choice as a case study. The first approach involves agent synchronization strategies, such as agent grouping by household, and mutual exclusive locks for car allocation in households. The second approach adopts model simplification to reduce inter-agent dependencies, enabling previously dependent agents to be computed in parallel. Performance evaluations were conducted on a simulated population of 1,871,371 individual persons in the city of Hamburg, Germany. The results demonstrate significant computation time reductions up to a speedup factor of 24.5 when using 32 cores. The model simplification approach, removing all inter-agent dependencies, achieved slightly higher speedups than agent synchronization strategies. Applying the simplified model is shown to produce similar simulation results with no significant differences. Additionally, we explore the transferability of these approaches to scenarios such as ride sourcing or car sharing, showing promising results for applications in this context.