Online intelligent dispatching strategy of housekeeping order based on multi-agent reinforcement learning

With the rise of internet services and higher living standards, online home services are booming. Online housekeeping order dispatch is often modeled as Multi-Depot Vehicle Routing with Time Windows (MDVRPTW), yet existing solutions are mostly offline and lack real-world optimization considerations. This paper formulates the online order-aunt matching problem as a dynamic mixed-integer programming model, incorporating service score, average travel time, and distance per order. It introduces ML2CO, an online matching strategy using multi-agent reinforcement learning (MARL). MARL elements, including reward mechanisms, state transition, and actions, are designed based on service state and location. The strategy uses a masked pointer network and is trained with the parameter-sharing MADDPG algorithm to enhance performance. ML2CO's generalization and effectiveness are validated through data analysis and experiments on “58 Home” platform data.