In this research, the team studied a dynamic on-demand food delivery system and proposed a rolling horizon-based optimization approach integrated with adaptive large neighborhood search (ALNS) to efficiently obtain high-quality solutions. Researchers then used a daily activity generation tosimulationol, CEMDAP, to create a simulation scenario of on-demand food delivery behaviors based on real-world roadway network, restaurant locations, and population demographics in the City of Riverside, California. Two delivery policies are proposed: One-R and Multi-R, which allow orders from one or multiple restaurants to be bundled in one driver’s delivery trip, respectively. The system-level evaluation shows that on-demand food delivery has great potential to reduce dining-related VMT, resulting in significant reductions of fuel consumption and emissions, especially with Multi-R delivery policy. Under 14%, 21% and 40% delivery penetration rate, the total dining-related VMT can be reduced by 5%, 10%, and 25%, respectively, compared to the baseline with no on-demand delivery, and the corresponding environmental impacts were also reduced significantly.
