Modeling stochastic service time for complex on-demand food delivery

Abstract Uncertainty is everywhere in the food delivery process, which significantly influences decision-making for complex on-demand problems, affecting efficiency and customer satisfaction. Especially, service time an indispensable part of process impacted by various uncertain factors. Due to simplicity high accuracy requirement, we model as a Gaussian mixture (GMM). In detail, transform distribution estimation problem into clustering determining probability each data belonging component (each cluster well). A hybrid algorithm proposed intelligently solve with criterion optimize quality simultaneously. First, simplicity, problem-specific encoding decoding methods are designed. Second, generate initial solutions good results, Chinese restaurant process-based initialization mechanism presented. Third, weighted-learning effectively guide update model. Fourth, local intensification based on maximum likelihood used exploit better solutions. The effect critical parameters performances investigated Taguchi design experimental method. To demonstrate effectiveness algorithm, carry out extensive offline experiments real-world historical data. Besides, employ GMMs obtained our platform, Meituan, assist order dispatching. results rigorous online A/B tests verify practical value introducing uncertainty real-life application.