Multi-Stage Vehicle Dispatch for Community Group-buying Logistics via Deep Reinforcement Learning

Community group-buying is an emerging shopping model in which individuals within the same community collectively purchase daily necessities at lower prices. The logistics transportation in this scenario requires higher turnover rates and timeliness compared to traditional online shopping. Therefore, making efficient vehicle dispatch decisions is necessary to improve delivery efficiency and reduce transportation costs. However, uncertainties in orders and complex dependencies between multi-stage decisions in the community group-buying scenario challenge the scheduling of delivery vehicles. In this paper, we propose a vehicle dispatch system for community group-buying logistics based on a predict-then-optimize process, which incorporates a demand prediction module and a vehicle dispatch module. In the prediction module, we categorize goods and predict the total quantity of each category using a multi-layer perceptron (MLP). In the vehicle dispatch module, we model the vehicle dispatch process as a reinforcement learning problem and use an action masking mechanism to prune the policy search space. Experimental results demonstrate that the vehicle dispatch system achieves state-of-the-art performance compared to several popular baselines across three different datasets. Our system reduces transportation costs by at least 8.89% and improves decision efficiency by at least 10%. Moreover, it behaves close to the approximate optimal solution in most cases.